2026-09-09·Daily digest

🔬 Research Pulse

Daily Digest

September 09, 2026

🤖 AI

🧠 LLMs

1. PlannerForge: LLM Agents for Scenario-Based Testing of Motion Planners in Autonomous Driving

Authors: Yuan Gao, Sebastian Müller, Mattia Piccinini... Published: 2026-09-08 | Citations: 0 arXiv | PDF

Research Question: How can the fragmented scenario-based testing pipeline for Autonomous Driving Systems (ADSs) — where scenario generation, retrieval, modification, execution, and analysis are handled by disconnected tools — be unified into a single LLM-agent framework that also enables ADS enhancement and benchmarking?

Summary: PlannerForge is the first end-to-end LLM-agent framework that unifies all stages of scenario-based testing for autonomous driving motion planners, adding LLM-driven ADS enhancement and benchmarking. It outperforms specialized baselines (Scenario Factory 2.0, BM25, From-Words-to-Collisions) across generation, retrieval, and modification tasks, and improves planner success from 50.4% to 70.2% via LLM-guided cost-tuning without fine-tuning.

Key Results: PlannerForge was evaluated with 10 off-the-shelf LLMs across 6 tasks under 5 prompt conditions, achieving best-per-task scores of 0.88-1.00. Open-source Qwen3.6:35B matches commercial APIs on 3 of 5 tasks. End-to-end chaining retains 83% (commercial) / 78% (open) of seed queries. It outperforms Scenario Factory 2.0 on NL generation (193 vs. 144 executable of 200), realizes 92-96% of requested attributes, beats BM25 at rank-1 selection (92.0% vs. 67.5%), and surpasses From-Words-to-Collisions on physically valid edits (≥94% vs. 31%). At N=400, cost-tuning lifts planner success from 50.4% to 70.2% and cuts collisions from 19.0% to 8.4%.

Key Findings:

Open-source 20-35B LLMs (Qwen3.6:35B) match commercial APIs on 3 of 5 pipeline tasks, making the framework viable without proprietary APIs
End-to-end chaining retains 78-83% of seed queries, showing modular LLM pipelines can compose without catastrophic drop-off
LLM-driven cost-function tuning at N=400 nearly halves collisions (19.0% → 8.4%) and boosts success by ~20 points without domain-specific fine-tuning

Technical Novelty: First unified LLM-agent framework covering the entire scenario-based testing pipeline (Generation → Selection → Modification → Module Routing → Planner Testing → Enhancement), plus two new LLM-driven stages (ADS Enhancement and ADS Benchmarking) not present in prior modular pipelines. Demonstrates that cost-function tuning via LLM feedback can improve planner performance without fine-tuning.

What's New: Prior LLM work on ADSs targeted isolated sub-systems (perception, planning, control); no prior framework unifies the full scenario-based testing lifecycle. PlannerForge additionally introduces LLM-based ADS enhancement, converting testing outputs into planner improvements.

Extension Opportunities:

Integrate closed-loop perception and sensor-noise models so scenarios stress the full perception-planning-control stack rather than motion planners alone
Add automated root-cause analysis agents that classify planner failures (e.g., prediction errors vs. cost-function misweighting) and propose targeted patches
Extend to multi-agent adversarial scenario generation where LLM agents co-evolve attacker/defender vehicle behaviors to surface long-tail edge cases

Replicability: Abstract does not explicitly mention a code release. Compute is moderate — inference on 20-35B open-source models (e.g., Qwen3.6:35B) requires a single high-memory GPU or multi-GPU node; commercial API access needed for baseline comparisons. A CommonRoad-style simulator is likely required for planner execution.

Research Gaps:

Scope is limited to motion planners in simulation — real-world sensor pipelines, weather, and hardware-in-the-loop are not addressed
17-22% of seed queries are lost across end-to-end chaining, indicating robustness gaps in inter-module handoffs

2. ReCite: Agentic Reasoning for Faithful Citation

Authors: Yuyang Huang, Bobo Li, Jiajia Song... Published: 2026-09-08 | Citations: 0 arXiv | PDF

Research Question: How can automated citation recommendation systems avoid misattribution—citing real but logically unsupportive papers—when semantic similarity retrieval fails to verify claim-evidence consistency?

Summary: ReCite reframes automatic citation from semantic-similarity retrieval to claim-level reasoning, using a decoupled agent that plans queries and reflectively verifies whether retrieved papers logically support the claim. Trained on synthesized reasoning trajectories, this lightweight system beats much larger generative models on strict citation accuracy.

Key Results: ReCite, a lightweight decoupled agentic framework trained on synthesized reasoning trajectories, outperforms state-of-the-art massive generative models on strict citation accuracy. Specific benchmark names, dataset sizes, and numerical deltas are not disclosed in the abstract.

Key Findings:

Semantic-similarity retrieval frequently misattributes claims to authentic but logically unsupportive papers, a failure mode distinct from hallucinated citations.
Decoupling location perception, intent-aware query planning, and reflective verification enables self-correction when candidates fail claim-evidence checks.
A small agent trained on synthesized reasoning trajectories outperforms state-of-the-art massive generative models on strict citation accuracy.

Technical Novelty: A decoupled agentic pipeline combining (1) location perception of where a citation is needed, (2) intent-aware query planning, and (3) reflective verification with self-correction loops triggered by claim-evidence inconsistency—trained on synthesized reasoning trajectories rather than relying on similarity retrieval alone.

What's New: Prior RAG citation systems focused on eliminating fabricated references; ReCite instead targets misattribution by introducing explicit claim-evidence verification and self-correction, shifting the paradigm from similarity search to verifiable logical grounding.

Extension Opportunities:

Extend the reflective verification loop to multi-hop citation chains where a claim requires composing evidence from multiple papers rather than a single source.
Apply the location-perception + intent-aware planning decomposition to related grounding tasks such as legal case citation or clinical guideline attribution.
Distill the synthesized reasoning trajectories into a smaller retrieval reranker that can be plugged into existing RAG writing assistants without a full agent loop.

Replicability: The abstract does not mention a code or data release. Reproduction would require synthesizing reasoning trajectories and fine-tuning a small agent LLM, which is feasible on modest GPU compute (single-node, few A100s) given the emphasis on the framework being 'lightweight'.

Research Gaps:

No reported evaluation of how the approach scales to multi-citation or multi-hop claims requiring composed evidence.
Unclear how robust the reflective verifier is to adversarial or subtly related papers that share terminology but not logical support.

👁️ Vision

1. Let It Go or Learn to Self-Correct: Continuous Diffusion for Constrained Discrete Tasks

Authors: Mariia Drozdova, Stéphane Liem Nguyen, François Fleuret Published: 2026-09-08 | Citations: 0 arXiv | PDF

Research Question: Why do standard DDPM samplers fail on globally constrained discrete tasks (Sudoku, graph connectivity, Latin squares, N-queens), and how can we better align continuous diffusion training/inference for discrete reasoning where early errors compound?

Summary: The paper shows that standard DDPM samplers preserve early discrete errors in globally constrained tasks like Sudoku, and that simply sampling from the model's clean x0 prediction dramatically improves validity (31%→95% on Sudoku). It further proposes self-correction training to close the train-test distribution gap so standard samplers also work well.

Key Results: Sampling directly from the model's clean prediction (x0-prediction) rather than the standard noisy-state-preserving reverse step improves Sudoku validity from 31% to 95% without retraining, with consistent gains on graph connectivity, Latin squares, and N-queens. Adding self-correction training (exposing model to own predictions) further boosts standard samplers' performance.

Key Findings:

Standard DDPM reverse sampling locks in early errors in globally constrained discrete tasks
Switching to direct x0-prediction sampling raises Sudoku validity from 31% to 95% with no retraining, with consistent gains across other constrained tasks
Self-correction training (exposing the model to its own predictions) closes the train-test mismatch and substantially improves standard samplers

Technical Novelty: Two contributions: (1) empirical demonstration that x0-direct sampling beats standard DDPM ancestral sampling for globally constrained discrete tasks, and (2) self-correction training that exposes the model to its own noisy predictions during training to reduce reverse-trajectory drift from the forward noising distribution.

What's New: Prior discrete diffusion work focuses on token-level noise processes; this paper instead diagnoses a specific failure mode of continuous diffusion on discrete constraint satisfaction — reverse-trajectory drift off the forward noising manifold — and offers both a training-free sampler fix and a training-time remedy.

Extension Opportunities:

Extend self-correction training to discrete diffusion models (D3PM, masked diffusion) for tasks like code generation or constrained text
Apply the clean-prediction sampler to combinatorial optimization problems (TSP, SAT solving, graph coloring) and benchmark against neural combinatorial baselines
Combine with reasoning-heavy LLM decoding: use self-correction training to fine-tune diffusion LMs for math/logic tasks where early token commitment hurts

Replicability: Abstract does not mention code release. Tasks (Sudoku, N-queens, Latin squares, graph connectivity) are standard synthetic benchmarks reproducible on modest single-GPU compute — likely small-to-mid transformer/UNet models trainable in hours.

Research Gaps:

Unclear how these techniques scale to larger, less-structured discrete reasoning (natural language, code) beyond synthetic constraint puzzles
No theoretical analysis quantifying when x0-direct sampling vs self-correction training is preferable

🦾 ROBOTICS

1. DeCAL: Towards Physically-Grounded Dexterous Vision-Language-Action Models via Contact-Aware Latent Co-Imagination

Authors: Yankai Fu, Ning Chen, Junkai Zhao... Published: 2026-09-08 | Citations: 0 arXiv | PDF

Research Question: How can vision-language-action (VLA) models handle contact-rich dexterous manipulation given severe visual occlusions and complex contact dynamics that homogeneous multimodal fusion approaches fail to capture?

Summary: DeCAL is a Mixture-of-Transformers VLA model for contact-rich dexterous manipulation that adaptively fuses vision and tactile signals via contact-aware gating and jointly imagines future visuo-tactile latents to encode physical dynamics. It reports 71% average success and 83.4% progress success across tasks with strong generalization to unseen scenarios.

Key Results: DeCAL achieved state-of-the-art results across all evaluated dexterous manipulation tasks with a 71% average success rate and 83.4% progress success rate, and demonstrated strong generalization to unseen scenarios. Specific baseline comparisons and per-task numbers are not detailed in the abstract.

Key Findings:

Adaptive, contact-aware tactile gating outperforms homogeneous multimodal fusion for occlusion-heavy dexterous tasks
Jointly imagining visual and tactile futures in latent space injects useful implicit physics knowledge into the action policy
A Mixture-of-Transformers with specialized understanding/imagination/action experts reaches SOTA (71% / 83.4%) while generalizing to unseen scenarios

Technical Novelty: Two specific contributions distinguish DeCAL from prior tactile-VLA work: (1) a contact-aware gating strategy inside Adaptive Visuo-Tactile Fusion that dynamically modulates tactile contribution based on contact state rather than static concatenation/attention fusion, and (2) Visuo-Tactile Latent Co-Imagination, a joint latent dynamics model that predicts future visual+tactile states, giving the policy an implicit physics prior. These sit on a Mixture-of-Transformers backbone with dedicated experts for understanding, imagination, and action.

What's New: Unlike prior tactile-augmented VLAs that concatenate or cross-attend tactile tokens statically, DeCAL gates tactile influence by contact state and adds a co-imagination objective over both modalities — turning tactile signal from an input feature into a driver of a learned physical world model.

Extension Opportunities:

Extend the Adaptive Visuo-Tactile Fusion with additional modalities (audio, proprioceptive force/torque) using the same contact-aware gating to broaden applicability to insertion/assembly tasks
Apply the Latent Co-Imagination objective as a self-supervised pretraining scheme on large-scale unlabeled teleoperation data with tactile sensors to bootstrap physical priors for other VLA backbones
Port the MoT expert design to bimanual or humanoid whole-body manipulation where contact events across multiple end-effectors need independent gating

Replicability: A project website is provided (https://aureleopku.github.io/DeCAL) but the abstract does not confirm code, dataset, or model weight release. Reproduction would require a dexterous hand with tactile sensors (e.g., DIGIT/GelSight or similar), teleoperated demonstrations, and multi-GPU training typical of MoT-based VLA models — likely 8+ high-memory GPUs for pretraining.

Research Gaps:

Abstract does not quantify how much each component (gating vs. co-imagination vs. MoT) contributes, nor compare against strong tactile-VLA baselines by name
No discussion of sim-to-real transfer, sensor generalization across tactile hardware, or long-horizon multi-stage dexterous tasks

2. Visible-Reachable Workspace for Perception-Aware Humanoid Design

Authors: Boxi Xia, Zijiang Yang, Ryan Shin... Published: 2026-09-08 | Citations: 0 arXiv | PDF

Research Question: How should humanoid robots be designed so that kinematic reachability actually coincides with visual observability, rather than treating perception as an afterthought that forces extra body motion to acquire views?

Summary: The paper argues that reachability alone is a misleading design metric for perception-driven humanoids and proposes the visible-reachable workspace (VRW), which measures reachable poses that are also visible from feasible sensing configurations. Using a 31-DoF humanoid with independently actuated cameras, they show camera articulation and a second actuated camera dramatically expand joint visible-reachable coverage, translating to measurably faster and more energy-efficient two-target manipulation.

Key Results: On a 31-DoF humanoid with independently actuated RGB-D cameras: camera articulation raised visible-reachable coverage from 38% to 97%; adding a second actuated camera lifted pairwise (two-region concurrent) coverage from 0.45 to 0.95, with a third camera yielding only marginal gains to 0.97. In a two-target reach-and-grasp benchmark, dual-actuated cameras cut mean completion time by 17% and mechanical energy by 19% vs. the same robot with fixed cameras. Hardware trials demonstrated simultaneous observation and manipulation of front/back and left/right target pairs without torso reorientation.

Key Findings:

Camera articulation alone increases single-region visible-reachable coverage from 38% to 97% on the same body.
A second actuated camera doubles concurrent two-region coverage (0.45 → 0.95); a third gives negligible marginal gain (→0.97), suggesting two is the sweet spot.
Downstream task benefits are real, not just geometric: 17% faster completion and 19% less mechanical energy on a controlled two-target reach-and-grasp, without torso reorientation.

Technical Novelty: Introduces the visible-reachable workspace (VRW) as a design-stage metric that conjoins reachability and visibility conditioned on the pose actually required to reach a target, and extends it to concurrent visibility of spatially separated regions — moving beyond classical reachability maps that ignore sensing, and beyond gaze-planning work that treats perception downstream of a fixed morphology.

What's New: Prior humanoid design copies human form factors and treats perception as a control-time problem; workspace analyses historically measure only kinematic reach. This work redefines the workspace metric itself to be perception-conditioned and shows it changes hardware decisions (camera count, actuation, placement) at design time.

Extension Opportunities:

Extend VRW to dynamic/temporal settings — evaluate visible-reachable coverage under moving targets, occlusions, or bimanual coordination where visibility must be maintained through a trajectory, not just a pose.
Use VRW as a differentiable objective inside co-design optimization loops (link lengths, joint placement, camera mount topology) to auto-generate humanoid morphologies from task specifications rather than hand-picking a 31-DoF layout.
Generalize the metric beyond RGB-D to sensor fusion (event cameras, tactile-visual, wrist-mounted cameras) and quantify diminishing returns for k>3 sensors across broader task distributions.

Replicability: Authors commit to open-sourcing all software and the full humanoid hardware design (DukeHumanoidv2). Reproducing the simulation-based VRW analysis is modest compute (workspace sampling); full hardware reproduction requires fabricating a 31-DoF humanoid with actuated RGB-D camera mounts — substantial mechanical/electrical investment.

Research Gaps:

VRW is evaluated on static targets and pose-level visibility; trajectory-level visibility, occlusion dynamics, and manipulation under clutter are not addressed.
The benchmark is a controlled two-target reach-and-grasp on one robot instance — generality across task distributions, morphologies, and comparison against learned active-perception baselines remains open.

3. Proxy Policy Steering

Authors: Chuanruo Ning, Tianrui Wang, Wei-Chiu Ma... Published: 2026-09-08 | Citations: 0 arXiv | PDF

Research Question: How can generalist robot policies be specialized to new tasks from limited demonstrations without degrading their broad pre-trained capabilities, especially when base model parameters may be inaccessible?

Summary: PPS introduces an inference-time adaptation method for generalist robot policies that trains two small proxy models whose velocity-space difference steers a frozen base sampler, avoiding any modification to the base. This preserves broad pre-trained capabilities while specializing to new tasks, achieving 53% absolute success rate gains over pi 0.5 across 12 manipulation tasks.

Key Results: PPS lifted the state-of-the-art pi 0.5 base policy by 53% absolute success rate on average across 8 real-world and 4 simulation manipulation tasks. Achieved zero-to-one gains on tasks the base never solves. Outperformed LoRA fine-tuning, from-scratch specialists, residual policies, and prior inference-time steering methods while preserving base capabilities like recovery from failure not shown in demonstrations.

Key Findings:

Velocity-space residual between reference and task proxies effectively isolates task-supervision-induced behavioral change under identified conditions
Frozen-base steering preserves emergent behaviors like failure recovery that are absent from the target demonstrations
PPS beats LoRA fine-tuning, from-scratch specialists, residual policies, and prior steering baselines, and works even without base parameter access

Technical Novelty: Training two lightweight proxy policies (reference + task) whose calibrated velocity-space difference forms a residual that steers a frozen diffusion/flow base sampler at every denoising step, requiring only forward velocity predictions from the base — no parameter access needed. Formalizes conditions under which the residual isolates task-induced behavioral change.

What's New: Unlike LoRA/fine-tuning (which modifies base weights and risks capability degradation) or residual policies (which typically add action-space corrections), PPS operates in the denoising velocity space of the base sampler and requires only forward passes through the base — enabling black-box adaptation of proprietary generalist policies.

Extension Opportunities:

Apply PPS to other generalist foundation models beyond pi 0.5 (e.g., RT-2, OpenVLA) to test generality of the velocity-space steering approach
Extend to multi-task adaptation where a single proxy pair steers behavior across a family of related tasks, sharing compute
Investigate proxy distillation to compress the two-proxy inference cost into a single model, or explore adaptive steering strength based on denoising step confidence

Replicability: The abstract does not mention code or data release. Reproduction would require access to pi 0.5 base policy (forward inference only), demonstration datasets for the 8 real-world + 4 simulation tasks, and modest compute for training two lightweight proxies (likely feasible on a single GPU given the 'lightweight' framing). Real-robot validation requires manipulation hardware.

Research Gaps:

Scalability and cost of running two proxies plus base at inference on latency-sensitive robots is not addressed in the abstract
Behavior when target task conflicts semantically with base priors, or when demonstrations are extremely scarce/noisy, is unexplored

💻 COMPUTE

1. Tools-CC-Bench: a Benchmark Suite for Collective Communication with Compression in HPC and AI Workloads

Authors: Haozhe Fan, Wei Wang, Xingchen Liu... Published: 2026-09-08 | Citations: 0 arXiv | PDF

Research Question: How can we systematically evaluate communication compression techniques for distributed HPC and LLM workloads when existing benchmarks lack support for diverse backends, realistic datasets, application-specific accuracy metrics, and overlap-induced resource contention?

Summary: CC-Bench is a lightweight, extensible benchmark suite for evaluating communication compression in distributed HPC and LLM workloads under realistic conditions. It uses declarative modeling to decouple profiling from backends, datasets, and fidelity metrics, and combines function-level interception with hardware counter monitoring to characterize latency, hardware utilization, numerical fidelity, and computation interference across three compression libraries on CPU and GPU clusters.

Key Results: The paper presents CC-Bench and uses it to evaluate three compression-enabled communication libraries across both CPU and GPU clusters, using representative datasets from HPC and LLM workloads. It characterizes per-phase latency, hardware utilization, numerical fidelity, and computation interference, revealing accuracy-performance trade-offs and deployment bottlenecks. Specific numerical results (throughput/compression ratios) are not disclosed in the abstract.

Key Findings:

Existing communication-compression benchmarks fail to capture overlap-induced resource contention, application-specific accuracy metrics, and cross-backend portability
Function-level interception combined with hardware counter monitoring exposes per-phase bottlenecks that black-box timing benchmarks miss
Evaluating three compression-enabled libraries on CPU and GPU clusters reveals concrete accuracy-performance trade-offs relevant for practical deployment

Technical Novelty: The declarative application-environment modeling that decouples profiling logic from communication libraries, datasets, and fidelity metrics — enabling portable cross-library evaluation — combined with function-level interception plus hardware counter monitoring to capture computation interference under overlap conditions. Prior benchmarks were tightly coupled to specific backends and ignored contention effects.

What's New: Unlike prior benchmarks tied to specific communication libraries or synthetic workloads, CC-Bench uses declarative application-environment modeling for portability across libraries, incorporates realistic HPC and LLM datasets, and measures both numerical fidelity and computation interference — dimensions typically ignored in collective-communication benchmarking.

Extension Opportunities:

Add support for emerging compression algorithms (e.g., learned/neural compressors, mixed-precision quantization schemes like FP8/FP4) and evaluate them against classical lossy/lossless methods in the CC-Bench framework
Extend the declarative application-environment modeling to cover multi-node heterogeneous topologies (e.g., NVLink+InfiniBand+Ethernet hybrids) and characterize how compression interacts with collective algorithms like ring vs. tree allreduce
Integrate CC-Bench profiling into training frameworks (DeepSpeed, Megatron, PyTorch FSDP) to auto-tune compression policies per layer/gradient based on real-time fidelity budgets

Replicability: The abstract does not mention code/data availability. Reproduction would require a multi-node CPU+GPU cluster (likely with NVIDIA GPUs and a high-speed interconnect like InfiniBand/NVLink), the three tested compression libraries, and representative HPC + LLM datasets. Moderate-to-high compute cost given multi-library, multi-cluster evaluation.

Research Gaps:

Lack of quantitative results in the abstract makes it hard to assess magnitude of the trade-offs discovered
Coverage limited to three compression libraries; emerging techniques like learned compression, sparsification variants, and low-precision collective algorithms are not confirmed to be included

2. HDA-MoE: Hybrid Parallelism and Dynamic, Adaptive Scheduling for Mixture-of-Experts with 3D Near-Memory Processing

Authors: Haochen Huang, Shuzhang Zhong, Shengxuan Qiu... Published: 2026-09-08 | Citations: 0 arXiv | PDF

Research Question: How can Mixture-of-Experts (MoE) inference be efficiently mapped onto 3D Near-Memory Processing (NMP) architectures given that existing parallelization strategies (TP, EP) suffer from high communication costs or unbalanced compute utilization, and MoE's dynamic routing further complicates deployment?

Summary: HDA-MoE is a framework for accelerating MoE LLM inference on 3D Near-Memory Processing architectures via a hybrid TP+EP offline mapping strategy combined with online adaptive scheduling to handle dynamic expert routing. It reduces communication overhead and improves compute utilization, delivering 1.1x-3.7x speedups over TP, EP, and hybrid baselines including the prior HD-MoE work.

Key Results: HDA-MoE achieves speedups of 1.1x-3.4x over Tensor Parallelism (TP), 1.1x-1.5x over Expert Parallelism (EP), 1.1x-3.7x over a Hybrid TP-EP compute-balanced baseline, and 1.1x-1.3x over HD-MoE (prior state-of-the-art). Evaluations are on MoE LLM inference workloads mapped to 3D NMP architectures with hybrid-bonded memory-compute integration.

Key Findings:

Pure TP and EP strategies are suboptimal for MoE on NMP due to communication costs and load imbalance respectively; hybrid mapping outperforms both
Static hybrid mapping alone is insufficient — dynamic routing requires runtime adaptive scheduling to maintain utilization
3D NMP's high internal bandwidth makes it a compelling substrate for MoE, but exploiting it requires workload-aware placement

Technical Novelty: Combining an offline hybrid TP+EP mapping algorithm (co-optimizing communication vs. compute balance) with an online adaptive scheduler that reacts to MoE's dynamic routing at runtime, specifically tailored for the distributed memory-compute organization of 3D NMP hardware. Prior work (HD-MoE) used hybrid parallelism but lacked adaptive runtime scheduling for dynamic routing.

What's New: First to combine offline hybrid parallel mapping with online dynamic/adaptive scheduling specifically for MoE on 3D NMP; extends prior HD-MoE by adding runtime adaptivity to the router's dynamic behavior rather than relying solely on static mapping.

Extension Opportunities:

Extend the offline hybrid parallel mapping algorithm to handle heterogeneous NMP dies with varying compute/memory capacities, enabling deployment on multi-vendor stacked systems
Integrate speculative expert prefetching with the online dynamic scheduler to further hide routing-induced latency for very sparse MoE models (e.g., DeepSeek-V3 style with 256+ experts)
Adapt the framework to training workloads (not just inference), which introduces gradient synchronization and optimizer state distribution challenges on NMP

Replicability: Code is available at https://github.com/PKU-SEC-Lab/HDA-MoE-TCAD26. Reproduction likely requires a 3D NMP architectural simulator (not commodity GPUs), so full replication needs simulation infrastructure rather than physical hardware — reasonable compute (single workstation) for simulation-based studies.

Research Gaps:

Evaluation is simulation-based on NMP architectures rather than real silicon, leaving open questions about physical thermal/yield constraints of 3D hybrid bonding at scale
Focus is on inference; training on NMP with MoE and its all-to-all communication remains unaddressed

3. Sample-Guided Exact Top-K Selection for Long-Context Sparse Attention

Authors: Siran Liu, Yang Xue, Theo Tang... Published: 2026-09-08 | Citations: 0 arXiv | PDF

Research Question: How to accelerate the exact Top-K selection stage in long-context sparse attention, given that production radix selectors require a full row-scale pass before finding actionable rank boundaries, doubling traversal cost on ragged score rows?

Summary: HPC-Ops Top-K accelerates exact Top-K selection for long-context sparse attention by using a fixed-stride sample of the score row to propose a coarse rank boundary, then certifying and refining it during the mandatory complete-row pass — avoiding a second row-scale traversal. The fused GPU kernel achieves a 1.55x geometric-mean speedup over the fastest external exact baseline while preserving exact correctness.

Key Results: HPC-Ops Top-K outperforms the fastest verified external exact baseline by 1.29–1.75x across 20 operator configurations (1.55x geometric-mean speedup) on Hy4-Preview indexer scores, and achieves 1.36x and 1.48x speedups on two framework-derived sparse-attention traces.

Key Findings:

Fixed-stride partial-row samples remain rank-calibrated across ragged lengths, enabling coarse boundary proposals without extra passes
Fusing complete-row certification with candidate formation eliminates the second row-scale traversal that standard radix selectors require
1.29–1.75x speedup across 20 configurations, plus 1.36x/1.48x on two end-to-end sparse-attention traces

Technical Novelty: A fixed-stride partial-view sample proposes a row-local coarse Top-K boundary that stays calibrated to the complete-row rank across ragged lengths, certified during the mandatory complete-row pass; a nested secondary boundary handles underfilled proposals, and the GPU kernel fuses certification + candidate formation with persistent/KV-split/direct-exact execution paths behind graph-capturable dispatch. Sampling accelerates the common path without ever sacrificing exactness.

What's New: Prior radix-based exact selectors discover their first actionable rank boundary only after a full row pass, then require a second refinement pass. This paper is the first to show that a cheap fixed-stride sample yields a boundary tight enough to fold refinement into the certification pass, with a nested fallback that preserves exactness on underfilled proposals.

Extension Opportunities:

Adapt the sample-guided coarse-boundary technique to other selection primitives (e.g., top-p/nucleus sampling, MoE routing) where ragged-row work dominates
Explore adaptive stride selection based on row-length statistics or score-distribution priors instead of fixed-stride sampling
Extend the ragged-row dispatch and graph-capturable execution model to non-NVIDIA accelerators (AMD/TPU/Ascend) and integrate with additional sparse-attention indexer families beyond Hy4-Preview

Replicability: Code is open-sourced at github.com/Tencent/hpc-ops as part of Tencent's HPC-Ops operator library. Reproduction requires GPUs supporting CUDA graph capture (likely NVIDIA H-class or similar) and access to Hy4-Preview indexer score traces; hardware specifics are not detailed in the abstract.

Research Gaps:

No evaluation on non-indexer Top-K workloads (e.g., generic ranking, retrieval) or non-GPU accelerators
Behavior under adversarial or highly skewed score distributions where fixed-stride sampling may frequently underfill is only briefly addressed via the secondary-boundary fallback, not quantified

⚡ ENERGY

1. Water, vacancies, and competing exchange interactions in Prussian blue analogues: a neutron diffraction study of field and dehydration-driven magnetic transitions

Authors: N. S. Dhami, C. V. Colin, V. Nassif... Published: 2026-09-08 | Citations: 0 arXiv | PDF

Research Question: How do water content, cyanide vacancies, and competing exchange interactions jointly govern the structural and magnetic behavior of Prussian blue analogue (PBA) frameworks, and what causes the anomalous magnetic response of MnFe-based PBAs?

Summary: The authors use temperature- and field-dependent neutron diffraction on a family of Prussian blue analogues to show that correlated cyanide vacancies and interstitial water produce a characteristic diffuse scattering feature and control both thermal expansion and magnetic ordering. They demonstrate that MnFe uniquely adopts a frustrated k=(1,0,0) magnetic structure that can be flipped to the common collinear ferrimagnetic k=(0,0,0) state either by a 1.3 T field or by dehydration, explained by near-compensation of Mn–Fe and frustrated Mn–Mn exchange.

Key Results: Neutron diffraction across 2–450 K on A4[Fe(CN)6]2.7 (A=Co, Mn, Ni), Rb2Ni4[Fe(CN)6]3.3, and Mn4[Cr(CN)6]2.7 shows: (1) a low-angle diffuse feature signaling correlated vacancies+interstitial water, absent in the cation-stabilized Rb2NiFe reference; (2) a continuous positive-to-negative thermal expansion crossover in CoFe and MnFe upon dehydration, while NiFe stays robust to 450 K; (3) collinear ferrimagnetic k=(0,0,0) ground states for all except MnFe, which orders with k=(1,0,0); (4) a critical field Bc=1.3 T drives MnFe into the k=(0,0,0) state, and dehydration alone reproduces the same transition; (5) a minimal Heisenberg model attributing the effect to near-compensation between AFM Mn–Fe coupling and geometrically frustrated AFM Mn–Mn on the fcc sublattice.

Key Findings:

A low-angle diffuse neutron scattering feature is a universal fingerprint of correlated vacancies + interstitial water in vacancy-bearing PBAs, absent in the cation-stabilized Rb2NiFe framework.
CoFe and MnFe undergo a continuous positive-to-negative thermal expansion crossover upon dehydration, whereas NiFe is structurally robust to 450 K.
MnFe sits near a magnetic compensation point: Bc=1.3 T or dehydration converts its frustrated k=(1,0,0) order to the collinear k=(0,0,0) ferrimagnetic state common to the family, explaining prior Fe K-edge XMCD anomalies.

Technical Novelty: First unified neutron diffraction study linking correlated vacancy/water disorder to a specific diffuse scattering signature across the PBA family, and first demonstration that dehydration and moderate field independently drive the same k=(1,0,0)→k=(0,0,0) spin reorientation in MnFe — resolving a longstanding Fe K-edge XMCD ambiguity via a minimal Heisenberg frustration model.

What's New: Unlike prior single-compound PBA studies, this work provides a comparative neutron picture across five compositions, ties the diffuse scattering to correlated vacancy-water order, and identifies water content — not just composition — as a control parameter for magnetic state, closing an XMCD interpretation gap.

Extension Opportunities:

Systematically vary Rb/alkali-cation stoichiometry to tune the vacancy-water diffuse scattering and correlate with magnetic ordering vector — build a phase diagram of vacancy order vs. spin structure.
Apply the same minimal Heisenberg model + in situ neutron diffraction protocol to Na/K-ion battery cycling of MnFe PBA electrodes to test whether the k=(1,0,0)→k=(0,0,0) transition is a marker of hydration state during operando charge/discharge.
Extend to mixed A-site (e.g., Mn1-xNixFe) PBAs to engineer the Mn–Mn frustration and locate the exact magnetic compensation point predicted by the model.

Replicability: No code/data availability mentioned in the abstract. Reproduction requires access to a neutron diffractometer with variable-temperature (2–450 K) and applied-field (>1.3 T) capability (e.g., ILL D1B/D20 or equivalent), plus PBA sample synthesis; modeling is a tractable Heisenberg mean-field/Monte Carlo calculation runnable on a laptop.

Research Gaps:

No operando measurements linking these hydration-driven magnetic transitions to actual battery cycling behavior in PBA electrodes.
The minimal Heisenberg model is qualitative; quantitative exchange constants and their dependence on local water/vacancy environment remain to be pinned down (e.g., via inelastic neutron scattering or DFT+U).

2. Capacitance of Undoped Thin-Film Diodes

Authors: Oskar J. Sandberg, Mathias Nyman, Stefan Zeiske... Published: 2026-09-07 | Citations: 0 arXiv | PDF

Research Question: How can the capacitance of undoped thin-film diodes—where injected carriers dominate over doping-induced carriers—be described analytically, and how can the built-in voltage be reliably extracted in such devices?

Summary: The paper derives the first analytical framework for capacitance in undoped thin-film diodes, correctly accounting for injected carriers and electrode charges rather than doping-induced depletion. It validates the theory with drift-diffusion simulations and organic solar cell experiments, and introduces a practical method for extracting built-in voltage from capacitance measurements.

Key Results: The authors derive a closed-form analytical description of capacitance in undoped thin-film diodes that fully accounts for injected charge carriers and electrode charge effects. They validate the framework via drift-diffusion numerical simulations and experimental measurements on organic solar cells, and demonstrate a method to extract the built-in voltage from capacitance data. Specific numerical values are not reported in the abstract.

Key Findings:

Capacitance in undoped thin-film diodes is dominated by injected carriers and electrode charges, not doping
A closed-form analytical expression captures this behavior and matches drift-diffusion simulations
Built-in voltage can be extracted directly from capacitance data using the new framework, validated on organic solar cells

Technical Novelty: Prior capacitance models for thin-film diodes assumed doping-dominated depletion physics (Mott-Schottky-like). This paper provides the first analytical treatment where injected carriers and electrode (geometric/image) charges are the dominant contributors, closing a long-standing gap in device physics for organic/undoped semiconductors.

What's New: Existing Mott-Schottky-style analyses assume doped semiconductors and fail for undoped organic/thin-film devices. This work provides an analytical (not just numerical) description tailored to the injection-dominated regime, unifying theory with experiment.

Extension Opportunities:

Apply the analytical framework to perovskite thin-film solar cells and LEDs, where undoped absorbers and injected-carrier dominance also occur, to test cross-material generality
Extend the model to include trap states, ionic motion, or frequency-dependent (impedance spectroscopy) effects to broaden its applicability beyond quasi-static capacitance
Develop an open-source fitting toolkit that automatically extracts built-in voltage, mobility, and injection parameters from measured C-V curves using the derived equations

Replicability: The abstract does not mention code or dataset availability. Reproduction would require a standard drift-diffusion solver (e.g., SETFOS, Oghma-Nano, or custom Python/MATLAB) and access to organic solar cell fabrication plus capacitance-voltage measurement equipment—modest compute (single workstation) but non-trivial experimental infrastructure.

Research Gaps:

The framework's applicability to devices with significant trap densities, ionic effects, or strong disorder is not established
Frequency-dependent capacitance and impedance spectroscopy extensions remain to be developed

3. Decorated electronic kagome lattice in twisted bilayer germanene

Authors: Dennis J. Klaassen, Rian A. M. Ligthart, Andrés R. Botello Mendez... Published: 2026-09-08 | Citations: 0 arXiv | PDF

Research Question: Can atomically controlled 2D materials realize artificial kagome lattices that host electronic flat bands and correlation-driven phases, which have remained scarce despite theoretical interest?

Summary: The authors show that commensurate large-angle twisted bilayer germanene on Ge2Pt produces two moiré phases distinguished by sublattice-exchange parity: odd-parity is semiconducting and even-parity is metallic, with the metallic phase hosting an empty-state resonance forming a C3-symmetric decorated kagome pattern. This establishes buckled twisted germanene as a materials platform for engineering kagome electronic states without requiring magic-angle precision.

Key Results: Using STM measurements combined with DFT calculations on commensurate twisted bilayer germanene grown on Ge2Pt, the authors demonstrate two electronically distinct large-angle moiré phases: sublattice-exchange odd bilayers are semiconducting, while even-parity bilayers are metallic and host an empty-state resonance forming a decorated kagome structure with C3 symmetry.

Key Findings:

Commensurate twisted bilayer germanene on Ge2Pt hosts two electronically distinct moiré phases selected by sublattice-exchange parity
Odd-parity bilayers are semiconducting; even-parity bilayers are metallic
The metallic phase exhibits an empty-state resonance forming a decorated kagome lattice with C3 symmetry, confirmed by both STM and DFT

Technical Novelty: Prior kagome realizations relied on molecular assembly or specific transition-metal compounds; this work engineers a decorated kagome electronic pattern via large-angle (not magic-angle) twisted bilayers of a buckled elemental 2D material, and identifies sublattice-exchange parity as the key selector between semiconducting and kagome-hosting metallic phases.

What's New: Introduces a large-angle (rather than magic-angle) twisted bilayer route to kagome physics using a buckled elemental 2D material, and identifies sublattice-exchange parity as an experimentally accessible symmetry selector between insulating and kagome-hosting metallic phases.

Extension Opportunities:

Perform low-temperature transport and STS measurements to search for flat-band correlated phases (superconductivity, magnetism) predicted for kagome geometries in the even-parity metallic phase
Extend the platform to other buckled group-IV 2D materials (silicene, stanene) on similar substrates to explore whether the decorated kagome emerges as a universal feature of large-angle twisted buckled bilayers
Apply gating or strain to tune the empty-state resonance energy toward the Fermi level, converting the emergent kagome pattern into an accessible flat-band system

Replicability: No code/data availability mentioned in the abstract. Reproduction requires MBE growth of germanene on Ge2Pt, low-temperature STM/STS instrumentation, and DFT compute (moderate HPC for moiré supercells at large twist angles).

Research Gaps:

No direct evidence yet that the decorated kagome resonance produces flat bands at the Fermi level or hosts correlation-driven phases
Role of the Ge2Pt substrate versus intrinsic bilayer physics in stabilizing the observed states is not disentangled

🏥 HEALTHCARE

1. Performance of Clinical AI System and Physicians and Frontier Language Models in primary care diagnostics

Authors: Andy Nkansah, Hanna Plotnitskaya, Stanislau Salavei... Published: 2026-09-08 | Citations: 0 arXiv | PDF

Research Question: How does a clinical AI system (Doctorina) compare to physicians and standalone frontier LLMs on end-to-end primary care diagnostics — including adaptive information gathering, diagnosis, workup, and initial treatment — rather than diagnosis alone?

Summary: The paper benchmarks Doctorina, a clinical AI system, against 8 physicians and 4 frontier LLMs on 150 synthetic Polish primary-care consultations covering adaptive history-taking, diagnosis, workup, and treatment. Doctorina beat physicians by 25 pp on Top-1 diagnostic concordance and led on workup/treatment scores, with frontier LLMs (notably Claude Opus 5 and Kimi K3) close behind on management.

Key Results: In 150 synthetic Polish-language primary-care consultations: Doctorina reached 82.0% Top-1 diagnostic concordance vs 57.0% for 8 physicians (Δ=25.0 pp, 95% CI 17.7-32.7) and 97.3% vs 85.0% primary-or-reference-differential concordance. Across 149 case pairs, normalized workup scores were 89.4 vs 66.9 and treatment scores 83.7 vs 61.2. Among 6 groups (Doctorina, physicians, 4 frontier LLMs), Doctorina led diagnostic point estimates, Kimi K3 second; Claude Opus 5 narrowly led management, with Opus, Doctorina, and Kimi tightly clustered. A second Doctorina run reproduced the advantage.

Key Findings:

Doctorina: 82.0% Top-1 vs physicians 57.0% (Δ 25.0 pp, CI 17.7-32.7); 97.3% vs 85.0% primary-or-differential concordance
Doctorina's advantage extends beyond diagnosis: workup 89.4 vs 66.9, treatment 83.7 vs 61.2 (normalized) across 149 case pairs
Among all six groups Doctorina led diagnostics with Kimi K3 second; Claude Opus 5 narrowly led management, with Opus, Doctorina, and Kimi tightly clustered — and a repeat Doctorina run reproduced the physician gap

Technical Novelty: Evaluates diagnosis + workup + treatment jointly after adaptive multi-turn information gathering, rather than the single-shot vignette scoring used in most prior clinical LLM benchmarks; also directly compares a purpose-built clinical AI system against both physicians and frontier LLMs in the same protocol and in a non-English (Polish) primary-care setting.

What's New: Combines three under-explored axes at once: end-to-end adaptive consultation (not static vignettes), joint scoring of diagnosis + workup + treatment, and a non-English (Polish) primary-care setting — plus a head-to-head between a bespoke clinical AI, generalist frontier LLMs, and human physicians in the same protocol.

Extension Opportunities:

Replicate the evaluation on real (non-synthetic) multilingual consultation transcripts with prospective patient outcomes, not just concordance to reference diagnoses
Ablate Doctorina's scaffolding (adaptive questioning, retrieval, tool use) against a frontier LLM baseline to isolate which component drives the 25-pp diagnostic gap
Build an open-source adaptive-consultation harness that wraps any frontier LLM with Doctorina-style information-gathering loops and benchmark it on the same 150-case set

Replicability: Abstract does not mention released code, weights, or the synthetic case set. Reproducing physician-comparison numbers would require the 150 Polish cases, the Doctorina system (proprietary), 8 licensed physicians, and API access to 4 frontier models (Claude Opus 5, Kimi K3, plus 2 others) — modest inference compute but high human-eval cost.

Research Gaps:

Cases are synthetic and Polish-only, with no linkage to real patient outcomes or safety events
Frontier LLMs are evaluated 'standalone' — the paper does not isolate how much of Doctorina's edge comes from scaffolding vs the underlying model, nor whether physicians had access to comparable tools

2. NOAH: Learning the Full Patient Journey. A Longitudinal Multimodal Time-Aware Model for Representation and Forecasting

Authors: Tobias Susetzky, Raphael Rehms, Dmitrii Seletkov... Published: 2026-09-08 | Citations: 0 arXiv | PDF

Research Question: How can AI models represent and forecast the full longitudinal, multimodal patient journey while handling irregular temporal dynamics, stochasticity, open-vocabulary events, and diverse modalities — gaps left by prior discriminative, modality-limited, or forecasting-limited approaches?

Summary: NOAH is a time-aware, task-agnostic generative transformer that models complete multimodal patient journeys — images, time-series, categorical events, and clinical text — using bidirectional time integration and a variational latent space. Trained on 559M events from the MIMIC family, it supports autoregressive forecasting, zero-shot classification, and counterfactual simulation from a single unified model.

Key Results: Built NOAH on 559M+ clinical events from 431,000 hospital visits of 299,000 patients across the MIMIC dataset family. Demonstrated strong probing performance for clinical outcomes across 15 ICD chapters and 29 comorbidities, plus time-to-event prediction, autoregressive forecasting with time control, zero-shot classification, and counterfactual intervention simulation.

Key Findings:

A single generative model can jointly handle images, time-series, categorical events, and structured/unstructured records without task-specific heads
Bidirectional time integration + variational latents outperform monotonic time-as-inductive-bias approaches for irregular clinical trajectories
The learned representations transfer to 15 ICD chapters, 29 comorbidities, and time-to-event tasks under probing

Technical Novelty: Novel bidirectional time integration (vs. monotonic time as inductive bias) combined with a variational latent space to capture stochastic trajectory evolution; unified generative transformer that natively ingests images, time-series, numeric signals, categorical events, and structured/unstructured text — moving beyond closed vocabularies and discriminative-only paradigms.

What's New: First holistic generative foundation model for patient journeys that combines open-vocabulary multimodal inputs, explicit stochasticity via variational latents, and controllable time — enabling forecasting and counterfactual reasoning that prior discriminative or modality-limited models cannot support.

Extension Opportunities:

Fine-tune NOAH on non-MIMIC EHR datasets (e.g., UK Biobank, eICU, private hospital systems) to test generalization beyond ICU-heavy MIMIC data
Build a clinical decision-support interface that exposes counterfactual intervention simulation to clinicians for treatment planning
Extend the variational latent space for uncertainty-aware risk stratification, surfacing prediction confidence intervals for high-stakes forecasts

Replicability: Abstract does not mention code/weights release. MIMIC data is publicly available with credentialed access. Reproduction would require substantial compute — training a multimodal transformer on 559M events likely needs multi-GPU clusters (tens of A100/H100-days at minimum).

Research Gaps:

Evaluation is anchored to MIMIC; external validation on diverse populations and non-ICU settings is unaddressed
Calibration, fairness across demographics, and clinical safety of counterfactual simulations are not covered in the abstract

3. MedGSSR: Generalizable Medical Image Super-Resolution 3D Reconstruction via Hierarchical Feed-forward Gaussian Splatting

Authors: Chengkai Wang, Luoyu Hong, Yiting Zhao... Published: 2026-09-06 | Citations: 0 arXiv | PDF

Research Question: How can medical 3D super-resolution be performed in a generalizable, feed-forward manner without per-subject optimization, pretrained priors, or coordinate-based implicit representations that compromise anatomical fidelity?

Summary: MedGSSR is an end-to-end feed-forward framework that represents medical volumes as explicit 3D Gaussian fields for super-resolution, replacing coordinate-based implicit representations (INRs) that dominate current Med3DSR. By decoupling reconstruction into structural preservation (Pyramid Anatomical Encoder) and textural refinement (Hierarchical Gaussian Projector), plus sub-voxel Gaussian decomposition for arbitrary-scale SR, it achieves generalizable, fast inference without per-subject optimization.

Key Results: MedGSSR demonstrates state-of-the-art performance on MRI and CT benchmarks for Med3DSR, with robust generalizability across unseen datasets without per-subject optimization. Specific numerical metrics (PSNR/SSIM) are not disclosed in the abstract but the paper claims to 'significantly outperform' SOTA methods on both modalities while enabling fast inference.

Key Findings:

Explicit 3D Gaussian representations preserve high-frequency anatomical detail better than implicit coordinate networks for medical volumes
Decoupling coarse structure from fine texture via hierarchical projection improves fidelity on both MRI and CT
The framework generalizes to unseen datasets without per-subject fine-tuning, a key barrier for clinical deployment

Technical Novelty: First application of explicit 3D Gaussian Splatting to medical volumetric super-resolution in a fully feed-forward manner, combined with a Pyramid Anatomical Encoder for coarse structure and Hierarchical Gaussian Projector for fine texture, plus a Differentiable Gaussian Voxelizer with sub-voxel decomposition enabling arbitrary-scale SR.

What's New: Prior Med3DSR relies on per-subject optimization (slow), pretrained diffusion priors (hallucination-prone), or INRs (limited high-frequency fidelity). MedGSSR is the first to port 3D Gaussian Splatting — proven in novel-view synthesis — to volumetric medical SR in a feed-forward, generalizable manner with explicit sub-voxel decomposition for arbitrary scaling.

Extension Opportunities:

Extend the Hierarchical Gaussian Projector to multi-modal fusion (e.g., PET-CT or MRI-CT co-registration super-resolution)
Apply the sub-voxel Gaussian decomposition to 4D dynamic imaging (cardiac MRI, perfusion CT) where temporal continuity matters
Integrate with a diffusion prior for uncertainty-aware SR in low-dose CT reconstruction to quantify hallucination risk

Replicability: Code is promised on the project webpage (william2ai.github.io/medgssr). Reproduction likely requires multi-GPU training (typical 3D medical SR uses A100/V100 class GPUs) with standard MRI/CT benchmarks (e.g., IXI, BraTS, LIDC). Exact compute not stated in abstract.

Research Gaps:

No reported quantitative metrics (PSNR/SSIM/LPIPS) or clinical validation (radiologist reader studies) in the abstract
Unclear how the method handles anisotropic acquisitions or severely undersampled slices common in real clinical protocols

🔬 MATERIALS

1. Signatures of nodal superconductivity in stoichiometric FeTe

Authors: Cequn Li, Zi-Jie Yan, Yang Ge... Published: 2026-09-08 | Citations: 0 arXiv | PDF

Research Question: What is the superconducting pairing symmetry of stoichiometric FeTe, the previously inaccessible endpoint of the Fe(Se,Te) phase diagram, and how does it compare to intermediate Se/Te compositions?

Summary: The paper resolves the pairing symmetry of stoichiometric FeTe by combining scanning SQUID microscopy, STM/S, transport, and mean-field theory on tunable thin films, finding evidence for a nodal superconducting gap (d-wave or nodal s-wave) that departs from the more isotropic gaps at intermediate Fe(Se,Te) compositions. It also uncovers a universal crossover from weak to rapid Tc suppression as superfluid stiffness decreases, linking FeTe to broader unconventional superconductor phenomenology.

Key Results: Using scanning SQUID microscopy, transport, STM/S, and mean-field calculations on tunable-stoichiometry FeTe thin films, the authors show: (1) micrometer-scale spatial variations in superfluid stiffness and Tc even in stoichiometric samples; (2) non-saturating London penetration depth down to 0.02 Tc following a power-law with exponent ~1–1.5; (3) V-shaped low-energy density of states; (4) two-gap modeling consistent with gap nodes or deep minima, pointing to d-wave or nodal s-wave pairing; (5) a crossover from weak to rapid Tc suppression as superfluid stiffness decreases.

Key Findings:

London penetration depth follows a power law with exponent ~1–1.5 down to 0.02 Tc, and DOS is V-shaped — hallmarks of gap nodes or deep minima
Micrometer-scale spatial inhomogeneity in both superfluid stiffness and Tc persists even at stoichiometric composition
FeTe exhibits a nodal pairing state distinct from the more isotropic gaps seen at intermediate Se/Te ratios, and shows a Tc-vs-stiffness crossover mirroring other unconventional superconductors

Technical Novelty: First combination of scanning SQUID microscopy (local superfluid stiffness), STM/S, and transport on stoichiometric FeTe thin films with tunable stoichiometry, resolving the long-standing pairing question at the FeTe endpoint where prior work relied on non-stoichiometric or excess-Fe samples that obscured intrinsic behavior.

What's New: First direct experimental characterization of the pairing state at the true FeTe endpoint using local probes; establishes FeTe as a distinct nodal regime rather than a continuous extension of the Fe(Se,Te) trend toward isotropic gaps.

Extension Opportunities:

Perform angle-resolved measurements (ARPES or quasiparticle interference STM) on the same tunable-stoichiometry FeTe films to directly map node locations and discriminate between d-wave vs nodal s±
Extend the mean-field two-gap framework into a full multi-orbital microscopic model incorporating the double-stripe magnetic order unique to FeTe, and test predictions against the observed power-law exponents
Systematically sweep across the full Fe(Se,Te) phase diagram with the same scanning SQUID + STM protocol to map how nodes evolve into the more isotropic gaps at intermediate compositions

Replicability: No code/data availability is mentioned in the abstract. Reproduction would require MBE/PLD thin-film growth facilities with stoichiometry control, a scanning SQUID microscope with mK-range base temperature (to reach 0.02 Tc), cryogenic STM/S, and standard DFT/mean-field computational resources — a well-resourced condensed matter experimental group, not casually reproducible.

Research Gaps:

Cannot definitively distinguish d-wave from nodal s-wave — requires phase-sensitive or momentum-resolved follow-up
Microscopic origin of the micrometer-scale spatial inhomogeneity in stoichiometric films is not identified

2. Controlling Hydrogen Isotope Retention at Helium Cavities through Radiation-Induced Segregation in Fusion Steels

Authors: Lihao Shi, Logan N. Clowers, Qing Peng... Published: 2026-09-08 | Citations: 0 arXiv | PDF

Research Question: How can hydrogen isotope (especially tritium) retention in fusion structural alloys be controlled at the atomistic level, given that irradiation-driven defects and solute redistribution govern trapping behavior in plasma-facing materials?

Summary: The paper identifies an irradiation-enabled mechanism where Cr and Ta segregation to helium cavity surfaces in F82H steel substantially enhances hydrogen isotope trapping. Combining triple-ion irradiation experiments with DFT, the authors show solute-tuned electronic structure (Ta hybridization, Cr d-band shift) strengthens H binding and raises migration barriers, offering an alloy-chemistry route to control tritium inventory in fusion reactors.

Key Results: Triple-ion irradiation experiments on reduced activation ferritic-martensitic steel F82H demonstrated pronounced radiation-induced segregation of Cr and Ta to helium cavity surfaces. DFT calculations showed these segregants substantially increase H binding energies and migration barriers at cavity surfaces, driven by Ta's strong H 1s-metal d orbital hybridization and Cr's favorable d-band shift toward bonding configurations, leading to elevated-temperature H retention.

Key Findings:

Triple-ion irradiation of F82H causes pronounced Cr and Ta segregation to cavity surfaces
Segregants markedly increase H binding energies and migration barriers at cavities, raising retention
Mechanism is electronic: Ta enables H 1s-d hybridization; Cr shifts surface d-band into a more bonding configuration

Technical Novelty: First atomistic link between radiation-induced solute segregation at helium cavities and hydrogen isotope trapping, using triple-ion irradiation of F82H combined with DFT to reveal electronic-structure origins (d-band shifts and orbital hybridization) of solute-enhanced H binding.

What's New: Prior work treated cavity-driven H trapping and solute segregation largely separately; this paper unifies them by showing radiation-induced segregation actively reprograms cavity surface chemistry to trap H, with a specific electronic-structure mechanism.

Extension Opportunities:

Design and test alloy compositions with tuned Cr/Ta ratios (or alternative d-band engineered solutes like W, V, Mo) to minimize tritium retention while preserving mechanical properties
Extend DFT+experiment framework to other RAFM steels (EUROFER97, CLAM) and tungsten-based plasma-facing materials to build a predictive segregation-retention database
Develop multiscale kinetic Monte Carlo or rate-theory models linking segregation kinetics, cavity evolution, and H trapping to predict tritium inventory over reactor lifetimes

Replicability: No code/data availability mentioned in abstract. Reproduction requires ion accelerator facility for triple-beam (He/H/heavy ion) irradiation, APT/TEM for segregation characterization, and moderate HPC resources for DFT calculations (VASP or similar) of Fe-Cr-Ta surfaces with H — accessible to established fusion materials groups.

Research Gaps:

No quantitative link yet between segregation dose/temperature dependence and tritium inventory over reactor-relevant fluences
Effect of other RAFM alloying elements (W, V, Mn) and multi-solute co-segregation on H trapping remains unexplored

3. Stimulated Brillouin Scattering in InGaP-on-Insulator Waveguides

Authors: Yuyang Xue, Lisa-Sophie Haerteis, Ryan L. Russell... Published: 2026-09-08 | Citations: 0 arXiv | PDF

Research Question: How can integrated photonic platforms achieve simultaneously high Brillouin gain AND narrow linewidth in a scalable, CMOS-compatible process? Existing chip-scale SBS demonstrations struggle to deliver both metrics on a platform that integrates cleanly with established photonic integrated circuits.

Summary: The authors introduce an InGaP-on-insulator waveguide platform and demonstrate backward stimulated Brillouin scattering with a high gain coefficient of 588 W^-1 m^-1 and a record-narrow 5.2 MHz linewidth at a 9.346 GHz shift. The wafer-scale CMOS-compatible fabrication makes InGaP a promising scalable substrate for Brillouin photonics.

Key Results: Demonstrated backward SBS in a single-crystalline InGaP-on-SiO2 waveguide near 1550 nm with a Brillouin gain coefficient of 588 W^-1 m^-1, a record narrow linewidth of 5.2 MHz, and a Brillouin frequency shift of 9.346 GHz. Fabrication is wafer-scale and CMOS-compatible.

Key Findings:

InGaP-on-SiO2 supports backward SBS near 1550 nm with gain coefficient of 588 W^-1 m^-1
Record narrow Brillouin linewidth of 5.2 MHz at 9.346 GHz frequency shift on an integrated platform
Process is wafer-scale and CMOS-compatible, enabling integration with existing PIC ecosystems

Technical Novelty: First demonstration of SBS in a single-crystalline InGaP-on-SiO2 waveguide platform. Prior chip-scale SBS work relied on chalcogenides (high gain but non-CMOS), silicon (requires suspension/pedestals), or Si3N4 (low gain, long devices). InGaP combines strong photoelastic response, a direct bandgap avoiding two-photon absorption at 1550 nm, and wafer-scale CMOS-compatible processing.

What's New: First SBS demonstration in single-crystalline InGaP-on-insulator; uniquely combines high gain, narrow linewidth, and CMOS-compatible wafer-scale fabrication — a trio prior platforms (chalcogenide, silicon, Si3N4) could not simultaneously offer.

Extension Opportunities:

Integrate the InGaP-on-insulator SBS waveguide with on-chip lasers/modulators to build a fully monolithic narrow-linewidth Brillouin laser or microwave photonic filter
Engineer waveguide geometry (suspended structures, phononic confinement, tapering) to further boost gain and reduce acoustic leakage into the SiO2 substrate
Explore forward intermodal SBS and cascaded Stokes generation for RF signal processing, delay lines, and non-reciprocal isolators on the same platform

Replicability: No code/data mentioned in the abstract. Reproduction requires III-V epitaxy (InGaP on GaAs), wafer bonding to SiO2, e-beam or DUV lithography, and a pump-probe SBS characterization setup with narrow-linewidth 1550 nm lasers and RF spectrum analyzers — a well-equipped integrated photonics cleanroom and photonics lab.

Research Gaps:

No demonstration yet of a full Brillouin laser or microwave photonic filter built on this InGaP platform
Acoustic confinement mechanisms and phonon lifetime limits in InGaP-on-SiO2 are not fully characterized; substrate leakage vs intrinsic material loss contributions to linewidth remain open

🔥 GitHub Trending

1. synthetic-sciences/openscience

⭐ 3535 stars | TypeScript

The open-source AI workbench for scientific research

agent ai ai-agent ai-scientist bioinformatics bun

2. spinabot/brigade

⭐ 3517 stars | TypeScript

Brigade — Your personal intelligence, built enterprise-grade

agent-runtime ai ai-crew autonomous-agents brigade brigade-agent

3. fuxicodex/Fuxi

⭐ 3397 stars | Python

FuXi is a fast, self-contained AI coding agent that lives in your terminal — edit code, run commands, and drive tools, with cost-aware routing across LLM providers.

agent ai ai-agent ai-coding autonomous-agent cli

4. Tencent-Hunyuan/AuK

⭐ 128 stars | Python

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

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

5. aminekago-web/Paradigm-Survival-Arena

⭐ 118 stars | HTML

Top 6 AI Paradigms Fighting for Survival in 2026

agario ai-simulation artificial-intelligence canvas educational evolutionary-algorithms

6. jaimasih05-commits/swarm-foraging-qlearn

⭐ 118 stars | HTML

Q-Learning Swarm Foraging 2026: Multi-Agent RL in Dynamic Grid Environments

ai-simulation artificial-intelligence autonomous-agents decision-making dynamic-environments exploration-exploitation

Generated by Research Pulse on 2026-09-09 06:05