2026-10-04·Weekly digest

🔬 Research Pulse

Weekly Digest

October 04, 2026

📈 Emerging Trends

🚀 Active Retrieval Replaces Passive Context in Agents

Frontier agents are shifting from stuffing everything into context to actively pulling raw past observations back in on demand. The architectural pattern is a lossless external store plus a learned policy that decides what to retrieve, when to retrieve it, and how to reorganize it mid-reasoning — a sharper formulation than the generic 'agent memory' framing we covered in August, now with multimodal and source-learning instances landing in the same week.

Signals:

VISTA drives Claude Opus 5.0 to 100% Relative Human Action Efficiency on ARC-AGI-3 by maintaining a lossless store of raw frames the model actively retrieves and reorganizes while reasoning
OmniSeek turns an Omni-LLM into an agent that picks a modality and temporal window per step, appending retrieved raw audio/video segments back into context, cold-started on 170K synthesized trajectories plus two-stage RL
SourceLearn builds a persistent source-specific model via self-directed revisiting, winning 13 of 15 benchmark-backend settings against Hybrid RAG by up to 22.6 points
AutoCompact learns when to compact rather than compacting on a fixed schedule — another signal that retrieval/forgetting is becoming a learned policy, not a config knob

🔨 Builder opportunity: An 'active memory' SDK for agent frameworks: a lossless observation log (frames, audio, tool outputs) plus a retrieval policy exposed as a tool call the LLM invokes mid-reasoning, with a cold-start trajectory synthesizer so teams can bootstrap the retrieval policy without a huge labeled dataset. Target customers already running LangGraph/CrewAI who are hitting context limits and getting lossy summarization.

🌱 Weight-Free Self-Improvement for Agents at Test Time

Several papers converge on a pattern where agents get better without gradient updates — by editing their own prompts, authoring reusable symbolic skills, or evolving reward programs under LLM supervision. This is distinct from latent/test-time reasoning: the artifact being improved is the agent's scaffolding (skills, prompts, reward specs), not activations, which makes the gains auditable, portable across model versions, and immune to catastrophic forgetting.

Signals:

RPG reaches 95% on 22 manipulation tasks and 30/30 real-robot trials by diagnosing failures with a privileged-state MLLM and authoring reusable symbolic skills plus system-prompt edits — explicitly branded 'weight-free'
InterEvolve has an LLM evolve reward program structure (staged rewards, completion conditions, tunable constants) at test time, verified across parallel sims, unlocking humanoid loco-manipulation competence that human-designed rewards miss
SourceLearn's persistent source model is refined by self-directed revisiting and task-guided refinement — again, improvement captured as structured artifacts rather than weight deltas

🔨 Builder opportunity: An 'agent scaffolding compiler' that treats prompts, skill libraries, and reward specs as a versioned, cross-task-validated artifact repo. Each production failure triggers an LLM-authored patch (new skill, prompt edit, reward tweak) that only lands after a gated eval against prior tasks. This gives teams the self-improvement loop without the ops burden of fine-tuning pipelines.

🌱 Benchmarks Are Catching VLAs Out

A wave of new robotics benchmarks is purpose-built to expose where current VLA/foundation-robot models fail, not to showcase wins. The common pattern: a known capability (tool use, dexterous retargeting, multi-robot coordination) decomposed into stages, with explicit measurement of where generalization collapses — and the results are unflattering for shipped VLAs like GR00T N1.7.

Signals:

HumanoidToolBench (18 tasks, 3.1k demos, Unitree G1 sim+real) shows GR00T N1.7 generalizes poorly to unseen tools and fails to ground instructions, with a wide gap between tool selection and task completion
FlashDexRetarget hits 90% on a 50-motion benchmark and scales to 1,000 motions — a dataset size calibrated to break prior retargeting pipelines
Watch, Infer, Coordinate releases a 3-environment benchmark for zero-shot coordination with a partner of unknown physical limitations, measured against oracle
DuoMind introduces RoboPoly for long-horizon coordinated manipulation and uses it to ablate orchestration hierarchy vs. semantic communication separately

🔨 Builder opportunity: A 'VLA capability report card' service: run a candidate robot foundation model through a standardized battery (tool generalization, dexterous retargeting scale, partner-constraint inference) and emit a per-axis score with failure traces. Robotics teams evaluating GR00T, OpenVLA, π0 etc. have no neutral comparison today — the vendors' own numbers don't decompose into these axes.

🌱 Capability Decomposition Replaces Aggregate Accuracy

Across math reasoning, protein biology, and tool use, researchers are rejecting single-number accuracy in favor of fine-grained diagnostic axes that reveal which sub-capability is bottlenecking a model. The framing is consistent: a seemingly monolithic skill is a composition of primitives, and progress requires measuring each one — then intervening on the specific gap, not the aggregate.

Signals:

'The Missing Primitive' splits math reasoning into Discovery/Generation/Digestion/Execution, identifies Discovery as the dominant bottleneck, and gets consistent gains from a primitive-privileged self-distillation targeted only at that gap
RipplePLM decomposes mutation effects into local vs. distal pathways via Direct-Distal Cross-Attention plus a Property Latent Chain supervising biochemical property changes separately
'Keyword Harnesses Fail Open' proposes a cheap diagnostic ladder for tool-use claims in small LMs — pointing out that current tool-use benchmarks fail-open and overstate competence
BenchTox benchmarks 20+ graph DL toxicity models under consistent preprocessing to correct for the inconsistent comparisons dominating the literature

🔨 Builder opportunity: Diagnostic eval harnesses as a product. Instead of selling 'our model scores X on MMLU,' sell teams a sub-capability profile for their domain (e.g., for math tutoring: Discovery vs. Execution scores; for drug discovery: local- vs. distal-effect prediction). The customer wins because they stop over-indexing on aggregate benchmarks that hide the specific failure that will bite them in production.

🚀 MoE Serving Breaks the VRAM Wall via Coordination

Memory-constrained MoE inference went from 'buy more H100s' to 'coordinate offloading, caching, and prefetching' in one week, with multi-order-of-magnitude TPOT improvements claimed. This genuinely advances the workload-aware heterogeneous inference theme — the new angle is cross-layer coordination (routing history → prefetch → cache replacement) rather than any single optimization.

Signals:

MoE-CORE reports ~27-33x TPOT improvements over vLLM Prefetch on DeepSeek-V4-Flash and GLM-5.2 quantized, layering four coordinated mechanisms (nonuniform caches, domain-informed init, routing-history replacement, cross-layer prefetch)
RapidMoE exploits cross-asymmetry via adaptive residual offloading specifically for large-scale MoE inference — a parallel attack on the same bottleneck
ePACT (energy-performance-aware commitment tracking for LLM serving) suggests the serving-layer research community has moved from 'optimize one knob' to 'commit-and-rollback across many knobs'

🔨 Builder opportunity: An open-source 'MoE serving coordinator' sidecar for vLLM/SGLang that implements routing-history-based cache replacement and cross-layer prefetch, tuned per-model. Target: teams running DeepSeek/GLM/Qwen MoE variants on 1-2 GPUs who currently can't fit the active expert set and would otherwise upgrade hardware. The research is fresh enough that the first open implementation will land before vendors ship it.

🤖 AI

🧠 LLMs

1. The Missing Primitive: Diagnosing and Repairing Mathematical Reasoning in Large Language Models

Authors: Shuo Xing, Zilin Dai, Chengyuan Qian... Published: 2026-10-01 | Citations: 0 arXiv | PDF

Research Question: Do LLMs possess structural mathematical understanding underlying their solutions, or does solution accuracy mask distinct capability profiles? Specifically, which mathematical reasoning sub-capabilities are the true bottlenecks, and can they be selectively repaired via post-training?

Summary: The paper argues that LLM solution accuracy on math problems hides distinct underlying capabilities, and introduces 'Mathematical Primitives' as a diagnostic lens across four axes (Discovery, Generation, Digestion, Execution). It identifies Discovery — recognizing which primitives to apply — as the dominant bottleneck, and proposes a primitive-privileged self-distillation method that targets this gap, yielding consistent gains across model scales.

Key Results: The paper introduces a benchmark evaluating four dimensions (Discovery, Generation, Digestion, Execution) of 'Mathematical Primitives.' It demonstrates that (1) models with similar solution accuracy have divergent capability profiles, (2) providing primitives unlocks substantial latent execution capacity, (3) Discovery is the dominant bottleneck, and (4) discovery-limited failures are most amenable to post-training repair. The proposed primitive-privileged self-distillation framework consistently improves math reasoning across model scales on challenging benchmarks (specific numbers not provided in abstract).

Key Findings:

Solution accuracy is a misleading aggregate: models with similar scores have very different Discovery/Generation/Digestion/Execution profiles
Supplying primitives at inference unlocks substantial latent execution capacity, implying execution is not the main bottleneck
Discovery (knowing which primitive to invoke) is the dominant failure mode, and discovery-limited failures respond well to targeted post-training

Technical Novelty: Two novel contributions: (1) the 'Mathematical Primitive' construct and a four-axis diagnostic benchmark (Discovery/Generation/Digestion/Execution) that decomposes reasoning rather than measuring end-to-end accuracy; (2) a primitive-privileged self-distillation framework that selectively transfers primitive-guided reasoning traces into the student, rather than full-trace distillation.

What's New: Prior math-reasoning benchmarks measure end-to-end accuracy; this work decomposes reasoning into four orthogonal primitive-centric capabilities, enabling targeted diagnosis. The post-training method is also novel in that it selectively distills primitive-guided traces rather than full reasoning chains, aligning the training signal with the identified bottleneck.

Extension Opportunities:

Apply the Discovery/Generation/Digestion/Execution decomposition to other reasoning domains (code, scientific reasoning, legal analysis) to see if 'primitive discovery' is a universal bottleneck
Build an inference-time scaffold that externally supplies candidate primitives (via RAG over a theorem/lemma library) to augment weak-discovery models without retraining
Extend the self-distillation framework to curriculum learning that progressively weans the student off primitive scaffolding, measuring whether discovery capability transfers

Replicability: Abstract does not mention code/data release. Reproducing would likely require (a) constructing or obtaining the primitive-annotated benchmark, (b) a teacher LLM capable of primitive-guided reasoning (likely frontier-scale), and (c) fine-tuning compute for student models at multiple scales — moderate-to-high GPU budget (multi-A100/H100 days).

Research Gaps:

No standardized way to catalog the universe of 'mathematical primitives' — the taxonomy's completeness and domain coverage are unclear
Unclear whether Discovery bottleneck persists under RL-based reasoning training (e.g., o1-style) or is specific to supervised/distilled models

2. From Knowledge Access to Source Learning: Developing Source-Specific Competence

Authors: Lucheng Fu, Kejing Xia, Yiyang Wang... Published: 2026-10-01 | Citations: 0 arXiv | PDF

Research Question: How can LLM agents move beyond repeated access of a persistent authoritative source to progressively build reusable, source-specific competence—capturing how the source's knowledge is structured, interpreted, and applied—rather than treating each interaction as isolated retrieval?

Summary: The paper reframes repeated agent interaction with an authoritative source as an opportunity to build persistent, source-specific competence rather than redundant retrieval. SourceLearn constructs and refines a persistent source model via self-directed revisiting (closing understanding gaps) and task-guided refinement (fixing representational gaps surfaced by downstream use), with updates always grounded in the authoritative source. It wins 13 of 15 benchmark-backend settings, beating Hybrid RAG by up to 22.6 points.

Key Results: SourceLearn was evaluated across 5 benchmarks and 3 LLM backends (15 total settings), achieving the best performance in 13 of 15 settings. It delivered gains of up to 22.6 points over Hybrid RAG and substantially outperformed static source representations and experience-based memory baselines.

Key Findings:

Treating a source as an object of learning (not just retrieval) yields substantial gains over Hybrid RAG and experience-based memory baselines
Combining self-directed (introspective) and task-guided (extrinsic) learning signals is complementary—each surfaces different kinds of gaps in source understanding
Anchoring updates to the authoritative source (rather than to accumulated experience) is key to avoiding drift while still benefiting from task feedback

Technical Novelty: Introduces 'source learning' as a distinct paradigm from RAG (access) and agent memory (experience). The persistent source model stores reusable understanding of a single authoritative source, refined via two complementary signals: Self-Directed Source Learning (identifies gaps and adaptively revisits the source) and Task-Guided Source Learning (uses downstream task experience to surface representational gaps). Critically, learning signals decide what to reconsider but updates are reconstructed from the authoritative source—preventing experiential drift away from ground truth.

What's New: Prior work splits into source-access improvements (RAG, indexing) and agent-memory systems (reusable experiential knowledge). This paper carves out a third axis—source competence—where the artifact being learned is a model of the source itself, grounded in the authoritative text but shaped by both introspective and task-driven signals.

Extension Opportunities:

Extend source models to multi-source settings where competence over one source informs learning on related sources (cross-source transfer and conflict resolution)
Apply SourceLearn to codebases as persistent sources—progressively building a reusable 'repo model' that improves agent coding performance over time instead of re-reading files per task
Investigate forgetting/decay and versioning when authoritative sources change over time, so source models stay faithful to updated ground truth

Replicability: The abstract does not mention a public code or data release. Reproduction would require the 5 benchmarks (unspecified in abstract), 3 LLM backends (likely API-based or open-weight models), and moderate compute for repeated source-reading loops over knowledge-intensive tasks—feasible on a single workstation if backends are API-hosted, or a small GPU cluster for open-weight serving.

Research Gaps:

No treatment (in the abstract) of multi-source or dynamically-updating sources—how source models evolve when the underlying authoritative text changes
Unclear how source-model size, storage, and compute scale with source complexity, or how to decide when a source is 'learned enough'

🤖 Agents

1. DuoMind: Enabling Distributed Multi-Robot Coordination with Semantic Communication

Authors: Hanchu Zhou, Dechen Gao, Hang Wang... Published: 2026-10-01 | Citations: 0 arXiv | PDF

Research Question: How can multiple robots coordinate long-horizon manipulation behaviors in a distributed fashion while preserving fine-grained execution reliability, given that current VLM/VLA progress has focused almost exclusively on single-robot settings?

Summary: DuoMind is a distributed hierarchical framework in which each robot runs a VLM orchestrator that reasons over its instruction, local observations, and semantic messages from peers, then dispatches low-level commands to a VLA action model. The authors also introduce RoboPoly, a new benchmark for long-horizon coordinated manipulation, and show DuoMind outperforms baselines on RoboPoly and RoboTwin with ablations validating both the orchestration hierarchy and the semantic communication channel.

Key Results: The paper introduces DuoMind, a distributed hierarchical framework pairing a VLM orchestrator with a VLA action model per robot, and demonstrates improved multi-robot task performance on two benchmarks: RoboPoly (a new benchmark they built for long-horizon coordinated manipulation under distributed control) and RoboTwin. Ablations reportedly confirm that both the hierarchical orchestration layer and the semantic inter-robot communication channel independently contribute to performance gains. Specific numerical results, success rates, and task counts are not provided in the abstract.

Key Findings:

Pairing a VLM orchestrator with a VLA executor per robot outperforms monolithic approaches on multi-robot long-horizon manipulation
Semantic (language-level) inter-robot communication is an effective coordination substrate — ablating it degrades performance
Existing multi-robot benchmarks are insufficient for closed-loop distributed coordination, motivating the new RoboPoly suite

Technical Novelty: The novel contribution is the explicit separation of concerns across two pretrained model families — VLMs for high-level reasoning and inter-agent natural-language-style semantic messaging, VLAs for low-level action generation — in a distributed (not centralized) multi-robot setting, plus the RoboPoly benchmark targeting closed-loop coordinated manipulation. Prior VLA work largely treats the robot as a single-agent system; prior multi-robot coordination work rarely leverages pretrained VLM reasoning as the coordination substrate.

What's New: First work (as framed) to combine pretrained VLMs and VLAs in a distributed, per-robot hierarchical stack where coordination happens through natural-language semantic messages rather than shared latent states or a centralized planner, plus a purpose-built benchmark for the setting.

Extension Opportunities:

Replace the free-form semantic message channel with a learned compressed communication protocol (e.g., via emergent communication or RL fine-tuning) to reduce latency and token cost in real deployments
Scale beyond two-robot coordination to heterogeneous fleets (ground + aerial, or arms + mobile bases) and test whether the hierarchical VLM-orchestrator pattern degrades gracefully with more peers
Add a failure-recovery / replanning layer that uses the semantic messages as an audit trail for debugging and credit assignment when a coordinated task fails mid-execution

Replicability: The abstract references a project page but does not confirm open-sourced code, weights, or the RoboPoly benchmark release. Reproducing would require access to VLA checkpoints (e.g., OpenVLA / π0-class models), a capable VLM backbone, and simulation infrastructure for RoboPoly/RoboTwin — likely multiple A100/H100-class GPUs for inference across robots plus simulator compute.

Research Gaps:

No stated analysis of communication bandwidth, latency, or robustness to dropped/corrupted messages — critical for real multi-robot deployment
Unclear how the approach scales past two robots or to heterogeneous embodiments, and whether orchestrator reasoning cost becomes prohibitive

👁️ Vision

1. VISTA: A Visual Harness for Reasoning in an Interactive World

Authors: Qiushi Han, Keya Hu, Linlu Qiu... Published: 2026-10-01 | Citations: 4 arXiv | PDF

Research Question: How can general-purpose multimodal models be equipped with long-horizon visual reasoning capabilities to effectively solve tasks across diverse interactive visual environments, rather than being limited by short-context perception and lossy memory representations?

Summary: VISTA is a visual harness that gives a frontier multimodal model (Claude Opus 5.0) long-horizon vision by maintaining a lossless memory of raw past observations the model can actively retrieve and reorganize while reasoning. It drives Claude Opus 5.0 to a perfect 100.00 Relative Human Action Efficiency on ARC-AGI-3 — completing all 25 public games with 57.4% fewer actions than first-time humans — and generalizes to other visual game and puzzle benchmarks with minimal adaptation.

Key Results: VISTA with Claude Opus 5.0 achieves a perfect 100.00 Relative Human Action Efficiency on ARC-AGI-3 (up from 40.68 baseline), completing all 25 public games using 57.4% fewer actions than first-time human participants. It also substantially outperforms minimal-harness baselines on three additional benchmarks covering visual games and puzzles, though specific numbers for those are not given in the abstract.

Key Findings:

A simple visual harness, without model fine-tuning, can lift Claude Opus 5.0 from 40.68 to a perfect 100.00 on ARC-AGI-3's Relative Human Action Efficiency metric
Preserving past observations in their original visual form (rather than summarizing) materially improves long-horizon interactive reasoning
The harness generalizes: the same design transfers to three additional visual game/puzzle benchmarks and beats minimal-harness baselines using the same underlying model

Technical Novelty: VISTA introduces a 'lossless visual memory' that preserves past observations in their original pixel form (rather than text summaries or compressed embeddings) combined with active retrieval, where the model itself chooses which past frames to re-materialize into its visual context as it reasons — a departure from prior agent harnesses that convert visual history into text or discard it.

What's New: Prior agent scaffolds typically convert visual history into text summaries or compressed representations. VISTA keeps observations lossless and lets the model actively reorganize its own visual context, treating visual memory as a first-class, retrievable medium rather than a transient perceptual input.

Extension Opportunities:

Apply VISTA's visual memory + active retrieval harness to real-world embodied robotics tasks (navigation, manipulation) where lossless observation history matters for long-horizon planning
Extend the harness to support multimodal memory beyond vision — integrating audio, proprioception, or structured world-state representations retrieved in parallel with visual frames
Investigate learned retrieval policies (e.g., fine-tuned or RL-trained) rather than relying on the model's zero-shot ability to decide which past observations to pull back into context

Replicability: The abstract does not mention code or data release. Reproducing results requires API access to Claude Opus 5.0 (significant inference cost given long-horizon multi-game evaluation across 25+ ARC-AGI-3 games plus three additional benchmarks) but no model training is implied — the harness is architectural scaffolding around an existing frontier model.

Research Gaps:

No evidence yet that lossless visual memory scales to open-world or real-world embodied settings where observation spaces are vastly larger than game frames
Unclear how much of the gain comes from the harness design vs. the underlying frontier model's capabilities — ablations against weaker base models would clarify generality

2. OmniSeek: Native Tool Integration for Multi-turn Audio-Visual Reasoning

Authors: Haibo Wang, Jiteng Mu, Jialu Li... Published: 2026-10-01 | Citations: 0 arXiv | PDF

Research Question: How can an Omni-LLM move beyond passive single-pass processing of long audio-visual sequences to actively and selectively retrieve sparse cross-modal evidence through multi-turn reasoning?

Summary: OmniSeek converts an Omni-LLM into an agent that actively decides which modality (audio or video) and which temporal window to inspect at each reasoning step, appending retrieved raw segments back into context. It is cold-started via SFT on a synthesized 170K-trajectory corpus (OmniTraj-170K) and refined with two-stage RL that includes a novel Audio-Visual Necessity reward to prevent single-modality shortcuts.

Key Results: The authors built OmniTraj-170K, a corpus of multi-hop Chain-of-Thought trajectories with interleaved audio/visual evidence, used it for SFT cold-start, then applied two-stage RL with verifiable rewards plus an Audio-Visual Necessity objective. The abstract claims consistent improvements in audio-visual reasoning across a wide range of benchmarks, but no specific numeric scores, baselines, or benchmark names are disclosed in the abstract.

Key Findings:

Agentic multi-turn tool use outperforms single-pass full-sequence processing for long audio-visual reasoning
SFT on synthesized multi-hop AV CoT trajectories is sufficient to cold-start cross-modal tool-use behavior before RL
An explicit Audio-Visual Necessity reward is needed to prevent the policy from collapsing to a single dominant modality

Technical Novelty: Treating evidence acquisition itself as a learned action within the Omni-LLM's reasoning loop (deciding when to look vs. listen and over which temporal window), combined with the explicit Audio-Visual Necessity reward that penalizes single-modality shortcuts — prior AV reasoning work typically processes the full stream in one pass or uses modality-agnostic retrieval.

What's New: First framework (per the claim) to make temporal + modality-selective evidence seeking a native, learned action inside an Omni-LLM's reasoning loop, with an RL objective explicitly designed to force genuine cross-modal grounding rather than visual-only or audio-only shortcuts.

Extension Opportunities:

Add a third modality (e.g., text documents, 3D/point-cloud, or sensor streams) to the tool-use action space and extend the Necessity objective to k-modality coverage instead of just audio+visual
Replace the fixed temporal window selection with a learned continuous-time attention/scheduling policy, or add a 'zoom' tool that recursively refines window granularity for very long videos (hours-scale)
Distill the agentic multi-turn behavior into a smaller edge model, or cache repeated tool-call sub-trajectories to cut inference cost — the current iterative protocol likely has high latency unsuited for real-time use

Replicability: No code, model weights, or dataset release is mentioned in the abstract. Reproduction would require: training an Omni-LLM base, generating ~170K multi-hop AV trajectories (expensive synthesis pipeline), SFT, and two-stage RL with verifiable rewards — likely tens to hundreds of H100-GPU-days. Replicability from abstract alone is low.

Research Gaps:

No numeric results, baselines, or named benchmarks are presented in the abstract, making magnitude of the claimed gains unverifiable
Unclear how the approach handles streaming/online settings or very long (hour-scale) content where tool-call budget matters

🦾 ROBOTICS

1. HumanoidToolBench: Benchmarking Humanoid Tool Use from Selection to Mobile Execution

Authors: Kyochul Jang, Seohyeon Park, Ohchul Kwon... Published: 2026-10-01 | Citations: 0 arXiv | PDF

Research Question: How can we jointly evaluate a humanoid robot's ability to select an appropriate tool AND coordinate manipulation plus locomotion to execute tool-use tasks — a combined capability no existing benchmark measures together on a humanoid platform?

Summary: HumanoidToolBench is an 18-task benchmark plus a 3.1k-demonstration dataset (ToolBook) that jointly evaluates tool selection, manipulation, and locomotion on a humanoid robot (Unitree G1), in both simulation and reality. Evaluating 7 sim and 3 real policies, the authors expose a wide gap between picking the right tool and actually completing the task, and show that current VLA models like GR00T N1.7 generalize poorly to unseen tools and fail to ground instructions properly.

Key Results: Introduced HumanoidToolBench (18 tasks across 3 scenarios, 3 execution levels, 2 tool-set modes) and ToolBook (3.1k demonstrations in simulation + real Unitree G1). Evaluated 7 policies in simulation and 3 on real hardware, demonstrating a substantial gap between tool selection and task completion. GR00T N1.7 probes showed degraded selection accuracy on unseen tools and problematic behavior of continuing execution even under unrelated instructions.

Key Findings:

There is a substantial performance gap between selecting the correct tool and successfully completing the task across evaluated policies
GR00T N1.7 tool-selection accuracy drops meaningfully on unseen tools, indicating weak generalization
Policies continue to execute tasks even when given unrelated instructions, revealing poor instruction grounding

Technical Novelty: First benchmark to jointly evaluate tool selection + mobile manipulation on a humanoid platform, structured along three axes (scenarios × execution levels × tool-set modes) with paired sim-and-real demonstration data on the Unitree G1 — prior tool-use benchmarks focus on either selection OR manipulation, and rarely on humanoids with locomotion in the loop.

What's New: Unlike prior work that treats tool selection and manipulation separately — or evaluates tool use on tabletop arms — HumanoidToolBench is the first benchmark to combine selection, manipulation, and locomotion on a humanoid, and pairs it with matched sim+real demonstrations on the same hardware.

Extension Opportunities:

Extend ToolBook with a larger, more diverse set of unseen tools (varied geometry, affordances, materials) to stress-test generalization and train selection heads with contrastive affordance learning
Add an instruction-grounding verification layer (e.g., VLM-based gating) that halts execution when the instruction is unrelated to the current tool/scene, directly targeting the 'continues under unrelated instructions' failure mode
Introduce multi-tool sequential tasks (pick tool A, use, swap to tool B) with explicit locomotion legs to probe long-horizon manipulation + navigation coupling beyond the current 18 single-tool tasks

Replicability: Code and data publicly available at snu-pi.github.io/HumanoidToolBench/. Reproduction requires a humanoid simulator plus optionally a Unitree G1 for real-robot evaluation; running GR00T N1.7-scale policies implies GPU compute suitable for VLA inference/fine-tuning (likely multi-GPU workstation for training, single modern GPU for eval).

Research Gaps:

Current humanoid VLA policies cannot reliably generalize tool selection to unseen tools
Policies lack instruction-grounding safeguards and will execute regardless of whether the instruction matches the context

2. Reconstruct, Practice, Go Real: Guided Self-Improvement for Embodied Agents

Authors: Yen-Jen Wang, Haozhe Jiang, Shuying Deng... Published: 2026-10-01 | Citations: 0 arXiv | PDF

Research Question: How can robot execution systems autonomously improve across diverse manipulation tasks without the heavy human effort of designing skills, rewards, and perception-control integration — and without updating model weights?

Summary: RPG is a weight-free self-improvement framework that turns an offline dataset into simulated practice tasks, uses an MLLM with privileged simulator state to diagnose failures and author reusable symbolic skills plus system-prompt edits, and gates each change via cross-task evaluation. The resulting frozen system reaches 95.0% on 22 manipulation tasks and 30/30 real-robot trials, substantially beating prior self-improving agent baselines.

Key Results: On 22 held-out manipulation tasks, RPG raised task success from 28.6% (round 1) to 95.0% (round 15), beating ASPIRE (75.5%) and CaP-Agent0 with GPT-6 Astra Pro (60.0%). After a shared calibration/hardware-adaptation step, the frozen system completed 30/30 physical trials (10 per task across 3 tasks).

Key Findings:

Iterative practice without weight updates can lift manipulation success from 28.6% to 95.0% over 15 rounds on 22 tasks
Cross-task evaluation of candidate skills/prompts before merging prevents regression and is critical to the compounding improvement
The improved symbolic skill library and system prompt transfer zero-shot to real hardware (30/30 trials) after only a shared calibration step

Technical Novelty: The novelty is a closed weight-free improvement loop that (a) reconstructs practice tasks in simulation from an offline dataset, (b) uses privileged simulator state plus dataset videos to let an MLLM diagnose failures and author/refine reusable symbolic skills and the system prompt, and (c) gates every candidate change through cross-task evaluation before it is merged. Prior self-improving agent work (ASPIRE, CaP) learns skills or edits prompts but does not combine simulator-grounded failure diagnosis, symbolic skill library growth, and cross-task regression testing in one frozen-weights loop.

What's New: Combines simulator-grounded failure diagnosis, growth of a reusable symbolic skill library, and system-prompt revision into a single autonomous loop that improves a frozen-weights MLLM policy — and validates every edit with cross-task tests before retention.

Extension Opportunities:

Extend the Reconstruct step to automatically mine practice tasks from in-the-wild human video (not just curated offline datasets), broadening the skill library without manual demonstrations
Add a sim-to-real residual policy or online calibration loop so the frozen symbolic skills self-adapt to new embodiments and lighting without the one-time hardware-adaptation procedure
Replace the symbolic skill representation with a hybrid neural-symbolic one that lets failure diagnoses mutate learned low-level controllers (not just prompt + symbolic code), enabling improvement on contact-rich tasks where symbolic skills bottleneck

Replicability: Project website (rpg-robot.github.io) is provided; code/data availability is not stated in the abstract. Reproducing would require a physics simulator, an offline manipulation dataset, a strong multimodal LLM for diagnosis and runtime coordination, and modest robot hardware for the 30-trial physical eval — compute is dominated by LLM API calls across 15 practice rounds rather than model training.

Research Gaps:

No evidence yet that the approach scales to contact-rich or deformable manipulation where symbolic skills are a poor abstraction
Reliance on an offline dataset to seed practice tasks limits applicability to domains lacking curated demonstrations

3. InterEvolve: Test-Time Evolution of Reward Programs for Humanoid Loco-Manipulation

Authors: Zhuo Lin, Sirui Xu, Liuyu Bian... Published: 2026-10-01 | Citations: 0 arXiv | PDF

Research Question: How can a pretrained humanoid loco-manipulation controller solve novel contact-rich, multi-stage tasks it was never trained for, without retraining — by repurposing existing skills via test-time adaptation and learning from its own execution feedback?

Summary: InterEvolve enables a humanoid robot to solve novel loco-manipulation tasks at test time without retraining, by combining an object-aware forward-backward behavioral foundation model with LLM-evolved reward programs that stage rewards, completion conditions, and tunable constants. An LLM revises program structure using execution feedback and a verified-skill library while a numerical optimizer tunes constants, with candidates verified across parallel simulations — unlocking competence that human-designed rewards miss and transferring to a physical Unitree G1.

Key Results: The authors demonstrate that (1) an object-aware forward-backward (FB) behavioral foundation model with object residuals on a frozen body prior enables test-time task specification via rewards; (2) LLM-evolved reward programs unlock loco-manipulation competence that human-designed rewards leave untapped, sometimes discovering novel strategies; (3) the system handles diverse tasks, complex scenes, and long-horizon compositions in simulation; and (4) evolved skills transfer to a physical Unitree G1 humanoid running autonomously from egocentric onboard perception. The abstract does not provide quantitative benchmark numbers or success-rate comparisons.

Key Findings:

Human-designed reward functions leave much of a pretrained FB model's loco-manipulation competence untapped; LLM-evolved reward programs release it, sometimes via novel strategies
Object residuals on a frozen body-prior FB model are sufficient to turn body-or-object rewards into executable loco-manipulation at test time, without full-model retraining
Co-evolving program structure (LLM) and constants (numerical optimizer) with parallel-sim verification produces behaviors for diverse tasks, complex scenes, and long-horizon compositions, and transfers sim-to-real on a Unitree G1 with onboard egocentric perception

Technical Novelty: Two coupled innovations: (1) an object-aware forward-backward behavioral foundation model that adds object residuals to a frozen body prior, letting new rewards about body OR objects compose into loco-manipulation behavior at test time — a departure from prior FB models that handle only body-centric tasks; and (2) reward programs as the planning–control interface: staged rewards with completion conditions and tunable constants, co-evolved by an LLM (structure) and a numerical optimizer (constants) using parallel-sim verification and a verified-skill library. Prior test-time task adaptation uses either flat reward functions or pure LLM code generation without the staged/verified program abstraction.

What's New: First work to combine behavioral foundation models (FB) with LLM-driven reward-program evolution for humanoid loco-manipulation. The object-aware FB residual decomposition is new, as is treating rewards as staged, verifiable programs co-optimized by LLM and numerical search rather than hand-engineered scalar functions or ad-hoc LLM-generated code.

Extension Opportunities:

Replace the LLM program-structure reviser with a learned policy that distills successful evolution traces, reducing per-task LLM calls and latency for faster test-time adaptation
Extend the reward-program interface to multi-robot or human-robot collaborative loco-manipulation, where staged rewards specify handoffs and shared contact events
Integrate tactile and force sensing into the object-aware FB residuals to improve contact-rich manipulation (e.g., deformable objects, tool use) beyond vision-only egocentric perception

Replicability: The abstract does not mention a code or data release. Reproducing would require: a humanoid simulator (likely Isaac Gym or MuJoCo) with parallel scenario rollout, a pretrained FB behavioral foundation model (nontrivial to train — typically hundreds of GPU-hours), an LLM API for program revision, and a Unitree G1 for hardware validation. Expect multi-GPU cluster compute for the FB pretraining phase; test-time evolution itself is cheaper but LLM-call-bounded.

Research Gaps:

No quantitative task-success numbers, baseline comparisons, or ablation metrics are reported in the abstract — making it hard to judge the magnitude of improvement over prior test-time adaptation or hand-designed-reward baselines
Scope of real-world validation (number of tasks on Unitree G1, failure modes, robustness to perception noise, latency of the LLM-in-the-loop evolution on hardware) is unspecified

4. FlashDexRetarget: Accelerating Dexterous Manipulation Data Generation through Multi-Motion Retargeting

Authors: Kyungmin Lee, Sibeen Kim, Dongyoon Hwang... Published: 2026-10-01 | Citations: 0 arXiv | PDF

Research Question: How can we retarget human hand-object demonstrations to dexterous robot hands with both high success rates and training efficiency, overcoming limitations of existing physics-based methods that trade one for the other?

Summary: FlashDexRetarget is an RL-based framework that retargets human hand-object demonstrations to dexterous robot hands with both high success and high training efficiency. By combining point-cloud observations, hand-object distance features, future trajectory encodings, complementary rewards, split left/right actor-critics, and the off-policy FlashSAC algorithm, it hits 90% success on a 50-motion benchmark — ~2.5x sampling baselines at up to 100x less compute than RL baselines — and scales stably to 1,000 motions.

Key Results: On a 50-motion benchmark (single and two-object interactions), FlashDexRetarget achieved 90% success rate — approximately 2.5x the evaluated sampling-based baselines — while using up to 100x less training compute than evaluated RL-based baselines. Consistent gains were shown across two robot hands (XHand and Sharpa Wave Hand), and the method remained stable scaling to 200, 500, and 1,000 motions, producing successful retargets more efficiently as the training set grew.

Key Findings:

90% retargeting success rate on 50-motion benchmark covering single- and two-object interactions, roughly 2.5x the sampling-based baselines
Up to 100x less training compute than evaluated RL-based baselines, driven by off-policy FlashSAC and the split bimanual actor-critic design
Scales stably to 200/500/1000 motions and gets more efficient per-motion as the training set grows, suggesting positive transfer across motions
Generalizes across two different robot hands (XHand and Sharpa Wave Hand) and works on real-world-captured human demonstrations

Technical Novelty: The combination of (1) a multi-modal observation design (object point clouds + hand-object distance features + future trajectory encodings) paired with complementary rewards supervising both object motion and reference hand-object relationships, (2) separate left- and right-hand actor-critic networks for bimanual decoupling, and (3) adapting FlashSAC — an off-policy SAC variant — specifically to dexterous motion tracking rather than using the on-policy methods common in prior retargeting RL work.

What's New: Prior physics-based retargeting work forced a tradeoff: sampling-based methods are general but low-success, while RL-based methods are high-success but motion-specific and compute-heavy. FlashDexRetarget is the first to combine off-policy RL (FlashSAC) with task-aware observations (point clouds + distance + future trajectory) and bimanual network decoupling to break this tradeoff, achieving both high success and dramatic compute reduction at scale.

Extension Opportunities:

Extend the framework to bimanual-plus-object manipulation with deformable or articulated objects (e.g., cloth, scissors), where point-cloud observations and distance features would need to encode articulation state
Integrate the retargeted trajectories into a downstream policy-learning pipeline (behavior cloning or RL fine-tuning) to measure how retargeting quality translates to real-world task success on physical hardware
Replace the hand-object distance features with learned contact-aware representations (e.g., signed-distance fields or contact graphs) to see whether this improves retargeting of contact-rich motions like in-hand reorientation

Replicability: Code and videos are available at the project page (davian-robotics.github.io/FlashDexRetarget/). The abstract does not quantify exact compute, but claims up to 100x less than RL baselines, suggesting it's reproducible on moderate GPU resources typical for Isaac Gym / IsaacLab-style simulation. Reproduction would require a GPU-accelerated simulator and the referenced motion dataset.

Research Gaps:

Real-world deployment: all reported results appear to be in simulation; sim-to-real transfer of the retargeted trajectories to physical dexterous hands is not demonstrated in the abstract
Interaction scope is limited to rigid single- and two-object manipulation; deformable objects, tool use, and long-horizon multi-stage interactions aren't evaluated

5. Watch, Infer, Coordinate: Inferring Robot Partner Constraints for Zero-Shot Coordination

Authors: Suyu Ye, Zheyuan Zhang, Vaishnav Tadiparthi... Published: 2026-10-01 | Citations: 0 arXiv | PDF

Research Question: How can a helper robot infer an unknown physical constraint (e.g., actuator faults, hardware degradation) of a partner robot by observing it coordinate with a third robot, and then leverage that inferred capability for zero-shot coordination on a new joint manipulation task?

Summary: The paper tackles zero-shot multi-robot coordination when one partner has unknown physical limitations, by inferring those constraints from a third-party demonstration of the partner cooperating with another robot. The core insight is that both robots' joint behavior — not just the constrained one's — reveals the hidden capability, since the helper compensates. The authors release a 3-environment benchmark and show their joint-behavior scoring approach approaches oracle-level coordination.

Key Results: The authors introduce the 'Watch, Infer, Coordinate' benchmark spanning three physically coupled manipulation settings and show their inference method substantially improves both constraint inference and zero-shot coordination performance across all three settings, approaching an oracle baseline that has access to the true constraints. Specific numeric gains vs. baselines are not stated in the abstract.

Key Findings:

Observing joint behavior of a cooperating pair is substantially more informative about one partner's physical constraints than observing that partner alone, because the other robot's compensatory actions encode the limitation
A candidate-scoring inference procedure over joint trajectories yields constraint estimates good enough to drive zero-shot coordination with a new helper on a new task
Across all three physically coupled manipulation settings in the proposed benchmark, the method approaches an oracle that has access to the true partner constraints

Technical Novelty: Prior constraint/capability inference methods typically reason only about the constrained agent's own behavior. The key new idea is that in physically coupled tasks, the unconstrained partner compensates for the limitation, so the joint behavior of both agents carries more information than either trajectory alone. The method scores candidate constraints by how well they explain the observed joint behavior, enabling zero-shot coordination without any interaction with the constrained partner.

What's New: Reframes partner-capability inference as a joint-behavior interpretation problem rather than a single-agent inverse problem, and ties it directly to zero-shot ad-hoc teamwork in physically coupled manipulation — a setting underexplored vs. the usual decoupled multi-agent RL coordination benchmarks like Overcooked or Hanabi.

Extension Opportunities:

Extend the inference approach to >2-robot teams or heterogeneous agents (robot + human) where multiple partners may each have unknown, possibly correlated constraints
Replace the discrete 'candidate constraint scoring' with a continuous latent-capability model (e.g., learned via variational inference or diffusion-based behavior cloning) to handle unseen fault types at test time
Transfer the method from simulation to real hardware with real actuator degradation (e.g., worn joints, reduced torque limits) and evaluate online adaptation as the partner's constraints drift during a task

Replicability: The abstract does not mention a code release or dataset URL. Replication would require implementing the three physically coupled manipulation environments (likely in a physics simulator such as MuJoCo or Isaac) plus multi-agent RL/planning baselines. Compute is likely modest by LLM standards — single-GPU multi-agent RL training is plausible — but exact requirements are not disclosed.

Research Gaps:

No real-robot evaluation reported in the abstract; sim-to-real transfer of inferred constraints under noisy sensing and real actuator faults remains open
Method appears to rely on a predefined candidate constraint set; handling open-ended or compositional fault spaces (e.g., simultaneous joint-limit + payload + latency constraints) is not addressed

💻 COMPUTE

1. Single-Particle Spectral Estimation

Authors: Adrian Chapman, Charles Derby, Steven T. Flammia... Published: 2026-10-01 | Citations: 0 arXiv | PDF

Research Question: How can one learn the single-particle energies $E_i$ of a non-interacting Hamiltonian when its structure is hidden by an unknown global unitary $U$ and entangled with an arbitrary 'spectator' Hamiltonian, without having to reconstruct $U$ itself?

Summary: The paper introduces SPICES, a hybrid quantum-classical algorithm that recovers the single-particle (quasiparticle) spectrum of a non-interacting Hamiltonian hidden inside a global unitary and coupled to an arbitrary spectator, without ever reconstructing the unitary. It proves the learning task is DQC1-hard and shows SPICES — using a Hadamard test plus a Fourier/contour-based classical postprocessor tied to zeros of the partition function at complex $\beta$ — efficiently recovers the spectrum in Wasserstein distance.

Key Results: The paper proves two main results: (1) the problem is DQC1-hard, establishing a classical-hardness lower bound even despite the non-interacting structure; and (2) their SPICES algorithm efficiently recovers the spectral distribution in Wasserstein distance under natural assumptions, by combining a Hadamard test with classical postprocessing that exploits the correspondence between quasiparticle energies and Fourier frequencies of partition-function zeros at complex inverse temperature. The abstract does not report concrete numerical benchmarks, dataset sizes, or qubit counts.

Key Findings:

Learning the hidden non-interacting weights $E_i$ under an unknown global unitary with a spectator is DQC1-hard, so even this 'simple' structure is not classically easy
SPICES recovers the spectral distribution to small Wasserstein error using only Hadamard-test measurements plus classical postprocessing, avoiding full tomography of $U$
Single-particle energies can be identified with Fourier frequencies governing zeros of the partition function at complex inverse temperature, giving a clean estimator

Technical Novelty: The novel element is the 'contour estimator' classical postprocessing: rather than trying to invert $U$ or do full Hamiltonian tomography, SPICES reads single-particle energies off the Fourier frequencies that govern zeros of the partition function along a complex-inverse-temperature contour, and pairs this with a standard Hadamard test. The DQC1-hardness result for this specific obfuscated non-interacting learning problem is also new.

What's New: Prior Hamiltonian-learning work typically assumes a known operator basis or reconstructs the unitary; this paper tackles a basis-obfuscated setting with a spectator, proves it is DQC1-hard, and introduces a contour-based estimator that targets only the spectrum (not the full Hamiltonian), which is both novel in formulation and well-matched to downstream materials-science workflows.

Extension Opportunities:

Implement SPICES on current superconducting or trapped-ion hardware for a small materials model (e.g., Hubbard dimer with spectator) and benchmark recovered spectra against exact diagonalization
Extend the contour-estimator postprocessing to handle noisy Hadamard-test data with realistic depolarizing/dephasing error models, and derive sample-complexity bounds under those noise channels
Couple SPICES as a subroutine inside a DFT or Green's-function workflow (e.g., feeding recovered quasiparticle energies into a self-consistent mean-field loop) to demonstrate end-to-end quantum-assisted materials simulation

Replicability: The abstract does not mention a public code release, dataset, or specific hardware requirements. Reproducing SPICES would require implementing Hadamard-test circuits for the system plus the contour-based classical postprocessing — likely feasible in simulation (Qiskit/Cirq) on commodity hardware for small system sizes; experimental reproduction would need a programmable quantum device supporting controlled time evolution under the obfuscated $H$.

Research Gaps:

No empirical validation (hardware or large-scale simulation) or concrete sample-complexity numbers are reported in the abstract
The 'natural assumptions' required for efficient Wasserstein recovery are not quantified in the abstract — unclear how restrictive they are for realistic materials Hamiltonians with strong spectator coupling or noise

2. MoE-CORE: Coordinated Expert Offloading and Residency for Memory-Constrained MoE Inference

Authors: Ke Yang, Yongji Gao, Xushi Li... Published: 2026-10-01 | Citations: 0 arXiv | PDF

Research Question: How can Mixture-of-Experts (MoE) models be served efficiently on memory-constrained AI appliances where total expert weights exceed device memory, without letting host-to-device transfers dominate inference latency?

Summary: MoE-CORE is an inference system that makes memory-constrained MoE serving practical by coordinating expert offloading, caching, and prefetching across prefill and decode. It layers four mechanisms (nonuniform caches, domain-informed init, routing-history replacement, cross-layer prefetch) plus optional score-based substitution, achieving ~27-33x TPOT improvements over vLLM Prefetch on DeepSeek-V4-Flash and GLM-5.2 quantized models.

Key Results: On DeepSeek-V4-Flash-W4A8, MoE-CORE achieves mean TPOT of 38.0-44.8 ms vs. 1268.9-1269.1 ms for vLLM Prefetch (~28-33x speedup). On GLM-5.2-W4A8C8, 206.6-220.5 ms vs. 5941.5-5941.8 ms (~27-29x speedup). Under an 84-GB NPU-memory cap, the best DeepSeek GSM8K configuration reaches 21.5 ms TPOT using approximate expert substitution plus multi-token prediction (MTP) at depth 2. Evaluated across 5 workloads per model.

Key Findings:

Coordinating four cache/transfer mechanisms yields order-of-magnitude TPOT wins over naive prefetching (38-45 ms vs. ~1269 ms on DeepSeek-V4-Flash-W4A8)
Allowing approximate expert substitution for low-score router misses, combined with multi-token prediction at depth 2, drives TPOT down to 21.5 ms on GSM8K under an 84 GB NPU cap
Alternating-buffer prefill staging decouples transfer latency from the inference critical path, and nonuniform per-layer cache allocation outperforms uniform allocation

Technical Novelty: The coordinated combination of four mechanisms rather than any single trick: (1) alternating-buffer staging of complete expert layers during prefill, (2) nonuniform per-layer cache capacity, (3) domain-informed cache initialization, (4) routing-history-aware replacement with cross-layer prefetching — plus an optional score-based substitution path that approximates low-score router misses instead of stalling on transfers. Prior offloading work (e.g., vLLM Prefetch) treats layers uniformly and does not exploit routing history or allow approximate substitution.

What's New: Prior offloading systems treat MoE experts as a uniform cache and stall on misses. MoE-CORE is novel in (a) treating router history as a prefetch signal with cross-layer lookahead, (b) allocating cache capacity nonuniformly across layers based on access patterns, and (c) introducing an optional score-gated substitution path that trades a bounded accuracy loss for eliminating miss stalls.

Extension Opportunities:

Extend coordinated residency policies to multi-NPU/multi-GPU distributed appliances with expert sharding across devices, exploring cross-device routing-history prediction
Replace the heuristic routing-history-aware replacement policy with a learned RL or small-transformer predictor that forecasts expert activation sequences conditioned on token embeddings
Investigate adaptive score-threshold substitution that trades off accuracy vs. latency dynamically per request type (e.g., stricter for math/code like GSM8K, looser for chat)

Replicability: Abstract states 'The code is here' (link not included in excerpt). Reproduction requires an NPU appliance with ~84 GB device memory and the specific quantized checkpoints (DeepSeek-V4-Flash-W4A8, GLM-5.2-W4A8C8); exact hardware (Ascend-class NPU implied by terminology) and host memory for offloaded experts are not specified in the abstract.

Research Gaps:

No reported accuracy/quality degradation numbers for the score-based substitution path — the latency-quality Pareto frontier is unquantified in the abstract
Comparison baseline uses asymmetric output caps (1K vs. 128 tokens), and only vLLM Prefetch is cited; comparisons against other recent MoE offloading systems (e.g., Fiddler, MoE-Infinity, Pre-gated MoE) are not mentioned

3. Spiking neural networks for streaming qubit readout

Authors: Barry M. Dillon, Aqib Javed, Jim Harkin... Published: 2026-10-01 | Citations: 0 arXiv | PDF

Research Question: How can qubit-state discrimination in superconducting quantum processors be made both high-accuracy and low-latency, while providing streaming (time-resolved) state estimates rather than only end-of-window classifications, so that readout can better capture transient nonidealities like crosstalk and relaxation events?

Summary: The paper introduces spiking neural network discriminators for superconducting qubit readout that process measurement traces in successive time chunks, enabling streaming time-resolved state estimates rather than end-of-window-only classification. The SNNs beat matched-filter discrimination, approach full-trace ANN accuracy, and — via quantization-aware training and hls4ml synthesis — fit per-chunk FPGA inference within the chunk arrival interval, pointing toward real-time, low-latency qubit readout for feedback and error correction.

Key Results: The authors demonstrate that spiking neural network (SNN) discriminators processing readout traces in successive time chunks outperform conventional matched-filter discrimination and approach the accuracy of full-trace ANNs on superconducting qubit readout. Using quantization-aware training and hls4ml synthesis, they show FPGA inference per chunk completes before the next readout chunk arrives, establishing feasibility for real-time streaming readout. Specific numeric accuracy/latency figures are not provided in the abstract.

Key Findings:

SNN discriminators outperform matched-filter readout and approach full-trace ANN accuracy on superconducting qubit traces
Chunk-wise processing yields a streaming, time-resolved qubit-state estimate that updates as the signal arrives, exposing transient effects like relaxation
Quantization-aware training plus hls4ml synthesis makes per-chunk FPGA inference latency smaller than the readout chunk interval, demonstrating hardware viability

Technical Novelty: First application of spiking neural networks to superconducting qubit-state discrimination with a streaming, chunk-by-chunk time-resolved classification output, combined with quantization-aware training and hls4ml synthesis to fit per-chunk inference inside the readout chunk interval on FPGA — unlike prior ANN-based readout that waits for the full trace, or matched filters that collapse the trace to a scalar.

What's New: Prior qubit readout work used matched filters (fast but ignoring temporal structure/crosstalk) or ANNs (accurate but operating on the full trace with end-of-window latency). This work is the first to bring SNNs to this problem, exploiting their inherently temporal, event-driven computation to deliver streaming classification compatible with FPGA real-time constraints.

Extension Opportunities:

Extend SNN discriminators to multi-qubit frequency-multiplexed readout where crosstalk is explicit, and benchmark joint-state classification vs. per-qubit matched filtering on a 5–10 qubit chip
Integrate the streaming SNN output into a closed-loop active-reset or mid-circuit measurement controller on FPGA, measuring end-to-end feedback latency and its impact on QEC cycle time
Apply the same quantization-aware + hls4ml pipeline to other streaming scientific-inference tasks (e.g., HEP trigger systems, neutrino detector pulse classification) to test transferability of the SNN+FPGA recipe

Replicability: The abstract does not mention a code or data release. Reproducing would require access to a superconducting qubit testbed (or published readout traces), a standard ML stack with hls4ml, and a target FPGA (e.g., Xilinx) for synthesis — all moderately accessible compute-wise, but the quantum hardware is the main barrier. Simulated readout traces could serve as a stand-in for the ML/FPGA portions.

Research Gaps:

No explicit numeric fidelity/latency figures or multi-qubit crosstalk benchmarks are summarized in the abstract, leaving the quantitative advantage over ANNs and matched filters unclear in absolute terms
The closed-loop utility — i.e., actually using the streaming estimate inside active reset, mid-circuit measurement, or QEC decoding pipelines — is implied but not demonstrated

4. CONFERM: Recurrence-Aware Temporal Mapping for Multi-Cycle Multi-Context CGRAs

Authors: Jun Yin, Jannes Willemen, Stef Cuyckens... Published: 2026-10-01 | Citations: 0 arXiv | PDF

Research Question: How can CGRA (Coarse-Grained Reconfigurable Array) mappers simultaneously handle loop-carried recurrences and multi-cycle compute nodes to maximize throughput, when existing mappers treat recurrence-aware scheduling and operator-level pipelining as separate concerns, limiting inter-iteration overlap and inflating routing pressure?

Summary: CONFERM is a recurrence-aware temporal mapper for multi-context CGRAs that unifies loop-carried recurrence scheduling with multi-cycle operator pipelining by letting the dominant temporal constraint shape the dataflow graph representation. By exploiting regular loop-carried offsets across interleaved iterations, it emits control sequences that repeat at a shorter cadence than the initiation interval, delivering 2.18x throughput improvement and 5.07x faster mapper convergence over state-of-the-art.

Key Results: Across ten benchmark kernels, CONFERM improves throughput by 2.18x over state-of-the-art CGRA mappers, shortens the emitted initiation interval by 46% through uniform iteration offsets, and converges 5.07x faster on average using the same heuristic mapper backend.

Key Findings:

Treating recurrence scheduling and operator pipelining jointly (rather than as separate passes) unlocks substantially more inter-iteration overlap on multi-context CGRAs
Uniform loop-carried offsets enable control sequences to repeat at a cadence shorter than the initiation interval, cutting emitted II by 46% and reducing configuration memory overhead
Guiding DFG representation by the dominant temporal constraint accelerates mapper convergence by 5.07x, suggesting bottleneck-prioritized scheduling is a strong heuristic for CGRA compilation

Technical Novelty: Unlike prior mappers that schedule a fixed DFG with recurrence-aware scheduling and operator-level pipelining as independent passes, CONFERM uses the dominant temporal constraint (longest recurrence vs. longest operator latency) to drive DFG representation itself, prioritizing bottleneck regions during scheduling and exploiting uniform loop-carried offsets so the emitted control sequence repeats at a shorter cadence than the original initiation interval.

What's New: Prior CGRA mappers schedule a fixed DFG and handle recurrence constraints and multi-cycle operator pipelining as orthogonal problems. CONFERM makes the DFG representation itself a function of the dominant temporal constraint, enabling joint optimization and exposing cross-iteration pipelining opportunities that fixed-DFG mappers cannot see.

Extension Opportunities:

Extend CONFERM to handle irregular/data-dependent control flow (e.g., conditional loops, early exits) where loop-carried offsets are not statically uniform
Integrate CONFERM's recurrence-aware DFG representation with ML compiler stacks (TVM, MLIR) to auto-tune CGRA mapping for transformer/attention kernels with mixed recurrence patterns
Explore co-design of CGRA microarchitecture (context memory depth, PE pipeline stages) tuned to CONFERM's shorter-cadence control sequences to further reduce configuration storage overhead

Replicability: The abstract does not mention a public code release or dataset. Reproduction would require implementing the mapper pass on top of an existing heuristic CGRA mapper backend and evaluating on the ten referenced DSP/ML kernels. Compute requirements are modest — CGRA mapping research is CPU-bound compiler work rather than GPU-intensive, so a workstation would suffice.

Research Gaps:

Existing CGRA mappers separate recurrence-aware scheduling from operator-level pipelining, missing joint optimization opportunities
Multi-context modulo scheduling has not been exploited for shorter-cadence control sequences that reduce both II and configuration storage pressure simultaneously

5. Is Your AI Fast Enough to Run a Fusion Reactor?

Authors: Nathaniel Chen, Andrew Rothstein, Ricardo Shousha... Published: 2026-10-01 | Citations: 0 arXiv | PDF

Research Question: How should ML practitioners select inference backends (CPU vs GPU) for neural network models embedded in real-time fusion reactor feedback control loops, where control-cycle latency budgets are strict and predictable timing is mission-critical?

Summary: The paper examines inference latency of neural networks deployed in DIII-D tokamak control loops, benchmarking ten fusion-relevant models across CPU and GPU backends. It finds that above ~5M parameters, CPU inference crosses from milliseconds into seconds, making GPU deployment essentially mandatory for larger models in real-time fusion control. The core argument is that backend selection must be co-designed with the model and the control-cycle budget, not treated as a downstream implementation detail.

Key Results: The authors built a benchmark covering ten neural networks and model components drawn from DIII-D tokamak fusion control and diagnostic pipelines, measuring inference latency across multiple backends. For models above 5M parameters, CPU backends take tens to thousands of milliseconds, while GPU inference is substantially faster — demonstrating a hard practical ceiling for CPU-oriented deployment in real-time fusion control.

Key Findings:

A 5M-parameter threshold exists beyond which CPU inference latency becomes incompatible with fusion control-cycle budgets, ranging from tens to thousands of milliseconds
GPU backends deliver substantially better latency and should be considered the default for non-trivial models in real-time fusion deployments
Deployment backend choice is not separable from model design — the control-cycle budget must drive both jointly, meaning ML model development for fusion cannot proceed CPU-first and hope to port later

Technical Novelty: First systematic benchmark targeted at fusion-control-specific neural networks (as opposed to generic ML benchmarks like MLPerf), coupled with codified lessons from actual DIII-D deployments. The novel framing is treating the inference backend as a first-class co-design variable alongside the model architecture when working against a hard control-cycle budget.

What's New: Prior fusion-ML work has focused on model accuracy and physics fidelity; this paper is unusual in treating inference-time systems engineering (backends, timing predictability, control-cycle budgets) as the central design constraint, backed by real deployment experience on an operating tokamak.

Extension Opportunities:

Add specialized accelerator backends (FPGAs, TPUs, or NVIDIA Jetson edge devices) to the benchmark since these are common in real-time physics control systems and have different latency/jitter profiles than general-purpose GPUs
Extend the benchmark to measure tail latency and timing jitter distributions (p99, p99.9) rather than just average inference time, since real-time control cares about worst-case predictability as much as mean speed
Build a model-selection advisor tool that, given a control-cycle budget (e.g., 1ms, 10ms) and available hardware, recommends which model architectures/sizes are feasible — turning the paper's measurements into actionable deployment guidance

Replicability: The abstract does not explicitly mention code/data release, but the paper is framed as 'a benchmark,' which strongly implies an open benchmark harness is intended. Reproduction would require access to the ten representative fusion models (or equivalents), standard CPU/GPU hardware, and ideally DIII-D-adjacent diagnostic data; the compute itself is modest (single-node inference benchmarking).

Research Gaps:

No treatment of timing jitter / worst-case latency bounds, which matter more than mean latency for real-time control stability
Limited coverage of alternative accelerators (FPGAs, custom ASICs, edge inference chips) that are relevant to in-vessel or near-plasma deployments where general-purpose GPUs may be impractical

⚡ ENERGY

1. Domain-engineered ferroelectric BiFeO$_3$ thin films for efficient bias-free THz emission

Authors: Z. Abdul Hadi, N. Rezi, M. Monti... Published: 2026-10-01 | Citations: 0 arXiv | PDF

Research Question: What is the microscopic origin of ultrafast photocurrent driving THz emission in ferroelectric BiFeO3, and can the ferroelectric domain configuration be used as a design parameter to enhance bias-free THz emitters?

Summary: The authors demonstrate that engineering ferroelectric domain patterns in epitaxial BiFeO3 thin films dramatically enhances bias-free THz emission, with stripe-domain films emitting 4× stronger fields than single-domain films and reaching half the output of a metallic spintronic reference. Using symmetry-resolved amplitude and phase measurements, they identify ultrafast photocarrier screening of ferroelectric polarization — enhanced by built-in fields at domain walls — as the dominant emission mechanism, resolving a long-standing debate about the microscopic origin of THz generation in ferroelectrics.

Key Results: In epitaxial BiFeO3 thin films, a periodic stripe-domain configuration emits a THz field 4× stronger than single-domain configurations, reaching ~50% of the output of a metallic spintronic reference emitter while remaining unsaturated at the highest excitation fluences. Polarization- and crystal-orientation-resolved amplitude/phase measurements separated coexisting photocurrents by symmetry, identifying ultrafast screening of ferroelectric polarization by photocarriers as the dominant mechanism for in-plane-polarized films (enhanced by built-in fields at domain walls), while out-of-plane polarization favors the bulk photovoltaic current.

Key Findings:

Stripe-domain BiFeO3 films emit a THz field 4× stronger than single-domain films and ~50% of a metallic spintronic reference emitter, without saturation at the highest fluences tested
Ultrafast screening of ferroelectric polarization by photocarriers is the dominant emission mechanism in in-plane-polarized films, boosted by strong built-in fields at domain walls
Out-of-plane polarization configurations instead favor the bulk photovoltaic current, showing that domain geometry selects between competing microscopic mechanisms

Technical Novelty: Combines deliberately engineered BiFeO3 domain architectures (single-domain vs. periodic stripe) with phase- and polarization-resolved THz emission spectroscopy to disentangle coexisting photocurrents by symmetry, directly attributing the dominant emission to ultrafast polarization screening at domain walls rather than the bulk photovoltaic effect — a mechanism assignment that previous ferroelectric THz emitter studies left unresolved.

What's New: Prior work established ferroelectrics as bias-free THz emitters but did not resolve which photocurrent (bulk photovoltaic, shift, screening) drives emission; this paper uses engineered domain configurations plus phase/polarization-resolved detection to both identify the mechanism and show that domain topology itself is a tunable design knob for emission strength.

Extension Opportunities:

Engineer finer or more complex domain patterns (e.g., sub-micron stripe periods, vortex domains, or 71°/109° wall configurations) via substrate miscut or lithographic templating to further boost domain-wall field enhancement and surpass the spintronic reference
Combine BiFeO3 stripe-domain films with heterostructure stacks (e.g., coupling to a spintronic bilayer or a plasmonic metasurface) to additively combine screening-current and spin-current THz emission mechanisms
Extend the symmetry-resolved amplitude/phase decomposition technique to other ferroelectrics (PbTiO3, HfO2-based, or 2D ferroelectrics) to map the universal relationship between domain topology and ultrafast photocurrent symmetry

Replicability: The abstract does not mention code or data availability. Reproduction requires: PLD/MBE growth of epitaxial BiFeO3 with controlled domain configurations on appropriate substrates (e.g., DyScO3, SrTiO3), ferroelectric domain characterization (PFM), and a femtosecond THz emission spectroscopy setup with electro-optic sampling and polarization control — specialized ultrafast condensed-matter lab equipment, not large compute.

Research Gaps:

No absolute quantification of the domain-wall built-in field or carrier dynamics is reported in the abstract — the microscopic model remains phenomenological at this stage
Comparison is limited to a single spintronic reference and does not benchmark against other ferroelectric or photoconductive emitters, nor explore the spectral bandwidth or temporal pulse shape in detail

2. Topo-Spectral Percolation Descriptors for Mechanistic Ion Transport Pathways from Static Crystal Structures

Authors: Diptendu Roy, Chiku Parida, Juan María García-Lastra... Published: 2026-10-01 | Citations: 0 arXiv | PDF

Research Question: How can we rapidly determine ion transport mechanisms (diffusion pathways, bottlenecks, dimensionality) from a single static crystal structure without expensive ab initio molecular dynamics (AIMD) simulations, while providing more than just a scalar transport number?

Summary: This paper introduces Topo-Spectral Percolation Descriptors (TSPD), a calibration-free method that extracts ion transport mechanisms from a single static crystal structure by analyzing the percolation topology and graph Laplacian spectrum of a physics-based migration network. Unlike surrogate models that output a single transport number, TSPD identifies pathways, bottleneck atoms, and transport dimensionality in seconds, validated on 8 cathode and electrolyte materials against AIMD and neutron diffraction.

Key Results: The authors validated TSPD against AIMD ion densities and neutron diffraction data across 8 materials spanning cathodes and solid electrolytes with 1D, 2D, and 3D transport. TSPD reproduced the established transport mechanism in every case (8/8) and correctly flagged structures where geometric openness does not translate to viable ion percolation. Runtime is on the order of seconds per structure vs. AIMD's prohibitive cost for screening.

Key Findings:

A barrier-weighted migration network's percolation thresholds and Laplacian spectrum together encode the transport mechanism of a crystalline ionic conductor without any system-specific calibration.
TSPD matched the established transport mechanism (1D/2D/3D) in all 8 benchmark materials and distinguished geometrically open structures that nonetheless fail to percolate ionically.
The method is most informative for pathway-limited, direction-dependent conductors and is complementary — not a replacement — for AIMD in strongly cooperative or near-isotropic systems.

Technical Novelty: The combination of (1) a periodic migration network constructed from a single static crystal with physics-based edge barrier energetics, (2) a barrier-threshold percolation analysis for pathway topology, and (3) spectral analysis of the barrier-weighted graph Laplacian — yielding a calibration-free, mechanism-level descriptor rather than a single scalar surrogate output typical of ML interatomic potentials or geometric void-radius descriptors.

What's New: Prior surrogate models reduce ion transport to a single scalar (e.g., predicted conductivity or activation energy), and geometric descriptors like free-sphere radius ignore energetic bottlenecks. TSPD is the first descriptor set to jointly expose pathway topology, dimensionality, and bottleneck identity from static structures using a physics-weighted graph without empirical calibration.

Extension Opportunities:

Integrate TSPD as a pre-screening filter in high-throughput materials discovery pipelines (e.g., Materials Project / OQMD) to triage candidate solid electrolytes before committing to AIMD/NEB calculations.
Extend the barrier-weighted graph Laplacian spectrum to predict quantitative conductivity (not just mechanism) by coupling it with temperature- and concentration-dependent kinetic Monte Carlo on the same migration network.
Adapt TSPD to disordered or amorphous electrolytes and glass-ceramic composites by generalizing the periodic migration network to local-cluster graphs built from MD snapshots or RMC-refined structures.

Replicability: The abstract does not mention a public code or data release. Reproduction would require: a crystal structure parser (pymatgen), a DFT/empirical engine for migration barrier estimates (NEB-like or bond-valence site energy), and standard graph/spectral libraries (networkx, scipy). Compute is modest — the method itself runs in seconds; validation against the 8 reference materials would require access to published AIMD trajectories or re-running AIMD (GPU-days per material).

Research Gaps:

No direct quantitative prediction of ionic conductivity or activation energy — TSPD is mechanistic rather than rate-predictive.
Limited applicability to strongly cooperative (concerted-hop, paddle-wheel) or nearly isotropic transport where static barriers poorly reflect dynamical correlations.

3. Programmable nonlinearity within nanophotonic waveguides

Authors: Benjamin A. Ash, Ryotatsu Yanagimoto, Mandar M. Sohoni... Published: 2026-10-01 | Citations: 0 arXiv | PDF

Research Question: How can we overcome the limitation that phase matching in nonlinear nanophotonic waveguides is fixed at fabrication time and vulnerable to fabrication imperfections, which restricts device reconfigurability and yield?

Summary: The paper introduces programmable χ^(2) nonlinearity in nanophotonic waveguides by projecting structured light onto photoconductive electrodes to create reconfigurable, electric-field-induced quasi-phase-matching patterns updatable in ~1 second. This enables post-fabrication control of spectral, modal, and polarization properties of three-wave mixing and closed-loop compensation of fabrication imperfections in waveguides over 10 cm long, while also enabling in-situ diagnostics like dispersion measurement and wavefront phase tomography.

Key Results: Demonstrated programmable χ^(2) holograms inside nanophotonic waveguides via structured optical illumination projected onto photoconductive electrodes, achieving reconfigurable quasi-phase matching updatable in ~1 second. Showed spectral, modal, and polarization control over three-wave mixing; used closed-loop in situ optimization to compensate phase mismatch in width-modulated and spiral waveguides exceeding 10 cm in length; and demonstrated on-chip dispersion measurements of transverse modes and wavefront phase tomography.

Key Findings:

Structured optical illumination on photoconductive electrodes can imprint spatially programmable χ^(2) holograms inside nanophotonic waveguides, enabling reconfigurable QPM in ~1 second.
Closed-loop in situ optimization recovers efficient nonlinear operation in challenging geometries including width-modulated and >10 cm spiral waveguides, mitigating fabrication-induced phase mismatch.
Programmable nonlinearity doubles as a diagnostic tool, enabling direct dispersion measurements across transverse modes and on-chip wavefront phase tomography.

Technical Novelty: Rather than poling with fixed-geometry electrodes at fabrication, the authors use structured light projected onto photoconductive electrodes to spatially pattern an electric-field-induced χ^(2) — effectively creating rewritable χ^(2) 'holograms' inside the waveguide. This decouples phase-matching design from fabrication and enables closed-loop optimization against actual device imperfections.

What's New: Prior QPM relied on permanent poling with fixed electrode geometries, locking in device function and leaving no recourse for fabrication variation. This work replaces static poling with light-defined, electrically-induced χ^(2) patterns, making the nonlinear response itself a software-defined, rewritable property of the chip.

Extension Opportunities:

Build cascaded multi-stage nonlinear photonic circuits (e.g., OPO networks, frequency comb synthesizers) that use the ~1s reconfiguration to retune each stage independently in situ.
Integrate the photoconductive-electrode QPM scheme with on-chip ML-based control loops for autonomous calibration of large waveguide arrays, enabling scalable yield recovery across wafers.
Extend programmable χ^(2) holograms to quantum photonics applications such as reconfigurable heralded single-photon sources, tunable squeezed-light generation, or programmable entangled-pair spectral shaping.

Replicability: No code/data availability mentioned in the abstract. Reproduction would require a nanophotonics fab-level stack: etched waveguides with photoconductive electrode layers, a structured-light projection system (SLM/DMD + imaging optics) aligned to the chip, tunable pump laser sources for three-wave mixing characterization, and feedback electronics/software for closed-loop optimization — i.e., a well-equipped integrated-photonics lab rather than compute.

Research Gaps:

Scalability and loss/thermal budgets of photoconductive electrode schemes in densely integrated cascaded nonlinear circuits are not addressed.
Long-term stability, drift, and endurance of the induced χ^(2) patterns under high optical powers (and in quantum/low-noise regimes) remain open.

4. Inertial Dynamics of a Skymeron

Authors: Mona Bhukta, Duc Minh Tran, Kilian Leutner... Published: 2026-10-01 | Citations: 0 arXiv | PDF

Research Question: Can layer-selective anisotropy engineering in synthetic antiferromagnets stabilize novel 3D composite spin textures with internal dynamical degrees of freedom beyond what intrinsic antiferromagnets (with identical sublattices) allow?

Summary: The paper introduces the 'skymeron,' a previously unobserved 3D composite spin texture in a synthetic antiferromagnet consisting of an out-of-plane skyrmion in one ferromagnetic layer coupled to an in-plane bimeron in the other, stabilized by layer-selective anisotropy tuning that drives a spin-flop into an orthogonal two-layer configuration. Time-resolved X-ray imaging reveals polarity-dependent inertial post-pulse motion followed by slower pinned-state return, with micromagnetic simulations tracing the inertia to finite-time internal reorientation relaxation.

Key Results: The authors demonstrated a spin-flop transition driven by tuning effective magnetic anisotropy difference between two ferromagnetic layers, producing a 'skymeron' — a 3D texture with an out-of-plane skyrmion in one layer coupled to an in-plane bimeron in the other. Element-specific X-ray microscopy resolved the texture layer-by-layer; time-resolved pump-probe X-ray microscopy captured polarity-dependent dynamics showing (a) short post-pulse inertia-like motion continuation and (b) slower return to the pinned initial state. Micromagnetic simulations attributed the inertia-like propagation to finite-time relaxation of skymeron reorientation. No specific benchmark numerics (velocities, timescales, anisotropy values) are quoted in the abstract.

Key Findings:

Tuning anisotropy difference between the two SAF layers drives a spin-flop that reorients one layer in-plane while the other remains out-of-plane, stabilizing a skyrmion–bimeron composite ('skymeron')
Element-specific X-ray microscopy resolves the composite texture layer-by-layer, directly confirming the orthogonal two-layer structure
Current pulses induce polarity-dependent dynamics with a short inertia-like post-pulse continuation and a slower return to the pinned state, originating from finite-time reorientation relaxation (an internal dynamical degree of freedom absent in intrinsic AFMs)

Technical Novelty: Prior work on synthetic antiferromagnet skyrmions assumed both layers remain collinear (both out-of-plane). This paper is the first to deliberately drive a spin-flop into an orthogonal two-layer configuration (one OOP, one in-plane) via layer-selective anisotropy engineering, producing a homotopic reconfiguration of an AFM skyrmion into a skyrmion–bimeron composite ('skymeron'), and the first to resolve its inertial, polarity-dependent current-driven dynamics layer-by-layer with element-specific X-ray microscopy.

What's New: First demonstration that synthetic antiferromagnets can host 3D composite spin textures (skyrmion + bimeron coupled across layers) via orthogonal spin-flop configurations — a regime inaccessible to intrinsic antiferromagnets with identical sublattices — and the first observation of inertial, polarity-dependent dynamics tied to an internal reorientation mode of such a composite texture.

Extension Opportunities:

Engineer multi-layer stacks (3+ layers) with graded anisotropy profiles to create higher-order composite textures (e.g., skymeron chains or hybrid hopfion-like 3D objects) with richer internal degrees of freedom
Exploit the polarity-dependent inertia for a racetrack-style memory or neuromorphic device where inertial continuation after a short current pulse enables lower-power, ballistic-like propagation between pinning sites
Combine skymeron dynamics with voltage-controlled magnetic anisotropy (VCMA) gating to electrically switch between skyrmion, bimeron, and skymeron states on demand for reconfigurable logic

Replicability: No code or data availability is mentioned in the abstract. Reproduction would require: (1) a thin-film deposition facility capable of synthetic antiferromagnet multilayer growth with precise per-layer anisotropy control (sputtering + interface engineering), (2) access to a synchrotron with STXM/XMCD for element-specific layer-resolved imaging and pump-probe timing (e.g., BESSY II, SLS, ALS), and (3) standard micromagnetic simulation infrastructure (MuMax3 or similar on a single GPU). The materials and synchrotron requirements make this a high-barrier experimental study; the simulation side is cheap.

Research Gaps:

Quantitative mapping of the anisotropy-difference phase diagram (where/when the skymeron is stable vs. collinear SAF skyrmion vs. in-plane bimeron) is not detailed in the abstract
Device-level performance metrics (velocity, pinning energy, write/read currents, thermal stability, scaling) needed for spintronic application are not provided

5. Robustness of the Verwey transition against remanent strain-induced defects in magnetite

Authors: M. A. Gala, K. J. Grzywa, K. Komędera... Published: 2026-10-01 | Citations: 0 arXiv | PDF

Research Question: How do remanent strain fields and extended line-like defects produced by uniaxial plastic deformation affect the Verwey transition in stoichiometric magnetite — a regime previously unexplored compared to the well-studied effects of point defects, doping, and oxygen non-stoichiometry?

Summary: The authors show that uniaxial-compression-induced extended defects and remanent strain fields in single-crystal magnetite — directly imaged by dark-field x-ray microscopy — leave the Verwey transition temperature and sharpness essentially unchanged, even after fracture. This cleanly separates the strong T_V suppression caused by point-like disorder from the near-invisibility of slip-related extended defects to the electronic ordering transition.

Key Results: Using dark-field x-ray microscopy (DFXM) and ac magnetic susceptibility on single-crystalline Fe3O4, the authors show that samples compressed along [011] to 200 MPa develop stable line-like defects and remanent strain fields, while loading past the fracture threshold produces denser defect networks and mechanical fracture. A separate sample was fractured by [001] compression. Despite clear real-space evidence of substantial structural disorder from DFXM lattice-orientation and residual strain mapping, neither T_V nor the sharpness of the Verwey transition shifted measurably in any compressed or fractured sample — in sharp contrast to the known strong suppression of T_V by point-like disorder.

Key Findings:

Compression to 200 MPa along [011] produces stable line-like defects and remanent strain fields visible in DFXM, while supra-fracture loading yields dense defect networks — yet T_V is unchanged in both
Fracture produced by [001] compression also leaves the ac-susceptibility signature of the Verwey transition intact, showing the result is not loading-geometry specific
Extended, slip-related defects behave qualitatively differently from point defects, which are known to strongly suppress T_V — identifying defect dimensionality as a key control variable

Technical Novelty: First combined use of dark-field x-ray microscopy (a synchrotron darkfield technique giving simultaneous real-space, orientation, and residual-strain mapping of bulk crystal grains) with ac susceptibility to isolate the effect of extended slip-related defects — as opposed to point defects — on the Verwey transition in stoichiometric magnetite.

What's New: Prior work established that point defects, doping, and off-stoichiometry suppress T_V, but the role of extended defects and residual strain from plastic deformation was largely unknown. This paper fills that gap and demonstrates a counterintuitive robustness, directly visualized at the sub-grain scale with DFXM.

Extension Opportunities:

Apply in-situ DFXM during cooling through T_V on strained samples to directly image how the monoclinic domain structure nucleates around line defects and residual strain fields
Extend to systematic combined loading-plus-irradiation experiments to test whether extended defects screen or amplify the T_V suppression caused by point defects when both are present
Repeat the protocol on doped magnetite (e.g., Zn-, Al-, or Ti-substituted Fe3O4) to see whether chemical disorder changes the robustness of T_V against extended defects, mapping a phase diagram of defect type vs transition sensitivity

Replicability: No code or data link is mentioned in the abstract. Reproduction requires access to a synchrotron DFXM beamline (e.g., ESRF ID06), stoichiometric single-crystal magnetite, a uniaxial compression rig calibrated to ~200 MPa and above fracture, and a low-temperature ac susceptometer spanning ~100–130 K. Non-trivial facility access, but computationally light.

Research Gaps:

Microscopic mechanism for why electronic ordering tolerates extended defects but not point defects remains unexplained — no theory offered
Only stoichiometric samples and two loading directions tested; interplay with chemical disorder, thermal cycling fatigue, or thin-film geometries is open

🏥 HEALTHCARE

1. Analysis of Quantized and Efficiently Adapted Protein Language Models

Authors: Ilan Yaniv Zeisler, Sebastian Clancy, Pouriya Bayat... Published: 2026-09-30 | Citations: 0 arXiv | PDF

Research Question: How do 4-bit quantization and parameter-efficient fine-tuning (QLoRA) affect the performance, internal representations, and generation quality of protein language models (PLMs) across prediction and generative tasks?

Summary: The paper systematically evaluates 4-bit quantization and QLoRA across a broad set of protein language models for both prediction and generation. It shows QLoRA recovers most full fine-tuning performance with ~90% memory savings and minimal early-layer representational drift, while 4-bit quantization of generative PLMs preserves structural metrics but shifts token-level output distributions in model-dependent ways.

Key Results: Across ESM-2, ESMC, ProtBERT, ProtT5, Ankh, Ankh3 and Profluent-E1, many model-task pairs retained >90% of full fine-tuning performance under QLoRA, with peak GPU memory savings approaching 90% on the largest models. Layer-wise probing showed QLoRA preserved early-layer representations while inducing task-specific adaptations in middle/late layers with smaller representational drift than full FT. For generation (ProLLaMA, ProtGPT2, ProGen2, ProteinGLM, ESM3), 4-bit quantization largely preserved predicted structural/sequence-level properties but token-level analysis exposed model-dependent shifts in autoregressive output distributions.

Key Findings:

QLoRA retains >90% of full fine-tuning performance on many protein prediction tasks with up to ~90% GPU memory reduction
Representationally, QLoRA preserves early-layer features and concentrates adaptation in middle/late layers, mirroring full FT with smaller drift
4-bit quantization of generative PLMs preserves sequence- and structure-level predictions but induces model-specific shifts in autoregressive token distributions, invisible to standard downstream metrics

Technical Novelty: First systematic cross-model study combining 4-bit quantization and LoRA adapters across seven encoder PLMs plus five generative PLMs, pairing standard downstream metrics with layer-wise representational analysis and token-level autoregressive distribution analysis — showing that aggregate sequence/structure metrics can mask quantization-induced distributional drift in generative models.

What's New: Unlike prior PLM efficiency studies that focus on one or two models or only prediction tasks, this work spans encoder and generative PLM families and complements task metrics with layer-wise and token-distribution analyses, surfacing quantization effects that aggregate metrics hide.

Extension Opportunities:

Benchmark newer quantization schemes (NF4 vs INT4 vs AWQ/GPTQ, 2-bit/3-bit) on the same PLM suite to map a Pareto frontier of memory vs. recovery per task family
Build a distributional-shift diagnostic toolkit (KL divergence over token distributions, perplexity drift, motif-level entropy) as a standard pre-deployment check for quantized generative PLMs
Investigate QLoRA + mixture-of-adapters routing across protein task families (binding, stability, localization) to see whether a single quantized base can host many task adapters without full FT fallback

Replicability: Abstract does not mention a code or data release. Reproduction would require loading each of the 12 public PLMs (ESM-2/3/C, ProtBERT, ProtT5, Ankh/Ankh3, Profluent-E1, ProLLaMA, ProtGPT2, ProGen2, ProteinGLM), running QLoRA fine-tuning on standard protein benchmarks, and generating sequences; feasible on a single modern GPU (24–80GB) per model thanks to 4-bit quantization, though the full sweep is multi-GPU-week scale.

Research Gaps:

Lack of standardized distributional-shift diagnostics for quantized generative PLMs beyond structure/sequence scorers
Unclear how training speed and power consumption scale with QLoRA across PLM architectures — abstract notes effects were 'varied' without a unified explanation

2. Fold'EM: Direct atomic structure inference from Cryo-EM particles

Authors: Advaith Maddipatla, Märt-Erik Mäeots, Marco Pegoraro... Published: 2026-10-01 | Citations: 0 arXiv | PDF

Research Question: Can atomic protein structures be inferred directly from a small number of cryo-EM particle images, bypassing the conventional two-stage pipeline of density map reconstruction followed by atomic model fitting — which is sample-hungry, costly, and degrades for heterogeneous or low-population conformational states?

Summary: Fold'EM is an inference-time framework that determines atomic protein structures directly from a small number of raw cryo-EM particle images by fusing protein-generative-model priors with the particle likelihood, skipping both density reconstruction and downstream model fitting. It works with known or jointly inferred particle orientations and can resolve multiple conformational states from mixed populations without per-state reconstructions.

Key Results: Fold'EM recovers accurate atomic structures from cryo-EM particle images across both synthetic and experimental datasets, in two regimes: (1) with known particle orientations, and (2) ab-initio where orientations are inferred jointly with structure. On heterogeneous datasets, it resolves distinct conformational states from mixed particle populations without separately reconstructing a density map per state. The abstract does not cite specific resolution numbers, particle counts, or benchmark comparisons.

Key Findings:

Direct image-to-atoms inference is feasible with far fewer particles than conventional reconstruction pipelines require
Ab-initio joint inference of structure and particle orientation succeeds without a pre-computed density map
Heterogeneous particle populations can be decomposed into distinct atomic conformational states in a single pass, avoiding per-state 3D classification + model building

Technical Novelty: Integrates protein generative-model priors directly at the particle-image likelihood level rather than post-hoc against a reconstructed density map, and jointly infers atomic coordinates with per-particle poses. This collapses the traditional reconstruction + model-building stack into a single inference-time procedure and natively handles conformational heterogeneity without per-state 3D classification.

What's New: Prior work injects sequence-based structural priors only after density reconstruction during model fitting. Fold'EM moves the prior upstream of reconstruction, making the generative model a first-class participant in the image likelihood and enabling structure determination in low-sample and heterogeneous regimes where density maps are poor or ill-defined.

Extension Opportunities:

Swap in newer / larger protein generative priors (e.g., AlphaFold3, Boltz, RFdiffusion) and quantify how prior strength trades off against required particle count
Extend to protein-ligand or protein-nucleic acid complexes, where current structure prediction priors are weakest and heterogeneity is most informative
Couple Fold'EM with on-the-fly microscope control to prioritize collection of particles that maximally disambiguate uncertain conformational states (active-learning cryo-EM)

Replicability: The abstract does not mention code, data, or compute availability. Reproduction would plausibly require GPU inference comparable to running a protein structure prediction model (AlphaFold-class) per particle batch, plus access to standard cryo-EM particle stacks (EMPIAR) — likely multi-GPU for ab-initio pose inference.

Research Gaps:

No reported quantitative comparison in the abstract against established cryo-EM pipelines (RELION, cryoSPARC, ModelAngelo) on resolution, accuracy, or runtime
Scope limited to proteins amenable to current structure prediction priors; behavior on ligands, nucleic acids, membrane complexes, and intrinsically disordered regions is unaddressed

3. RipplePLM: Structural and Property Decoupling for Protein Mutation Effect Generation

Authors: Liuzhenghao Lv, Yuyang Liu, Yuyang Gao... Published: 2026-10-01 | Citations: 0 arXiv | PDF

Research Question: How can protein-to-text generation systems produce mutation-effect descriptions that properly disentangle structural (local vs. distal residue context) and biochemical (property-change) evidence, rather than collapsing mutation information into undifferentiated representations?

Summary: RipplePLM is a mutation-aware protein-to-text framework that stops treating mutation context as a single blob. It uses Direct-Distal Cross-Attention over a residue-level perturbation field (split by predicted contact maps into local and distal pathways) plus a Property Latent Chain that supervises biochemical property changes inside the LLM, yielding large ROUGE-L gains on MutaDescribe and better expert-rated biological accuracy.

Key Results: On the MutaDescribe benchmark, RipplePLM beats mutation-specific baselines on both temporal and structural splits. Under a matched-backbone comparison, average structural-split ROUGE-L jumps from 22.23 to 35.65 (a ~60% relative gain). Expert evaluation reports a higher share of biologically accurate/relevant descriptions than the mutation-specific baseline, and additional ablations plus low-N fitness regression experiments corroborate the quality of the learned mutation-aware representations.

Key Findings:

Structurally decoupling mutation context into direct-contact and multi-hop distal pathways via contact-map-guided cross-attention substantially improves generation quality (structural-split ROUGE-L 22.23 → 35.65 at matched backbone).
Injecting latent property tokens for biochemical attributes like thermostability and optimal pH into the LLM hidden state measurably improves biological accuracy of generated mutation descriptions per expert evaluation.
The learned mutation-aware representations transfer: they help in low-N fitness regression, suggesting the decoupled representation captures functionally meaningful mutation signal, not just linguistic patterns.

Technical Novelty: Two coupled ideas: (1) Direct-Distal Cross-Attention (DDCA) that routes a residue-level Mutation Perturbation Field — derived from pretrained PLM embeddings — through two separate attention pathways defined by predicted contact maps (immediate contact neighborhood vs. multi-hop distal context), explicitly structurally decoupling mutation evidence; and (2) a Property Latent Chain (PLChain) that injects expert-guided biochemical property-change supervision as latent tokens inside the LLM hidden-state pathway. Prior protein-to-text systems encoded mutation context as a single undifferentiated vector and lacked explicit property-level supervision.

What's New: First work (per the abstract) to explicitly decompose mutation representations into structural (local vs. distal, driven by contact maps) and biochemical (property-change latent tokens) streams for text generation, rather than fusing them into one undifferentiated embedding — bridging mechanistic structural biology intuition with LLM-based protein captioning.

Extension Opportunities:

Extend DDCA from single-point mutations to multi-residue/epistatic variants by generalizing the Mutation Perturbation Field to multi-site perturbations and multi-hop graphs over several mutation anchors.
Swap the predicted contact maps for AlphaFold2/ESMFold 3D structures or AF3 complex structures so the direct/distal pathways operate on real geometric neighborhoods, including protein–ligand or protein–protein interfaces.
Broaden the Property Latent Chain beyond thermostability/pH to include catalytic efficiency (kcat/Km), binding affinity (ΔΔG), solubility, and expression, then test whether richer property supervision transfers to zero-shot fitness prediction on ProteinGym.

Replicability: Code is released at https://github.com/Lyu6PosHao/RipplePLM. The MutaDescribe benchmark is public. Compute requirements aren't specified in the abstract, but the recipe (pretrained PLM + predicted contact map + LLM backbone with cross-attention and latent tokens) is standard for a mid-sized multimodal fine-tune — plausibly reproducible on a single 40–80GB GPU for inference and a small multi-GPU node for training, pending the repo's actual configs.

Research Gaps:

Reliance on predicted contact maps rather than full 3D structure means the 'distal' pathway inherits contact-prediction errors and ignores side-chain geometry and ligand/interface context.
Property supervision is limited to a small set of biochemical attributes (thermostability, pH); kinetic, binding, and dynamical properties — and multi-mutation epistasis — remain uncovered.

4. Benchmarking graph-based models for in-silico toxicity prediction in drug discovery

Authors: Noel Suarez-Barro, Manuel Lama, Juan C. Vidal Published: 2026-09-29 | Citations: 0 arXiv | PDF

Research Question: How can graph deep learning models for molecular toxicity prediction be fairly compared, given that current literature suffers from inconsistent datasets, preprocessing pipelines, and evaluation protocols that make performance claims difficult to interpret?

Summary: The paper introduces BenchTox, an open-source standardized benchmarking framework that evaluates 20+ graph deep learning models for molecular toxicity prediction under consistent datasets, preprocessing, and evaluation protocols. By combining this empirical benchmark with a structured literature analysis, the authors provide a reproducible assessment of the GDL-for-toxicity field that corrects for the inconsistent comparisons currently dominating the literature.

Key Results: The authors built a unified benchmarking framework (BenchTox) and systematically evaluated 20+ representative graph deep learning approaches under consistent experimental conditions across multiple toxicity datasets and partitioning strategies. The empirical study is complemented by a structured literature analysis to contextualize methodological trends. Specific numeric results (AUC/accuracy per model/dataset) are not stated in the abstract.

Key Findings:

Current literature comparisons of GDL toxicity models are unreliable due to inconsistent datasets, preprocessing, and evaluation protocols
Over 20 representative graph-based models can be fairly compared when run under a unified framework across multiple datasets and partitioning strategies
A structured literature review alongside empirical benchmarking reveals discrepancies between reported performance claims and standardized measurements

Technical Novelty: The novelty lies not in a new model but in a standardized, reproducible benchmarking infrastructure (BenchTox) that fixes dataset preprocessing, splits, and evaluation protocols across 20+ GDL models — addressing a methodological gap rather than an algorithmic one. Combined with a structured literature review, this exposes which reported gains are real vs. artifacts of inconsistent evaluation.

What's New: Prior toxicity-prediction papers propose individual models and self-report gains on inconsistently preprocessed datasets. This work is the first (per the authors) to apply a unified, reproducible benchmarking methodology — in the spirit of MoleculeNet or OGB but focused specifically on toxicity endpoints with GDL — to over 20 approaches with multiple partitioning strategies.

Extension Opportunities:

Integrate 3D-aware or conformer-based GNN architectures (e.g., equivariant networks like SchNet, DimeNet, SE(3)-Transformers) into the benchmark to test whether geometric priors improve toxicity prediction over 2D graph approaches
Extend the framework with scaffold-based and time-split partitioning to measure out-of-distribution generalization, which is more realistic for drug discovery than random splits
Add uncertainty quantification and calibration metrics (e.g., conformal prediction, Bayesian GNNs) to the benchmark, since toxicity decisions in drug prioritization require confidence estimates, not just point predictions

Replicability: Code is open-source at https://gitlab.citius.gal/noel.suarez/benchtox. Toxicity datasets (likely Tox21, ToxCast, ClinTox, SIDER from MoleculeNet) are public. Reproducing the full benchmark requires moderate GPU compute — a single modern GPU should suffice for individual model runs, though exhaustive evaluation across 20+ models × multiple datasets × partition strategies likely needs a cluster or multi-day single-GPU runs.

Research Gaps:

No standardized evaluation infrastructure existed for GDL-based toxicity prediction, making it unclear which architectural advances translate to real gains
Limited analysis of how dataset partitioning strategy (random vs. scaffold vs. time-based) affects reported model performance in toxicity tasks

5. Can LLMs Reliably Annotate Bioassay Metadata to Improve Data Readiness?

Authors: Laura van Weesep, Riccardo Tedoldi, Jens Sjölund... Published: 2026-10-01 | Citations: 0 arXiv | PDF

Research Question: Can LLMs reliably predict and audit missing BioAssay Ontology (BAO) metadata annotations (assay format, detection technology) directly from assay text descriptions, to address the critical sparsity of structured metadata in public bioassay repositories?

Summary: The paper documents a severe metadata annotation gap in PubChem (>99.9% of ~2M bioassays lack BAO-mapped terms) and demonstrates that both open-source and proprietary LLMs can reliably predict assay format and detection technology from raw assay text, with recall ≥0.96 on common classes. Beyond annotation, LLMs function as effective auditors: expert curators revised their own labels after seeing LLM-generated evidence, suggesting a path toward scalable, semi-automated metadata curation for molecular property prediction pipelines.

Key Results: Quantified annotation gaps across ~2M PubChem bioassays: 36% lack assay format, 89% lack BioAssay type, and >99.9% lack any BAO-mapped assay format or detection technology term. Evaluated 7 open-source and proprietary LLMs against silver labels from PubChem and ChEMBL, achieving recall ≥0.96 for biochemical and cell-based assay formats and similar performance for detection technology. Manual inspection revealed many LLM-silver label disagreements stem from inconsistencies between silver sources rather than LLM errors. In a qualitative study, LLM-generated evidence caused a senior industrial curator to revise some of their own labels. Performance differences between proprietary and open-source models were small.

Key Findings:

PubChem metadata is critically sparse: 36% missing assay format, 89% missing BioAssay type, >99.9% missing BAO-mapped format or detection technology
LLMs achieve ≥0.96 recall on biochemical and cell-based assay formats; disagreements concentrate in under-represented classes
Proprietary vs open-source LLM performance gap is small, enabling cost-effective open-source deployment
Many LLM-silver label disagreements trace to silver source inconsistencies, not LLM errors — LLMs can flag mislabeled assays and even prompt expert curators to revise their own labels

Technical Novelty: First systematic quantification of BAO annotation sparsity across all of PubChem combined with a head-to-head LLM evaluation (7 models, open-source vs proprietary) specifically for bioassay metadata curation. Novel use of LLMs as auditors — not just annotators — where disagreement with silver labels is treated as a signal of potential mislabeling rather than LLM error, validated through expert curator review.

What's New: Prior work on bioassay curation has focused on rule-based or ontology-matching approaches; this paper is distinctive in treating LLMs as both scalable annotators AND auditors of silver labels, quantifying the full-scale BAO annotation gap, and validating LLM utility through a qualitative study with a senior industrial curator who revised their own labels.

Extension Opportunities:

Build an automated BAO annotation pipeline that bulk-fills missing metadata for the >99.9% of PubChem bioassays lacking BAO-mapped terms, with confidence scoring and human-in-the-loop review for low-confidence predictions
Extend the LLM auditing approach to detect and flag mislabeled entries in industrial screening databases (e.g., ChEMBL) by surfacing LLM-silver disagreements as curation candidates — essentially a 'metadata diff' tool for curators
Train or fine-tune a specialized small model on LLM-generated annotations plus curator corrections to produce a cheap, deployable per-class classifier with calibrated reliability estimates for downstream ML foundation model training

Replicability: Abstract does not explicitly mention code/data release. Evaluation sets are derived from publicly available PubChem and ChEMBL, so dataset reconstruction is feasible. Reproduction requires inference-only compute for 7 LLMs (both open-source and proprietary API access) over evaluation subsets — modest compute if evaluation sets are small, but full-scale PubChem annotation (~2M assays) would need meaningful API/GPU budget.

Research Gaps:

Per-class reliability estimates and calibration are still needed before LLM annotations can safely feed downstream ML pipelines, especially for under-represented assay classes
No clear automated protocol for integrating LLM annotations into curation workflows at scale — the human-review bottleneck and governance for silver-vs-LLM disagreements remain open

🔬 MATERIALS

1. Selective suppression of electronic orders via interlayer coupling in superconducting bilayer nickelate thin films

Authors: Ziao Han, Lifen Xiang, Tianren Wang... Published: 2026-10-01 | Citations: 0 arXiv | PDF

Research Question: How can the robust spin-density-wave (SDW) order in bilayer nickelates be selectively suppressed to understand its interplay with superconductivity, given that conventional doping and oxygen annealing fail to weaken it (unlike in cuprates)?

Summary: The paper demonstrates that combining oxygen annealing with isovalent rare-earth A-site substitution acts as effective c-axis uniaxial pressure in bilayer nickelate thin films, suppressing T_SDW from 150 K to 70 K and completely eliminating a secondary charge-like anisotropy order while realizing superconductivity. XAS and electronic structure calculations attribute this to strengthened interlayer coupling reconstructing the Fermi surface, supporting a spin-spinless stripe ground state.

Key Results: By combining oxygen annealing with isovalent rare-earth A-site substitution to mimic c-axis uniaxial pressure, the authors fabricated superconducting bilayer nickelate thin films where T_SDW was suppressed from 150 K to 70 K. A secondary charge-like anisotropy order was completely eliminated in the superconducting state while SDW persisted in weakened form. Polarization-resolved O K-edge XAS plus electronic structure calculations confirmed that strengthened interlayer coupling reconstructs the Fermi surface, consistent with a spin-spinless stripe ground state.

Key Findings:

T_SDW suppressed from 150 K to 70 K via combined annealing + A-site substitution, which emulates c-axis uniaxial pressure
A secondary charge-like anisotropy order is completely eliminated in the superconducting state, while SDW persists in weakened form — indicating selective rather than uniform suppression of electronic orders
Polarization-resolved O K-edge XAS and DFT calculations show strengthened interlayer coupling reconstructs the Fermi surface, consistent with a spin-spinless stripe ground state

Technical Novelty: The combination of oxygen annealing with isovalent rare-earth substitution to emulate c-axis uniaxial pressure in thin films — a chemical analog to the hydrostatic pressure experiments that originally revealed SC in bulk La3Ni2O7 — enabling access to the superconducting state at ambient pressure while isolating and selectively eliminating the charge-like anisotropy order from the SDW.

What's New: Unlike cuprates where doping uniformly suppresses magnetic order, this work shows nickelates host two distinct intertwined orders that can be decoupled: interlayer coupling (via chemical pressure) selectively eliminates the charge-like anisotropy order while only partially weakening the SDW. It also provides the first thin-film realization of superconducting bilayer nickelates without applied hydrostatic pressure via this chemical-pressure route.

Extension Opportunities:

Systematically map the A-site substitution phase diagram (varying rare-earth ionic radius) against T_SDW and T_c to identify the optimal interlayer coupling strength for maximizing superconductivity
Apply the same chemical-pressure-plus-annealing protocol to other bilayer/trilayer nickelate families (e.g., La4Ni3O10) to test whether selective suppression of charge-like order is a universal route to SC
Use ultrafast pump-probe or RIXS on these films to directly probe the dynamics of the spin-spinless stripe ground state and clarify whether the charge-like anisotropy order competes with or is distinct from SDW

Replicability: No code/data availability is mentioned in the abstract. Reproduction requires PLD or MBE thin-film growth of rare-earth-substituted bilayer nickelates, high-pressure oxygen annealing, and access to synchrotron polarization-resolved O K-edge XAS plus DFT/DFT+U electronic structure capability — significant experimental infrastructure rather than compute.

Research Gaps:

The microscopic nature of the charge-like anisotropy order and its direct relationship to superconducting pairing remains unresolved
The spin-spinless stripe ground state is inferred rather than directly imaged — real-space probes (STM, neutron scattering on films) are still needed

2. Highly anisotropic collective modes of altermagnetic superconductors with Bogoliubov Fermi surfaces

Authors: Huaisong Zhao, Peng Zou, Xia-Ji Liu... Published: 2026-10-01 | Citations: 0 arXiv | PDF

Research Question: How do emergent Bogoliubov Fermi surfaces (BFSs) in altermagnetic superconductors with d-wave spin-split bands reshape collective-mode dynamics, and can spin response serve as a probe of these modes given broken time-reversal symmetry?

Summary: This theoretical paper shows that altermagnetic superconductors with d-wave spin-split bands host collective modes with pronounced momentum-space anisotropy, and that applying a magnetic field creates Bogoliubov Fermi surfaces which fundamentally reshape mode dynamics through anisotropic Landau damping. The broken time-reversal symmetry couples phase/amplitude modes to spin fluctuations, making the spin dynamical structure factor a novel probe of in-gap Higgs excitations.

Key Results: The authors demonstrate theoretically that in charge-neutral altermagnetic superconductors: (1) BCS collective modes inherit d-wave symmetry producing pronounced momentum-space anisotropy; (2) broken time-reversal symmetry couples collective modes to spin-density fluctuations, making them visible in the spin dynamical structure factor; (3) applying a magnetic field creates BFSs that introduce gapless particle-hole excitations and strongly anisotropic Landau damping, strongest along the momentum-space diagonal; (4) this diagonal damping suppresses the low-energy phonon mode at small momenta while simultaneously generating an in-gap structure in the Higgs mode with enhanced spin response signature.

Key Findings:

BCS collective modes in altermagnetic superconductors inherit d-wave symmetry with strong momentum-space anisotropy
Bogoliubov Fermi surfaces induced by magnetic fields produce anisotropic Landau damping that is strongest along the momentum-space diagonal, suppressing the low-energy phonon mode at small momenta
An in-gap Higgs mode structure emerges concurrent with phonon suppression, producing a pronounced enhancement in the spin response that serves as an experimental signature

Technical Novelty: First systematic treatment of collective modes (both phonon/phase and Higgs/amplitude) in altermagnetic superconductors hosting BFSs, uniquely identifying that broken time-reversal symmetry couples these modes to the spin channel — enabling spin dynamical structure factor as a probe where traditional density probes would be insensitive. Prior work on BFSs focused on thermodynamics and spectroscopy; prior altermagnet work treated normal-state properties, not superconducting collective dynamics.

What's New: Combines three active research threads — altermagnetism, Bogoliubov Fermi surfaces, and superconducting collective modes — to show that BFSs are a generic mechanism for highly anisotropic collective dynamics, and establishes spin response (rather than conventional density/optical response) as the sensitive probe owing to broken time-reversal symmetry mixing amplitude/phase modes with spin.

Extension Opportunities:

Extend the analysis from charge-neutral to charged superconductors to include long-range Coulomb interactions, converting the phonon mode to a plasmon and examining how BFSs modify the plasmon dispersion in realistic solid-state altermagnets
Perform numerical simulations with ultracold Fermi gas platforms (given the cond-mat.quant-gas tag) to experimentally observe the anisotropic Landau damping and in-gap Higgs mode via Bragg spectroscopy of the spin structure factor
Generalize to other altermagnet symmetries (g-wave, i-wave) or combine with spin-orbit coupling to map how different nodal structures of BFSs modify collective-mode anisotropy patterns

Replicability: No mention of code/data availability in the abstract. The work is theoretical (RPA/mean-field BdG framework implied for collective modes); reproduction requires modest compute — standard momentum-space diagonalization and susceptibility calculations on a workstation, likely doable in days with Python/Julia implementations of BdG + bubble diagrams.

Research Gaps:

No treatment of long-range Coulomb interactions (charge-neutral assumption limits direct applicability to real solid-state altermagnets where the phase mode becomes a gapped plasmon)
No concrete experimental proposal or material candidate identified for observing the predicted in-gap Higgs signature in the spin channel

3. Benchmarking average atom potentials derived from atomic cluster expansions

Authors: Deepak Somani, Lorenzo Piersante, Anirudh Raju Natarajan Published: 2026-10-01 | Citations: 0 arXiv | PDF

Research Question: Can an exact average-atom (A-atom) potential derived directly from a linear atomic cluster expansion (ACE) accurately describe chemically disordered alloys using small training datasets, and which chemical site basis works best?

Summary: The paper benchmarks a new formalism for deriving exact average-atom potentials directly from linear atomic cluster expansions (ACE), bypassing the traditional two-step process of fitting an interatomic potential then averaging. Using Fe-W (vs EAM reference) and Mo-Nb (vs DFT reference with only ~1500 small training structures), the A-atom potentials reproduce elastic constants, lattice parameters, mixing enthalpies, Bain paths, and — with the occupation basis — surface energies, demonstrating that disordered alloy properties can be recovered from small datasets when the chemical site basis is chosen carefully.

Key Results: Benchmarked on two systems: (1) Fe-W: linear ACE fit to EAM-generated data produced an A-atom potential that reproduces disordered-phase properties from both explicit random supercells and conventionally averaged EAM potentials. (2) Mo-Nb: ACE fit to DFT energies/forces from ~1500 small structures (avg 7 atoms each) reproduced DFT elastic constants, lattice parameter, mixing enthalpy, and Bain path of SQS. Among three chemical site bases, only the occupation basis also reproduced DFT surface energies. The derived A-atom potential predicts ideal solution entropy outweighs destabilizing vibrational contributions to finite-temperature free energy in disordered Mo-Nb.

Key Findings:

An A-atom potential derived from a linear ACE fit to ~1500 small Mo-Nb DFT structures reproduces elastic constants, lattice parameter, mixing enthalpy, and SQS Bain path
Of three chemical site bases tested, only the occupation basis also reproduces DFT surface energies — basis choice is critical for transferability to broken-symmetry environments
For Mo-Nb, the A-atom potential predicts ideal solution entropy outweighs destabilizing vibrational contributions, stabilizing the disordered phase at finite temperature
Small training datasets suffice when the ACE+A-atom formalism is used correctly, lowering the data burden for disordered alloy modeling

Technical Novelty: Prior A-atom potentials were derived by averaging existing interatomic potentials, inheriting their errors and requiring a well-trained parent potential. This work benchmarks a recently developed formalism that computes the exact A-atom potential analytically and directly from a linear ACE's basis functions, enabling direct fitting from small DFT datasets without needing a converged many-body parent potential first. Novel comparison of three chemical site basis choices (including occupation basis) and demonstration that basis choice critically affects transferability to surfaces.

What's New: First systematic benchmark of the exact-from-ACE A-atom derivation against both EAM (controlled test) and DFT (real test) references, and the first direct comparison of chemical site basis choices showing that occupation basis uniquely preserves surface-energy accuracy. Decouples A-atom quality from parent-potential quality by fitting directly to reference data.

Extension Opportunities:

Extend the formalism to ternary and higher-order alloys (e.g., refractory high-entropy alloys like Mo-Nb-Ta-W) where A-atom potentials would be most impactful for screening compositions
Couple the A-atom ACE potential to CALPHAD-style free energy models or Monte Carlo simulations to predict order-disorder transitions and phase diagrams directly from small DFT datasets
Investigate why only the occupation basis reproduces surface energies by systematically analyzing how different chemical site basis choices affect transferability to broken-symmetry environments (grain boundaries, interfaces, defects)

Replicability: The abstract does not mention code or data availability. Reproduction would require: DFT calculations on ~1500 small Mo-Nb structures (~7 atoms each, feasible on a modest HPC cluster in days), ACE fitting infrastructure (e.g., pacemaker, ACE.jl, or similar open-source ACE libraries), and implementation of the A-atom derivation formalism. Compute is modest by DFT standards — tractable for a single researcher with cluster access.

Research Gaps:

Limited to binary alloys (Fe-W, Mo-Nb) — behavior in multicomponent/HEA systems with competing chemistries is untested
Only linear ACE is examined; nonlinear ACE or other many-body expansions may yield different basis-choice sensitivities and are not benchmarked
Finite-temperature predictions rely on harmonic/ideal-solution approximations; anharmonic and short-range-order effects remain outside scope

4. Enhanced Critical Currents and Irreversibility Fields in YBa$_2$Cu$_4$O$_8$ Films through Ca-Substitution

Authors: Jiangteng Liu, Shuhei Funaki, Yuki Ogimoto... Published: 2026-09-28 | Citations: 0 arXiv | PDF

Research Question: Can chemical substitution (Ca for Y) optimize hole doping in intrinsically underdoped YBa2Cu4O8 (Y124) superconducting films to simultaneously enhance Tc, critical current density Jc, and irreversibility field, without disrupting the dominant vortex pinning mechanism?

Summary: The authors show that Ca substitution for Y in epitaxial YBa2Cu4O8 films acts as a chemical doping knob that raises hole density toward optimal, improving Tc and dramatically enhancing Jc (up to 10× at 6 T, 40 K) and the irreversibility field, while leaving the dominant pinning by Y-125 planar stacking faults unchanged. They further benchmark Jc and vortex creep against the fundamental limits Jd and Gi, revealing systematic correlations that generalize across superconductor families.

Key Results: In epitaxial (Y1-xCax)Ba2Cu4O8 films with x ≤ 0.10, Ca substitution increased hole concentration and drove the system toward optimal doping. At 40 K, 10% Ca doping boosted Jc by 2.4× at 0.03 T and 10× at 6 T versus undoped films. Jc follows a Jc ∝ B^(-0.5) field dependence independent of Ca content, consistent with pinning by planar Y-125 stacking-fault intergrowths (confirmed by electron microscopy). Thermally activated creep slowed at T ≳ 15 K with Ca doping; at lower T, creep rate S showed an unusual decrease with increasing field. Positive correlations observed between Jc and depairing current Jd, and between S and Ginzburg number Gi.

Key Findings:

10% Ca substitution yields Jc enhancements of 2.4× at 0.03 T and 10× at 6 T at 40 K, with simultaneous Tc and irreversibility-field increases.
Jc ∝ B^(-0.5) across all Ca contents, indicating planar Y-125 stacking-fault intergrowths (seen in TEM) remain the dominant pinning centers.
Ca doping slows vortex creep above ~15 K, but at lower T the creep rate S decreases with increasing field — an unusual behavior for cuprates.
Positive correlations hold between Jc and the depairing current Jd, and between creep rate S and the Ginzburg number Gi, consistent with trends spanning diverse superconductors.

Technical Novelty: First systematic demonstration that aliovalent Ca-for-Y substitution can be used to tune carrier density in epitaxial Y124 thin films (rather than bulk/polycrystals), combined with the finding that this chemical tuning enhances Jc and shifts the irreversibility line while leaving the planar stacking-fault pinning mechanism intact. Also introduces cross-material correlation analysis linking measured Jc and creep S to the fundamental bounds Jd and Gi.

What's New: Unlike oxygen-stoichiometry tuning (which is largely locked in the stable 124 phase), this work uses cation (Ca) substitution to overcome the intrinsic underdoping of Y124, and does so in epitaxial films rather than bulk. The combined demonstration that doping shifts Jc and the irreversibility field while preserving a specific defect-based pinning mechanism, plus the connection to universal Jd/Gi bounds, distinguishes it from prior single-material optimization studies.

Extension Opportunities:

Explore higher Ca concentrations (x > 0.10) or co-dopants to push past optimal doping and map the full dome of Jc/Tc versus hole concentration in Y124.
Engineer artificial Y-125 stacking-fault densities (via growth-rate or stoichiometry control) to deliberately tune the planar-defect pinning landscape and test whether Jc(B) scaling can be shifted away from B^(-0.5).
Investigate the anomalous low-T, high-field decrease of creep rate S using systematic magnetization relaxation measurements and vortex-glass/collective-creep models to identify the crossover mechanism.

Replicability: No code/data availability is indicated in the abstract. Reproduction would require a pulsed-laser-deposition or similar epitaxial oxide thin-film growth facility, high-pressure oxygen or appropriate O2 processing for the 124 phase, SQUID/VSM magnetometry for Jc and creep, and TEM for stacking-fault imaging — i.e., a well-equipped experimental superconductivity lab, not computational compute.

Research Gaps:

The microscopic origin of the anomalous low-temperature decrease of creep rate S with increasing field is not resolved.
The study is limited to x ≤ 0.10; the overdoped regime and the full Jc/Tc dome versus Ca content remain unexplored, as does co-engineering of pinning-site density alongside doping.

5. Electronic structure and two-orbital model of the quadlayer La$_5$Ni$4$O${13}$

Authors: Jian-Xiang Sun, Haokan Xiao, Cui-Qun Chen... Published: 2026-09-30 | Citations: 0 arXiv | PDF

Research Question: How does strain (isotropic and c-axis uniaxial) modify the electronic structure, Fermi surface topology, and spin-fluctuation landscape of the quadlayer Ruddlesden-Popper nickelate La5Ni4O13, and could it provide a pathway to superconductivity analogous to bilayer/trilayer nickelates?

Summary: The authors perform DFT and RPA calculations on the quadlayer Ruddlesden-Popper nickelate La5Ni4O13 under ambient pressure, 5% isotropic compressive strain, and 4% c-axis uniaxial strain, and construct a Wannier-downfolded two-orbital (eg) tight-binding model that captures the low-energy physics. They show that c-axis compression selectively lifts the dz2 bonding band to create a new γ hole pocket and enhances a (π,π) spin fluctuation, suggesting uniaxial strain as a route to superconductivity in n=4 nickelates analogous to the n=2 and n=3 members.

Key Results: Using DFT + Wannier downfolding + RPA on La5Ni4O13: (1) 5% isotropic compressive strain broadens Ni-eg bands and drives O-p → Ni-d charge transfer; (2) 4% c-axis uniaxial strain selectively pushes the dz2 bonding band upward, generating a new γ hole pocket with dz2 character while leaving dx2-y2 nearly unchanged; (3) Fermi surface at ambient and isotropic strain contains two electron pockets (α, δ) plus three hole pockets (β, β', β''); (4) leading RPA spin susceptibility peak shifts from q ≈ (2π/3, 2π/3) at ambient pressure to q ≈ (π, π) under both strain conditions, with enhanced magnitude under c-axis compression; (5) a quadlayer two-orbital tight-binding model reproduces the low-energy eg manifold.

Key Findings:

Isotropic 5% strain broadens Ni-eg bandwidth and transfers charge from O-p to Ni-d, but preserves the ambient-pressure Fermi surface topology (α, δ electron pockets; β, β', β'' hole pockets)
4% c-axis uniaxial strain selectively shifts the dz2 bonding band upward with little change to dx2-y2, producing a new γ hole pocket of dz2 character — the signature associated with bilayer/trilayer nickelate superconductivity
RPA leading spin susceptibility shifts from q ≈ (2π/3, 2π/3) at ambient pressure to q ≈ (π,π) under both strain conditions, and is enhanced under c-axis compression, indicating a favorable pairing environment

Technical Novelty: First Wannier-downfolded two-orbital effective model for the quadlayer (n=4) RP nickelate La5Ni4O13, plus the first systematic DFT+RPA comparison of ambient vs isotropic vs c-axis uniaxial strain showing that uniaxial c-axis compression selectively creates a dz2 γ pocket and shifts the dominant spin fluctuation to (π,π) — mirroring the mechanism believed to drive superconductivity in bilayer La3Ni2O7 and trilayer La4Ni3O10.

What's New: Extends the DFT+RPA + two-orbital Wannier framework — previously applied to bilayer La3Ni2O7 and trilayer La4Ni3O10 — to the quadlayer n=4 member, and distinguishes isotropic vs c-axis uniaxial strain effects, isolating c-axis compression as the specific tuning knob that reproduces the dz2-pocket + (π,π) spin fluctuation fingerprint of the known RP superconductors.

Extension Opportunities:

Perform FLEX or functional RG on the derived two-orbital quadlayer model to compute explicit superconducting gap symmetry (s± vs d-wave) and Tc estimates under each strain condition
Extend the DFT+RPA pipeline to arbitrary n-layer Lan+1NinO3n+1 series (n=5,6,...) to map how spin-fluctuation wavevector and dz2 pocket emergence scale with layer count
Add Hubbard U via DFT+DMFT on the Wannier model to capture correlation-driven mass enhancement and compare the Mott/Hund physics across the ambient, isotropic, and uniaxial strain regimes

Replicability: No code or data release mentioned in the abstract. Reproduction requires a standard DFT package (VASP/Quantum ESPRESSO/WIEN2k) with Wannier90 for downfolding and a custom RPA susceptibility code on the two-orbital model. Feasible on a modest HPC cluster (tens to low hundreds of CPU cores for the DFT structural relaxations and k-mesh convergence; RPA on the 8-band model is cheap).

Research Gaps:

Pairing symmetry and Tc estimates are not computed — only the normal-state spin susceptibility is reported, so the suggested superconductivity remains inferential
Strong electronic correlations (Hubbard U, Hund's J) are not treated beyond DFT+RPA; a DFT+DMFT or t-J treatment would be needed to confirm the two-orbital model's quantitative predictions

🔥 GitHub Trending

1. NandhaKishorM/laya

⭐ 30620 stars | Python

Non-autoregressive System 1 decision engine. Typed choice, score and yes/no decisions over any text in a single forward pass, in 100+ languages, with a router that picks the right checkpoint per reque

calibration classification decision-model huggingface jev modernbert

2. hypit-ai/hypit

⭐ 19182 stars | TypeScript

Clone any viral video with AI agents. Not just a script, the whole workflow: swap the face, the words, the B-roll, ship 100 variants in one command, and get your 100M views.

agentic-ai ai ai-agents ai-video compiler dsl

3. FareedKhan-dev/kimi-k3-in-c

⭐ 8875 stars | C

A 2.78-trillion-parameter Kimi K3 running inference on a single CPU in 8.24 GB of RAM. Portable C99: no BLAS, no framework, no GPU.

avx2 c99 cpu-inference deep-learning from-scratch inference-engine

4. deeplethe/utopia

⭐ 8057 stars | Rust

World's first open-source enterprise world model.

agent-memory bitemporal graphrag knowledge-base knowledge-graph llm

5. Human-Agent-Society/reef

⭐ 7485 stars | Python

Infrastructure for continually self‑improving agents

agent-infrastructure ai-agents continual-learning inference llm llm-training

6. jev-chat/jev-chat-jarvis

⭐ 7324 stars | Kotlin

装在手机上的对话副驾：在 QQ / X / 飞书里读懂对方、给出候选回复、一键填入输入框，发不发由你。非侵入，只读屏幕，不 hook 不改包。

accessibility-service ai-assistant android chat-assistant chat-copilot llm

7. drumih/turbo-fieldfare

⭐ 6858 stars | Swift

Gemma 4 26B-A4B inference in ~2 GB of RAM on any M-series MacBook

apple-silicon gemma gemma4 gemma4-26b-a4b gpgpu llm

8. mizorewww/laya-mlx

⭐ 6752 stars | Python

Native MLX runtime for Laya typed decision models — 7–14 ms short decisions on M3 Max. No text generation, PyTorch, or cloud API.

apple-silicon decision-model inference laya local-ai machine-learning

9. bojieli/ai-infra-book

⭐ 5860 stars | Python

《深入理解 AI Infra：量化分析与系统设计》（李博杰著）开源书稿：从硬件约束和模型架构出发，量化推导 LLM 推理与训练系统设计。含全书正文、PDF、配套计算工具与实验

accelerator ai-infra ai-infrastructure book datacenter-network deepseek

10. KKKKhazix/AIHOT

⭐ 5697 stars | TypeScript

一个自己找热点、自己写日报的网站框架。把信源和精选标准换成你的，它就是你的行业热点站。

ai chinese content-curation daily-digest docker-compose llm

11. yetone/magpie

⭐ 4740 stars | Go

Every agent's model. One place. Codex on DeepSeek, Claude Code on Kimi, from the menu bar.

claude-code codex deepseek gemini-cli llm macos

12. kirodotdev/KiroCrew

⭐ 4272 stars | Python

A persistent workspace for development work that self-improves and continues beyond one session.

agent agentic-ai agents ai-agent ai-agents automation

13. syv-ai/HyperQwen

⭐ 1850 stars | Python

Serve large Qwen models fast on the GPUs you actually own. Qwen3.8-27B on a single 24 GB card with vLLM: 127 tok/s single-user (381 when the answer quotes the prompt), ~1,035 tok/s at 64 concurrent, 1

consumer-gpu cuda inference-optimization kv-cache llm-inference local-llm

14. Tencent-Hunyuan/AuK

⭐ 1435 stars | Python

AuK: An Open-Source Foundational Model for Speech Generation and Editing

deep-learning foundation-models python pytorch separation speaker-extraction

15. datawhalechina/zero-to-sglang

⭐ 1373 stars | Python

Official SGLang x Datawhale course on LLM inference: understand inference, build a mini-sglang from scratch, then read the real SGLang source and land your first PR. Available in English and Chinese.

cuda datawhale inference-engine infra kv-cache llm

Generated by Research Pulse on 2026-10-04 08:09