2026-09-02·Daily digest

🔬 Research Pulse

Daily Digest

September 02, 2026

🤖 AI

🧠 LLMs

1. From Production Traffic to Post-Training: Building a Self-Hosted LLM That Covers the Corporate Request Mix

Authors: Olga Tsymboi, Dmitrii Stoianov, Ramil Latypov... Published: 2026-09-01 | Citations: 0 arXiv | PDF

Research Question: How can enterprises consolidate fragmented LLM serving fleets (driven by data-residency constraints and continuous model adoption) onto a single self-hosted model that covers the full corporate request mix without sacrificing quality on instruction following, function-calling, and internal task distribution?

Summary: The paper presents a production-grounded recipe for consolidating enterprise LLM traffic onto a single self-hosted model by training separate GRPO experts per quality axis and merging them via two-stage SLERP, avoiding the reward interference of joint multi-objective optimization. The resulting model beats a ~7x larger baseline across three enterprise-relevant benchmarks and serves 116M requests/month, demonstrating that carefully targeted post-training can outperform brute-force scale for the corporate request mix.

Key Results: In non-reasoning mode, the consolidated model surpasses a ~7x larger baseline on in-house Arena (69.6 vs 65.8), instruction following (0.85 vs 0.83), and function-calling (0.79 vs 0.77), while lifting general dialogue benchmarks. It absorbs 50% of platform traffic (116M requests/month) at a fraction of serving cost. Each of the three GRPO experts exposed distinct failure modes: semantic collapse, over-calling, and verbosity hacking.

Key Findings:

Per-axis GRPO experts merged via two-stage SLERP outperform joint multi-objective training by avoiding cross-domain reward interference
The consolidated model beats a ~7x-parameter baseline on in-house Arena (69.6 vs 65.8), IF (0.85 vs 0.83), and function-calling (0.79 vs 0.77)
Each reward axis induces a characteristic reward-hacking failure (semantic collapse, over-calling, verbosity hacking) that requires an axis-specific mitigation

Technical Novelty: Training separate GRPO experts per quality axis (instruction following, function-calling, task distribution) to avoid cross-domain reward interference, then merging via two-stage SLERP — rather than joint multi-objective RL. Also novel: diagnosing distinct reward-hacking failure modes (semantic collapse, over-calling, verbosity hacking) per axis and applying domain-specific fixes.

What's New: Unlike prior work that either scales model size or optimizes a single objective, this paper introduces a production-informed decomposition: stratifying benchmarks from real traffic across 200+ apps, training axis-isolated GRPO experts, and composing them post-hoc via SLERP. The explicit taxonomy of per-axis reward-hacking pathologies is also a novel diagnostic contribution.

Extension Opportunities:

Extend the two-stage SLERP merging recipe to more than three axes (e.g., add safety, multilingual, or code-generation experts) and study whether merge quality degrades sublinearly with number of experts
Apply the production-traffic-stratified benchmark construction methodology to open-source model evaluation to build reproducible enterprise-grade benchmarks
Investigate adaptive routing between the merged model and specialized experts based on per-request classification, potentially recovering the individual-expert peak quality on hard queries

Replicability: The abstract does not mention public code or data release. Reproduction would require substantial compute for GRPO training on a large base model plus a proprietary corpus of production traffic from 200+ internal apps, making full reproduction infeasible outside a comparable enterprise setting. The methodology (per-axis GRPO + SLERP merge) is replicable in principle on open benchmarks.

Research Gaps:

No public release of the production-stratified benchmarks or SLERP merge weights, limiting external validation
Unclear how the recipe generalizes to reasoning-mode workloads or to domains beyond the three studied axes

2. HarnessDev: Can LLMs Create and Evolve Their Own Agent Harness?

Authors: Yuhao Wu, Jingyuan Zhang, Jiajun Shi... Published: 2026-09-01 | Citations: 0 arXiv | PDF

Research Question: Can LLMs create and evolve their own agent harnesses (execution infrastructure) rather than just perform tasks within human-designed harnesses, and how do we evaluate this capability?

Summary: HarnessDev is a benchmark that evaluates whether LLMs can build and iteratively refine their own agent execution infrastructure, not just solve tasks within one. Across 6 creator models and 4 domains, generated harnesses match humans on writing/ML but lag on code and search, and Evolution-based improvements transfer poorly across held-out tasks and different runtime models.

Key Results: Evaluated 6 creator LLMs across 4 domains and 5 downstream benchmarks (2,207 unique instances). Generated harnesses lag human-engineered references on code and search/research tasks, but match or exceed references on writing and ML experimentation. Evolution stage yields unstable gains that only partially transfer to held-out tasks, with performance strongly dependent on the runtime execution model.

Key Findings:

Generated harnesses substantially trail human-engineered references on code and on search/research domains
Generated harnesses match or exceed references on writing and ML experimentation, with wide token-cost variation
Evolution-stage gains are unstable and depend heavily on the runtime model, indicating poor cross-model transfer

Technical Novelty: Reframes agent evaluation from task-output performance to infrastructure-creation capability. Introduces a two-stage Creation + Evolution protocol with held-out hidden tasks, jointly measuring capability (task success) and efficiency (token cost) of the generated harness itself.

What's New: Prior agent benchmarks measure downstream task performance under a fixed harness; HarnessDev is the first to treat the harness itself as the unit of evaluation and measure both its capability and efficiency across creation and iterative evolution stages.

Extension Opportunities:

Extend HarnessDev to additional domains (e.g., robotics, data engineering, security) and measure cross-domain harness transfer
Study harness co-evolution where the creator and runtime models are jointly optimized, addressing the observed cross-model transfer limitation
Develop meta-learning approaches that learn harness design priors from successful human-engineered references to close the code/search gap

Replicability: Abstract doesn't mention code/data release. Reproducing would require API access to 6 LLMs across 2,207 downstream instances plus iterative Evolution runs — likely substantial API cost (thousands of dollars) but no specialized hardware.

Research Gaps:

Limited generalization of evolved harnesses across runtime models and to held-out tasks
Domain-dependent performance gap vs. human harnesses (especially code and search) is not yet explained mechanistically

👁️ Vision

1. TempCloze: Can Video-LLMs Identify the Missing Middle?

Authors: Wenqi Pei, Henry Hengyuan Zhao, Yilai Liu... Published: 2026-09-01 | Citations: 0 arXiv | PDF

Research Question: How can we evaluate visual temporal reasoning in Video-LLMs without linguistic shortcuts from option wording, answer correlations, or language priors that contaminate existing text-mediated temporal benchmarks?

Summary: TempCloze reframes video temporal reasoning evaluation as a cloze task: given start and end clips, models pick the true middle from four same-source distractors constructed along Semantic, Alignment, and Progression axes. Across 31 Video-LLMs, temporal Alignment (when events occur) emerges as the dominant failure mode, while semantic and local-progression reasoning are relatively intact.

Key Results: Introduces TempCloze, a video cloze benchmark with 1,521 filtered videos from 7 sources (mainly long-take and egocentric). Evaluated 10 proprietary + 21 open-source Video-LLMs across three distractor dimensions (Semantic/Alignment/Progression). Empirically demonstrates Alignment as the primary bottleneck — models handle semantic plausibility and local progression but fail at temporal 'when' reasoning. Includes error analysis on TempCloze-Mixed and TempCloze-Hard splits across four representative models, probing candidate order, context direction, visible span, frame density, and test-time scaling.

Key Findings:

Alignment (temporal 'when') is the primary bottleneck across both proprietary and open-source Video-LLMs
Models handle semantic plausibility and local event progression reasonably well — the gap is specifically temporal grounding
Behavioral sensitivity to candidate order, context direction, visible span, and frame density reveals shortcut behaviors even in the visual-only setup

Technical Novelty: The video-cloze framing (predict missing middle given start+end) removes text-mediation entirely, and the three-axis distractor construction (Semantic=what, Alignment=when, Progression=how) with same-source distractors sharing scenes/objects specifically ablates appearance cues — isolating temporal reasoning in a way prior MCQ-over-text benchmarks cannot.

What's New: First video-cloze benchmark that eliminates language mediation for temporal reasoning; introduces a principled three-axis distractor taxonomy (Semantic/Alignment/Progression) with same-source construction to control for appearance shortcuts.

Extension Opportunities:

Design a temporal-alignment-focused fine-tuning objective (e.g., contrastive middle-clip prediction) targeting the identified Alignment bottleneck, then measure lift on TempCloze-Hard
Extend the cloze format to multi-gap prediction (arbitrary missing spans) or to audio-visual cloze to test cross-modal temporal grounding
Build an automated distractor generation pipeline using the three-axis taxonomy (Semantic/Alignment/Progression) to scale the benchmark to new domains like surgical or industrial video

Replicability: Abstract does not mention released code or data explicitly. Reproduction would require access to the curated 1,521 videos + distractor sets, plus inference compute for 31 Video-LLMs (proprietary API costs + GPU inference for open-source models — moderate but not extreme scale since it's evaluation-only, not training).

Research Gaps:

No mechanism proposed to fix the identified Alignment bottleneck — the benchmark diagnoses but does not remediate
Coverage skewed to long-take and egocentric video; generalization to edited/multi-shot content or specialized domains (medical, scientific) is untested

🦾 ROBOTICS

1. A System for Fast, Resilient, and Adaptable Loco-Manipulation Behaviors on Humanoid Robots

Authors: Duncan Calvert, Luigi Penco, Dexton Anderson... Published: 2026-09-01 | Citations: 0 arXiv | PDF

Research Question: How can humanoid robots achieve fast, reliable loco-manipulation in human spaces without requiring lengthy retraining for each new task or environment variation?

Summary: The paper presents a runtime-editable behavior authoring system for humanoid loco-manipulation that combines Affordance Templates, behavior trees, and live perception editing. It demonstrates fast task execution (door in 34s, ball sorting in 45s) on Unitree H1-2 and Alex robots and argues that behavior architecture—not just learning—can be a primary driver of humanoid capability.

Key Results: Demonstrated six task variants on Unitree H1-2 and Alex humanoids: push door traversal in 34 seconds, sorting six colored balls in 45 seconds under human disturbance. Timed authoring sessions showed new loco-manipulation behaviors can be created from scratch or adapted in hours. Claims competitive performance vs. recent learned systems.

Key Findings:

Runtime-editable behavior authoring enables new loco-manipulation behaviors to be created or adapted in hours, not weeks
Object-centric Affordance Templates plus behavior trees can match learned systems on complex tasks like door traversal and disturbance-robust sorting
A continuously-synchronized operator interface allows monitoring, authoring, and repair without stopping the robot

Technical Novelty: Combines object-centric Affordance Templates with a behavior tree structure and runtime-editable perception via 'behavior scene' and 'primitive scene actions'—all synchronized live to a continuously connected operator interface. This runtime editability during execution differentiates it from static behavior trees and end-to-end learned policies that require retraining.

What's New: Positions classical behavior architecture (Affordance Templates + trees) as competitive with end-to-end learned humanoid policies, with the key differentiator being runtime editability of both behaviors and perception scenes during live operation.

Extension Opportunities:

Integrate LLM-based behavior tree generation to auto-author Affordance Templates from natural language task descriptions
Combine the runtime-editable behavior scene with learned perception modules (e.g., VLMs) to enable zero-shot object grounding for new affordances
Extend the whole-body controller to support contact-rich manipulation like tool use or bimanual assembly beyond simple pick/place primitives

Replicability: Abstract does not mention code/data release. Reproduction requires access to Unitree H1-2 or Alex humanoids (~$100K+ hardware), a whole-body controller supporting concurrent walking and manipulation, and the custom behavior authoring UI. Compute for runtime is modest (robot-local), but hardware barrier is high.

Research Gaps:

Scalability of hand-authored behaviors to open-world task diversity vs. learned generalization is not addressed
No systematic comparison of authoring time vs. data collection time for equivalent learned policies

2. Facet-0: A Robotic Foundation Model for Contact-Rich Precise Manipulation

Authors: Haoyuan Deng, Haichao Liu, Wenkai Guo... Published: 2026-09-01 | Citations: 0 arXiv | PDF

Research Question: How can robotic foundation models achieve sub-millimeter precision in contact-rich assembly tasks where prior approaches fail due to inadequate handling of contact dynamics, force feedback, and part-specific variability?

Summary: Facet-0 is a robotic foundation model that jointly predicts actions and their expected contact wrenches via flow matching, then uses RL with an Action-Wrench Critic to refine policies based on contact outcomes rather than just task progress. On sub-millimeter computer assembly, it achieves 82% success vs 15% for the strongest baseline, demonstrating that explicitly modeling force consequences is critical for precision manipulation.

Key Results: Facet-0 achieves 82% mean success on five sub-millimeter computer-assembly tasks vs 15% for the strongest baseline (5.5x improvement), with 0.5 mm placement accuracy and 50 ms command latency. Trained on ManuFacet-1K, a 1,000-hour force-synchronized corpus spanning three embodiments and multiple manufacturing cells.

Key Findings:

Explicitly coupling action prediction with wrench prediction via flow matching enables policies to reason about contact consequences before execution
A distributional critic that distinguishes contact outcomes (not just task progress) concentrates RL improvement on decisive interactions and yields 5.5x success improvement over baselines
A lightweight bounded actor over a frozen representation enables per-part on-robot adaptation while maintaining 50ms latency and 0.5mm precision

Technical Novelty: Joint action-wrench proposal via flow matching that predicts both the action chunk AND the induced future wrist-wrench profile together, coupled with a distributional Action-Wrench Critic that discriminates between motions with similar task progress but different contact outcomes. The bounded actor reusing frozen representations for on-robot adaptation while preserving non-commanded action-contact coupling via auxiliary wrench head is architecturally distinctive.

What's New: Prior VLA foundation models (RT-2, OpenVLA, Pi-0) treat manipulation as vision-language-action prediction without explicit force reasoning. Facet-0 makes wrench a first-class prediction target aligned with vision-language-kinematic representations, and introduces phase-aware, contact-selective RL credit assignment — a significant departure from pure imitation or task-progress-based RL for assembly.

Extension Opportunities:

Extend the joint action-wrench flow matching to bimanual or multi-arm assembly where coordinated contact profiles must be predicted across multiple end-effectors
Apply the Action-Wrench Critic framework to soft/deformable object manipulation where contact outcomes are even more variable (fabric assembly, cable routing)
Investigate transfer of the frozen representation + bounded actor pattern to non-manufacturing domains like surgical robotics or micro-manipulation where similar precision-contact tradeoffs exist

Replicability: Abstract does not mention code/data release. Reproduction would require: (1) the ManuFacet-1K dataset (1,000 hours of force-synchronized multi-embodiment data — likely proprietary), (2) three physical robot embodiments with wrist F/T sensors, (3) manufacturing cells for computer assembly, (4) significant GPU compute for VLA pretraining + RL post-training. Practically non-reproducible outside a well-funded robotics lab.

Research Gaps:

Generalization to non-rigid or highly-deformable parts and to novel manufacturing domains outside the training embodiments remains unaddressed
The tradeoff between frozen-representation adaptation speed and long-term skill acquisition is not fully characterized

3. Accelerating Reinforcement Learning via MPC Solver-Gradient Guidance for Weights-varying MPC

Authors: Baha Zarrouki, Arslan Thobani, Jasper Hoffmann... Published: 2026-09-01 | Citations: 0 arXiv | PDF

Research Question: How can online adaptation of MPC cost-function weights be learned efficiently when RL is sample-inefficient and gradient-based policy learning via differentiable MPC is biased under model mismatch?

Summary: The paper proposes SG-RL, a hybrid approach that augments PPO-based RL for online MPC cost-weight adaptation with bounded gradients derived from a differentiable MPC solver. By injecting solver-gradient guidance at four modular points in PPO (actor scaling, policy loss, advantage, value), it retains RL's unbiased closed-loop return objective while gaining the sample efficiency of gradient-based methods, validated on two autonomous racing platforms.

Key Results: On two full-scale autonomous racing platforms with intentional model mismatch, SG-RL matched PPO's best closed-loop return using up to 70.6% fewer samples, outperformed gradient-based policy learning baselines by at least 54% in closed-loop return, and generalized zero-shot to unseen environments.

Key Findings:

Solver-gradient guidance reduces PPO sample requirements by up to 70.6% while matching best closed-loop return.
SG-RL outperforms pure gradient-based policy learning (GB-PL) by at least 54% in closed-loop return under intentional model mismatch.
Learned weight-adaptation policies generalize zero-shot to unseen environments on full-scale racing platforms.

Technical Novelty: Rather than choosing between pure RL (unbiased but high-variance) or pure differentiable-MPC gradients (low-variance but biased), SG-RL uses bounded solver-derived gradients as auxiliary guidance while keeping sampled closed-loop return as the true objective, and instantiates this in PPO via four modular injection points (actor-update scaling, policy loss, advantage estimation, value learning).

What's New: Prior work forced a choice between sample-inefficient RL and bias-prone differentiable-MPC gradient methods; this work explicitly reframes the tradeoff as bias-variance and uses solver sensitivities as auxiliary, bounded guidance within RL rather than as the primary optimization signal, with four distinct PPO injection strategies.

Extension Opportunities:

Apply SG-RL to other safety-critical domains with differentiable MPC (quadrotor control, legged locomotion, manipulation) to test generality beyond autonomous racing.
Extend the four solver-gradient injection points (actor scaling, policy loss, advantage, value) to off-policy algorithms like SAC or TD3 where solver gradients could further reduce sample complexity.
Investigate adaptive weighting of the solver-gradient guidance term based on measured model-mismatch magnitude, so guidance strength decays as the surrogate becomes less reliable.

Replicability: No code/data availability is mentioned in the abstract. Reproduction would require differentiable MPC infrastructure, a PPO implementation, and access to full-scale autonomous racing platforms (or high-fidelity simulators) — likely a nontrivial hardware/simulation stack plus standard RL compute (single-GPU workstation scale).

Research Gaps:

Only demonstrated on autonomous racing; generality across broader MPC application domains (aerial, manipulation, process control) remains unproven.
The four solver-gradient injection algorithms are presented as modular but the abstract does not indicate which combinations dominate under what conditions, leaving guidance for practitioners unclear.

💻 COMPUTE

1. LLM Inference on IMC-NoC Architecture with Balanced Dataflow and Fine-Grained Parallelism

Authors: Yimin Wang, Yue Jiet Chong, Xuanyao Fong Published: 2026-09-01 | Citations: 0 arXiv | PDF

Research Question: How can we design a scalable hardware-software co-design architecture that overcomes memory bandwidth, compute density, and on-chip communication bottlenecks for LLM inference, particularly handling the heterogeneous data dynamicity (static weights vs. dynamic intermediate data) and the distinct compute intensities of prefill vs. decode phases?

Summary: LEAP is a hardware-software co-designed IMC-NoC architecture that fuses in-memory compute (static weights), near-memory compute (dynamic activations), and in-network compute (partial-sum reduction) into a unified fabric for LLM inference. A partitioning/mapping/scheduling framework with prefill-decode disaggregation dynamically reconfigures PEs per phase, delivering ≥1.52× throughput and 24.91× energy efficiency vs. GPUs.

Key Results: Demonstrated the LEAP architecture achieves ≥1.52× throughput improvement and 24.91× energy efficiency improvement over commercial GPU platforms for LLM inference. Validated a three-tier heterogeneous fabric (IMC PE for static weights, NMC PE for dynamic data, INC for partial-sum reduction) combined with prefill-decode disaggregation that dynamically reconfigures PE organizations.

Key Findings:

Heterogeneous three-tier compute (IMC/NMC/INC) is necessary because LLMs mix static weights with dynamic intermediate tensors that pure IMC cannot efficiently serve
On-chip data movement, not compute, is the dominant bottleneck at scale-up — motivating in-network reduction
Prefill and decode have sufficiently different arithmetic intensity that runtime PE reorganization (disaggregation) materially improves utilization
Achieves ≥1.52× throughput and 24.91× energy efficiency gains over commercial GPUs

Technical Novelty: The tri-tier integration of IMC + NMC + In-Network Compute (INC) on a single NoC fabric — where each layer is specialized (static weights, dynamic tensors, reductions) — combined with runtime prefill-decode PE reconfiguration. Prior IMC accelerators typically handle only static weights and struggle with dynamic intermediate data; prior disaggregation work is typically at the datacenter/GPU level, not intra-chip PE reorganization.

What's New: Unlike prior IMC accelerators focused on static-weight matrix-vector multiply, LEAP treats compute, memory, and communication as a single processing-communication fabric with in-network compute, and applies prefill-decode disaggregation as an intra-chip PE reconfiguration primitive rather than a cluster-level scheduling decision.

Extension Opportunities:

Extend the prefill-decode disaggregation scheduler to handle MoE (mixture-of-experts) models where expert routing adds another axis of dynamic data movement
Integrate speculative decoding or continuous batching (vLLM-style) into the partitioning/mapping framework to further exploit the reconfigurable PE organization
Explore KV-cache compression or quantization-aware mapping on the NMC tier to reduce dynamic memory pressure during long-context decode

Replicability: No code/data availability mentioned in the abstract. Reproduction would require an architectural simulator (likely custom cycle-accurate or extended from Timeloop/Accelergy/Booksim), IMC device models, and NoC modeling — a substantial systems-research effort rather than commodity compute.

Research Gaps:

No evaluation across model scales, sequence lengths, or batch regimes disclosed in the abstract — GPU speedup claims need context on which workloads
Analog IMC non-idealities (ADC overhead, device variation, accuracy degradation) are not addressed in the abstract

2. MakoXC: Rearchitecting DFT Exchange-Correlation with Matrix-Aligned and Knowledge-Organized Sparsity

Authors: Haozhi Han, Fusong Ju, Jing Bai... Published: 2026-09-01 | Citations: 0 arXiv | PDF

Research Question: How can DFT exchange-correlation (XC) evaluation—a cubic-scaling bottleneck—be restructured so that linear-scaling nearsightedness-induced sparsity produces regular, dense workloads suitable for modern AI accelerators (GPUs) rather than irregular sparse patterns that waste throughput?

Summary: MakoXC rearchitects the exchange-correlation step of DFT by reshaping electronic nearsightedness sparsity into dense, matrix-aligned computations that fully exploit GPU throughput. It delivers 67.8x speedup over standard XC and 4.7x over state-of-the-art linear-scaling baselines, enabling 1,231-atom DFT calculations in under five minutes on 64 GPUs.

Key Results: MakoXC achieves 67.8x average speedup over standard XC evaluation and 4.7x over state-of-the-art linear-scaling methods. Integrated into a production commercial DFT package, it scales XC evaluation to ubiquitin (1,231 atoms, def2-SVP basis) on 64 GPUs, completing end-to-end DFT in under 5 minutes.

Key Findings:

Nearsightedness-induced sparsity can be reorganized into accelerator-aligned dense clusters without sacrificing numerical correctness
67.8x average speedup vs standard XC and 4.7x vs SOTA linear-scaling methods
Enables ubiquitin-scale (1,231 atoms, def2-SVP) DFT with end-to-end runtime under 5 minutes on 64 GPUs

Technical Novelty: Three co-designed techniques: (1) Matrix-Aligned Cells that reorganize irregular nearsightedness interactions into dense clusters aligned to accelerator tile geometry; (2) Sparsity-Guided Activation that converts implicit sparsity into numerically-correct structured execution rather than dynamic sparse dispatch; (3) Kernel-Fused Pipeline unifying fragmented small kernels into a single compute-intensive path. Prior linear-scaling methods reduce asymptotic complexity but leave GPUs underutilized due to irregular workloads.

What's New: Prior linear-scaling XC methods lower asymptotic complexity but generate irregular sparse workloads that starve modern accelerators. MakoXC is novel in co-designing sparsity structure with hardware tile geometry, translating implicit sparsity into structured dense execution rather than relying on sparse kernels—achieving genuine linear scaling AND high hardware utilization simultaneously.

Extension Opportunities:

Apply the Matrix-Aligned Cells + Sparsity-Guided Activation pattern to other cubic/quartic-scaling quantum chemistry kernels (e.g., ERI evaluation, MP2, CCSD) that also exhibit nearsightedness sparsity
Port the Kernel-Fused Pipeline to non-NVIDIA accelerators (TPUs, AMD MI300, Cerebras) to test the generality of the accelerator-friendly reformulation
Combine MakoXC with ML-based XC functionals (neural network functionals) where the structured sparse execution path could accelerate training on large molecular datasets

Replicability: Abstract does not mention open-source release; integration is into a 'production-grade commercial DFT package,' suggesting proprietary code. Reproduction requires access to that package plus a multi-GPU cluster (up to 64 GPUs demonstrated). Independent reimplementation would be non-trivial without source.

Research Gaps:

Extension to correlated post-HF methods (MP2, coupled cluster) that share nearsightedness properties is not addressed
Behavior for metallic/delocalized systems where nearsightedness weakens is unclear; results focus on molecular systems

3. KORD: Breaking the Key-Generation Bottleneck in Dealerless FSS via Protocol--Hardware Co-Design

Authors: Yijing Peng, Lin Liu, Yujie Xue... Published: 2026-08-31 | Citations: 0 arXiv | PDF

Research Question: How can we eliminate the trusted dealer in Function Secret Sharing (FSS) without incurring the prohibitive communication and interaction-round overhead (which grows linearly with input bit-width) that makes existing dealerless key generation the dominant bottleneck in privacy-preserving computation?

Summary: KORD attacks the key-generation bottleneck in dealerless FSS by co-designing a cryptographic protocol with special-purpose chips that mutually attest to form a distributed root of trust, eliminating the trusted dealer while collapsing linear-round interaction into one round. Combined with a cross-key GGM-tree scheduler, it achieves 4–5 orders of magnitude less communication and shifts key generation from dominating (>96%) to a minor fraction (11.9%) of private ResNet-18 inference time.

Key Results: KORD demonstrates: (1) 7,633–70,274× reduction in per-key-generation communication vs. state-of-the-art distributed FSS across FSS building blocks; (2) throughput of 12.75 million 32-bit DPF keys/second at 204 MHz using only 21.5K LUTs post-route; (3) 99.8% AES lane utilization; (4) on private ResNet-18 inference, key generation's share of end-to-end time drops from >96% to 11.9%; (5) collapses multi-round interaction (linear in GGM depth) into a single round.

Key Findings:

Hardware-anchored mutual attestation can safely replace a trusted dealer while enabling a single-round key-generation protocol independent of GGM tree depth
Cross-key scheduling of independent GGM traversals sustains 99.8% AES lane utilization, yielding 12.75M 32-bit DPF keys/sec at 204 MHz with only 21.5K LUTs
Communication cost per key drops by 7,633–70,274× vs. the SOTA distributed FSS protocol, transforming key generation from the dominant end-to-end cost to a small fraction on realistic ML workloads

Technical Novelty: Prior dealerless FSS relies purely on cryptographic protocols with O(bit-width) rounds of interaction. KORD introduces a protocol–hardware co-design: dedicated chips establish a mutually-attested root of trust that acts as a distributed 'virtual dealer,' enabling a restructured single-round protocol independent of GGM depth. The cross-key scheduling that interleaves independent GGM-tree traversals to keep AES lanes ~100% utilized is also new — prior software protocols cannot exploit this pipelining.

What's New: First work to break the dealer dependency in FSS through protocol–hardware co-design rather than purely cryptographic means, using mutual attestation between dedicated chips as a shared trust boundary and exploiting hardware parallelism via cross-key GGM scheduling — a combination absent from prior dealerless FSS literature.

Extension Opportunities:

Extend the co-designed chip + mutual-attestation root-of-trust approach to other MPC primitives beyond FSS (e.g., garbled circuits, OT extension, silent VOLE) where trusted setup is similarly a bottleneck
Adapt KORD's cross-key GGM-tree scheduling to GPU/ASIC accelerators for cloud-scale privacy-preserving ML inference serving multiple tenants concurrently
Integrate KORD hardware into an end-to-end private LLM inference stack (beyond ResNet-18) to see whether transformer workloads with different FSS gate mixes still benefit, and quantify the trust-model implications for regulated deployments

Replicability: The abstract does not mention open-sourced code, RTL, or datasets. Reproduction would require FPGA/ASIC design expertise, access to the special-purpose chip design (or equivalent RTL), an AES-hardened FPGA (post-route figures suggest a mid-range FPGA), and an FSS/MPC software stack for ResNet-18 evaluation. Without published artifacts, full replication is likely infeasible outside the authors' lab; partial protocol-level replication in simulation is plausible.

Research Gaps:

Trust and threat model of the mutually-attested chips (supply-chain, side-channel, and fault-injection resistance) is not fully characterized in the abstract
Generalization beyond FSS building blocks and ResNet-18 — e.g., to transformer-based private inference or multi-party (>2) settings — is unaddressed

⚡ ENERGY

1. Science Of Nuclear Fusion: Insights and Ideas

Authors: Johann Rafelski, Andrew Steinmetz Published: 2026-09-01 | Citations: 0 arXiv | PDF

Research Question: How can advances across plasma physics, laser technology, and particle physics enable smaller-scale, higher-energy-density fusion systems — particularly aneutronic pathways — that overcome the practical limits (neutron damage, tritium scarcity, size) of conventional D-T magnetic/inertial confinement fusion?

Summary: A survey and perspective piece arguing that aneutronic fusion pathways — especially those using 3He, 11B, and Be — enabled by laser-plasmon field enhancement and muon catalysis, offer credible small-scale alternatives to D-T tokamaks. The authors introduce two original reaction schemes (self-sustaining p-11B-14N and Be+3He) and reframe 'limitless energy' claims against real isotopic abundances, motivating aneutronic fusion for terrestrial power and space propulsion.

Key Results: The paper is a survey/perspective rather than an experimental result, so it does not report new measurements. It qualitatively synthesizes: (1) recent laser-driven proton-boron (p-11B) fusion achievements as demonstrated by prior groups; (2) natural isotopic abundance arguments showing '3He and 11B constraints make 'limitless energy' claims context-dependent; (3) a proposed self-sustaining p-11B-14N aneutronic reaction cycle; (4) a proposed Be + 3He aneutronic target concept; and (5) analyses of plasmonic field-enhancement and muon-catalyzed fusion pathways.

Key Findings:

Aneutronic reactions (p-11B, d-3He, Be+3He) sidestep the neutron-damage and tritium-breeding burdens that dominate D-T reactor engineering, making them attractive for civilian and space-propulsion use
Natural isotopic abundance of 3He and 11B is a real constraint on the 'limitless energy' narrative — fuel-cycle sustainability, not just plasma physics, gates scalability
A proposed p-11B-14N reaction cycle could yield self-sustaining, nearly fully aneutronic operation, extending recent laser-driven p-11B results
Plasmonic opto-electric field enhancement and muon-catalyzed fusion are credible table-top routes to lowering the effective Coulomb barrier

Technical Novelty: Two proposed reaction schemes are the paper's original contributions: (1) extending laser-driven p-11B into a self-sustaining, nearly fully aneutronic p-11B-14N (proton-boron-nitride) cycle, and (2) a beryllium + 3He aneutronic target mixture. Framing 3He as the pivotal fuel across multiple table-top concepts is also a synthesizing contribution.

What's New: Prior surveys focus on tokamak/ICF roadmaps; this paper deliberately elevates table-top and aneutronic schemes, introduces two new reaction concepts (p-11B-14N cycle, Be+3He target), and integrates plasmonic field enhancement, muon catalysis, and isotope-supply realism into a single fusion-physics narrative that also flags space propulsion as a near-term aneutronic application.

Extension Opportunities:

Simulate the proposed p-11B-14N cycle in a 1D plasma/reaction-network code (e.g., FLASH or a custom Bateman-equation solver) to quantify ignition thresholds, energy gain, and neutron leakage vs. pure p-11B
Build a plasmonic nanostructure design study (FDTD in Lumerical/MEEP) targeting the field enhancement factors required to lower Coulomb-barrier tunneling for p-11B or d-3He at table-top laser intensities
Develop an isotope-supply techno-economic model for 3He (lunar regolith mining vs. tritium-decay harvesting vs. Be+3He breeding) to test whether the Be/3He scheme is materials-limited before it is physics-limited

Replicability: No code, datasets, or experimental apparatus are released — this is a theoretical/survey paper. 'Reproducing' the conceptual proposals would require plasma reaction-network simulations (modest CPU, single-node) plus, for validation, access to petawatt-class laser facilities (ELI, NIF-scale) for p-11B experiments, which is not table-top in practice.

Research Gaps:

No quantitative reactor-scale modeling of the proposed p-11B-14N or Be+3He cycles — ignition criteria, gain, and confinement times are not derived
3He supply chain (lunar mining, breeding pathways) is invoked but not analyzed as a hard constraint on any of the proposed schemes

2. Accelerating dynamic simulations of photoexcited materials and their evolution by electron-informed machine learning

Authors: Yunzhe Jia, Fankai Xie, Yunfei Bai... Published: 2026-09-01 | Citations: 0 arXiv | PDF

Research Question: How can we perform accurate, large-scale nonadiabatic excited-state molecular dynamics simulations of photoexcited materials without the prohibitive cost of rt-TDDFT, while still capturing excited-state electron-phonon couplings and photoinduced phonon anharmonicity that standard MLMD misses?

Summary: The paper introduces EMLMD, a machine-learning molecular dynamics framework whose potential is conditioned on nonequilibrium electronic descriptors (notably electron temperature) calibrated against rt-TDDFT, giving first-principles accuracy for excited-state atomic evolution at MLMD cost. Applied to photoexcited bismuth and selenium, it resolves coherent-vs-thermal phonon competition during a photoinduced phase transition and both atomic-scale and global dynamics of photoamorphization.

Key Results: The authors demonstrate EMLMD by calibrating nonequilibrium electronic information (e.g., electron temperature) from rt-TDDFT benchmarks to reconstruct excited-state PES. They resolve time- and momentum-resolved phonon dynamics in photoexcited bismuth, revealing competition between coherent and thermal phonons during photoinduced phase transition, and simultaneously capture atomic-scale microscopic dynamics and global structural rearrangement in selenium photoamorphization — two systems previously intractable at this scale/accuracy combination.

Key Findings:

An electron-temperature-conditioned ML potential trained from rt-TDDFT reproduces excited-state PES with electron-phonon coupling and photoinduced anharmonicity intrinsically included
In photoexcited bismuth, EMLMD resolves the competition between photogenerated coherent phonons and thermal phonons driving the photoinduced phase transition
In selenium photoamorphization, EMLMD simultaneously captures atomic-scale microscopic dynamics and global structural rearrangement — a scale gap standard methods cannot bridge

Technical Novelty: Prior MLMD models fit ground-state PES only; excited-state ab initio MD (rt-TDDFT/Ehrenfest) is accurate but limited to small cells and short times. EMLMD is novel in rigorously calibrating an electronic-temperature (or nonequilibrium electronic descriptor) from rt-TDDFT and conditioning the ML potential on it, so the learned PES natively reproduces excited-state electron-phonon coupling and photoinduced anharmonicity rather than treating photoexcitation as a perturbation on a ground-state model.

What's New: It closes the accuracy-vs-scale gap between rt-TDDFT and standard MLMD by making the ML potential an explicit function of a rigorously calibrated nonequilibrium electronic state, rather than a ground-state-only surrogate.

Extension Opportunities:

Apply EMLMD to photovoltaic absorber materials (perovskites, chalcogenides) to study hot-carrier cooling and defect formation under realistic excitation densities
Couple EMLMD with explicit electron dynamics (surface hopping or Ehrenfest) to handle systems where a single electron-temperature approximation breaks down, e.g., strongly localized excitons
Build a transferable foundation model of excited-state PES across many materials by training on a curated rt-TDDFT dataset, analogous to MACE-MP or M3GNet but electron-temperature-conditioned

Replicability: The abstract does not mention a public code or dataset release. Reproduction would require rt-TDDFT benchmark trajectories (GPU-accelerated DFT code such as Quantum ESPRESSO/SIESTA/PWmat, tens of thousands of CPU/GPU-hours for a few hundred atoms) plus MLMD training infrastructure (Deep Potential, NequIP, or similar). EMLMD production runs should then be tractable on a single GPU node.

Research Gaps:

The framework assumes an electron-temperature description of the excited state; strongly non-thermal or highly localized excitations may require richer electronic descriptors
Transferability across materials families and excitation regimes, and coupling to explicit nonadiabatic transitions (surface hopping) beyond mean-field PES, remain open

3. Geometry-Controlled Magnetic and Electronic Landscapes in Anisotropic van der Waals Materials

Authors: Maciej Śmiertka, Ewelina Cybula, Oliwia Janikowska... Published: 2026-09-01 | Citations: 0 arXiv | PDF

Research Question: How can one create programmable magnetic and electronic landscapes in van der Waals materials without altering composition, strain, gating, or heterostructure assembly?

Summary: The paper introduces 'geometronics,' a design principle in which shaped substrates locally reorient anisotropic van der Waals crystals so a uniform external stimulus generates programmable landscapes. Using bilayer CrSBr on an inverted pyramidal nanoindentation, the authors realize coexisting AFM and FM phases in a single crystal and a 10–12 meV switchable excitonic potential well.

Key Results: Demonstrated that bilayer CrSBr transferred onto an inverted pyramidal nanoindentation produces coexisting antiferromagnetic and ferromagnetic phases within one continuous crystal under a uniform external magnetic field, yielding a switchable excitonic potential well of 10–12 meV visualized via spatially resolved spectroscopy.

Key Findings:

Substrate topography can locally rotate an anisotropic vdW crystal so a homogeneous magnetic field produces spatially distinct magnetic phases.
Bilayer CrSBr on an inverted pyramid hosts coexisting antiferromagnetic and ferromagnetic domains within one continuous crystal.
The resulting magnetic landscape creates a switchable excitonic potential well of 10–12 meV, directly imaged by spatially resolved spectroscopy.

Technical Novelty: Introduces 'geometronics' — using substrate topography to locally reorient an anisotropic crystal so a uniform external field produces spatially varying phases — instead of the standard toolbox of composition, strain, gating, or heterostructure engineering.

What's New: Prior work engineered vdW electronic structure via composition, strain, gating, or heterostructures; this work turns substrate geometry itself into the tuning knob by exploiting intrinsic crystal anisotropy.

Extension Opportunities:

Apply the geometronics concept to other anisotropic vdW materials (e.g., NiPS3, CrPS4, MnPS3) to map their orientation-dependent phase diagrams on shaped substrates.
Engineer arrays of nanoindentation geometries (cones, ridges, gratings) to create programmable 2D lattices of excitonic potential wells for exciton trapping or moiré-like superlattices.
Couple the geometry-induced excitonic potential wells to optical cavities to explore polariton confinement or build reconfigurable exciton-based logic devices.

Replicability: Abstract does not mention code/data availability. Reproduction requires CrSBr crystal growth, FIB or nanolithography to fabricate inverted pyramid nanoindentations, dry-transfer capability for bilayer exfoliation, and cryogenic magneto-optical spatially resolved spectroscopy (µ-PL) — substantial specialized experimental infrastructure.

Research Gaps:

Generality across other anisotropic vdW magnets and non-magnetic anisotropic systems remains unproven.
Scaling to dense, deterministic arrays and quantifying reproducibility, thermal stability, and switching dynamics are not addressed.

🏥 HEALTHCARE

1. FLaG: Frequency-Domain Latent-attention Gated Pooling for Token Aggregation

Authors: Kewei Li, Rongying Zhang, Xueli Wang... Published: 2026-09-01 | Citations: 0 arXiv | PDF

Research Question: Most token pooling methods for aggregating encoder outputs into fixed-dimensional representations operate only in the original token space, ignoring frequency-domain structure that could reveal complementary signals for downstream prediction across modalities.

Summary: FLaG is a plug-in token-aggregation module that transforms encoder outputs into the Fourier domain, summarizes the rFFT spectrum with learnable latent queries, applies a sample-conditioned channel gate, and reconstructs modulated tokens for pooling. It delivers consistent gains over standard pooling across protein (AMP), vision (CIFAR), and language (RoBERTa) benchmarks, demonstrating that a frequency-domain inductive bias transfers across modalities.

Key Results: FLaG achieves the best macro-averaged Spearman correlation, RMSE, and Recall@50 across four AMP backbone-species settings (ESM2), the highest top-1 accuracy on CIFAR-10 with ResNet18, and best mean results on 5 of 7 RoBERTa language metrics (mean pooling still wins on STSBenchmark). Mechanistic analysis shows low-frequency prediction sensitivity across most encoder layers with increased high-frequency sensitivity in the final layer.

Key Findings:

A single frequency-domain pooling architecture transfers across protein, vision, and text encoders without modality-specific tuning
Predictions depend mostly on low-frequency spectral components across encoder layers, but the final layer shifts toward higher-frequency sensitivity
The residual gate broadly amplifies spectral channels while preserving low-frequency energy dominance, and latent cross-attention allocates spectral capacity in a sample- and species-specific way
Gains are backbone- and task-dependent — mean pooling still wins on STSBenchmark, indicating FLaG is a bias not a universal upgrade

Technical Novelty: Combines four ideas into one plug-in pooler: (1) rFFT-based re-expression of encoder tokens with concatenated real/imaginary channels, (2) latent-query cross-attention over spectral tokens, (3) sample-conditioned residual channel gating that modulates spectral channels, and (4) inverse reconstruction back to token space before final aggregation. Prior frequency-domain modules typically replace attention layers rather than acting as a drop-in pooling head.

What's New: Prior work has explored Fourier-based token mixing (FNet, GFNet) as replacements for attention within encoders. FLaG instead uses the frequency domain specifically for the pooling/aggregation step, combining latent-query attention over spectra with a residual channel gate — and validates transferability across three unrelated modalities with the same module.

Extension Opportunities:

Apply FLaG to audio/speech encoders (Wav2Vec, Whisper) where frequency-domain priors are natively meaningful and likely yield larger gains
Replace rFFT with learnable transforms (wavelets, DCT, or learned unitary bases) to test whether Fourier is optimal or just a convenient inductive bias
Extend to multimodal fusion by aggregating spectra from different modality encoders in a shared latent-query space with cross-modal channel gating

Replicability: Code and data are available at github.com/Kewei2023/AMPCliff/tree/FLaG plus supplementary at healthinformaticslab.org/supp/. Backbones used (ESM2, ResNet18, RoBERTa) are standard and reproducible on a single modern GPU; AMP, CIFAR-10/100, and the RoBERTa language tasks are all public benchmarks.

Research Gaps:

No ablation reported here isolating the contribution of latent attention vs channel gate vs rFFT re-expression
Unclear why mean pooling beats FLaG on STSBenchmark — no theory for when frequency bias helps vs hurts, limiting principled adoption

2. Science sandboxes measure the scientific capability of AI agents

Authors: Arya S. Rao, Rodrigo I. Castro, Sager J. Gosai... Published: 2026-08-31 | Citations: 0 arXiv | PDF

Research Question: How can we systematically measure whether AI agents actually understand the underlying rules of scientific systems, rather than just optimizing metrics without genuine scientific reasoning?

Summary: Introduces 'science sandboxes,' a framework that evaluates AI agents through iterative experiment-feedback-hypothesis cycles across environments ranging from physical experiments to invented rule systems. Instantiated in regulatory genomics and protein fitness settings, the framework reveals that frontier agents often optimize metrics without learning the underlying rules, and their reasoning degrades on systems that violate familiar biological priors.

Key Results: The authors instantiated the science sandbox framework in two biological domains (regulatory genomics and protein fitness prediction) and demonstrated that frontier agents can successfully optimize quantitative metrics without grasping underlying rules. Scientific reasoning notably deteriorated when agents faced systems with rules outside familiar biological priors. Specific numerical benchmarks are not provided in the abstract.

Key Findings:

Frontier agents can hit quantitative optimization targets while failing to infer the rules that govern the system
Scientific reasoning quality collapses when the sandbox's rules fall outside biological priors the agents were pretrained on
A wet/damp/dry spectrum of environments enables controlled comparison of empirical vs. abstract scientific capability in a single evaluation framework

Technical Novelty: The 'wet/damp/dry' spectrum — organizing sandbox environments by empirical verifiability (physical experiments → data-trained predictors → invented rule systems) — combined with a unified experimentation-feedback-hypothesis loop that evaluates reasoning quality, not just endpoint metrics. Prior agent benchmarks typically score final task success without probing rule discovery.

What's New: Shifts agent evaluation from single-shot task success toward iterative rule-discovery, and formalizes a verifiability spectrum (wet/damp/dry) that lets the same agent loop be graded on both metric performance and qualitative scientific reasoning.

Extension Opportunities:

Instantiate science sandboxes in non-biological domains (chemistry, materials science, physics) to test whether the observed failure modes generalize beyond biology
Build training curricula using sandboxes with rules that deliberately violate common priors, to improve agents' rule-inference capability
Develop automated evaluation harnesses that score qualitative scientific reasoning traces alongside quantitative performance, usable as a benchmark for new frontier models

Replicability: The abstract does not mention code or data release. Reproduction would likely require access to the regulatory genomics and protein fitness predictors (some public, e.g., Enformer, ESM-based models) plus API access to frontier LLM agents; compute would be modest for the sandbox loop itself but non-trivial for many agent trajectories.

Research Gaps:

No demonstrated method for improving agents' rule-inference when priors mislead them
Coverage is limited to two biological domains — generality across scientific disciplines is untested

3. PopPert: Population-level Joint-Distribution Modeling for Single-Cell Perturbation Prediction

Authors: Handong Wang, Jiaxin Qi, Haochen Feng... Published: 2026-09-01 | Citations: 0 arXiv | PDF

Research Question: How can we accurately predict transcriptional responses to perturbations from single-cell RNA-seq data when control and perturbed cell populations are inherently unpaired (since scRNA-seq destroys measured cells), given that existing methods incorrectly assume cell-to-cell correspondence?

Summary: PopPert reframes single-cell perturbation prediction as learning population-level joint gene expression distributions rather than per-cell mappings, sidestepping the false cell-correspondence assumption made by prior methods on inherently unpaired scRNA-seq data. It uses a low-rank Gaussian Copula to capture cross-gene dependencies and predicts how perturbations shift distribution parameters, achieving superior performance across genetic and chemical perturbation benchmarks on DE recovery, effect estimation, and distribution matching.

Key Results: PopPert was evaluated across multiple single-cell benchmarks spanning both genetic and chemical perturbations, achieving superior overall performance on three axes: (1) differential expression recovery, (2) perturbation effect estimation, and (3) population-level distribution matching. The abstract does not report specific numeric deltas vs. baselines, but claims consistent gains across the benchmark suite.

Key Findings:

Population-level joint distribution modeling outperforms cell-level correspondence-based methods across multiple benchmarks for both genetic and chemical perturbations
Low-rank Gaussian Copulas are sufficient to capture the cross-gene co-expression structure needed for realistic perturbation response prediction
The distributional framework retains the ability to sample synthetic perturbed single cells, so it does not sacrifice single-cell-level utility despite operating at the population level

Technical Novelty: Reframing perturbation prediction from single-cell-to-single-cell mapping to population-level joint distribution parameter prediction, combined with a low-rank Gaussian Copula to explicitly model cross-gene statistical dependencies. This eliminates the incorrect cell-correspondence assumption baked into methods like scGen/CPA/GEARS while still permitting sampling of synthetic single cells from the learned joint distribution.

What's New: Prior work (scGen, CPA, GEARS, etc.) predicts perturbation responses cell-by-cell and implicitly assumes control cells can be matched to perturbed counterparts — an assumption violated by scRNA-seq's destructive measurement. PopPert is the first to explicitly parameterize the population-level joint distribution and predict perturbation-induced parameter shifts, aligning the modeling formulation with the actual unpaired data-generating process.

Extension Opportunities:

Replace the low-rank Gaussian Copula with more expressive dependency models (e.g., vine copulas, normalizing flows, or diffusion-based joint distributions) to capture nonlinear/non-Gaussian gene co-expression
Extend the framework to multi-modal perturbation prediction (scATAC-seq, spatial transcriptomics, proteomics) using the same population-distribution paradigm to model joint chromatin+expression response
Use the synthetic perturbed cell sampling capability to build an in-silico screening pipeline for drug discovery — predicting distributional responses to unseen compound/genetic perturbations and prioritizing candidates by distributional divergence from disease states

Replicability: Code is publicly available at https://github.com/whd1125/PopPert. Datasets are established public single-cell perturbation benchmarks (likely Norman, Replogle, sci-Plex families). Compute requirements are not specified in the abstract but Gaussian Copula + low-rank parameterization on population-level statistics is substantially lighter than per-cell generative models — likely single-GPU trainable.

Research Gaps:

Abstract does not quantify improvements with specific numbers or name specific baselines beaten
Gaussian Copula assumes underlying Gaussian marginals after transformation — may struggle with highly zero-inflated or multi-modal gene expression distributions common in scRNA-seq

🔬 MATERIALS

1. Understanding the superconducting proximity effect in semiconductors through quantum oscillations

Authors: Milo Coombs, Teun A. J. van Schijndel, Yu Wu... Published: 2026-09-01 | Citations: 0 arXiv | PDF

Research Question: How can one measure the normal-state semiconductor parameters (density, mass, g-factor, mobility, subband occupation) of a buried InAs quantum well beneath a superconducting film, when the metal shunts DC transport and prevents conventional Hall/magnetotransport characterization of the hybrid?

Summary: The paper introduces a shunt-corrected Shubnikov–de Haas / Dingle analysis that recovers the normal-state parameters of a buried InAs quantum well underneath six different superconducting films (Al, Sn, V, Nb, Ta, Re), which conventional transport cannot access because the metal short-circuits the semiconductor. It shows the buried well's mass and g-factor are preserved to within 10%, quantum lifetimes are not degraded (and are enhanced under Al and Sn), and each metal contributes a single interface subband whose hybridization is bounded to 2–4 meV.

Key Results: Using Shubnikov-de Haas oscillations with a shunt-aware Dingle analysis, the authors extract normal-state parameters of a buried InAs quantum well under six superconductors (Al, Sn, V, Nb, Ta, Re). Findings: (1) every metal introduces one interface subband, occupation clustering into two classes; (2) buried-well effective mass and g-factor change by <10%; (3) no film shortens the quantum lifetime — Al and Sn actually lengthen it; (4) interface-subband hybridization is bounded to 2–4 meV.

Key Findings:

Every superconducting film (Al, Sn, V, Nb, Ta, Re) adds exactly one interface subband, with occupations partitioning into two distinct classes.
The buried InAs well's effective mass and g-factor are unchanged to within 10% across all six metals, and quantum lifetimes are preserved or enhanced (notably by Al and Sn).
Quantum-lifetime measurements bound the semiconductor–interface-subband hybridization to 2–4 meV, providing the normal-state input that tunneling spectroscopy renormalizes but cannot itself measure.

Technical Novelty: A Dingle analysis of SdH oscillations that explicitly accounts for the parallel conduction (shunt) of the overlying superconducting film in its normal state, allowing quantitative extraction of buried-semiconductor quantum lifetime, mass, g-factor and subband occupation — parameters previously inaccessible because the metal short-circuits standard magnetotransport.

What's New: Prior work on superconductor–semiconductor hybrids relied on tunneling spectroscopy of the proximitized state, leaving normal-state hybrid parameters unmeasured because the metal shunts DC transport. This paper is the first systematic, cross-material extraction of density, mass, g-factor, mobility, quantum lifetime, and subband structure of the buried semiconductor under a superconductor via shunt-aware SdH analysis.

Extension Opportunities:

Apply the same shunt-corrected SdH/Dingle methodology to other 2DEG platforms (InSb, Ge/SiGe, HgTe) under superconductors to build a cross-material parameter database for Majorana/gatemon device design.
Correlate the measured 2–4 meV hybridization and interface-subband class with induced superconducting gap and hard-gap quality from tunneling spectroscopy on identical stacks, closing the loop between normal-state and proximitized parameters.
Extend to gate-tunable and finite-magnetic-field regimes to map how hybridization, g-factor renormalization, and subband occupation evolve toward the topological phase transition in nanowire/2DEG Majorana devices.

Replicability: No code/data availability stated in the abstract. Reproducing requires MBE-grown InAs quantum wells with in-situ deposited Al/Sn/V/Nb/Ta/Re films, dilution-refrigerator magnetotransport up to fields sufficient to resolve SdH oscillations above each film's upper critical field, plus standard Dingle-plot analysis — accessible to any mesoscopic-transport group with hybrid-materials growth capability.

Research Gaps:

Direct microscopic mechanism distinguishing the two observed interface-subband occupation classes across the six metals is not resolved.
The bound on hybridization (2–4 meV) is set by mobility-limited uncertainty; a tighter determination and its explicit connection to induced-gap magnitude in the same devices remains open.

2. Resolving unconventional gap structure in kagome superconductors with hybrid microwave circuits

Authors: Yejin Lee, Haolin Jin, Sushmita Chandra... Published: 2026-09-01 | Citations: 0 arXiv | PDF

Research Question: How can one probe the superconducting pairing symmetry (gap structure) of kagome superconductor thin flakes, which are too small and delicate for conventional electrodynamic measurements?

Summary: The paper introduces a hybrid microwave circuit platform that noninvasively probes the electrodynamic response of fragile kagome superconductor flakes via contactless coupling, overcoming the volume/damage limitations of conventional techniques. By suppressing TLS losses through device geometry engineering, the authors resolve a linear-in-temperature superfluid density at low T, providing evidence for a nodal superconducting gap structure in kagome superconductors.

Key Results: The authors integrated superconducting flakes into hybrid microwave resonator circuits via contactless coupling and measured the temperature-dependent superfluid density. They observed linear-in-T behavior at low temperatures, which is a signature consistent with a nodal superconducting gap structure (as opposed to a fully gapped s-wave state). They also demonstrated that engineered device geometry successfully suppressed parasitic two-level-system (TLS) losses that typically dominate microwave circuit dissipation, allowing isolation of the intrinsic material response.

Key Findings:

Contactless microwave coupling enables measurement of superfluid response in thin superconducting flakes while preserving structural integrity
Device geometry engineering can suppress parasitic two-level-system losses sufficiently to isolate the intrinsic material electrodynamic response
Kagome superconductor flakes exhibit linear-T superfluid density at low temperatures, consistent with a nodal (not fully gapped) pairing symmetry

Technical Novelty: The core novelty is a noninvasive, contactless coupling scheme that integrates delicate superconducting flakes into microwave resonator circuits without contacts or patterning, combined with device geometry engineering that suppresses TLS losses to a level where the intrinsic superfluid response can be resolved. Prior electrodynamic gap probes required bulk crystals or invasive lithography incompatible with fragile flakes.

What's New: Bridges quantum-device engineering (superconducting circuit QED techniques for TLS mitigation) with correlated-matter physics, enabling a class of electrodynamic measurements previously inaccessible for exfoliated/thin-flake superconductors and delivering symmetry-resolving evidence in the kagome family.

Extension Opportunities:

Apply this hybrid microwave platform to other fragile 2D/flake superconductors (e.g., twisted bilayer graphene, magic-angle systems, other kagome variants like CsV3Sb5 family members) to systematically map pairing symmetries across correlated materials
Combine the contactless microwave probe with in-situ tuning knobs (gate voltage, strain, magnetic field, pressure) to map how the nodal gap structure evolves with control parameters and correlate with charge-density-wave or topological transitions
Extend the resonator architecture to multi-mode or multi-frequency measurements to extract the full complex conductivity spectrum, enabling direct identification of node locations (line vs point nodes) and pairing channel (d-wave, p-wave, etc.)

Replicability: The abstract does not mention code or data availability. Reproduction would require a dilution refrigerator (mK temperatures), microwave measurement chain (VNA, cryogenic amplifiers, circulators), cleanroom fabrication of superconducting resonators (likely NbTiN or Al on low-loss substrate), and access to high-quality exfoliated kagome superconductor flakes (e.g., CsV3Sb5). Compute needs are modest — mainly resonator simulation (Sonnet/HFSS) and superfluid density fitting.

Research Gaps:

The pairing symmetry of kagome superconductors has remained experimentally elusive due to inability of standard electrodynamic probes to handle small, fragile flake samples
Microwave circuit measurements of intrinsic material response have been limited by parasitic TLS losses that mask the signal of interest

3. Thickness-Dependent Orbital-to-Spin Torque Signatures in Cr/Gd/Co Thin Films

Authors: Tiago de Oliveira Schneider, Michel Heidkamp, Luana Caron... Published: 2026-09-01 | Citations: 0 arXiv | PDF

Research Question: How does a Gd interlayer of varying thickness (0-5 nm) between Cr and Co mediate orbital-to-spin torque conversion, and does stack ordering affect damping-like versus field-like torque contributions?

Summary: The paper investigates orbital-torque generation in Cr/Gd/Co and inverted Co/Gd/Cr multilayers as a function of Gd interlayer thickness, using harmonic Hall measurements combined with structural and magnetometric characterization. It demonstrates that the field-like torque reverses sign upon stack inversion while the damping-like torque does not, and interprets thickness dependence using both a naive-layer and an interfacial-alloy model to account for observed Gd-Co intermixing.

Key Results: Using harmonic Hall measurements on Cr(10nm)/Gd(t)/Co(3nm)/TaOx and inverted Co(3nm)/Gd(t)/Cr(10nm) stacks with t_Gd from 0 to 5 nm, the authors extracted E-field-normalized damping-like (ξ_DL^E) and field-like (ξ_FL^E) torque efficiencies as a function of Gd thickness. Cross-sectional chemical mapping revealed pronounced Gd-Co intermixing, and temperature-dependent magnetic compensation points were observed. Critically, reversing the stack order flipped the sign of the FL torque but not the DL torque.

Key Findings:

Field-like torque efficiency changes sign upon reversing the Cr/Gd/Co stack order, while damping-like torque efficiency retains its sign
Cross-sectional chemical mapping reveals pronounced Gd-Co intermixing, motivating an alloy-based interpretation of the torque-vs-thickness data alongside the standard naive-layer model
Magnetometry shows temperature-dependent ferrimagnetic compensation points consistent with intermixed Gd-Co regions rather than a discrete Gd layer

Technical Novelty: Combines harmonic Hall torque metrology with cross-sectional chemical mapping and an explicit 'alloy model' accounting for interfacial Gd-Co mixing, rather than the standard naive-layer analysis — and exploits stack-order reversal as a symmetry probe to separate FL (sign-changing) from DL (sign-preserving) contributions in an orbital-torque source (Cr) coupled through a rare-earth spacer.

What's New: Prior orbital-torque studies of Cr-based sources have focused on Cr/ferromagnet bilayers; this work introduces a rare-earth (Gd) spacer, systematically varies its thickness, and uses stack-order reversal plus explicit chemical mapping to separate intrinsic orbital-to-spin conversion from interdiffusion effects.

Extension Opportunities:

Fabricate atomically sharp Cr/Gd/Co interfaces via low-temperature MBE or insert diffusion-barrier layers (e.g., thin Cu, Ta) to disentangle intrinsic orbital-torque physics from Gd-Co alloying artifacts
Extend the Gd thickness sweep with light rare-earths (Tb, Dy, Ho) or replace Cr with heavier orbital-Hall metals (Nb, Ru, Zr) to build a systematic map of orbital-to-spin conversion efficiency versus 4f spin-orbit coupling
Perform temperature-resolved harmonic Hall across the compensation point to test whether the DL/FL sign asymmetry tracks the ferrimagnetic net moment reversal, which would discriminate bulk orbital-Hall from interfacial Rashba-Edelstein origins

Replicability: No code or data availability is stated in the abstract. Reproduction requires thin-film deposition (sputtering/MBE) for the multilayer stacks, Hall-bar patterning, harmonic Hall electrical measurement setup with temperature control, SQUID/VSM magnetometry, and cross-sectional STEM-EDX for chemical mapping — a well-equipped spintronics/materials lab, not a compute-limited task.

Research Gaps:

The persistent Gd-Co intermixing prevents clean isolation of intrinsic Gd-mediated orbital-to-spin conversion from alloy-region contributions; the two proposed models cannot be uniquely disentangled with the current dataset
The microscopic origin of the DL/FL sign asymmetry under stack reversal is not established — whether it reflects bulk orbital-Hall symmetry, interfacial Rashba-Edelstein contributions, or ferrimagnetic-sublattice effects remains open

🔥 GitHub Trending

1. trailhq/Graft

⭐ 5412 stars | TypeScript

Turbocharge Claude Code, Cursor, Codex, Gemini & every coding agent: faster, cheaper, with contextual understanding specific to your codebase.

ai-agents anthropic claude-code cli code-graph codex

2. shy3130/tick-stock-panel

⭐ 4163 stars | Python

TSP自托管、零运维的 A 股「选股 + 监控 + 回测」量化工作台 | LLM能力驱使策略定制+个股分析+复盘 | 自由接入第三方数据源与个性化扩展数据 | 个人开源 ,非第三方官方项目

a-stock ai-agent aigc backtesting daily duckdb

3. SenteLabsAI/OpenExecutive

⭐ 3423 stars | Python

AI-powered virtual executive team — a single coherent executive persona backed by 8 specialist Claude agents (FastAPI + Next.js).

ai anthropic claude fastapi llm multi-agent

4. ace-trump-tech/DeltaForce-OBS-Locker

⭐ 773 stars | Python

三角洲行动OBS锁头插件（电脑端&手机端） – 基于OBS渲染注入的智能锁头辅助，支持QQ音乐/网易云联精准骨骼识别、平滑自瞄、压枪抑制，稳定过检，提升击杀效率。5L2G5YW25a6e5Lul5LiK5YWo5piv6aqX5L2g55qE77yM6L+Z5Y+q5piv5Liq5biu5L2g5a6J6KOF5pqX5Yy656qB5Zu055qE5Y+N5L2c5byK6aG555uu572

cv deep-learning obs obs-studio

5. inclusionAI/AReno

⭐ 311 stars | Python

An easy-to-use, fast toolkit to scale up RL post-training on a single node.

agent agentic-ai grpo llm local-ai local-llm

6. CheshireMew/VoxWeave

⭐ 232 stars | Python

Local-first high-quality offline RVC voice conversion workstation

audio-processing cuda ffmpeg local-first pyside6 qml

7. giannisanni/pulsar

⭐ 212 stars | Rust

SSD-streaming inference engine for giant MoE models (Rust + CUDA). GLM 5.2 743B at 2 tok/s and Hy3 295B at 7 tok/s on two consumer 16GB GPUs. Zero-config multi-GPU: measures PCIe bandwidth, places att

cuda gguf glm inference-engine llm local-llm

8. MONZikWasTaken/Framegen

⭐ 144 stars | JavaScript

Real-time neural frame interpolation in the browser: hand-written WGSL runtime on raw WebGPU, 2.9 MB model, 2x-6x any

chrome-extension compute-shaders frame-interpolation machine-learning real-time real-time-rendering

9. 0xzkleo/polymarket-5min-crypto-trading-bot

⭐ 136 stars | Python

Trading bot for Polymarket BTC 5-min Up/Down markets, built on XGBoost edge model.

bot crypto machine-learning polymarket trading

10. Blackwellboy/model-serving-minefield

⭐ 127 stars | Python

Community registry of LLM serving-path traps that produce confidently wrong measurements: templates, tool parsers, reasoning fields, quant kernel paths, CUDA toolchains, KV allocation, eval harnesses,

benchmarking chat-template cuda debugging llama-cpp llm-serving