2026-11-03·Daily digest

🔬 Research Pulse

Daily Digest

November 03, 2026

🤖 AI

🧠 LLMs

1. LLMs Depend on Specialist Tools for Out-of-Distribution Generalization in MS/MS Structural Elucidation

Authors: Gabriel Asher, Cailum M. K. Stienstra, Tomo Oga... Published: 2026-11-02 | Citations: 0 arXiv | PDF

Research Question: Can frontier LLMs perform de novo MS/MS structural elucidation without tools, and does their reasoning generalize to out-of-distribution (undocumented) analytes versus merely recalling well-documented ones?

Summary: The paper tests whether a frontier LLM's reported tool-free structural elucidation success generalizes from 1D NMR to de novo MS/MS. Using a 100-analyte benchmark split by literature documentation, the authors show the LLM's apparent capability is largely memorization (52% ID vs 2% OOD) and that genuine OOD performance only emerges when the agent is granted specialist de novo structure generator tools (up to 38% OOD).

Key Results: Using a 100-analyte benchmark from MassSpecGym split by textual footprint, a frontier LLM without tools achieved 52% Top-1 accuracy on in-distribution analytes but collapsed to 2% Top-1 out-of-distribution. Adding de novo structure generator tools raised OOD Top-1 to 38%, and the resulting agent outperformed standalone de novo models both ID and OOD.

Key Findings:

A frontier LLM without tools scores 52% Top-1 ID but only 2% Top-1 OOD, revealing heavy reliance on literature recall rather than spectral reasoning
Access to de novo structure generators lifts OOD Top-1 accuracy to as high as 38%, the only intervention that meaningfully improved OOD performance
The tool-augmented LLM agent outperforms the standalone specialist de novo models in both ID and OOD regimes, indicating useful orchestration value from the LLM

Technical Novelty: The 'textual footprint' split that cleanly separates memorization from reasoning by measuring literature documentation per analyte, combined with a controlled ablation that holds model/harness/output constant while varying only tool access — isolating the contribution of specialist tools to agent capability.

What's New: Prior claims of tool-free LLM structural elucidation did not disentangle recall from reasoning; this work introduces a documentation-based ID/OOD split that exposes the memorization confound and demonstrates that specialist tools — not the LLM alone — drive OOD scientific generalization.

Extension Opportunities:

Expand the benchmark beyond 100 analytes and apply the textual-footprint ID/OOD split methodology to other scientific domains (crystallography, IR, protein NMR) to systematically audit recall-vs-reasoning in frontier LLMs
Build a tool-router agent that dynamically selects among multiple specialist de novo MS/MS generators (e.g., MIST, MSNovelist, SpecTUS) conditioned on spectrum properties, aiming to close the ID-OOD gap further
Fine-tune or RL-train an LLM on tool-use traces from the de novo generator agent to internalize chemistry-aware spectrum reasoning rather than relying on external tools at inference time

Replicability: Benchmark is derived from the public MassSpecGym dataset, suggesting data is accessible; code availability is not stated in the abstract. Compute is modest — frontier LLM API inference over 100 analytes plus runs of de novo MS/MS generator tools, feasible on a workstation with API access.

Research Gaps:

Lack of literature-aware benchmark splits across other scientific-reasoning tasks where frontier LLMs are claimed to 'reason' from raw signals
No principled framework for which specialist tools yield the largest marginal gains per domain, or how to co-train LLMs with such tools for better internal reasoning

2. Can LLM Agents Be Forward-Deployed Engineers? Requirement Discovery and Delivery for Enterprise Text-to-SQL

Authors: Marianne Menglin Liu, Rongguang Wang, Zheng Wang... Published: 2026-11-02 | Citations: 0 arXiv | PDF

Research Question: Can LLM coding agents function as forward-deployed engineers (FDEs) — autonomously discovering ambiguous customer requirements and translating them into post-training deployment decisions — in a realistic enterprise text-to-SQL engagement?

Summary: The paper evaluates whether LLM agents can perform the forward-deployed engineer role on a realistic enterprise text-to-SQL engagement, split into requirement discovery and post-training delivery phases. While top agents discover >80% of hidden requirements, only 25% of delivery runs pass all acceptance criteria, revealing systemic gaps between benchmark-ready models and production-ready deployment.

Key Results: On a simulated enterprise text-to-SQL engagement across 15 models: top models discover >80% of withheld requirements (acceptance criteria, data permissions, hardware constraints), while open-weight models discover only 40-50%. In the delivery phase, only 25% of runs across 8 coding-agent configurations pass all acceptance criteria. 40% of withheld requirements are missing or misrepresented post-discovery.

Key Findings:

Discovery capability is highly model-dependent: frontier models exceed 80% requirement recovery while open-weight models stall at 40-50%
Delivery is the bottleneck — only 25% of 8 agent configurations pass all acceptance criteria, driven by unreliable execution and underused compute budgets
Agents fail to systematically compare base models and post-training methods, and benchmark accuracy does not translate to production readiness

Technical Novelty: First evaluation framework that decomposes the FDE role into discovery (requirements elicitation from a simulated grounded customer) and delivery (post-training recipe planning + execution) phases, measured against withheld ground-truth requirements rather than just task completion.

What's New: Moves beyond coding-benchmark-style evaluation to measure the full FDE workflow (ambiguous intake → requirement elicitation → post-training delivery) on a real enterprise use case, with a grounded simulated customer and withheld requirements as ground truth.

Extension Opportunities:

Build a human-in-the-loop requirement review layer that flags the ~40% missing/misrepresented requirements before delivery phase begins
Design an explicit experiment-environment specification protocol (hardware, budget scope, production checks) and benchmark whether it closes the 75% delivery failure gap
Extend the FDE simulation framework to other enterprise verticals beyond text-to-SQL (e.g., RAG deployment, fine-tuning for domain classification) to test generalizability

Replicability: Abstract does not mention code/data release. Reproduction would require: a simulated customer harness grounded in enterprise requirements, 15 model API/weights access, and post-training compute for text-to-SQL (likely multi-GPU for the open-weight models tested). Non-trivial to reproduce without released artifacts.

Research Gaps:

No mechanism for agents to self-detect when discovered requirements are incomplete or misrepresented
Lack of standardized experiment-environment specification covering hardware, budget, and production readiness checks for post-training agents

👁️ Vision

1. Reviso: Not All Visual Evidence Deserves Equal Attention in Video Captioning

Authors: Tengfei Liu, Yang Shi, Yuran Wang... Published: 2026-11-02 | Citations: 0 arXiv | PDF

Research Question: How can video captioning systems avoid overemphasizing minor details and missing essential context by learning to selectively allocate visual attention based on video content priorities?

Summary: Reviso is an agentic video captioning framework that learns priority-guided active observation, using a hierarchical Context-Focus-Points plan from a low-FPS preview to selectively revisit segments with specialized tools. It is trained via a verification-guided trajectory pipeline and GRPO with custom rewards for plan alignment and tool-call efficiency, outperforming open-source baselines across four benchmarks.

Key Results: Reviso outperforms evaluated open-source baselines across four video captioning benchmarks, with generated captions better supporting downstream video question answering. Specific numerical scores are not provided in the abstract.

Key Findings:

Not all visual evidence deserves equal attention — content-dependent priority allocation improves caption quality
Hierarchical CFP planning combined with goal-directed revisiting outperforms uniform-attention baselines
Captions generated via priority-guided observation better support downstream video QA tasks

Technical Novelty: Introduces three novel components: (1) hierarchical Context-Focus-Points observation planning from global previews, (2) verification-guided training trajectory construction with teacher refinement, and (3) GRPO training with three custom rewards — key-fact reward, Plan Alignment Reward (PAR), and Unproductive Tool-Call Penalty (UTCP) that scales cost based on caption quality.

What's New: Shifts video captioning from passive uniform-frame processing to agentic active perception with explicit planning, tool use, and plan-adherence/cost-aware RL rewards — unlike prior MLLM captioners that treat all frames uniformly.

Extension Opportunities:

Apply the Context-Focus-Points (CFP) hierarchical planning approach to other multimodal tasks like video question answering or long-form video summarization
Extend the verification-guided trajectory construction pipeline to other agentic MLLM training scenarios beyond captioning
Develop adaptive tool libraries beyond temporal revisiting (e.g., object tracking, OCR, audio analysis) to expand active perception capabilities

Replicability: Abstract does not mention code/data availability. Reproducing would require significant compute: GRPO-based RL training on an MLLM backbone, trajectory generation with teacher models, and evaluation across four benchmarks — likely multiple GPU-days to weeks on A100-class hardware.

Research Gaps:

Abstract does not quantify gains over baselines or report absolute benchmark scores
Scaling behavior and generalization to very long videos (hours) or streaming settings is unaddressed

🦾 ROBOTICS

1. Pixel: A Co-Designed Full-Stack Robot System for Learning General-Purpose Mobile Manipulation

Authors: Tony Tao, Jason Jingzhou Liu, Ruslan Salakhutdinov... Published: 2026-11-02 | Citations: 0 arXiv | PDF

Research Question: How can robot hardware, data collection interfaces, whole-body control, and policy learning be co-designed as an integrated full-stack system — rather than developed independently — to enable general-purpose mobile manipulation?

Summary: Pixel is a co-designed full-stack mobile manipulation system that jointly develops a low-cost wheeled humanoid, multiple teleop and robot-free data collection interfaces, a unified whole-body controller, and a hierarchical policy learning pipeline. By aligning these components rather than developing them in isolation, it enables language-conditioned long-horizon mobile manipulation requiring dexterity, active perception, and whole-body coordination.

Key Results: The paper demonstrates Pixel, a low-cost wheeled humanoid with human-aligned kinematics and compliant grippers, performing mobile-manipulation behaviors requiring dexterity, active perception, navigation, and whole-body coordination. The abstract reports qualitative demonstrations across hardware, data collection, and policy learning evaluations but does not cite specific benchmark numbers, success rates, or dataset sizes.

Key Findings:

Human-aligned kinematics plus compliant grippers let the same whole-body controller serve both teleoperation and autonomous policy rollout, reducing data-to-deployment mismatch
Robot-free human demonstrations can be unified with robot teleop data in a single policy learning pipeline when the embodiment gap is minimized by design
A hierarchical stack with language conditioning extends short-horizon visuomotor control to long-horizon mobile manipulation tasks

Technical Novelty: The novelty is the co-design philosophy itself: a unified whole-body controller shared across teleoperation and autonomous policies, plus a common policy-learning pipeline that treats human-collected and robot-collected demonstrations as a single data distribution — enabled by human-aligned kinematics that reduce the embodiment gap.

What's New: Prior mobile manipulation work typically optimizes hardware, data pipelines, controllers, and learning algorithms in silos. Pixel's contribution is treating the full stack as a joint design problem, with explicit interface alignment (human kinematics ↔ robot kinematics ↔ controller ↔ policy) as the organizing principle.

Extension Opportunities:

Scale the robot-free human data-collection interfaces to crowdsource large-scale demonstration datasets and study transfer efficiency from human-only to robot deployments
Replace the hierarchical language-conditioned policy with a VLA foundation model fine-tuned on Pixel's co-designed data to test whether hardware/data alignment improves VLA generalization
Extend the unified whole-body controller to bimanual coordination with contact-rich tasks (e.g., opening doors while carrying objects) using the compliant grippers as a tactile feedback channel

Replicability: The project page (pixel-robot.com/pixel) is referenced but code/data release is not confirmed in the abstract. Reproducing requires fabricating the custom wheeled humanoid hardware (described as low-cost but still non-trivial), building teleoperation rigs, and training visuomotor policies — likely multi-GPU training over days plus substantial hardware engineering effort.

Research Gaps:

No quantitative benchmark numbers in the abstract — unclear how Pixel compares to prior mobile manipulation systems (e.g., Mobile ALOHA, TidyBot, HomeRobot) on standardized tasks
The scalability of robot-free human data collection and its measurable contribution to policy performance vs. teleop-only data is not characterized

2. ForceLDM: A Force-aware Latent Dynamics Model for Contact-Rich Manipulation

Authors: Peiyuan Tang, Yao Li, Xiaodong Zhang... Published: 2026-11-02 | Citations: 0 arXiv | PDF

Research Question: How can robot policies for contact-rich manipulation better exploit force/tactile signals not just as auxiliary observations, but to model future interaction dynamics that guide action generation—without requiring privileged future sensor data at deployment?

Summary: ForceLDM is a force-aware latent dynamics model for contact-rich manipulation that uses knowledge distillation: a teacher with privileged future force/torque and optical flow signals teaches a student—using only current observations—to anticipate future contact dynamics in a compact latent space. A progressive noise-injection curriculum prevents over-reliance on distilled future features, and the method reportedly beats SOTA across five real-world tasks while generalizing to new objects and perturbations.

Key Results: ForceLDM was evaluated on five real-world contact-rich manipulation tasks and reportedly outperforms state-of-the-art baselines, generalizes to novel objects, and remains robust under environmental perturbations. The abstract does not disclose specific success-rate numbers, baseline names, or quantitative deltas.

Key Findings:

Distilling future dynamics into a latent representation outperforms direct future-observation prediction, which entangles task-relevant dynamics with irrelevant sensory details.
Privileged future force/torque + optical flow are useful teacher signals even though they are unavailable at deployment time.
Curriculum-based progressive noise injection on the distilled features improves robustness and generalization to novel objects and environmental perturbations.

Technical Novelty: Prior force/tactile-augmented policies treat these signals as additional inputs. ForceLDM instead uses privileged future F/T and optical flow as a teacher to distill a compact latent representation of future contact/motion dynamics into a student that sees only current observations—avoiding pixel-level future prediction and sidestepping deployment-time dependence on future sensors. A curriculum-based progressive noise-injection scheme prevents the student from over-relying on the distilled future features.

What's New: Shifts force/tactile sensing from being an auxiliary input to being the training-time supervision for a future-dynamics latent, using an asymmetric teacher-student setup that is force-aware rather than purely visual, plus a noise-curriculum regularizer tailored to distilled future features.

Extension Opportunities:

Replace optical flow + F/T teacher signals with tactile-array or event-camera streams to distill richer local contact dynamics for deformable or fine-grained assembly tasks.
Apply the teacher-student future-dynamics distillation paradigm to bimanual or whole-body loco-manipulation, where privileged future contact signals are even harder to obtain at deployment.
Combine ForceLDM's latent dynamics with model-based planning (e.g., MPC in the learned latent) rather than only using it to condition a reactive policy.

Replicability: Project page and code are advertised at https://tpy001.github.io/ForceLDM/. Reproducing requires a real robot with a wrist-mounted F/T sensor, RGB cameras (for optical flow), and the five contact-rich task setups—compute for training a latent dynamics model + policy is modest (likely a single multi-GPU workstation), but hardware is the limiting factor.

Research Gaps:

No reported comparison with tactile-array-based policies or diffusion-policy variants that already consume force; unclear how much of the gain comes from the teacher signals vs. the latent-dynamics formulation.
Scalability to long-horizon, multi-stage contact-rich tasks (e.g., full assembly sequences) and transfer across robot embodiments is not addressed.

3. PanoTac: A Dense Three-Axis Tactile Sensor for Robust Touch-Centric Manipulation

Authors: Binglin, Wang, Xiaole... Published: 2026-11-02 | Citations: 0 arXiv | PDF

Research Question: How can we provide robots with low-cost, dense three-axis tactile feedback (capturing both contact geometry and shear forces) for contact-rich manipulation tasks where vision alone is insufficient?

Summary: PanoTac is a low-cost (<$50), 60 Hz compliant tactile fingertip that outputs a dense 12×32×3 taxel map with per-taxel normal and shear forces. Integrated into a visuo-tactile policy, it boosts contact-rich manipulation success by 46.7 points over vision-only baselines across three tasks.

Key Results: PanoTac delivers a 12×32×3 tactile representation per fingertip (normal + 2 shear components per taxel) at 60 Hz for under $50. Visuo-tactile policies outperform vision-only baselines by an average of 46.7 percentage points across test-tube preparation, grain scooping, and salt pouring. Ablations confirm that removing either spatial contact information or shear feedback degrades performance.

Key Findings:

Dense three-axis tactile sensing can be achieved at commodity cost (<$50) and real-time rates (60 Hz)
Visuo-tactile policies beat vision-only by 46.7 percentage points on contact-rich tasks (test-tube prep, grain scooping, salt pouring)
Both spatial contact distribution and shear feedback are individually necessary — ablating either degrades performance

Technical Novelty: A compliant fingertip that yields a dense 3-axis (normal + shear) taxel grid at every contact point — most prior low-cost tactile skins give either sparse force arrays or vision-based normal-only depth (e.g., GelSight) rather than dense shear at every taxel, and at a $50/60 Hz price-performance point.

What's New: Combines density (12×32 taxels), three-axis sensing (normal + shear), low cost, and real-time operation in a single fingertip — a combination not jointly achieved by existing GelSight-style, piezoresistive, or capacitive tactile sensors.

Extension Opportunities:

Scale the sensor array to full-hand or whole-body coverage for in-hand manipulation and bimanual tasks
Combine PanoTac signals with diffusion/transformer-based policies trained on larger contact-rich datasets, or use the shear data for closed-loop slip detection and grasp force regulation
Use the dense three-axis signal as supervision for learning differentiable contact/friction models that transfer to simulation

Replicability: Project page at https://panotac.github.io is referenced; code/CAD availability not confirmed in the abstract. Hardware reproduction requires fabrication of the compliant fingertip (<$50/unit) plus a standard robot arm and GPU for policy training — modest compute, non-trivial hardware build.

Research Gaps:

Sensor durability, drift, and recalibration over long-horizon use are not characterized in the abstract
Generalization beyond the three evaluated tasks and to deformable/dynamic objects remains open

💻 COMPUTE

1. Rose: Efficient Embodied RL through Graph Optimization and GPU Compute Resource Control

Authors: Wen Zeng, Ke Hong, Xu Fu... Published: 2026-11-02 | Citations: 0 arXiv | PDF

Research Question: How can embodied RL systems eliminate fine-grained GPU underutilization within subcomponent phases and nested loops, which coarse component-level pipelining (e.g., overlapping inference and simulation) fails to address, while also preventing MPS-based concurrency from degrading due to SM contention on the critical path?

Summary: Rose is a runtime-aware system that automates fine-grained GPU overlap and resource management for embodied RL while preserving synchronous training semantics. It combines a window-guided graph optimizer that finds and exploits low-activity GPU windows inside nested loops with an MPS-based slack-aware SM-quota manager that mitigates contention on the critical path, yielding 1.20–1.92× throughput gains.

Key Results: Rose achieves 1.20–1.92× training throughput improvement and increases average SM activity by 23.8–39.3 percentage points over state-of-the-art embodied RL systems across diverse policies, simulators, RL algorithms, and GPU platforms. Gains are driven by a two-stage pipeline: window-guided graph optimization for fine-grained overlap plus slack-aware MPS SM-quota tuning.

Key Findings:

Coarse-grained component pipelining leaves substantial low-activity GPU windows inside subcomponent phases and nested loops that fine-grained overlap can reclaim.
Naive concurrent scheduling under MPS can be negated by cross-process GPU time slicing and SM contention — explicit per-process SM quotas informed by branch synchronization slack are needed.
Combining local profiling with end-to-end iteration validation lets the system guarantee throughput gains with minimal tuning overhead across heterogeneous policies, simulators, algorithms, and GPUs.

Technical Novelty: Prior systems overlap work only at coarse component boundaries (simulator-group pipelining of inference/sim/training). Rose introduces (1) a lightweight programmable interface exposing execution hierarchy and state accesses so a runtime can detect fine-grained low-activity GPU windows inside nested loops, (2) a window-guided graph optimizer that applies transformations to overlap complementary work across components AND loop levels, and (3) a slack-aware MPS SM-quota tuner that uses branch synchronization slack to protect the critical path from SM contention — combining local profiling with end-to-end iteration validation.

What's New: First embodied RL system to automate fine-grained overlap below the component boundary via a programmable hierarchy/state-access interface, and to couple that with slack-aware MPS SM-quota control rather than treating concurrency and resource partitioning as separate concerns.

Extension Opportunities:

Extend the slack-aware resource manager beyond NVIDIA MPS to AMD/ROCm MPS equivalents or multi-GPU/multi-node settings where cross-device slack and NVLink contention become the critical path.
Integrate Rose's window-guided graph optimizer with asynchronous or off-policy RL algorithms (e.g., IMPALA, SEED RL) to see whether relaxing synchronous semantics compounds the fine-grained overlap gains or exposes new scheduling windows.
Apply the programmable execution-hierarchy interface to LLM post-training RLHF/RLVR pipelines (actor–critic–reward–rollout), where similar heterogeneous GPU phase imbalance exists but at much larger model scale.

Replicability: The abstract does not mention a public code or artifact release. Reproduction would require a multi-GPU NVIDIA system supporting MPS, GPU-accelerated simulators (likely Isaac Gym / Isaac Lab or similar), and the specific RL algorithms/policies benchmarked — nontrivial but within reach of a single-node 1–8 GPU workstation for most embodied RL baselines.

Research Gaps:

Scope is limited to synchronous RL on single-node NVIDIA GPUs with MPS; async RL, multi-node, and non-NVIDIA accelerators remain open.
Requires developers to annotate execution hierarchy and state accesses through the programmable interface — annotation burden and correctness guarantees on complex pipelines are not addressed in the abstract.

2. Sonar-MoE: In-Backward Expert Updates for Memory-Efficient Mixture-of-Experts Training

Authors: Avni Mittal, Avinash Anand, Ashutosh Kumar... Published: 2026-11-02 | Citations: 0 arXiv | PDF

Research Question: How can we reduce the GPU memory required for training Mixture-of-Experts (MoE) models, where expert gradients dominate memory usage and grow with model scale, without sacrificing training correctness or throughput?

Summary: Sonar-MoE ports optimizer-in-backward from dense models to batched MoE expert tensors, updating each expert tensor during the backward pass and immediately freeing its gradient. On OLMoE and Granite MoE fine-tuning it cuts peak memory 19-22% and step time 17-23% while remaining bitwise identical to standard training (without expert gradient clipping).

Key Results: On OLMoE-1B-7B and Granite-3.1-3B-A800M fine-tuned in bf16 across two expert-kernel backends, Sonar-MoE reduced peak memory by 19-22% and cut step time by 17-23% versus standard training with global-norm clipping. Updates are bitwise identical to standard training (absent expert gradient clipping), and held-out loss stays within 0.003 of all baselines. Per-expert (non-batched) variants incur 20-72% slowdowns at the same memory, which Sonar-MoE avoids.

Key Findings:

In-backward updates on batched expert tensors cut peak GPU memory 19-22% during bf16 MoE fine-tuning.
Step time drops 17-23% versus standard training with global-norm clipping, with held-out loss within 0.003 of baselines.
Updating per-expert instead of per batched tensor induces a 20-72% slowdown, showing the batched granularity is essential.

Technical Novelty: Prior optimizer-in-backward work targets dense parameters per-tensor; Sonar-MoE lifts that idea to batched MoE expert tensors so entire grouped-expert weight tensors are updated and their gradients freed inside the backward pass, holding at most one layer of expert gradients. The batched (not per-expert) update is the specific trick that preserves speed, while dense parameters still receive global-norm clipping.

What's New: Unlike prior optimizer-in-backward methods aimed at dense weights, Sonar-MoE operates on grouped/batched expert tensors so expert gradients never accumulate across layers, achieving bitwise-identical updates to standard training while retaining dense-parameter global-norm clipping.

Extension Opportunities:

Extend the in-backward batched-tensor update to pretraining-scale MoE runs (hundreds of billions of params) and quantify memory/throughput at larger expert counts and parallelism regimes (EP/TP/PP).
Combine Sonar-MoE with low-precision optimizer states (FP8/INT8 Adam) or ZeRO-style sharding to compound memory savings, and investigate interaction with expert gradient clipping approximations that preserve bitwise equivalence.
Adapt the batched-tensor in-backward update to other 'grouped' parameter structures — e.g., LoRA/DoRA adapter banks, grouped-GEMM MLPs, or vision MoEs — to generalize beyond language MoE experts.

Replicability: The abstract does not mention a code or data release. Reproduction would require the published OLMoE-1B-7B and Granite-3.1-3B-A800M checkpoints, bf16 fine-tuning on a single modern GPU (A100/H100-class, likely 40-80GB), and integration into one of the two expert-kernel backends used; effort is moderate for a team familiar with MoE training stacks.

Research Gaps:

Does not demonstrate compatibility with expert gradient clipping (bitwise equivalence requires skipping it).
Results are limited to fine-tuning of two mid-scale MoEs in bf16; pretraining scale, larger expert counts, and interactions with EP/TP/PP sharding remain unaddressed.

3. Scaling global tractography with ExaTract: from large cohorts to ultra-high-resolution datasets

Authors: Simon Legeay, Felix Matuschke, Bastien Herlin... Published: 2026-11-02 | Citations: 0 arXiv | PDF

Research Question: How can global tractography—a theoretically rigorous but computationally prohibitive method for reconstructing white matter connectivity from diffusion MRI—be scaled to handle both large population cohorts (HCP) and ultra-high-resolution ex vivo mesoscopic datasets?

Summary: ExaTract is an open-source MPI+Kokkos parallelization of spin-glass global tractography that uses spatial domain decomposition with overlapping regions and boundary synchronization to preserve the method's global convergence guarantees. The framework is validated at two previously inaccessible scales: the full HCP cohort for population-scale analysis, and the Chenonceau ex vivo dataset, yielding the first mesoscopic structural connectome from global tractography.

Key Results: The authors built ExaTract as an open-source MPI+Kokkos application that partitions the tractography domain into spatially overlapping regions across distributed memory nodes with boundary synchronization preserving global convergence guarantees. They validated it on (1) the complete HCP dataset, demonstrating population-scale throughput, and (2) the Chenonceau ex vivo ultra-high-field dataset, producing what they claim is the first mesoscopic structural connectome derived from global tractography. The abstract does not provide specific speedup numbers, node counts, or runtime benchmarks.

Key Findings:

A domain-decomposition strategy with overlapping regions and periodic boundary synchronization makes spin-glass global tractography tractable on distributed-memory HPC without sacrificing global convergence
ExaTract processes the complete HCP dataset at population scale, which was previously impractical for global tractography methods
The Chenonceau ultra-high-resolution ex vivo dataset yields the first mesoscopic structural connectome reconstructed via global tractography

Technical Novelty: Prior spin-glass / global tractography (e.g., Reisert, Fillard) is inherently sequential due to its MCMC-like stochastic optimization. ExaTract's novelty is a domain-decomposition scheme with overlapping spatial regions and periodic boundary synchronization that preserves the global convergence properties of the underlying spin-glass sampler while enabling distributed-memory parallelism—implemented portably via MPI+Kokkos for HPC architectures.

What's New: First HPC-portable (MPI+Kokkos) implementation of global tractography that scales to both large cohorts and mesoscopic ex vivo resolutions, breaking the sequential bottleneck of prior spin-glass tractography while retaining its convergence guarantees.

Extension Opportunities:

Integrate ExaTract outputs into downstream connectome analysis pipelines (e.g., graph-theoretic metrics across HCP subjects) to study population variability of global-tractography-derived connectivity
Port the Kokkos backend to GPU-dense architectures (e.g., Grace Hopper, MI300) and benchmark against the CPU/MPI baseline to characterize scaling on heterogeneous exascale systems
Extend the spin-glass energy model with microstructure-aware priors (e.g., NODDI, fiber orientation dispersion) to leverage the mesoscopic resolution of ex vivo data

Replicability: The abstract describes ExaTract as open-source, which supports replicability. Reproduction requires HPC infrastructure (multi-node MPI cluster with Kokkos-supported backends), the HCP diffusion MRI dataset (publicly available with registration), and the Chenonceau ex vivo dataset (likely restricted-access). Full-cohort HCP global tractography likely requires hundreds-to-thousands of core-hours per subject even with ExaTract.

Research Gaps:

The abstract omits quantitative scaling metrics (strong/weak scaling efficiency, node counts, wall-clock speedups) needed to compare against local tractography or prior parallelization attempts
Biological validation—whether the mesoscopic connectome aligns with known neuroanatomy or histology—is not addressed in the abstract

⚡ ENERGY

1. Discovering hierarchical governing equations for community-structured complex network dynamics with LLM agents

Authors: Jinlei Zhang, Zixuan Zhang, Hao Qiu... Published: 2026-11-02 | Citations: 0 arXiv | PDF

Research Question: How can we discover governing equations for complex network dynamics when community-level shared dynamics coexist with individual-level variations, and when relevant functional structures may be missing from the initial candidate library?

Summary: SAGE is an LLM-agent framework that discovers governing equations for community-structured networks by hierarchically separating community-shared dynamics from node-specific residual variations, using an iterative Library-Discovery-Review loop that adaptively expands the candidate function space. It outperforms baselines across six synthetic and real-world networks and can transfer community-level laws to nodes with no observations.

Key Results: SAGE was evaluated across two synthetic networks and four real-world networks (mobile communication, brain electrophysiology, power grids, urban rail transit). It recovers compact interpretable governing equations, reproduces observed dynamics, and outperforms representative equation-discovery baselines. Demonstrated robustness under incomplete observations and successful transfer of community-level laws to nodes with entirely missing observations. Specific numerical benchmarks are not provided in the abstract.

Key Findings:

Hierarchical decomposition into community-level and individual-level dynamics yields more compact and interpretable equations than flat per-node discovery
Self-reflective LLM review enables adaptive expansion of the candidate function library when initial terms are insufficient
Community-level laws learned from observed nodes transfer successfully to nodes with entirely missing observations, enabling imputation-free prediction

Technical Novelty: The hierarchical two-stage decomposition (community-shared law first, then node-specific residuals) orchestrated by an LLM-guided self-reflective agent that adaptively expands the candidate function library when existing terms are insufficient. Prior sparse-regression methods (SINDy-style) use fixed libraries and treat each node independently.

What's New: First framework to formulate network equation discovery as a hierarchical problem exploiting community structure, combined with LLM-driven adaptive library refinement — contrasting with fixed-library sparse-regression approaches (SINDy, SR) that treat nodes independently and cannot introduce missing functional forms.

Extension Opportunities:

Apply SAGE to biological signaling networks (gene regulatory or metabolic networks) where cell-type-level shared dynamics and cell-specific variations are known to coexist
Extend the Library-Discovery-Review loop to handle temporally non-stationary community structures where communities themselves evolve over time
Integrate SAGE with causal discovery methods to not only recover governing equations but also infer directed interaction mechanisms between communities

Replicability: The abstract does not mention code or data release. Reproduction would require an LLM API (likely GPT-4 class), the six benchmark datasets (two of which are synthetic and reproducible, four real-world including public power grid, brain, and mobility data), and modest compute for the symbolic regression components — primarily LLM inference costs rather than GPU training.

Research Gaps:

Abstract does not quantify performance margins over baselines or specify which baselines were used
Scalability to very large networks (thousands+ of communities) and sensitivity to community detection quality are not addressed

2. Characterisation of surface layers formed on plasma-facing components in controlled fusion devices: Role of heavy ion elastic recoil detection

Authors: Petter Ström, Per Petersson, Marek Rubel... Published: 2026-11-02 | Citations: 0 arXiv | PDF

Research Question: How can material migration, deposition patterns, and surface layer composition on plasma-facing components in fusion tokamaks be characterised after controlled tracer injection experiments (N-15 and MoF6) to understand erosion/re-deposition dynamics?

Summary: The paper uses time-of-flight heavy ion elastic recoil detection analysis to characterise surface layers on carbon-based limiter tiles retrieved from TEXTOR tokamak after N-15 and MoF6 tracer injection experiments. It maps the depth distribution and co-deposition behaviour of molybdenum, nitrogen isotopes, fluorine, boron, and other species, concluding that Mo migration proceeds via erosion and prompt re-deposition with N-15 trapped alongside Mo.

Key Results: Using time-of-flight heavy ion elastic recoil detection analysis (ToF-HIERDA) on TEXTOR toroidal limiter tiles (CFC and fine grain graphite), the authors demonstrated: (1) Mo migration dominated by erosion and prompt re-deposition, with highest Mo concentration in the outermost 100 nm layer; (2) N-15 co-trapped with deposited Mo; (3) depth-resolved profiling of the top ~400 nm (resolution limited by surface roughness); (4) identification of multiple species — D fuel, C-12, B-10/B-11 (from boronisation), C-13 (prior tracer campaigns), N-14 (edge cooling), and Inconel-eroded metals; fluorine and N-15 showed irregular depth profiles unlike Mo.

Key Findings:

Molybdenum deposition is concentrated in the outermost ~100 nm layer, consistent with an erosion/prompt-redeposition migration mechanism
N-15 tracer is co-trapped with deposited Mo, while fluorine and N-15 display irregular (non-monotonic) depth profiles unlike Mo
Multiple species identifiable in a single ToF-HIERDA measurement: D, C-12/C-13, B-10/B-11 (boronisation), N-14 (edge cooling), and Inconel-eroded metals; depth resolution within top 400 nm is limited by surface roughness

Technical Novelty: Application of time-of-flight heavy ion elastic recoil detection analysis (ToF-HIERDA) to simultaneously quantify light (D, B, C, N, F) and heavy (Mo, Inconel-derived metals) species on the same tokamak limiter samples, combined with dual tracer injection (isotopically-labeled N-15 and MoF6) to disentangle migration pathways of multiple elements in a single campaign.

What's New: Combines dual isotopic tracer injection (N-15 + MoF6) with ToF-HIERDA's simultaneous multi-element light-and-heavy-species sensitivity to directly correlate heavy metal migration with nitrogen trapping on real tokamak limiter samples — going beyond single-element post-mortem studies typical of prior fusion wall analyses.

Extension Opportunities:

Combine ToF-HIERDA with complementary techniques (SIMS, XPS, or nuclear reaction analysis) to overcome the surface-roughness resolution limit and resolve finer depth distributions below 100 nm
Apply the same tracer methodology (MoF6 + N-15) to ITER-relevant tungsten or beryllium plasma-facing components to map high-Z material migration under reactor-like conditions
Build a predictive migration model calibrated against these Mo/N-15/F depth profiles to simulate long-term wall erosion/redeposition in next-generation devices like DEMO

Replicability: No code/dataset mentioned. Replication requires access to a tokamak with controlled tracer injection capability (e.g., TEXTOR-class device, now decommissioned), post-mortem limiter tiles, and a ToF-HIERDA beamline — making exact reproduction infeasible for most groups, though the analytical technique itself is available at several accelerator labs (e.g., Uppsala's tandem accelerator).

Research Gaps:

Surface roughness of real plasma-facing components severely limits depth resolution of ion-beam techniques — complementary methods or smoother witness samples are needed
Mechanistic coupling between heavy-metal re-deposition and light-species (N, F) trapping is only phenomenologically observed; underlying chemistry and binding states remain uncharacterised

3. PaddleMaterials: An End-to-End Deep Learning Toolkit for Materials Science

Authors: Xuwei Liu, Zhimin Zhang, Xiaoguang Hu... Published: 2026-11-02 | Citations: 0 arXiv | PDF

Research Question: How can fragmented materials-learning models and workflows (property prediction, structure generation, interatomic potentials, electronic structure, spectra, microscopy) be unified into a single verifiable, performant framework without losing task-specific scientific semantics?

Summary: PaddleMaterials is an open-source framework that unifies six classes of materials-ML tasks (property, generation, MLIPs, electronic structure, spectra, microscopy) behind semantically-typed task interfaces and a shared compiler backend. It demonstrates broad compiled-execution speedups (up to 336%) and sub-10 meV/atom agreement with DFT on LiMnO2 at <5% of DFT cost, validating verifiable cross-task workflows.

Key Results: Across 65 model instances in 11 groups, compiled execution accelerated every instance in 8 groups with group median speedups of 11%–224% and max speedups of 11%–336%. On LiMnO2, predicted formation energies for generated/relaxed structures were within ≤7 meV/atom of same-geometry PBE+U DFT references, at <1/20th the cost of a DFT single-point calculation. Case studies validated end-to-end chaining on LiMnO2 and toluene.

Key Findings:

Compiled execution accelerated every model instance in 8 of 11 groups, with median speedups of 11%–224%
LiMnO2 formation energies predicted by the chained pipeline deviated ≤7 meV/atom from same-geometry PBE+U DFT
Full ML pipeline steps cost <1/20 of a single-point DFT calculation, enabling practical workflow scaling
Unified task interfaces encoding units and aggregation conventions enable verifiable cross-task composition across six heterogeneous task families

Technical Novelty: A task-interface abstraction that records physical quantities, units, aggregation conventions, and references to make cross-task composition verifiable, combined with a single-flag dynamic→compiled execution switch (graph optimization + operator fusion + hardware codegen) spanning six heterogeneous materials task classes — broader scope than prior single-task frameworks (e.g., MatBench, JARVIS, OCP).

What's New: Prior toolkits typically target a single task class (potentials, property prediction, or generation). PaddleMaterials is the first to unify six classes with semantic task interfaces plus a one-switch compiled backend, enabling verifiable chaining demonstrated on real DFT-referenced case studies.

Extension Opportunities:

Add new task modules (e.g., defect energetics, phase diagram construction, or inverse design via Bayesian optimization) that reuse the shared unit/semantics interface
Benchmark the compiled backend on non-Paddle hardware (TPUs, AMD MI300) or integrate ONNX/Triton kernels to broaden hardware portability
Build an active-learning loop that uses MLIP uncertainty to trigger DFT labels and feeds them back through the unified data interface for autonomous materials discovery

Replicability: Described as open-source. Reproducing the full 65-model benchmark and DFT reference calculations likely requires a multi-GPU cluster (PaddlePaddle environment) plus a DFT code (VASP/QE) for PBE+U references; individual model inference should run on a single modern GPU.

Research Gaps:

No reported comparison against competing frameworks (OCP, MatBench, JARVIS, MACE-based stacks) on identical benchmarks
Case studies limited to two systems (LiMnO2, toluene) — generalization to complex alloys, interfaces, or disordered systems is unverified

🏥 HEALTHCARE

1. DiabAgent: Toward Agentic Physiological Control

Authors: Chengxi Li, Shaoxuan Li, Deyao Hong... Published: 2026-11-02 | Citations: 0 arXiv | PDF

Research Question: How can LLM-based agents be systematically evaluated and deployed for sustained physiological control (specifically insulin dosing in Type 1 diabetes), beyond static clinical QA benchmarks?

Summary: DiabAgent reframes LLM evaluation in medicine from clinical QA to sustained physiological control, using Type 1 diabetes insulin dosing as a concrete testbed. The paper contributes a tool-using agent framework (APC), a patient simulator (DiabetSim), and a 212-task benchmark (SweetBench), showing that Claude Opus 5.5 beats the OpenAPS closed-loop baseline (COGI 73.5 vs 71.2).

Key Results: Using Claude Opus 5.5 with a 1-hour decision interval, the APC framework achieved a COGI (Composite Glucose Index) of 73.5, outperforming OpenAPS (open-source closed-loop baseline) at 71.2. Evaluated on SweetBench — 212 glucose-control tasks instantiated from 32 scenario designs via the DiabetSim simulator. Ablations confirmed that individual adaptation and contextual information both improved control.

Key Findings:

An LLM agent making hourly insulin decisions can outperform a mature rule-based closed-loop system (OpenAPS) on composite glucose metrics
Individual patient adaptation and contextual observations (meals, exercise, illness) are both ablation-verified contributors to performance
A hybrid design — rule-based high-frequency control with lower-frequency LLM oversight — is a promising cost/performance compromise

Technical Novelty: First framework to cast physiological control as an agentic tool-use problem with persistent per-patient memory, rather than one-shot clinical QA. Introduces three paired artifacts (APC framework + DiabetSim simulator + SweetBench benchmark) to move medical LLM evaluation from static Q&A to sustained closed-loop intervention.

What's New: Prior medical LLM work optimizes for single-turn clinical reasoning; this is one of the first systematic treatments of LLMs as sustained physiological controllers with memory, tool use, and quantitative comparison to a deployed closed-loop system.

Extension Opportunities:

Apply the APC framework to other chronic conditions requiring sustained dosing (anticoagulation, chemotherapy, hypertension, dialysis fluid management)
Explore the hybrid architecture explicitly suggested — pair high-frequency rule-based inner loop with lower-frequency LLM strategic decisions — to quantify the cost/performance Pareto frontier
Extend DiabetSim with adversarial patient variability (sensor dropout, pump failures, misreported carbs) to stress-test agent robustness before clinical trials

Replicability: Abstract does not explicitly commit to open-sourcing code/data, though the 'research suite' framing suggests release intent. Reproduction requires access to Claude Opus 5.5 (API cost non-trivial given hourly decisions across 212 tasks) plus the DiabetSim simulator. Compute is modest — bottleneck is LLM API calls, not training.

Research Gaps:

No real-patient validation — results are entirely in-simulator, and sim-to-real transfer for metabolic control is historically brittle
Hourly decision cadence is coarse relative to real CGM/pump systems (5-min intervals); the hybrid architecture is motivated but not deeply characterized

2. The DynaCell Evaluation Suite for Dynamic 3D Virtual Staining of Live Cells

Authors: Alexandr A. Kalinin, Dihan Zheng, Taylla Milena Theodoro... Published: 2026-11-02 | Citations: 0 arXiv | PDF

Research Question: How do virtual staining models perform on volumetric 3D time-lapse live-cell imaging, and what trade-offs exist between structural fidelity and phenotypic fidelity that single-metric evaluations obscure?

Summary: DynaCell is an evaluation suite for 3D time-lapse virtual staining of live cells, pairing label-free and fluorescent volumetric movies with baseline models and a three-tier metric framework. Its central finding is that regression models win on structural localization while generative models win on phenotype distribution — exposing a trade-off that any single metric would hide and that should guide model choice per downstream task.

Key Results: Built an evaluation suite spanning 3 cell types, 2 microscopes, 4 organelles, and 3 perturbation states with three metric tiers (pixel-level fidelity, organelle segmentation performance, single-cell phenotype similarity). Demonstrated that regression baselines yield better spatial organelle localization while generative baselines better capture cross-cell phenotype distributions — a trade-off invisible under single-metric scoring.

Key Findings:

Regression baselines better predict spatial localization of target organelles in 3D time-lapse data
Generative baselines better capture the distribution of phenotypes across cells
Single-metric evaluation obscures the structural-vs-phenotypic trade-off, implying task-specific model selection for localization, segmentation, tracking, or phenotypic profiling

Technical Novelty: First standardized evaluation suite specifically for 3D time-lapse (volumetric + temporal) virtual staining with a tiered metric design that separately measures pixel fidelity, downstream segmentation utility, and phenotypic distribution similarity across cells — moving beyond PSNR/SSIM-only comparisons that dominated prior 2D virtual staining work.

What's New: Prior virtual staining benchmarks focus on 2D static images and pixel-level metrics; DynaCell is the first to standardize 3D + time-lapse evaluation and to formalize phenotype-distribution fidelity as a first-class metric tier alongside pixel and segmentation metrics.

Extension Opportunities:

Train hybrid regression-generative models (e.g., diffusion with pixel-regression auxiliary loss) to recover both localization accuracy and phenotype distribution fidelity
Extend the three-tier metric framework to include temporal consistency metrics for tracking (cell lineage preservation, motion coherence) that time-lapse specifically enables
Benchmark foundation-model-scale virtual staining (e.g., fine-tuned vision transformers or video diffusion models) against the released baselines on DynaCell to see if scale closes the structural-vs-phenotypic gap

Replicability: Authors explicitly release curated paired datasets, baseline model checkpoints, and evaluation code as part of DynaCell. Reproduction compute is not quantified in the abstract, but 3D time-lapse training over multiple cell types and microscopes typically requires multi-GPU workstations (e.g., 1–4 A100s) over days.

Research Gaps:

No model yet dominates both structural and phenotypic metric tiers — hybrid or multi-objective training is unexplored
Temporal-specific metrics (tracking consistency, lineage fidelity) are not yet part of the suite despite time-lapse being the setting

3. Contrastive Adapter Training and Consensus Knowledge Distillation for Multi-Source-Free Domain Adaptation in Skin Cancer Diagnosis

Authors: Ali Abedi, Q. M. Jonathan Wu, Ning Zhang... Published: 2026-11-02 | Citations: 0 arXiv | PDF

Research Question: How can we perform multi-source-free unsupervised domain adaptation (MSFDA) for skin cancer diagnosis without accessing source data, while avoiding catastrophic forgetting, overfitting on small clinical datasets, and noisy supervision from hard pseudo-labels?

Summary: CAT-CKD is a parameter-efficient MSFDA method for skin cancer diagnosis that combines contrastive adapter training on a frozen ViT backbone with consensus knowledge distillation from multiple source models. It outperforms existing MSFDA baselines (86.1% average AUROC across five skin lesion datasets) while training only 4.3M parameters, enabling privacy-preserving adaptation.

Key Results: CAT-CKD achieves an average AUROC of 86.1% across five publicly available skin lesion datasets, outperforming existing MSFDA methods while requiring only 4.3M trainable parameters (vs full backbone fine-tuning in prior work).

Key Findings:

Establishing a shared domain-invariant feature space via supervised contrastive learning BEFORE source-specific training improves downstream MSFDA performance
Soft consensus distillation via KL divergence over multi-source logits produces cleaner supervision than hard pseudo-labels or confidence-weighted aggregation under domain shift
Lightweight ConvPass adapters in a frozen ViT (4.3M params) outperform full-backbone fine-tuning baselines, mitigating catastrophic forgetting on small clinical datasets

Technical Novelty: Two-stage approach that (1) establishes a shared domain-invariant feature space BEFORE source-specific training using ConvPass adapters in a frozen ViT with supervised contrastive learning, and (2) uses soft logit-based consensus distillation via KL divergence instead of hard pseudo-labels or confidence-weighted aggregation. Unlike prior MSFDA, domain shift is addressed at the feature-learning stage rather than only at aggregation.

What's New: Prior MSFDA methods fine-tune the full backbone and address domain shift only at aggregation; CAT-CKD is the first to combine (a) frozen-backbone adapter training with supervised contrastive learning for pre-emptive domain-invariant representation, and (b) logit-level consensus distillation for noise-robust target adaptation.

Extension Opportunities:

Apply the CAT-CKD framework to other medical imaging tasks with privacy constraints (e.g., histopathology, radiology) where multi-institutional data cannot be centralized
Explore alternative lightweight adapter architectures (LoRA, prefix tuning) in place of ConvPass to further reduce the 4.3M parameter footprint or improve cross-domain generalization
Extend the consensus knowledge distillation to incorporate uncertainty-aware aggregation (e.g., evidential deep learning) rather than simple logit averaging to better handle disagreeing source models

Replicability: Code is publicly available at https://github.com/A-Abedi/CAT_CKD. Five publicly available skin lesion datasets used. With only 4.3M trainable parameters and a frozen ViT backbone, reproduction is feasible on a single consumer GPU (e.g., RTX 3090/4090).

Research Gaps:

Existing MSFDA methods cause catastrophic forgetting and overfitting on small clinical datasets due to full backbone fine-tuning
Prior approaches rely on hard pseudo-labels or confidence-weighted aggregation, introducing noisy supervision under domain shift, and do not establish a shared feature space prior to aggregation

🔬 MATERIALS

1. High-Entropy A-Site Disorder Enhances Oxygen Evolution Reaction Activity and Stability in Nickelate Thin Films

Authors: Min-Ju Choi, Le Wang, Krishna Prasad Koirala... Published: 2026-11-02 | Citations: 0 arXiv | PDF

Research Question: How does epitaxial strain jointly with A-site high-entropy configurational disorder influence defect formation, Ni valence, electronic structure, and OER activity in rare-earth nickelate thin films?

Summary: The paper shows that an epitaxial high-entropy rare-earth nickelate (La,Pr,Nd,Sm,Eu)NiO3 under tensile strain forms Ruddlesden-Popper faults, lowers Ni valence, and shifts RE 4f states toward the Fermi level, producing stronger 4f-O 2p-Ni 3d hybridization. These coupled effects yield earlier OER onset, higher activity than compressively strained and single-A-site counterparts, and improved resistance to OER-induced degradation vs LaNiO3.

Key Results: Epitaxial (La0.2Pr0.2Nd0.2Sm0.2Eu0.2)NiO3 thin films under tensile strain exhibit Ruddlesden-Popper fault formation, lowered average Ni oxidation state, increased low-energy unoccupied states (via O K-edge XAS), earlier OER onset, stronger pre-catalytic Ni redox features, higher OER activity than compressively strained films and single-A-site nickelates, and enhanced structural stability vs LaNiO3. DFT shows tensile strain shifts RE 4f states toward E_F, enhancing RE 4f-O 2p-Ni 3d hybridization. Specific current density/overpotential numbers not provided in abstract.

Key Findings:

Tensile strain in high-entropy nickelate films promotes Ruddlesden-Popper fault formation and lowers average Ni oxidation state
O K-edge XAS and DFT reveal strain-driven shift of RE 4f states toward E_F, enhancing 4f-O 2p-Ni 3d hybridization while leaving occupied DOS near E_F largely unchanged
High-entropy A-site disorder plus tensile strain produces earlier OER onset, higher activity than compressive films and single-A-site nickelates, and better structural stability than LaNiO3

Technical Novelty: First systematic use of A-site high-entropy configurational disorder combined with epitaxial strain in rare-earth nickelates to control RE 4f-derived electronic states near E_F, linking strain-induced Ruddlesden-Popper defects to enhanced OER activity via 4f-O 2p-Ni 3d hybridization — a mechanism not accessible in single-A-site nickelates.

What's New: Prior nickelate OER work focused on single-A-site cation chemistry, strain, or oxygen stoichiometry in isolation; this work introduces A-site configurational entropy as an independent tuning knob and identifies RE 4f-mediated hybridization as a strain-activatable activity descriptor.

Extension Opportunities:

Explore larger high-entropy A-site compositions (6+ rare earths, or mixing with alkaline earths) to further tune 4f-2p-3d hybridization and benchmark against this 5-cation baseline
Combine A-site high-entropy design with B-site substitution (Fe, Co, Mn) to co-tune eg occupancy and defect chemistry for alkaline OER
Develop in-operando XAS/STEM protocols to track Ruddlesden-Popper fault evolution during OER cycling to establish structure-stability design rules

Replicability: No code/data availability mentioned in abstract. Reproduction requires PLD/MBE epitaxial thin-film growth on strain-tuning substrates, synchrotron O K-edge XAS, STEM for RP-fault imaging, electrochemical OER cells, and DFT (likely VASP/QE with Hubbard-U or hybrid functionals for 4f states) — specialized materials-science infrastructure, not commodity compute.

Research Gaps:

No quantitative OER metrics (overpotential at 10 mA/cm², Tafel slope, turnover frequency) are stated in the abstract to compare against state-of-the-art benchmarks
Mechanistic role of Ruddlesden-Popper faults as active sites vs degradation nucleators remains correlative rather than causal

2. Imaging the Current Distribution in Superconducting Transition-Edge Sensors with Scanning SQUID Microscopy

Authors: Samantha Walker, Austin Kaczmarek, Douglas A. Bennett... Published: 2026-11-02 | Citations: 0 arXiv | PDF

Research Question: How do spatial variations in the superconducting state and transport current distribution within transition-edge sensors (TESs) — arising from device geometry, superconducting leads, and normal-metal proximity features — affect device behavior under bias, which existing TES models largely neglect?

Summary: The paper uses scanning SQUID microscopy to directly image the local superconducting state and transport current distribution in biased MoAu X-ray transition-edge sensors, revealing that current flow and superconducting regions reconfigure with bias in ways that depend strongly on device geometry and proximity effects. This provides the first spatially resolved evidence that mesoscale non-uniformities — ignored in current TES models — are important for understanding TES physics and informing device design.

Key Results: Using scanning SQUID microscopy (SSM), the authors directly imaged local diamagnetic response (to map superconducting regions) and Oersted fields (to reconstruct transport current distribution) in two MoAu X-ray TES designs voltage-biased within the resistive transition. They demonstrated bias-dependent reconfigurations of the current distribution that correlate with current-induced changes in superconducting regions. The abstract reports qualitative imaging results rather than specific quantitative benchmarks (no resolution, SNR, or current-density numbers given).

Key Findings:

Superconducting regions within a voltage-biased TES are spatially inhomogeneous and shaped by proximity effects from leads and normal-metal features
Transport current distribution reconfigures as a function of bias point within the resistive transition, tracking the evolution of superconducting regions
Standard TES models that assume spatially uniform resistive transitions miss physically significant mesoscale effects that influence device operation

Technical Novelty: Direct, spatially resolved imaging of both local superconducting regions (via diamagnetic response) and transport current distribution (via Oersted field reconstruction) in biased MoAu TESs — prior TES characterization is largely bulk/transport-based and treats the device as spatially uniform.

What's New: First application of scanning SQUID microscopy to simultaneously map both the diamagnetic (superconducting) response and the Oersted-field-derived current distribution in voltage-biased TESs, exposing bias-dependent spatial structure that prior non-imaging characterization could not resolve.

Extension Opportunities:

Incorporate the measured mesoscale current/superconducting-region maps into updated TES electrothermal models (e.g., extensions of the two-fluid or RSJ-based models) and benchmark predicted noise and responsivity against measurements
Apply SSM imaging across a wider library of TES geometries (varying leads, normal-metal bars, aspect ratios) to derive design rules that minimize non-uniform current flow and associated excess noise
Combine SSM with time-resolved measurements or correlate with X-ray pulse response to directly link spatial current reconfiguration to pulse-shape variation and energy-resolution degradation

Replicability: No code or data availability is mentioned in the abstract. Reproduction requires access to a scanning SQUID microscope with sub-micron pickup loop, a dilution/ADR cryostat capable of reaching TES transition temperatures (~100 mK), fabricated MoAu TES devices, and current-reconstruction inversion software — specialized equipment largely limited to a few labs (e.g., NIST, Stanford).

Research Gaps:

Existing TES models do not incorporate spatial inhomogeneity in the superconducting state or current flow driven by geometry and proximity effects
The connection between imaged mesoscale current reconfigurations and macroscopic TES performance metrics (noise, energy resolution, pulse shape) remains to be quantified

3. Coherent 1-μm-Band Co-Transmission with C+L-Band Channels over 50-km Hollow-Core Fiber

Authors: Siyu Chen, Zichen Qian, Mingming Zhang... Published: 2026-11-02 | Citations: 0 arXiv | PDF

Research Question: Can the 1-μm band serve as a viable additional transmission window alongside C+L bands in hollow-core fiber (HCF) systems, bypassing gas-absorption losses that limit C/L-band performance in HCFs?

Summary: The paper experimentally demonstrates simultaneous coherent transmission in the 1-μm band alongside C+L band channels over 50 km of hollow-core fiber, carrying 54×125-GBaud PDM-16QAM C+L channels and 2×15-GBaud PDM-QPSK 1-μm channels. It positions the 1-μm window as a viable additional transmission band for HCF systems, circumventing gas-absorption losses that constrain C/L-band performance.

Key Results: Experimentally demonstrated simultaneous coherent co-transmission over 50-km HCF carrying 54×125-GBaud PDM-16QAM channels in C+L bands plus 2×15-GBaud PDM-QPSK channels in the 1-μm band, showing feasibility of combined multi-band HCF transmission.

Key Findings:

Coherent transmission across 1-μm and C+L bands can be achieved simultaneously over 50-km HCF without fundamental incompatibility
The 1-μm band avoids gas-absorption constraints affecting C/L bands in HCFs, offering complementary spectral capacity
High-order modulation (PDM-16QAM at 125 GBaud) in C+L and PDM-QPSK at 15 GBaud at 1 μm coexist in a shared HCF link

Technical Novelty: First demonstration of coherent co-transmission combining the 1-μm band with C+L bands over hollow-core fiber, exploiting HCF's low-loss windows outside the gas-absorption regions that constrain conventional C/L operation in HCFs.

What's New: Prior HCF research focused predominantly on C/L-band operation; this work pioneers tri-band coherent co-transmission that explicitly leverages the 1-μm window to sidestep gas-absorption penalties unique to hollow-core designs.

Extension Opportunities:

Scale up the 1-μm channel count and symbol rate (beyond 2×15-GBaud PDM-QPSK) toward higher-order modulation like 16/64QAM to match C+L spectral efficiency
Extend transmission distance beyond 50 km and characterize wavelength-dependent attenuation, dispersion, and inter-band crosstalk over longer HCF spans with amplification
Develop integrated multi-band amplifiers (e.g., Yb-doped for 1-μm plus EDFA/Raman for C+L) and ROADM architectures tailored to tri-band HCF networks

Replicability: No code/data mentioned; replication requires specialized optics hardware — 50-km HCF, Yb-based 1-μm transceivers, C+L coherent transceivers at 125 GBaud, and multi-band WDM/amplification infrastructure. Not computationally reproducible.

Research Gaps:

Limited 1-μm channel count and lower symbol rate/modulation order relative to C+L, leaving spectral efficiency parity unresolved
Only 50 km demonstrated — long-haul behavior, inter-band nonlinear interactions, and amplification strategies for 1-μm in HCF remain open

🔥 GitHub Trending

1. Ebony-Vinyl/dsh-our-free-model

⭐ 58823 stars | JavaScript

在 dsh 里装上这个插件即可，无需登录、注册或填 API Key，就能使用包括 DeepSeek V4.1 Flash、Kimi K3 在内的前沿模型——完全免费，不限量。 All you do is install this plugin in dsh: no login, no sign-up, no API key — the frontier models are just there,

ai-agents deepseek deepseek-harness developer-tools dsh dsh-plugin

2. yetone/magpie

⭐ 29010 stars | Go

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

claude-code codex deepseek gemini-cli llm macos

3. google/elephant-goldfish-model

⭐ 11158 stars | Python

Open-source skills for designing and reviewing software with AI agents

agent-skills ai-agents antigravity claude-code claude-code-plugin claude-skills

4. KKKKhazix/AIHOT

⭐ 9369 stars | TypeScript

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

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

5. deeplethe/utopia

⭐ 8665 stars | Rust

World's first open-source enterprise world model.

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

6. jev-chat/jev-chat-jarvis

⭐ 8227 stars | Kotlin

The chat decision assistant: before you reply, Jev reads the chat, judges intent and risk, and drafts replies you fill in with one tap. You press send. Android · Windows · macOS · iOS | 聊天辅助决策工具：先看懂对方

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

7. Human-Agent-Society/reef

⭐ 7790 stars | Python

Infrastructure for continually self‑improving agents

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

8. ccch1mneyyy/dsh-Go

⭐ 7211 stars | Dart

dsh-go — DeepSeek Harness 装进口袋：手机本机运行，也能遥控你的电脑。

ai-agent android android-app coding-agent deepseek deepseek-harness

9. elie222/rakazo

⭐ 6830 stars | TypeScript

Open-source Grok Bot alternative. Choose your own model and sandbox.

ai-agents chatgpt docker electron expo grok

10. wangzan101/Mika

⭐ 5992 stars | Kotlin

An on-device AI agent for Android that operates apps and the web to get things done for you.

accessibility ai-agent android automation jetpack-compose kotlin

11. syv-ai/HyperQwen

⭐ 2510 stars | Python

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

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

12. mw00/project-maya

⭐ 2464 stars | Cuda

Project Maya - run GLM-5.3-Flash (321B MoE), DeepSeek-V4.1-Flash-1017 (748B MoE) and MiniMax-M3.1-Flash (428B MoE) privately on your own GPU(s), from one card to 16, NVIDIA or AMD, Linux or Windows. E

cuda deepseek deepseek-v4 glm glm-5 llm

13. timoncool/YuE2-Studio

⭐ 1841 stars | TypeScript

Local AI song generator with an editable score — YuE2 on your GPU: full songs with vocals, sheet music, covers, exact replay. Native Windows app, no Python, installer with auto-update.

abc-notation ai-music covers cuda ggml local-ai

14. datawhalechina/zero-to-sglang

⭐ 1694 stars | Python

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

cuda datawhale inference-engine infra kv-cache llm

15. Tencent-Hunyuan/AuK

⭐ 1573 stars | Python

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

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

Generated by Research Pulse on 2026-11-03 09:05