2026-11-04·Daily digest

🔬 Research Pulse

Daily Digest

November 04, 2026

🤖 AI

🧠 LLMs

1. MDSKILL: Learning to Internalize Skill Memory through Multi-View Distillation for Efficient Dynamic Agent Self-Evolution

Authors: Dian Jin, Jinluan Yang, Kairong Han... Published: 2026-11-03 | Citations: 0 arXiv | PDF

Research Question: How can self-evolving RL agents efficiently internalize skill memories when self-generated memories contain noisy, redundant, entangled, or conflicting knowledge that makes direct on-policy self-distillation (OPSD) ineffective under limited training budgets?

Summary: MDSKILL addresses the noise and entanglement problem in self-generated skill memories by splitting them into three functional views — state grounding, action planning, and execution control — and distilling view-specific reasoning branches into a shared policy via on-policy self-distillation. It beats SOTA self-distillation/RL baselines by up to 7.1% on ALFWorld and 8.1% on WebShop without stronger teachers or added compute.

Key Results: MDSKILL improves upon SOTA by up to 7.1% on ALFWorld and 8.1% on WebShop (evaluated on three agent benchmarks) against RL, skill-augmented, and self-distillation baselines, without additional computational overhead and without requiring stronger external teacher models. Ablations validate contributions of each functional view and the dynamic skill-selection metrics.

Key Findings:

Functional decomposition of skill memories into three views reduces conflicting supervision and improves distillation stability
Dynamic quality metrics for skill selection outperform static filtering when constructing distillation branches
Multi-view OPSD enables continuous self-evolution without reliance on larger external teacher models

Technical Novelty: Decomposing skill memory into three functional views (state grounding, action planning, execution control) and running multi-branch on-policy self-distillation that consolidates these complementary reasoning branches into one shared policy, plus dynamic quality metrics for skill selection — all without an external teacher.

What's New: Prior self-distillation methods treat skill memory as a monolithic prompt source; MDSKILL is the first to decompose memory into functional views and distill them as complementary reasoning branches into a single shared policy.

Extension Opportunities:

Apply the three-view decomposition (state grounding, action planning, execution control) to multimodal/vision-language agents where grounding and planning are more cleanly separable
Replace the static three-view taxonomy with a learned, task-conditioned decomposition that discovers functional views from trajectory data
Extend to multi-agent settings where complementary skill memories across agents are distilled into a shared policy, or combine with retrieval-augmented skill libraries for lifelong learning

Replicability: Code is available at anonymous.4open.science/r/MDSKILL-CF53. Benchmarks (ALFWorld, WebShop) are open. Compute not stated, but OPSD on agent benchmarks of this scale is typically feasible on a small multi-GPU node (A100-class).

Research Gaps:

No comparison with methods that use external stronger teachers — unclear how much of the gap is closable without them
The three functional views are hand-designed; generalization to domains where this taxonomy doesn't fit (e.g., open-ended coding, scientific reasoning) is untested

🤖 Agents

1. GeoFaham: Domain-Specialized Multi-Agent Geospatial Analytics for Disaster Response

Authors: Muhammad Sohail Danish, Gilles Q. Hacheme, Akram Zaytar... Published: 2026-11-03 | Citations: 0 arXiv | PDF

Research Question: How can LLM-based coding agents reliably execute multi-step geospatial disaster-response workflows that require preserving spatial scope, temporal context, and intermediate state across heterogeneous sources (vector DBs, map features, imagery, rasters)?

Summary: GeoFaham is a multi-agent geospatial system that routes disaster-response queries to specialist agents (vector ops, map features, imagery discovery, raster analysis) under an Orchestrator with dual-ledger planning, using an external Response Store to keep large artifacts out of the LLM context. On a new 40-task benchmark grounded in real disasters, it outperforms strong general coding agents (0.940 vs 0.879) with 7.5× fewer tokens.

Key Results: GeoFaham scores 0.940 on GeoFaham-Bench (40 tasks grounded in 17 real-world disaster records) vs. 0.879 for the next-best general-purpose coding agent (Codex, Claude Code, OpenCode), while using 7.5× fewer tokens.

Key Findings:

Domain-specialized multi-agent decomposition beats general-purpose coding agents on multi-step geospatial tasks (0.940 vs 0.879)
Externalizing large geospatial artifacts via a Response Store (exchanging IDs + summaries) yields a 7.5× token reduction without sacrificing accuracy
Dual-ledger orchestration enables dependency-aware replanning that recovers from invalid geometric ops and lost spatial/temporal metadata — common LLM failure modes in this domain

Technical Novelty: A dual-ledger Orchestrator control loop (dependency-aware planning + progress/validation tracking with replanning) combined with a Response Store that persists full geospatial artifacts and typed metadata outside the LLM context, letting specialist agents exchange compact IDs instead of blowing context on large rasters/vectors.

What's New: Prior work treats geospatial analytics as either single-step retrieval or generic code generation; GeoFaham is the first to pair a geospatial-specialist agent taxonomy with artifact-centric state management and a disaster-grounded benchmark that evaluates non-rigid ground truth and recovery behavior.

Extension Opportunities:

Expand GeoFaham-Bench beyond 40 tasks/17 disasters to cover long-tail hazards (wildfire smoke plumes, compound events) and non-English geographies to test generalization
Add a specialist agent for time-series/change-detection over Sentinel-1/2 stacks, since the current decomposition emphasizes single-timestep raster analysis
Replace the Response Store with a typed artifact graph (e.g., STAC-backed) so agents can reason over provenance and lineage, not just opaque IDs

Replicability: Code and benchmark are public at github.com/microsoft/GeoFaham. Reproduction needs API access to frontier LLMs (for both GeoFaham and baselines like Codex/Claude Code) plus geospatial data sources (satellite imagery, map feature APIs); compute is modest since work is orchestration/API-bound, not training.

Research Gaps:

Benchmark is small (40 tasks, 17 disasters) and skewed toward retrospective records — real-time ingestion and streaming updates aren't evaluated
No ablation reported in the abstract isolating gains from the Response Store vs. the dual-ledger orchestrator vs. the specialist decomposition

2. PyRUA-DataFarm: LLM Agent as a Cost-Efficient Teleoperator for Real-Ready Robot Data

Authors: Kewei Zhang, Juncheng Ma, Guodong Lv... Published: 2026-11-03 | Citations: 0 arXiv | PDF

Research Question: Can an LLM coding agent replace human teleoperators to generate real-ready robot training data at a fraction of the cost, while matching or exceeding policy performance trained on real teleoperated demonstrations?

Summary: DataFarm uses an LLM coding agent to replace human teleoperators: given one calibration episode, it writes a simulation task, expert controller, and success judge, then renders rollouts as photo-real multi-view video with exact action labels. Policies trained only on this synthetic data match or beat teleoperation-trained policies on real robots (93 vs 84 successes) and dramatically outperform under visual distribution shift (87 vs 46), at roughly a third the per-hour cost of North American teleoperation labour.

Key Results: On five real-robot tasks, policies trained only on DataFarm-generated data achieved 93/100 successes vs 84/100 for teleoperation-trained policies. Under distribution shift (unseen tablecloth + coloured light), DataFarm policies retained 87/100 successes while teleoperation-trained policies dropped to 46/100. Cost: $21 one-time API per task + $5 render contract + $0.80/episode. One hour of three-view data costs $107 avg, vs $117 (China teleoperation quote) and $290 (North America labour alone).

Key Findings:

Agent-generated synthetic data matches teleoperation on in-distribution evaluation (93 vs 84 / 100)
Synthetic data generalizes far better under visual perturbation (87 vs 46 / 100), suggesting diversity benefits over narrow human demos
Cost structure is dominated by one-time task setup ($21 API), making per-episode marginal cost ($0.80) nearly free vs human labour

Technical Novelty: The full pipeline integration is novel: an LLM agent that writes simulation scene code, expert controller, and success judge from a single calibration episode, combined with a real-image-conditioned video model that renders multi-view consistent photographic rollouts. Independent LLM judges gate every episode before admission. Prior sim2real work required hand-authored simulators and separate domain randomization; prior video-model data generation lacked exact action labels.

What's New: First end-to-end demonstration that an LLM agent can author the entire data-generation stack (scene, policy, judge) from a single calibration episode, with video-model rendering supplying real-ready visuals and exact action labels — collapsing sim2real, data collection, and task specification into one automated pipeline.

Extension Opportunities:

Scale the agent-generated task library to hundreds of tasks and train a generalist foundation policy, testing whether DataFarm's robustness advantage compounds with scale
Replace the single calibration episode with zero-shot scene reconstruction from a phone video, removing the last human-in-the-loop step
Extend to bimanual or mobile manipulation where teleoperation cost scales worse, and benchmark the cost-performance gap there

Replicability: No code/data availability mentioned in the abstract. Reproduction would require: a capable coding LLM (Claude/GPT-class), a multi-view consistent video diffusion model conditioned on real images, a physics simulator, and a real robot setup for the calibration episode and evaluation. Estimated non-trivial compute for video rendering; API costs explicitly listed (~$21+$0.80/episode).

Research Gaps:

Abstract does not report task diversity ceiling — unclear which task categories (deformables, long-horizon, contact-rich) the agent can specify correctly
No analysis of failure modes when the judge or expert controller is wrong, or how often the independent certification rejects episodes

🦾 ROBOTICS

1. HLO-AMP: Hierarchical Loco-Manipulation from Object-Aware Motion Priors

Authors: Jean-Pierre Sleiman, Bilal Hammoud, He Li... Published: 2026-11-03 | Citations: 0 arXiv | PDF

Research Question: How can hierarchical humanoid loco-manipulation pipelines produce natural, human-like motion while faithfully handling the distributed multi-contact wrenches that real payloads impose — particularly near platform payload limits — rather than relying on task-agnostic WBCs trained with randomized end-effector force proxies and manual reward shaping?

Summary: HLO-AMP replaces the standard task-agnostic whole-body controller in hierarchical humanoid loco-manipulation with a single object-aware motion prior trained by imitating retargeted MoCap references while the manipulated object is physically present in simulation, so realistic multi-contact payload wrenches are absorbed at the low level. A reference-free high-level policy is then trained on top, with its commands anchored to the empirical envelope of the LL training data to prevent off-distribution queries. On a Unitree G1, the resulting stack generalizes bimanual box pick-and-place across object masses, dimensions, and poses, and reactively recovers from mid-carry drops.

Key Results: The authors demonstrate on a Unitree G1 humanoid that a single object-aware motion-prior low-level policy, trained via motion imitation on retargeted MoCap references with the manipulated object physically present in the scene, combined with a reference-free high-level policy anchored to the empirical envelope of LL training data, generalizes bimanual box pick-and-place across varying object masses and dimensions, varied initial and goal poses, and reactively recovers from intentional mid-carry drops — all with the same deployed policy. The paper reports qualitative/video evidence rather than specific quantitative benchmark numbers in the abstract.

Key Findings:

Training the low-level controller with the manipulated object in the scene — rather than with randomized end-effector force disturbances — is critical for capturing distributed multi-contact wrenches near the humanoid's payload limit.
MoCap-based motion imitation at the LL yields visibly more human-like coordination than reward-shaped WBCs while still exposing a task-space command interface usable by an HL policy.
Anchoring the HL action channel to the empirical envelope of LL training commands structurally keeps the LL in-distribution, enabling a single deployed policy to generalize across object properties, poses, and to recover from mid-carry drops.

Technical Novelty: Prior hierarchical loco-manipulation stacks train a task-agnostic WBC via reward shaping without the object, approximating payload effects via randomized end-effector force disturbances. This paper's novelty is training the low-level controller itself as an object-aware motion prior via MoCap imitation with the object physically simulated in the scene, so the policy absorbs realistic distributed contact wrenches at the source, while still exposing a WBC-compatible task-space command interface. A second novelty is anchoring the HL action channel to the empirical envelope of LL training commands, structurally preventing out-of-distribution queries to the LL.

What's New: The combination of (1) training the LL as an AMP-style object-aware motion prior with the payload simulated in-scene, and (2) HL action-space anchoring to the LL training envelope is new — prior work either trained task-agnostic WBCs without the object or used motion priors without hierarchical composition for long-horizon loco-manipulation at payload limits.

Extension Opportunities:

Extend the object-aware motion prior to non-rigid or articulated payloads (e.g., deformable bags, liquid-filled containers, doors/drawers) where contact wrench distributions are even harder to approximate with end-effector force proxies.
Replace the hand-selected MoCap primitives with a learned library of object-aware priors covering more manipulation classes (push, lift-carry-stack, hand-over to a human), then condition the HL policy over the discrete prior index to scale to broader long-horizon tasks.
Investigate automatic estimation of the 'empirical envelope' of LL commands — e.g., a learned safety critic or density model over LL training commands — so HL anchoring becomes automatic rather than hand-tuned, enabling safe composition across multiple LL priors.

Replicability: The abstract advertises only a video (https://youtu.be/viR_ULnRDQQ) — no code, datasets, retargeted MoCap, or pretrained weights are mentioned. Reproducing it would plausibly require a Unitree G1 (or high-fidelity sim), a MoCap pipeline with retargeting for a humanoid, and multi-GPU RL training for motion imitation (AMP-style training typically needs on the order of 1–8 GPU-days). Full physical reproduction would also require a G1 robot (~$16K+) and the exact MoCap references, which are not disclosed.

Research Gaps:

No quantitative benchmarks or ablations are cited in the abstract (e.g., success rate vs. a force-disturbance WBC baseline, payload-limit comparisons), so the magnitude of the improvement over prior WBC-based stacks is unclear.
The approach relies on hand-curated MoCap primitives covering the downstream task; scaling to tasks without readily available MoCap references or to novel object classes beyond boxes is unaddressed.

2. Should the Action Expert Train the VLM? Action-Token Supervision and Gradient Routing in Mixture-of-Transformer VLAs

Authors: Asher James Hancock, Boyi Li, Yuxiao Chen... Published: 2026-11-03 | Citations: 0 arXiv | PDF

Research Question: In Mixture-of-Transformer VLA models, does blocking gradients from the flow-matching action expert to the VLM ('insulation') genuinely protect VLM capabilities, or is it redundant when action-token supervision and vision-language co-training are already applied? And when is coupling (allowing gradient flow) actually better?

Summary: The paper challenges the common VLA training recipe that blocks gradient flow from a flow-matching action expert into the VLM ('insulation'). Through factorial ablations across autonomous driving and robot manipulation, the authors show that coupling (allowing gradients) substantially outperforms insulation when action-token supervision is absent, and still helps meaningfully in closed-loop and manipulation settings even when it is present — undermining insulation as a default.

Key Results: The authors disentangle action-token supervision from gradient coupling/insulation across autonomous driving and manipulation. Findings: (1) Open-loop driving: insulation matches coupling when action tokens are supervised; without action-token supervision, coupling reduces trajectory-prediction error by ~25%. (2) Closed-loop driving (joint training from scratch): coupling improves route progress by ~17% with action-token supervision and ~40% without. (3) Robot manipulation: coupling improves performance under both supervision settings. Conclusion: insulation is not a safe default — its value depends on supervision regime and evaluation (open vs closed-loop).

Key Findings:

Insulation and coupling are roughly equivalent in open-loop driving only when action-token supervision is present; removing that supervision makes coupling ~25% better on trajectory error
In closed-loop driving trained jointly from scratch, coupling increases route progress by ~17% (with action tokens) to ~40% (without)
Coupling consistently improves robot manipulation regardless of action-token supervision, suggesting insulation's benefits are setting-dependent rather than universal

Technical Novelty: Prior VLA recipes (e.g., π0-style) bundle insulation + action-token supervision + VL co-training as a fixed recipe. This paper is the first to factorially ablate gradient routing (coupling vs. insulation) against action-token supervision across both driving and manipulation, isolating which protective mechanism actually matters and showing the dominant recipe is often suboptimal.

What's New: First systematic disentanglement of the two mechanisms (gradient routing vs. action-token supervision) that prior VLA recipes conflated, reframing insulation as a conditional rather than default choice.

Extension Opportunities:

Explore partial/selective gradient routing — e.g., gate expert gradients to specific VLM layers or heads rather than all-or-nothing coupling vs insulation, potentially combining preservation with action-relevance
Extend the ablation to larger VLM backbones and additional embodiments (humanoid, dexterous hands) to test whether the coupling-beats-insulation result scales, and whether vision-language benchmark degradation emerges at scale
Investigate curriculum schedules that start insulated (to preserve VL capabilities) then gradually enable coupling once the expert is warm, quantifying the trade-off on held-out VL benchmarks vs. control tasks

Replicability: Abstract does not mention a code release. Reproduction requires a pre-trained VLM backbone plus a flow-matching action expert, and both an autonomous-driving benchmark (open- and closed-loop, likely NAVSIM/nuPlan or similar) and a manipulation benchmark. Expected compute: multi-GPU (8+ A100/H100-class) training runs per ablation cell — non-trivial but within a well-funded lab's reach.

Research Gaps:

No measurement reported in the abstract of whether coupling degrades the VLM's original vision-language capabilities — the main justification for insulation
Limited to one VLM/expert architecture family (Mixture-of-Transformer); whether findings transfer to dense or MoE VLAs is open

3. RoboTryOn: Synthesizing Dexterous Robot Demonstrations from Human Videos via Virtual Try-On

Authors: Seungyong Lee, Suhyeok Jang, Minyoung Kim... Published: 2026-11-03 | Citations: 0 arXiv | PDF

Research Question: How can we scalably generate robot training demonstrations from egocentric human videos for high-DoF embodiments, bridging the visual and action gap that plagues existing robot-overlay and pose-retargeting approaches?

Summary: RoboTryOn reframes human-video-to-robot-demonstration translation as a virtual try-on task: it preserves the demonstrated motion, scene, and objects while swapping the human for a target robot specified by a reference image. Using autoregressive keyframe translation with hand-skeleton motion conditioning plus a depth-aware inverse dynamics model, it synthesizes training data that lifts novel-task success rates from 2% to 60% across humanoid embodiments without real robot demos.

Key Results: Policies trained on RoboTryOn synthetic data achieved 60% average success rate on novel tasks (including bimanual manipulation) versus 2% baseline — without any real robot demonstrations. Training with synthetic data matched teleoperation-only policy performance at ~1/3 the data collection cost. Validated across multiple humanoid embodiments.

Key Findings:

Virtual try-on formulation preserves task semantics and background better than overlay/retargeting approaches, closing visual gaps for high-DoF robots
Depth-conditioned IDM reliably extracts action labels from synthesized videos, enabling end-to-end policy learning without teleoperation
Synthetic-data-trained policies match teleoperation-trained policies at ~33% of the data collection cost, including on challenging bimanual tasks

Technical Novelty: Framing human-to-robot translation as a virtual try-on problem (keep motion/scene, swap embodiment) rather than overlay or retargeting. Combines autoregressive keyframe translation for visual consistency, hand-skeleton-conditioned video synthesis, and a depth-aware inverse dynamics model that extracts action labels from synthesized robot videos.

What's New: Prior work either renders robots over human videos (visual artifacts) or retargets human poses to robot actions (action-space mismatch). RoboTryOn is the first to apply a conditional video synthesis try-on paradigm with autoregressive keyframes plus a depth-aware IDM, specifically targeting high-DoF humanoid embodiments including bimanual manipulation.

Extension Opportunities:

Extend the virtual try-on paradigm to full-body humanoid locomotion tasks, not just manipulation, using leg skeleton conditioning
Replace the hand-skeleton motion condition with a learned 3D contact/affordance representation to improve manipulation fidelity on deformable or occluded objects
Combine with sim-to-real residual policies — use RoboTryOn synthetic data for pre-training and small real-robot rollouts for fine-tuning to close remaining dynamics gaps

Replicability: Abstract does not mention code/data release. Reproducing would require a video diffusion backbone, hand-pose estimation pipeline, depth estimator, IDM training infrastructure, and multiple humanoid robot platforms for evaluation — likely requiring multi-GPU training clusters and significant robotics hardware access.

Research Gaps:

Fidelity to physical dynamics — synthesized videos may be visually convincing but encode kinematically impossible or non-contact-faithful trajectories that an IDM cannot fully correct
Generalization to non-humanoid embodiments (quadrupeds, mobile manipulators) and tasks beyond manipulation (locomotion, long-horizon tasks) is untested

💻 COMPUTE

1. What Sets the Serving Speed of Looped Language Models? Measurements and Optimizations Graded by Exactness

Authors: Dezhen Lu, Runguo Li, Zixiang Ni... Published: 2026-11-03 | Citations: 0 arXiv | PDF

Research Question: What determines serving throughput for looped language models (which run one block of layers many times per token), and which engine-level optimizations can accelerate them without compromising output exactness?

Summary: The paper dissects why looped language models (Ouro, Huginn-0125) are slow to serve — thousands of tiny per-layer kernels fire once per loop — and presents optimizations graded by how exactly they preserve model outputs. Bit-exact kernel fusion yields 1.11-1.59x at batch 1, while engine-level changes (prefix reuse, PD disaggregation with a sized KV transfer budget) reach 1.77-3.3x with accuracy preserved within one point on 1319 test questions.

Key Results: Measured Ouro and Huginn-0125 in an open-source looped engine: at batch 1, fusing 5689 small per-token kernels behind rounding fences gives 1.11-1.59x throughput with 1014/1014 requests bitwise-identical; an approximate fusion variant was no faster and failed the GSM8K 1-point rule (61/62 vs 64 on 87 questions). Huginn defaults reach 1.77-3.16x, Ouro 1.65-2.15x, agent sessions 1.91-3.21x, with Huginn's noise/prefix-reuse changes staying within one point on all 1319 test questions (packed prefill: -0.61). Under prefill-decode disaggregation, ExactLoop hits 3.3x base throughput (peer error handling off), 3.1x on; a tuned KV transfer budget cuts commit tail from ~1s to 16ms (base 2.8s).

Key Findings:

The dominant cost in looped-model serving is per-kernel overhead from ~5689 small kernels per token, not the loop submission itself (naive fusion only hits 0.99x).
Bit-exact fusion with rounding fences delivers 1.11-1.59x while preserving all 1014 requests identically; the approximate variant gains no speed and loses accuracy, so it is rejected.
End-to-end optimizations (noise, prefix reuse, 64-graph cache, PD disaggregation with a one-request-sized transfer budget) reach 1.65-3.3x across Ouro/Huginn/agent workloads, cutting commit tail from 2.8s base to 16ms.

Technical Novelty: Grading serving optimizations by exactness guarantees (bitwise-identical, statistical-tolerance, approximate) rather than just speedup; kernel fusion behind rounding fences that preserves bit-exactness across 5689 per-token kernels; a KV transfer budget sized to one in-flight request that eliminates engine polling rounds in PD disaggregation for looped models specifically.

What's New: First systematic serving-perf study for looped language models that explicitly grades each optimization by its exactness guarantee, rejecting approximate variants that break a stated accuracy rule even when they would otherwise ship.

Extension Opportunities:

Apply the exactness-graded optimization taxonomy (bitwise-identical vs within-tolerance vs approximate) to other architectures with heavy per-layer overhead like MoE or diffusion LMs
Investigate the 20/128 GPU full-hit divergence from position-keyed noise — build deterministic kernels that make GPU prefix cache reuse bitwise-exact
Extend ExactLoop's sized transfer budget idea to general PD-disaggregated serving to eliminate polling-interval tails for non-looped models

Replicability: Paper references an open-source looped-model engine and specific test sets (1014 requests, 1319 Huginn questions, 87 GSM8K questions); code availability is implied but not explicitly stated. Reproduction needs GPU serving infrastructure capable of PD disaggregation plus the Ouro and Huginn-0125 checkpoints — likely a single multi-GPU node.

Research Gaps:

GPU prefix-cache reuse remains non-deterministic (20/128 full hits differ) despite position-keyed noise fixing the CPU side.
Packed prefill for Huginn only meets the one-point accuracy rule on the point estimate, not robustly — deterministic packing for looped models is unresolved.

2. EvoCG: Self-Evolving Precision Control for Faster Conjugate Gradient

Authors: Jiaqi Zhai, Haozhi Han, Kaiyi Huang... Published: 2026-11-03 | Citations: 0 arXiv | PDF

Research Question: How can conjugate gradient (CG) solvers exploit GPU low-precision throughput advantages when the tolerable precision varies dynamically with the evolving Krylov subspace state?

Summary: EvoCG is a mixed-precision conjugate gradient solver that uses a reinforcement-learning policy to dynamically select per-tile numerical precision during iteration, co-designed with a unified Tile-CSC SpMV kernel to exploit GPU low-precision throughput. On 47 SPD matrices it achieves 1.93x–6.99x geometric-mean speedups over state-of-the-art solvers (cuSPARSE, PETSc, Ginkgo, Mille) while matching FP64 final residual accuracy.

Key Results: Across 47 SPD matrices from 8 scientific domains, EvoCG achieved geometric-mean speedups of 2.52x over cuSPARSE, 3.06x over PETSc, 2.48x over Ginkgo, 1.93x over Mille-FP64, and 6.99x over Mille-Mixed. Vs EvoCG-FP64 baseline: 1.28x faster time-to-solution with only 11.4% more iterations and 1.09% runtime overhead, with final true-residual norms nearly identical to FP64. Extends to Jacobi-preconditioned CG without retraining for additional 1.15x speedup.

Key Findings:

Online RL-based precision control outperforms static mixed-precision schedules by adapting to the evolving Krylov state
Precision adaptation incurs only 1.09% runtime overhead and 11.4% iteration overhead while delivering 1.28x wall-clock speedup over FP64
The learned policy generalizes to Jacobi-preconditioned CG without retraining, suggesting the state representation captures solver-agnostic convergence signals

Technical Novelty: Treating per-tile precision as an online RL control action within the CG iteration loop (rather than a fixed static schedule), co-designed with a Tile-CSC unified mixed-precision SpMV kernel that makes tile-level precision switching profitable at low runtime overhead.

What's New: Prior mixed-precision CG work uses fixed precision schedules or coarse-grained switching; EvoCG is the first to frame tile-wise precision as an online control problem solved by RL, and the first to co-design the SpMV kernel (Tile-CSC) specifically to make fine-grained precision switching cheap enough to be profitable.

Extension Opportunities:

Apply the RL precision-control framework to other Krylov methods (GMRES, BiCGStab, MINRES) where precision tolerance also evolves with iteration state
Extend to non-SPD systems and explore whether the policy transfers across matrix classes without retraining, similar to the Jacobi-preconditioner generalization shown
Integrate with hardware-specific tensor core formats (FP8, MXFP, block-floating-point) on Hopper/Blackwell GPUs to push throughput gains further

Replicability: Abstract does not mention a code release. Reproduction would require a modern NVIDIA GPU (likely A100/H100-class for the FP64 vs lower-precision throughput gap), the SuiteSparse-style SPD matrix collection, and reimplementation of the Tile-CSC SpMV kernel plus RL policy training infrastructure. Medium-to-high engineering effort.

Research Gaps:

Generalization beyond SPD matrices and beyond CG to other Krylov solvers is untested
No analysis of policy transferability across GPU architectures or how the policy degrades on matrices outside the training distribution of 8 domains

3. ForgeKernel: Self-Evolving Kernels Catching and Surpassing Experts and Agents

Authors: Yaojian Chen, Dazhi Peng, Yiyun Zheng... Published: 2026-11-03 | Citations: 0 arXiv | PDF

Research Question: How can GPU kernel-generation agents accumulate knowledge across runs and continue evolving kernel architectures, rather than restarting from fixed initial knowledge and locking into an early-chosen architecture?

Summary: ForgeKernel is a self-evolving GPU kernel-generation system that accumulates architectural knowledge across runs by distilling hardware-mechanism insights from reference kernels and sealing validated kernels as references for subsequent optimization passes. It outperforms both expert-written kernels and prior agent systems on NVIDIA Hopper/Blackwell and Ascend 910C/950 benchmarks, holding top rank on four public leaderboards.

Key Results: On 27 operators/29 configs across NVIDIA Hopper and Blackwell, ForgeKernel matches or beats its expert reference on 25 configs with a 10% geomean gain. On Ascend 910C, it beats the per-operator public best on all 8 level-4 CANN-Bench Season 1 final operators with 25.1% geomean gain and ranks #1 on the A3 full-suite leaderboard. As of 2026-10-30 it holds #1 on all three CANN-Bench 2.0 leaderboards (910C, 950DT, 950PR) and the SOL-ExecBench FlashInfer-Bench track.

Key Findings:

Cross-run knowledge accumulation plus architectural re-evaluation yields a 10% geomean speedup over expert references on NVIDIA Hopper/Blackwell across 27 operators
On Ascend 910C the system beats per-operator public bests on all 8 level-4 CANN-Bench Season 1 operators with 25.1% geomean gain
Distilling hardware-mechanism knowledge from references outperforms imitating source code and generalizes to new GPU architectures (Blackwell, 950DT, 950PR)

Technical Novelty: Prior kernel agents (KernelBench-style) re-run optimization from scratch and get stuck in the first-chosen architecture. ForgeKernel's novelty is twofold: (1) distilling a high-performance reference into reusable hardware-mechanism knowledge instead of imitating source code, and (2) sealing a validated kernel as the next run's reference so architecture itself can be revisited across runs.

What's New: Unlike prior kernel agents that reset knowledge per run and lock architectures early, ForgeKernel treats each validated kernel as a reusable reference and distills validated architectural decisions into portable hardware-mechanism knowledge, enabling architecture-level evolution across runs.

Extension Opportunities:

Port the architectural-knowledge distillation loop to AMD CDNA/RDNA or Intel Gaudi backends to test cross-vendor transfer of hardware-mechanism knowledge
Apply the seal-and-reuse reference mechanism to full fused-model compilation (e.g. end-to-end attention+MoE pipelines) rather than per-operator kernels
Add a cross-run knowledge-pruning/forgetting policy so accumulated knowledge doesn't bias against newer hardware generations when the reference pool grows stale

Replicability: The abstract references CANN-Bench, SOL-ExecBench, and FlashInfer-Bench public leaderboards for evaluation, but does not mention open-sourced code or weights. Reproduction would require access to NVIDIA Hopper (H100/H200), Blackwell (B100/B200), and Ascend 910C/950 hardware plus the respective vendor toolchains — substantial multi-vendor compute.

Research Gaps:

No discussion of knowledge-base scaling limits or interference as more operators and hardware generations accumulate
Unclear how the system handles correctness regressions when reusing sealed kernels as references across hardware generations with different numerical behaviors

⚡ ENERGY

1. Periodically poled lithium niobate with submicrometer periodicity using electric-field poling technique

Authors: Yuchen Zhang, Sanbing Li, Shengbo Song... Published: 2026-11-03 | Citations: 0 arXiv | PDF

Research Question: How can periodically poled lithium niobate (PPLN) structures with submicrometer periodicity be reliably fabricated in thin-film LN on silicon, with full-thickness polarization reversal and controllable lateral domain broadening, to enable scalable domain engineering for integrated nonlinear photonics?

Summary: The authors fabricate monocrystalline lithium niobate films on low-resistivity silicon (LNOS) via chemical-mechanical polishing and use electric-field poling to produce PPLN gratings with periods from 6 μm down to 200 nm, achieving full-thickness polarization reversal. They characterize how coercive field (Kay-Dunn scaling) and lateral domain broadening depend on film thickness and electrode linewidth, providing design rules for submicrometer domain engineering in thin-film LN.

Key Results: The authors fabricated monocrystalline LN films on low-resistivity Si (LNOS) with thicknesses from micrometers down to tens of nanometers via CMP of pristine crystal films, then demonstrated electric-field-poled PPLN with periods ranging from 6 μm down to 200 nm. They achieved full-thickness polarization reversal under the poling electrodes with well-defined depth profiles, showed the coercive field obeys the Kay-Dunn scaling law vs. thickness, and found that lateral domain broadening scales with film thickness but is approximately independent of electrode linewidth at fixed thickness.

Key Findings:

Submicrometer-period PPLN down to 200 nm was demonstrated in thin-film LN on Si with full-thickness, well-profiled domain reversal
Coercive field follows the Kay-Dunn scaling law as film thickness decreases from micrometers to tens of nanometers
Lateral domain broadening increases with film thickness but is approximately independent of electrode linewidth at a fixed thickness, enabling simple pre-compensation design

Technical Novelty: Prior electric-field poling work on LN thin films has struggled to simultaneously achieve submicrometer periods, full-thickness domain reversal, and well-controlled depth profiles, especially on conductive Si substrates. This paper uses CMP of pristine crystal (rather than ion-sliced) LN on low-resistivity Si as a new substrate stack and empirically maps coercive field (Kay-Dunn) and lateral broadening vs. thickness/linewidth, reaching a 200 nm period — a regime typically accessed only by less scalable techniques like calligraphic/AFM or ion-beam poling.

What's New: Combines CMP-thinned pristine-crystal LN on conductive low-resistivity Si with electric-field poling to reach 200 nm periods — an order of magnitude finer than typical E-field-poled bulk PPLN (~μm) — while retaining full-thickness reversal and providing quantitative broadening/coercive-field scaling laws usable as design rules.

Extension Opportunities:

Integrate the 200 nm-period PPLN films into ridge/rib waveguides and measure SHG/SPDC conversion efficiency for counter-propagating or backward quasi-phase-matching schemes that require sub-μm periods
Develop a lateral-broadening pre-compensation design tool: parameterize electrode linewidth shrinkage vs. film thickness from the measured broadening data and verify sub-100 nm period PPLN fabrication
Extend the CMP + E-field poling process to thin-film LN on insulator (LNOI) or hybrid Si3N4/LN platforms to combine sub-μm PPLN with low-loss waveguiding and on-chip resonators

Replicability: No code or dataset is mentioned in the abstract. Reproduction requires specialized fabrication infrastructure: pristine LN crystal bonding to low-resistivity Si, chemical-mechanical polishing to tens-of-nm thickness, electron-beam lithography for submicrometer electrodes, and high-voltage electric-field poling plus domain characterization (e.g., PFM/SHG microscopy). No significant compute is required; the barrier is cleanroom/materials-lab access.

Research Gaps:

No demonstration of actual nonlinear device performance (conversion efficiency, phase-matching bandwidth, loss) using the 200 nm-period gratings
Low-resistivity Si substrate may be lossy at optical frequencies; integration path to low-loss waveguiding (e.g., LNOI, Si3N4-on-LN) is not addressed

2. Ta-based S-I-S junctions for integrated superconducting Electronics

Authors: Trent Borman, Cameron Bullock, Cole Chason... Published: 2026-11-03 | Citations: 0 arXiv | PDF

Research Question: Can α-Ta based S-I-S Josephson junctions with TaOx barriers be fabricated at industrial scale with sufficient uniformity to serve as a viable material platform for classical and quantum superconducting electronics, replacing conventional Nb/AlOx junctions?

Summary: This paper demonstrates all-α-Ta Josephson junctions with TaOx tunnel barriers fabricated using industry-standard 150 mm foundry processing, achieving Jc uniformity as low as 2.0% across 20x20 mm chips for sub-μm² junctions. It establishes seed-layer engineering (Nb or TaNx, with a mandatory re-seed for the top electrode) as the enabling process step for nucleating α-Ta on both electrodes.

Key Results: Demonstrated fabrication of α-Ta/TaOx/α-Ta junctions on 150 mm wafers using standard foundry tooling. Achieved critical current density uniformity (Jc spread) as low as 2.0% across a 20x20 mm chip for 0.4 μm² junctions. Showed both Nb and TaNx work as seed layers, and established that a re-seed layer is required at room temperature to nucleate α-Ta for the top electrode. Used Ambegaokar-Baratoff physics with room-temperature probing of variable junction areas to characterize wafer-scale Jc.

Key Findings:

α-Ta can be nucleated for both base and top electrodes using Nb or TaNx seed layers, but a re-seed layer is required at room temperature for the top electrode
Junction critical-current spread as low as 2.0% was achieved for 0.4 μm² junctions across a 20x20 mm chip — competitive with mature Nb/AlOx processes
Seed layer choice has subtle but measurable effects on electrical tunneling behavior, meaning seed selection is a design parameter not just a process convenience
The full process is compatible with standard 150 mm foundry tooling, demonstrating scalability

Technical Novelty: The all-Ta junction stack — using α-phase Ta for both electrodes with a native TaOx tunnel barrier — fabricated end-to-end on 150 mm wafers with foundry tooling. Prior work has focused on Ta as a capacitor pad material (e.g., in transmon qubits) while retaining Al/AlOx junctions. The identification of a mandatory re-seed layer for top-electrode α-Ta nucleation at room temperature is a specific process insight.

What's New: First demonstration of an all-α-Ta/TaOx/α-Ta junction stack fabricated with industrial 150 mm foundry tooling at competitive uniformity. Differs from prior Ta work that used Ta only as a qubit pad material while keeping Al-based junctions, and from prior Ta junction work that was not foundry-compatible.

Extension Opportunities:

Characterize two-level-system (TLS) loss and coherence times (T1/T2) of qubits built using these α-Ta/TaOx/α-Ta junctions to validate quantum advantages over Nb/AlOx baselines
Investigate the subtle seed-layer-dependent tunneling behavior systematically — map how Nb vs TaNx seed choice affects subgap leakage, IcRn product, and barrier homogeneity
Scale the process to 200/300 mm wafers and integrate with multi-layer wiring/airbridges to demonstrate full quantum processor fabrication flows

Replicability: No code/data mentioned (materials science process paper). Reproduction requires a 150 mm wafer fab line with sputtering (Nb, Ta, TaN), controlled oxidation, lithography, and cryogenic probing infrastructure — not reproducible outside well-equipped superconducting electronics cleanrooms.

Research Gaps:

No cryogenic coherence/loss measurements are reported — the paper demonstrates process viability but not the hypothesized quantum coherence benefits that motivate replacing AlOx barriers
The mechanism behind seed-layer-dependent tunneling behavior is noted but not explained, leaving an open materials question about Ta/TaOx interface chemistry

3. Controlling nitrogen incorporation in (Eu,Ba,Ca)TiO$_3$-derived perovskite-type oxynitrides through Fermi Level Engineering

Authors: Margarida Barroso, Marie Neumann, Rachele Sciotto... Published: 2026-11-03 | Citations: 0 arXiv | PDF

Research Question: How can nitrogen incorporation in perovskite oxynitrides be rationally controlled during ammonolysis to overcome intrinsic self-limiting nitridation caused by defect-induced Fermi level shifts?

Summary: The paper introduces Fermi Level Engineering as a predictive framework for controlling nitrogen incorporation in perovskite oxynitrides, combining DFT with experiments on (Eu,Ba,Ca)TiO3 systems. By exploiting the Eu2+/3+ redox couple to pin the Fermi level at ~2.9 eV above the VBM, the authors overcome the intrinsic self-limiting nitridation of pure BaTiO3 (0.03 N/f.u.) to achieve 0.65 N/f.u.

Key Results: Demonstrated that pure BaTiO3-derived oxynitrides self-limit at only 0.03 N per formula unit due to Fermi level downshift making further N substitution energetically unfavorable. By A-site substitution with Eu in Eu(1-x)Ba(x)TiO3 and Eu(1-x)Ca(x)TiO3, the Eu2+/3+ redox couple pins the Fermi level at ~2.9 eV above the VBM, enabling nitrogen uptake up to 0.65 N per formula unit (measured via hot gas extraction) — a >20x increase. Further incorporation is limited by intra-Ti(O,N)6 octahedra distortion from internal bond-angle variance.

Key Findings:

Pure BaTiO3-derived oxynitrides are self-limited to 0.03 N per formula unit due to Fermi level downshift during nitridation
Eu2+/3+ redox couple pins Fermi level at ~2.9 eV above VBM, enabling controlled substitution and up to 0.65 N per formula unit
Upper limit of N incorporation in Eu-containing systems is set by intra-Ti(O,N)6 octahedral distortion from bond-angle variance, not charge compensation

Technical Novelty: First application of Fermi Level Engineering as a unified descriptor framework for oxynitride synthesis, using redox-active A-site cations (Eu2+/3+) as a Fermi level pinning mechanism to decouple charge compensation from the self-limiting N incorporation barrier — moving beyond empirical trial-and-error ammonolysis.

What's New: Shifts oxynitride synthesis from empirical trial-and-error to a defect-thermodynamic framework where the Fermi level serves as a single unifying descriptor connecting defect formation enthalpies, charge transition levels, and achievable stoichiometry.

Extension Opportunities:

Apply Fermi Level Engineering framework to other redox-active A-site cations (Ce3+/4+, Sm2+/3+, Yb2+/3+) to achieve even higher N incorporation or tunable band gaps for photocatalysis
Extend the concept to B-site or mixed anion systems (oxyfluorides, oxysulfides) where similar Fermi-level-driven self-limiting phenomena may apply
Build a high-throughput DFT screening pipeline using charge transition levels as descriptors to predict optimal cation combinations for targeted N:O ratios

Replicability: No code/data mentioned in abstract. Reproduction would require: ammonolysis furnace with NH3 flow, perovskite precursor synthesis capability, hot gas extraction for N quantification, XRD/neutron diffraction for structure, and DFT compute (likely VASP/Quantum ESPRESSO) for defect formation energies and charge transition levels — moderate but specialized materials-science infrastructure.

Research Gaps:

Does not explore kinetic effects of ammonolysis (temperature, NH3 partial pressure) within the Fermi Level Engineering framework
Functional property outcomes (e.g., photocatalytic, dielectric, magnetic) of the achieved 0.65 N/f.u. phases are not quantified in the abstract

🏥 HEALTHCARE

1. Predicting protein structures at any folding stage from sequence

Authors: Lisheng Zhang, Qingsong Li, Zinan Xie... Published: 2026-11-03 | Citations: 0 arXiv | PDF

Research Question: Protein structure predictors like AlphaFold identify native folds but don't describe the conformational continuum from unfolded to folded states. How can we generate sequence-conditioned structural ensembles at any specified folding fraction Q without per-protein simulation?

Summary: FoldScape is a sequence-conditioned deep learning framework that generates Cα protein conformational ensembles at any specified folding fraction Q, using hierarchical sliding-window distance regression plus a rectified-flow decoder trained on folding and mechanical-unfolding trajectories. It covers ~10x more of the folding coordinate than AlphaFold 3 seed ensembles across 20 human proteins while maintaining constant accuracy across folding stages, filling the ensemble-generation gap that native-only predictors leave open.

Key Results: FoldScape was evaluated on 20 human proteins (36-857 residues). At matched sample size, it covers ~10x more of the folding coordinate than seed-varied AlphaFold 3 ensembles, robust across proteins, order parameters, and binning choices. Reconstruction accuracy stays approximately constant along Q (from collapsed coil to near-native globule). A stride-one crop-and-superpose scheme extends inference to chains several times longer than a single forward pass. Training trajectories resolve mechanical unfolding at residue resolution; elongation-conditioned ensembles show reproducible continuous observables across pulling velocities.

Key Findings:

FoldScape covers ~10x more of the folding coordinate than seed-varied AlphaFold 3 on 20 human proteins (36-857 residues)
Reconstruction accuracy is approximately constant along Q, not degrading away from the native state — unusual for structure predictors
FoldScape and AlphaFold 3 are complementary: AF3 anchors the immediate native vicinity, FoldScape covers the rest of the folding landscape
Stride-one crop-and-superpose enables inference on chains several times longer than the base window; partially unfolded chains form geometrically coherent segments

Technical Novelty: Sequence-conditioned generation of Cα ensembles at a user-specified folding fraction Q, via hierarchical sliding-window regression of 1D inter-residue distances followed by rectified-flow chain assembly. Training combines multi-scale coarse-grained folding trajectories with all-atom mechanical-unfolding trajectories — a departure from training only on native PDB structures. Stride-one crop-and-superpose extends to arbitrarily long chains.

What's New: Prior work (AlphaFold, ESMFold, RoseTTAFold) targets a single native structure. Boltzmann-generator and diffusion ensemble methods typically require per-protein simulation or training. FoldScape is the first to produce sequence-conditioned ensembles parameterized by an explicit folding-progress variable Q, trained on folding/unfolding trajectories rather than static PDB structures, with no per-protein simulation at inference.

Extension Opportunities:

Combine FoldScape (broad Q coverage) with AlphaFold 3 (tight native anchoring) in a hybrid ensemble generator that samples both the folding funnel and native basin
Extend from Cα-only to all-atom ensembles by adding a side-chain reconstruction head conditioned on Q, enabling interaction-site analysis during folding
Apply the Q-conditioned ensembles to predict folding kinetics (barrier heights, Φ-values) or IDP/partially-disordered protein behavior where native-only predictors fail

Replicability: Abstract does not mention code/data availability. Training requires multi-scale CG folding trajectories and all-atom mechanical-unfolding trajectories, which are computationally expensive to generate but may be reused from published MD datasets. Model training likely needs multi-GPU compute (comparable to mid-size diffusion/flow models); inference is a single forward pass per sample and should be tractable on a single GPU.

Research Gaps:

Cα-only representation omits side chains, limiting direct use for binding-site, allostery, or drug-design analyses during folding
Does not reach the immediate native vicinity (where AF3 excels); fidelity of the ensemble distribution to true Boltzmann statistics along Q is not directly validated

2. Fairness Metrics based on the AUC in Clinical Predictive AI

Authors: João Matos, Ben Van Calster, Richard D. Riley... Published: 2026-11-03 | Citations: 0 arXiv | PDF

Research Question: How should AUC-based fairness metrics for clinical prediction models be interpreted and used, given that AUC is non-collapsible (subgroup and overall AUCs are not directly comparable)?

Summary: The paper critically examines eight AUC-based fairness metrics used in clinical prediction, exposing how AUC non-collapsibility undermines several commonly used ones. It recommends reporting a full AUC matrix (overall plus within- and cross-group components) and favors within-group AUC parity, while cautioning against gap-to-overall, equity-scaled, and minimum-based metrics.

Key Results: The paper critically appraises eight AUC-based fairness metrics (within-group AUC parity, xAUC parity, gap-to-overall AUC, equity-scaled AUC, minimum-based metrics, xAUC^1, xAUC^0, and xAUC^1/xAUC^0 parity) by characterizing their estimands and relating them to the decomposition of overall AUC into within- and cross-group components. Behavior is illustrated via simulations based on EuroSCORE II, a clinical prediction model for in-hospital mortality. Finds that gap-to-overall AUC, equity-scaled AUC, and minimum-based metrics compare non-comparable quantities and are not fit for purpose.

Key Findings:

AUC is non-collapsible: subgroup and overall AUC values are not directly comparable, invalidating a common implicit assumption in fairness audits
Within-group AUC parity is the recommended primary fairness metric; cross-group AUC parity is supplementary but can conflict with calibration
Gap-to-overall AUC, equity-scaled AUC, and minimum-based metrics compare incomparable quantities and are not fit for purpose

Technical Novelty: First systematic treatment unifying eight AUC-based fairness metrics under the lens of AUC non-collapsibility, with estimand-level characterization tying each metric to the within-/cross-group AUC decomposition rather than treating them as interchangeable.

What's New: Prior work proposed these metrics in isolation; this paper is the first to systematically compare them through the estimand lens of AUC decomposition and non-collapsibility, producing concrete reporting recommendations.

Extension Opportunities:

Develop an open-source toolkit implementing the recommended 'AUC matrix' reporting standard (overall + within/cross-group AUCs) for clinical ML pipelines
Extend the analysis to multi-class or survival (time-to-event) outcomes, where AUC generalizations (e.g., C-index) also face non-collapsibility issues
Investigate the interplay between cross-group AUC parity and calibration empirically across real clinical datasets to quantify the tradeoff the authors flag

Replicability: No explicit code/data link mentioned in the abstract. Simulations are based on EuroSCORE II (publicly described clinical prediction model). Compute requirements are minimal — statistical simulations, no deep learning — reproducible on a standard laptop.

Research Gaps:

No consensus reporting standard for subgroup AUC in clinical ML — the proposed AUC matrix needs community adoption
Tension between cross-group AUC parity and calibration fairness is identified but not fully resolved

3. ProteoCraft: Protein Design as Proposal-Verification Search

Authors: Fang Wu, Hongbin Lin, Bingchen Duan... Published: 2026-11-02 | Citations: 0 arXiv | PDF

Research Question: How can we efficiently design proteins with desired protein-protein interactions when the sequence space is combinatorially large and structural evaluation (folding model inference) is expensive?

Summary: ProteoCraft reframes protein binder design as a tree-search optimization problem that uses approximate structural gradients from a diffusion folding model as a cheap proposal policy and full folding inference as an independent verifier. A coarse-to-fine schedule first explores distinct sequence basins via multi-site edits, then refines verified candidates with single-site edits, yielding higher design quality in silico and strong wet-lab validation (82.9% hit rate on PD-L1, 0.879 nM KD).

Key Results: ProteoCraft was evaluated across 14 de novo binder-design settings covering 8 target proteins, 5 TCR-facing immune targets, and 4 localized CDR redesign tasks, improving design quality and high-confidence yield under both AF2-IG and Rosetta-based interface evaluation. Wet-lab validation achieved an 82.9% binding hit rate for PD-L1 miniproteins with a best measured KD of 0.879 nM, and confirmed binding for a designed IFNAR2 VHH.

Key Findings:

Decoupling proposal (gradient-based) from verification (gradient-free folding) outperforms direct-gradient and pure-search baselines across 14 binder-design settings.
A coarse-to-fine multi-site then single-site edit schedule effectively allocates the expensive verification budget.
Gains persist under independent AF2-IG and Rosetta evaluators and translate to experimental success (82.9% PD-L1 hit rate, 0.879 nM best KD, confirmed IFNAR2 VHH binding).

Technical Novelty: Framing protein design as proposal-verification tree search that decouples a cheap gradient-based proposal policy (structural gradients from a diffusion folding model) from an independent gradient-free verifier (full folding inference), plus a coarse-to-fine schedule that uses multi-site edits to explore basins and single-site edits to refine verified candidates — in contrast to prior work that either backprops directly through folding models (expensive, biased) or uses pure MCMC/evolutionary search without structural proposals.

What's New: First to cast binder design as explicit proposal-verification tree search with diffusion-model structural gradients as the proposal policy and gradient-free folding as the verifier, combined with a basin-exploration-then-refinement budget schedule.

Extension Opportunities:

Apply the proposal-verification search framework to other expensive-evaluation biomolecular design tasks (e.g., enzyme active-site design, RNA aptamer design, small-molecule-binding pocket design) where gradient-based proposals and no-gradient verifiers can be decoupled.
Replace AlphaFold2-based verification with faster/cheaper surrogate models (ESMFold, Boltz, distilled folding networks) and study the Pareto frontier between verification cost and final hit rate.
Extend the coarse-to-fine multi-site to single-site schedule into a learned adaptive scheduler (e.g., bandit or RL controller) that dynamically allocates verification budget based on current basin diversity and confidence.

Replicability: Code is released at https://github.com/smiles724/ProteoCraft_v1. Reproduction requires GPU resources to run diffusion-based structure predictors and full AlphaFold2-initial-guess inference across many candidates per target; experimental validation of binders further requires wet-lab capability (miniprotein/VHH expression and SPR/BLI for KD measurement).

Research Gaps:

Prior binder-design methods either backprop through folding networks (biased/expensive) or use gradient-free search without structural guidance, leaving verification budget poorly allocated.
Limited work integrates diffusion-based structure predictors as proposal distributions rather than as end-to-end generators.

🔬 MATERIALS

1. Intra-unit-cell Josephson scanning tunneling spectroscopy of the cuprate Bi${1.2}$Pb${0.8}$Sr$_2$CaCu$2$O${8+δ}$

Authors: Fengyu Yao, Jiong Mei, Wenjin Mao... Published: 2026-11-03 | Citations: 0 arXiv | PDF

Research Question: How does the sign-changing d_{x^2-y^2}-wave pairing symmetry of cuprate superconductors manifest at the atomic scale in local Cooper-pair (Josephson) tunneling within a single CuO2 unit cell?

Summary: The authors perform Josephson scanning tunneling spectroscopy on Pb-doped Bi-2212 with a conventional s-wave Pb tip, resolving intra-unit-cell modulation of Cooper-pair tunneling with enhanced response at oxygen sites and suppression at copper and nodal-diagonal sites. Microscopic bilayer + Wannier-based calculations reproduce the pattern as interference between opposite-sign d-wave pairing contributions, providing real-space, phase-sensitive atomic-scale evidence for d_{x^2-y^2} pairing in cuprates.

Key Results: Using Josephson STS with an s-wave Pb tip on Bi_{1.2}Pb_{0.8}Sr_2CaCu_2O_{8+δ}, the authors measured a zero-bias conductance enhancement with approximately quadratic scaling vs. junction conductance (signature of phase-diffusive Josephson tunneling). Atomic-resolution maps show intra-unit-cell modulation: enhanced Josephson response at oxygen sites, suppression at copper sites and along diagonal (nodal) directions. Microscopic calculations using bilayer electronic structure and Wannier-based tunneling matrix elements reproduce the pattern via interference between opposite-sign pairing contributions.

Key Findings:

Zero-bias conductance enhancement scales approximately quadratically with junction conductance, consistent with phase-diffusive Josephson tunneling rather than quasiparticle leakage
Josephson response exhibits pronounced intra-unit-cell structure: strong at O sites, suppressed at Cu sites and along the nodal (diagonal) directions
Spatial pattern is quantitatively reproduced by a bilayer-electronic-structure + Wannier tunneling-matrix-element model that invokes interference between opposite-sign lobes of the d-wave gap

Technical Novelty: First atomic-resolution, site-resolved Josephson STS maps of a cuprate that directly resolve sub-unit-cell (O vs Cu site) Cooper-pair tunneling modulation, combined with Wannier-based tunneling-matrix-element theory that explains the spatial pattern as destructive interference between opposite-sign d-wave lobes — a real-space phase-sensitive signature complementary to momentum-space ARPES or corner-junction phase experiments.

What's New: Prior cuprate STM work primarily measured single-particle (quasiparticle) density of states; corner-junction/tricrystal experiments established d-wave symmetry only at mesoscopic scales. This paper delivers the first atomic-scale, site-resolved Cooper-pair tunneling map in a cuprate, exposing the sign structure of the order parameter inside a single CuO2 plaquette.

Extension Opportunities:

Apply the same intra-unit-cell Josephson STS protocol to other unconventional superconductors (e.g., Fe-based, nickelates, kagome, heavy-fermion) to directly image sign-changing order parameters (s±, p-wave, chiral)
Combine Josephson STS with a vector magnetic field or phase-sensitive SQUID-tip geometry to map the full complex order parameter (amplitude + phase) at atomic resolution, including near vortices and impurities
Extend measurements across the doping phase diagram (underdoped pseudogap → overdoped) to track how intra-unit-cell Josephson modulation evolves with the pseudogap, charge order, and nematicity

Replicability: No code/data availability is stated in the abstract. Reproduction requires a cleaved Bi-2212 (Pb-doped) single crystal, a mK-range dilution-refrigerator STM with superconducting Pb tip, vibration isolation sufficient for Josephson-regime junction stability, and DFT+Wannier90 post-processing for the tunneling-matrix-element theory — accessible only to specialized cuprate STM groups.

Research Gaps:

No systematic doping or temperature dependence reported — relationship to the pseudogap, charge order, or nematicity remains open
Josephson response in the vicinity of defects, dopant atoms, and vortex cores (where sign-sensitive bound states should appear) is not addressed

2. Effect of molybdenum content on phase transformations during continuous heating of metastable β Ti-Mo alloys

Authors: Veronika Kočiščáková, Anna Veverková, Josef Stráský... Published: 2026-11-03 | Citations: 0 arXiv | PDF

Research Question: How does molybdenum content affect β-phase stability and the sequence of phase transformations (athermal ω → isothermal ω → α) during continuous heating of metastable β Ti-Mo alloys?

Summary: The paper investigates how Mo content (12, 15, 18 wt.%) governs phase transformations in metastable β Ti-Mo alloys during linear heating to 750 °C. By combining electrical resistance, dilatometry, DSC, and electron microscopy, the authors show that increasing Mo progressively suppresses ω-phase evolution and subsequent α precipitation, and demonstrate the value of complementary techniques for mapping complex transformation pathways.

Key Results: Using Ti-12Mo, Ti-15Mo, and Ti-18Mo (wt.%) heated at 5 °C/min up to 750 °C, the authors demonstrated via combined electrical resistance, dilatometry, DSC, and electron microscopy (on water-quenched specimens) that all three alloys follow the sequence: athermal ω dissolution → isothermal ω stabilization/growth → ω dissolution → α precipitation. Higher Mo content progressively suppressed ω evolution and α precipitation, and narrowed the ω-related transformation temperature interval.

Key Findings:

All three alloys share the same transformation sequence: athermal ω dissolution → isothermal ω growth → ω dissolution → α precipitation
Increasing Mo content suppresses both ω evolution and α precipitation and narrows the ω-related temperature interval
Electrical resistance, DSC, dilatometry, and electron microscopy show excellent cross-agreement, validating a consistent transformation description

Technical Novelty: The systematic combination of three complementary in-situ techniques (electrical resistance, dilatometry, DSC) with ex-situ electron microscopy across a controlled Mo compositional series, enabling consistent cross-validation of each transformation step in the ω → α pathway.

What's New: Rather than studying a single composition or technique in isolation, this work systematically varies Mo content and cross-validates each transformation step using four complementary methods, producing a coherent compositional picture of ω suppression in binary Ti-Mo.

Extension Opportunities:

Extend the Mo composition range (e.g., Ti-8Mo to Ti-20Mo) and vary heating rates to map a full TTT/CHT diagram parameterized by β-stabilizer content
Apply the same multi-technique protocol to other β-stabilizer systems (Nb, V, Ta) or ternary Ti-Mo-X alloys for biomedical/aerospace applications where ω suppression is desirable
Couple the in-situ measurements with synchrotron XRD or atom-probe tomography to quantify ω/α volume fractions and Mo partitioning in real time, enabling calibrated CALPHAD/phase-field models

Replicability: No code/data repository mentioned. Reproduction requires arc/induction-melted Ti-Mo ingots, standard metallography, a dilatometer, DSC, four-point resistivity rig, and SEM/TEM — feasible for any materials science lab with β-Ti processing capability; no significant compute required.

Research Gaps:

Quantitative phase fractions and Mo partitioning between β and ω/α are not resolved — only qualitative suppression trends
Only one heating rate (5 °C/min) is explored, leaving kinetic rate dependence and isothermal hold behavior uncharacterized

3. Nonreciprocal charge transport in quantum materials: From superconductors to magnetic systems

Authors: Boning Yu, Yalei Huang, Alltrin Dhanarajagopal... Published: 2026-11-03 | Citations: 0 arXiv | PDF

Research Question: How to unify the understanding of nonreciprocal charge transport across diverse quantum materials (superconductors, magnets, vdW systems), and disentangle intrinsic mechanisms (band structure, SOC) from extrinsic artifacts (vortex motion, geometry, heating)?

Summary: A review of nonreciprocal charge transport across bulk crystals, heterostructures, vdW systems, high-Tc superconductors, and magnets, with central focus on the superconducting diode effect. It separates intrinsic mechanisms (band structure, SOC, symmetry breaking) from extrinsic artifacts (vortex motion, geometry, heating) and flags the lack of a unified mechanistic picture.

Key Results: This is a review paper — it synthesizes existing literature rather than presenting new measurements. No specific benchmarks, datasets, or numerical results are cited in the abstract. The review catalogs the superconducting diode effect (different critical currents for opposite current directions) as the central application-driven focus.

Key Findings:

Superconducting diode effect — asymmetric critical currents under broken inversion + time-reversal symmetry — is the dominant application-driven theme
Many reported nonreciprocal signals conflate intrinsic band-structure effects with extrinsic artifacts (vortex dynamics, device asymmetry, thermal gradients), motivating stricter experimental controls
Symmetry breaking, spin-orbit coupling, and magnetic order are the three recurring ingredients across disparate material systems, but their relative weights remain contested

Technical Novelty: As a review, novelty lies in the synthesis: explicitly separating intrinsic (band-structure, inversion/time-reversal breaking, SOC) from extrinsic (vortex dynamics, device geometry, Joule heating) contributions, and bridging superconducting and magnetic nonreciprocity literatures that are often treated separately.

What's New: Unifying treatment spanning superconducting and magnetic nonreciprocity in one review, with explicit attention to extrinsic confounds that are often downplayed in individual experimental papers.

Extension Opportunities:

Build a standardized experimental protocol/benchmark suite to distinguish intrinsic nonreciprocity from heating and geometric artifacts — the review explicitly calls for better experimental standards
Design and fabricate a superconducting diode device integrated into a cryogenic logic circuit (e.g., rectifier for SFQ or qubit readout lines) using a vdW heterostructure platform
Develop a unified theoretical/simulation framework (e.g., Ginzburg-Landau + SOC + magnetic order) that predicts the sign and magnitude of the diode efficiency across material classes, testable against published datasets

Replicability: No code or dataset — it's a review. Reproducing the underlying experiments requires dilution refrigerators (mK), lock-in amplifiers for second-harmonic resistance measurements, and access to clean-room fabrication for vdW heterostructures or thin-film superconductors.

Research Gaps:

No unified theoretical framework predicting diode efficiency across material classes
Lack of community-standard experimental protocols to rule out heating, vortex, and geometric artifacts

🔥 GitHub Trending

1. NandhaKishorM/laya

⭐ 34999 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. whitecircle/halo

⭐ 842 stars | Python

Halo is an open-source framework built by White Circle for training large language and multimodal models

agent async-rl distributed-training fine-tuning grpo large-language-models

3. Calix-L/DanKS

⭐ 834 stars | Python

RL‑Empowered Small‑Scale Competitive Guandan Agent

card-games game-ai guandan pytorch reinforcement-learning

4. simplexai-labs/PolyStrata

⭐ 826 stars | C++

Run large AI models on your own PC: Qwen3.8-Flash-Next on Strata's engine, and GLM 5.3 Flash, DeepSeek V4 Flash, LFM2.5, Qwen3.6, Ornith 1.5, Gemma 4, Tiel Coder and MiniCPM-V on a second engine writt

anthropic-api cuda gemma gguf glm inference-engine

5. kanaworksai-hue/audio-studio-mcp

⭐ 584 stars | Python

Local AI BGM and sound effects for Claude Code and any MCP client. Stable Audio 3 on Apple Silicon (MLX) and Windows NVIDIA GPUs (PyTorch). No cloud, no API key.

apple-silicon audio-generation claude-code cuda game-audio mcp

6. Indras-Mirror/Xolotl

⭐ 458 stars | C++

Xolotl (formerly Strata-DSV4.1): run DeepSeek-V4.1-Flash, all 552B, as a local coding agent on one gaming PC (RTX 4090 + 90 GB RAM) at up to ~27 tok/s, 512K context, October 1017 builds included. Plus

cuda deepseek deepseek-v4 gguf glm llm-inference

7. ekzhang/resin

⭐ 401 stars | Python

🧱 Experimental library for building inference engines

cuda deepseek gpt-oss gpu inference inkling

8. alexturkey/voxcpm2-turkce-ses-tasarimi

⭐ 383 stars | Python

Turkish voice design with VoxCPM2 on a 4 GB GPU: describe a voice, pick a variant, read text in 30 languages. Patched audio.cpp, local UI in 30 languages, 14 ready-made voices.

cuda ggml low-vram speech-synthesis text-to-speech tts

9. xingjian-bai/opal

⭐ 338 stars | Python

OPAL: a reactive optimizer. A coding agent rewrites the optimizer while your model trains.

deep-learning llm-agents optimizer pytorch

10. fermionresearch/phonon

⭐ 328 stars | Python

Phonon: open speech recognition models (Phonon-2, Phonon-1) — CLI, CPU and CUDA images

asr cuda docker low-bit mlx pytorch

11. CerealAxis/Nola-Translator

⭐ 326 stars | TypeScript

Lightweight, powerful real-time speech translation with multilingual support and a multilingual UI.

asr-model audio-to-text crispasr livecaptions llama-cpp pytorch

12. LINJIANG12/video2book-courses

⭐ 297 stars | Unknown

六十七门真实网课的模块合辑全书 + 思维导图复习笔记 + 逐字稿，全部由 Video2Book 从课程音频重构：683 册模块全书、921 篇复习笔记、3,522 份逐字稿，从逐讲精读到考前速查，笔记支持 Markmap / XMind 脑图导入，可按课点播定制。

8086 ai assembly course-notes css3 database

13. Interactive-Active-AI/NextTransformer

⭐ 272 stars | Python

:sunrise: A comprehensive, cutting-edge open-source framework for next-generation Transformers, with compute-matched benchmarking and a platform for autonomous architecture research.

auto-research benchmark deep-learning language-models linear-attention llm

14. ceyyy427/finathink

⭐ 247 stars | Python

Finathink — Financial Research Optimizer for auditable, point-in-time research, mathematical learning, and constrained portfolio analysis.

backtesting data-provenance deep-learning finance financial-data financial-research

15. Fazmin/AILearningGuide

⭐ 246 stars | TypeScript

Interactive desktop app with 58 labs covering ML fundamentals, transformers, RAG, agents, fine-tuning and quantization and more.

ai ai-education deep-learning educational educational-resources educational-tool

Generated by Research Pulse on 2026-11-04 09:05