2026-09-10·Daily digest

🔬 Research Pulse

Daily Digest

September 10, 2026

🤖 AI

🧠 LLMs

1. ConvMem: Convolutional Memory for Long-Context Reasoning

Authors: Hongming Zhang, Zhaozhen Gu, Fengshuo Bai... Published: 2026-09-09 | Citations: 0 arXiv | PDF

Research Question: How can LLMs perform long-context reasoning beyond fixed context limits without incurring the high latency, sequential bottlenecks, and RL training overfitting risks of iterative memory-update approaches like MemAgent?

Summary: ConvMem is a training-free framework that treats an LLM-with-query as a convolutional kernel and applies it hierarchically over text segments, turning long-context reasoning into a logarithmic-depth tree instead of a linear chain. It adds Configurable Strides, Skip Connections, and Multi-Kernel Convolution to preserve evidence and decompose queries, and beats training-free baselines on RULER-HotpotQA and RULER-2WikiMultiHopQA while avoiding the OOD overfitting seen in RL-trained memory agents.

Key Results: ConvMem outperforms training-free baselines on RULER-HotpotQA and RULER-2WikiMultiHopQA benchmarks, and avoids the parametric-prior overfitting failure mode observed in RL-trained sequential memory models on out-of-distribution tasks. The paper demonstrates that hierarchical convolution reduces the reasoning path from linear O(N) chains to logarithmic O(log N) trees. Specific accuracy numbers are not stated in the abstract.

Key Findings:

Hierarchical convolution shortens the reasoning path from linear to logarithmic depth, mitigating error accumulation across long contexts
The framework is massively parallelizable across both text segments (siblings) and reasoning threads (kernels), unlike sequential MemAgent-style memory
RL-trained sequential memory models tend to overfit to parametric priors and degrade on out-of-distribution tasks — a failure mode training-free ConvMem sidesteps

Technical Novelty: Reframing long-context reasoning as hierarchical CNN-style convolution over text — the LLM+query acts as a shared convolutional kernel applied to segments, with Configurable Strides (overlap control), Skip Connections (evidence propagation across levels), and Multi-Kernel Convolution (decomposing a query into disentangled semantic channels). This replaces the sequential recurrent-memory paradigm (MemAgent) with a training-free, parallel, logarithmic-depth tree.

What's New: Prior long-context memory methods (MemAgent and similar) are sequential, recurrent, and require RL fine-tuning. ConvMem borrows the CNN inductive bias — a shared kernel applied hierarchically with strides and skip connections — and maps it onto LLM prompting, yielding a training-free, parallel alternative with logarithmic reasoning depth.

Extension Opportunities:

Apply ConvMem's convolutional kernel abstraction to multimodal long-context tasks (video, long documents with figures) where hierarchical summarization of heterogeneous segments is natural
Learn or auto-tune the Configurable Stride and Multi-Kernel decomposition per-query using a lightweight router LLM, rather than treating them as fixed hyperparameters
Combine ConvMem's parallel hierarchical tree with KV-cache reuse across sibling segments to further cut inference cost, since sibling summarizations share the same kernel prompt

Replicability: The abstract states the method is training-free, which removes the largest cost barrier (no RL infrastructure needed). Reproduction requires only inference-time compute on a capable LLM plus the public RULER benchmarks. No code/data availability is mentioned in the abstract; would need to check the arXiv listing.

Research Gaps:

The abstract does not quantify latency/throughput gains vs MemAgent, nor report absolute accuracy numbers, making the magnitude of the improvement hard to assess
Evaluation is limited to RULER multi-hop QA variants; performance on tasks requiring true global aggregation (long-doc summarization, code repo QA) is untested

⚡ CUDA/Systems

1. Avatar: Toward Autonomous End-to-End Orchestration of Scientific Workflows using LLMs

Authors: Suman Raj, Hai Duc Nguyen, Haochen Pan... Published: 2026-09-09 | Citations: 0 arXiv | PDF

Research Question: Where should agentic LLM reasoning be introduced into scientific workflow management systems (WMSs), how can its risk be bounded, and does it actually improve orchestration over fixed hand-tuned rules?

Summary: Avatar is an actor-based scientific workflow orchestration architecture that separates decision policy (rule-based or LLM-backed) from execution mechanism through a validated action catalog. It demonstrates that agentic reasoning can be safely bounded and, when applied, meaningfully improves resource efficiency — cutting compute wastage 55% and GPU-busy time 40% — while a single core reproduces native execution across three workflows.

Key Results: Avatar's rule mode reproduces native execution across three workloads using a single unchanged core. LLM-backed Avatar reduces compute wastage by 55% and cuts GPU-busy time by 40% compared to rule-based orchestration.

Key Findings:

A single pluggable core reproduces native rule-based execution across three distinct workloads without modification
LLM-backed orchestration reduces compute wastage by 55% versus rule-based baseline
LLM-backed orchestration cuts GPU-busy time by 40%, showing agentic reasoning yields tangible efficiency gains, not just flexibility

Technical Novelty: An actor-based architecture (orchestrator, executor, provenance monitor) where each actor's decision policy is pluggable — rule-based or LLM-backed — through a single adapter-validated action catalog, letting conventional and agentic control share one core across different WMSs. Prior WMSs bake orchestration policy into the engine; Avatar decouples policy from mechanism.

What's New: First WMS design that treats agentic reasoning as a swappable policy over a shared, validated action catalog, allowing head-to-head comparison of rule-based and LLM-based orchestration on the same substrate rather than as separate systems.

Extension Opportunities:

Plug in additional decision policies (e.g., RL-based or hybrid symbolic-LLM policies) via the adapter-validated action catalog to compare against pure LLM and rule modes
Extend the provenance monitor with formal safety guarantees or verification, using it to detect and roll back unsafe LLM-proposed actions in real time
Integrate Avatar with additional WMSs (Parsl, Pegasus, Nextflow) to test whether the single core generalizes beyond the three initial workloads, and benchmark on heterogeneous multi-cluster deployments

Replicability: Built on the Academy actor framework; the abstract does not explicitly mention a code release. Reproduction would require the Academy framework, an LLM inference endpoint (likely GPU-hosted), and cluster resources to run the three scientific workloads — moderate-to-high compute given GPU workload orchestration measurements.

Research Gaps:

Unclear which specific orchestration decision points benefit most from LLM reasoning versus rules — the paper motivates but the abstract does not resolve a taxonomy
Bounding risk of agentic decisions in production workflows (e.g., cost blowups, incorrect scheduling) remains an open concern the adapter-validated catalog only partially addresses

👁️ Vision

1. Advanced Brain Tissue Imaging with Data-Consistent Diffusion Priors in Laminographic X-Ray Nanoimaging

Authors: Wenxuan Fang, Abraham L. Levitan, Ana Diaz... Published: 2026-09-09 | Citations: 0 arXiv | PDF

Research Question: How to reconstruct high-fidelity 3D nanoscale brain volumes from X-ray laminography data when the tilted acquisition geometry creates a missing-cone in Fourier space that conventional methods cannot recover, and direct 3D deep learning is blocked by data scarcity and compute cost?

Summary: LUCID is a reconstruction framework for X-ray laminography that couples multi-view diffusion priors with strict projection-domain data consistency to fill the missing-cone artifact inherent to tilted-geometry nanoimaging. Trained only on fully sampled tomographic volumes, it generalizes to experimental laminography data and recovers unmeasured Fourier components — improving spatial fidelity for downstream brain connectomics.

Key Results: On simulated datasets, LUCID substantially improves spatial fidelity and restores missing Fourier components, outperforming baseline methods (specific numeric benchmarks not stated in abstract). On experimental laminography data, LUCID — trained exclusively on fully sampled tomographic volumes — generalizes robustly and recovers unmeasured Fourier information, demonstrating cross-modality transfer from tomography training to laminography inference.

Key Findings:

Multi-view 2D diffusion priors can effectively impose 3D structural constraints without requiring scarce 3D training data
Enforcing projection-domain data consistency against the laminography forward model prevents the diffusion prior from hallucinating and preserves measured signal
A model trained on fully sampled tomography generalizes to a different acquisition geometry (laminography) at inference — a useful cross-modality transfer result

Technical Novelty: LUCID combines multi-view 2D diffusion priors with projection-domain data consistency enforcement against the laminography forward model, sidestepping 3D deep learning's data-scarcity problem while still recovering true 3D structure. Novel to prior work: training exclusively on fully sampled tomographic volumes yet generalizing to a different (laminography) acquisition geometry at inference time.

What's New: Prior missing-cone approaches either use handcrafted regularizers (blur artifacts) or direct 3D deep networks (data/compute-limited). LUCID uniquely fuses generative diffusion priors with a physics-consistent forward model, and demonstrates train-on-tomography → apply-to-laminography transfer.

Extension Opportunities:

Apply the multi-view diffusion + data-consistency framework to other missing-cone modalities (cryo-ET, limited-angle CT, electron tomography) where similar Fourier gaps distort reconstructions
Extend LUCID to joint reconstruction across multiple laminography tilt series or multi-modal inputs (e.g., X-ray + electron microscopy) to further constrain 3D structure for connectomics pipelines
Fine-tune the diffusion prior on domain-specific brain tissue distributions (e.g., specific species, cortical regions, or pathological samples) and benchmark segmentation/tracing accuracy downstream in connectomics workflows

Replicability: Abstract does not mention code, data release, or a training-set size. Reproduction would require: a diffusion model training pipeline (multi-GPU, likely tens of GPU-days for high-resolution 3D volumes), access to fully sampled tomographic brain datasets for training, and synchrotron/lab-source laminography data for evaluation. Compute is moderate-to-high; data access is the harder barrier.

Research Gaps:

Abstract omits quantitative metrics (PSNR/SSIM/FSC), runtime, and comparisons to specific baselines — hard to judge magnitude of improvement
Downstream biological validation (does improved reconstruction actually improve neuron tracing or synapse detection in connectomics pipelines?) is not addressed

🦾 ROBOTICS

1. Learning Terrain-Adaptive Humanoid Locomotion on Granular Terrain

Authors: Junnosuke Kamohara, Feiyang Wu, Andy Ningan Zong... Published: 2026-09-09 | Citations: 0 arXiv | PDF

Research Question: How can humanoid robots achieve reliable locomotion on real-world granular terrain (sand, basalt) where rigid contact models fail and existing simplified granular models rely on ad-hoc heuristics?

Summary: The paper introduces a 3D Resistive Force Theory-based granular contact solver embedded in RL training, combined with a VAE-encoded terrain latent trained via teacher-student distillation, enabling humanoids to walk on real sand and basalt. It claims the first demonstration of agile humanoid locomotion on real-world granular terrain, validated in MPM simulation and on hardware.

Key Results: Demonstrated first agile humanoid locomotion on real-world granular terrains including basalt, dry sand, and beach sand. In MPM (NVIDIA Newton) simulation, the method achieved significantly higher success rate than baselines on unseen granular terrains and showed zero-shot terrain identification/adaptation. Specific numerical benchmarks are not quantified in the abstract.

Key Findings:

3D RFT contact modeling produces sim-to-real transfer to granular terrain where rigid contact models fail
VAE-based latent terrain encoding enables zero-shot identification and adaptation to unseen granular surfaces
Hardware validation succeeds across basalt, dry sand, and beach sand — a first for humanoid platforms

Technical Novelty: First integration of physics-grounded 3D Resistive Force Theory as a differentiable/efficient granular contact solver directly inside an RL simulator, replacing rigid or heuristic tangential contact models, coupled with a VAE-based terrain latent for teacher-student adaptive control.

What's New: Prior work either used rigid contact (ignoring granular physics) or normal-force models with heuristic tangential drag; this work uses full 3D RFT for physically grounded intrusion + drag inside RL, and is the first humanoid demonstration on real granular media.

Extension Opportunities:

Extend the 3D RFT contact solver to heterogeneous or layered terrains (e.g., wet sand, snow, mud) that violate homogeneous granular assumptions
Combine the VAE terrain latent with exteroceptive vision (depth/LiDAR) for anticipatory gait adaptation before foot contact
Apply the teacher-student RFT-in-the-loop RL pipeline to quadrupeds, wheeled-legged hybrids, or manipulation tasks (digging, scooping) on granular media

Replicability: Project page provided (humanoid-gm-locomotion.github.io/HUMANOID-GM/); code availability not explicitly stated in the abstract. Reproduction would require a humanoid platform, GPU cluster for RL, NVIDIA Newton/MPM for validation, and a real granular testbed.

Research Gaps:

No quantitative benchmark numbers (success rates, velocity, energy) surfaced in the abstract
Scope limited to dry homogeneous granular media; wet/cohesive/layered terrains and dynamic terrain changes remain open

2. SwingBot: Learning Whole-Body Brachiation for Humanoid Robots

Authors: Yujie Xiong, Peng Zhai, Taixian Hou... Published: 2026-09-09 | Citations: 0 arXiv | PDF

Research Question: How can high-DoF humanoid robots perform continuous brachiation (overhead-bar swinging) when the controller must discover long-horizon release-swing-capture sequences, coordinate whole-body momentum across alternating contacts, and operate without reliable measurements of segment-relative displacement or hook-contact state?

Summary: SwingBot is a learning framework that enables high-DoF humanoid robots to perform continuous brachiation using passive wrist hooks. It combines biomimetic keyframes to guide exploration through rare release-swing-capture transitions with recurrent privileged-state estimation for sensor-free deployment, and is validated on hardware across varying payloads, disturbances, and bar spacings.

Key Results: Hardware experiments demonstrated continuous bar traversal on a real humanoid with passive wrist hooks, showing robustness to payload variation, external disturbances, and varying bar spacings. The abstract does not report specific numerical benchmarks (success rates, traversal speeds, or payload magnitudes).

Key Findings:

Structuring RL around biomimetic keyframes makes long-horizon release-swing-capture sequences reachable during early exploration where naive exploration fails
Recurrent privileged-state estimation can encode segment-relative displacement and hook-contact state into latents usable at deployment without those direct measurements
The learned policy transfers to hardware and remains robust to payload changes, external disturbances, and different bar spacings

Technical Novelty: Two coupled innovations: (1) biomimetic keyframes that make the rare release-swing-capture transitions reachable during early RL exploration by anchoring the policy to primate-inspired posture waypoints, and (2) recurrent privileged-state estimation that distills segment-relative displacement and hook-contact information into compact latents deployable without those sensors. Prior brachiation work targeted low-DoF or purpose-built swinging robots; this addresses whole-body humanoid coordination.

What's New: First learning-based framework to demonstrate continuous whole-body humanoid brachiation on hardware, addressing the coordination challenge of high-DoF systems rather than purpose-built two-link swingers, and solving the sensing gap around hook-contact state via recurrent privileged estimation.

Extension Opportunities:

Replace passive wrist hooks with active grippers to enable brachiation on irregular or non-parallel supports (branches, pipes, ledges)
Extend the keyframe-based curriculum to hybrid locomotion that transitions between bipedal walking and brachiation for cluttered environments
Integrate onboard vision to autonomously localize bars and plan multi-step swing trajectories rather than assuming known bar geometry

Replicability: The abstract does not mention code, model, or dataset release. Reproduction would require a torque-controlled humanoid with passive wrist hooks, a physics simulator (likely Isaac Gym/MuJoCo) for policy training, and GPU compute typical of humanoid RL work (single-node multi-GPU, days of training). Hardware access is the primary barrier.

Research Gaps:

No autonomous perception loop — bar locations appear to be assumed known rather than visually detected
Passive hooks constrain the system to parallel horizontal bars; grasping irregular or angled supports is unaddressed
Abstract lacks quantitative benchmarks (success rate, traversal speed, energy) making comparison to future work difficult

3. Show-Harness: Just a VLM Agent Can Play Robots

Authors: Yanzhe Chen, Zechen Bai, Zhijun Cao... Published: 2026-09-09 | Citations: 0 arXiv | PDF

Research Question: How can foundation VLMs' broad world knowledge be translated into effective robot control without costly embodiment-specific pretraining or specialized teleoperation hardware?

Summary: Show-Harness introduces an 'embodied harness' — a semantic action interface plus embodiment-specific interpreters — that lets both frontier and small open-source VLMs control robots without embodiment-specific pretraining. A companion GUI tool (GUMI) allows humans to demonstrate tasks across embodiments without teleoperation hardware, and experiments show the approach outperforms agentic and VLA baselines on generalization.

Key Results: Show-Harness enables (1) zero-shot robot control using closed-source frontier VLMs via a semantic action interface, and (2) adaptation of small open-source VLMs with just a few GPU-hours of fine-tuning. Experiments show the harness-equipped VLM agents generalize across tasks, embodiments, and environments, outperforming representative agentic and VLA (vision-language-action) paradigms. Specific benchmark numbers are not stated in the abstract.

Key Findings:

Zero-shot robot control from closed-source frontier VLMs is feasible with the right semantic interface
Small open-source VLMs reach competitive performance after only a few GPU-hours of fine-tuning
GUI-based demonstration collection via GUMI removes the need for specialized teleoperation hardware and generalizes across embodiments

Technical Novelty: A compact discrete semantic action interface that decouples high-level intent (handled by the VLM) from low-level embodiment grounding (handled by deterministic interpreters), while keeping the VLM responsible for fine-grained physical decisions — contrasted with monolithic VLA models that fuse perception, reasoning, and control into one trained policy.

What's New: Rather than training end-to-end VLA models, Show-Harness treats the VLM as the primary decision-maker and shifts the burden of generalization onto a well-designed discrete semantic interface — inverting the current trend of scaling embodiment-specific pretraining.

Extension Opportunities:

Extend GUMI's GUI-based demonstration interface to crowdsource large-scale multi-embodiment datasets from non-expert users, then distill into smaller VLMs
Add bimanual/dexterous manipulation semantic units and test whether the same interface generalizes to high-DOF hands without retraining the VLM
Integrate closed-loop error recovery by feeding execution-time interpreter feedback back into the VLM's semantic reasoning trace

Replicability: Abstract does not confirm code/data release. Reproducing the fine-tuned open-source variant requires only a few GPU-hours per the authors; the zero-shot path needs API access to a frontier VLM. GUMI is described as hardware-free, easing data collection reproduction.

Research Gaps:

Abstract lacks concrete benchmark numbers, success rates, or comparison tables against named baselines
Unclear how the semantic action vocabulary scales to contact-rich, dexterous, or long-horizon tasks beyond pick-and-place-style manipulation

💻 COMPUTE

1. Python in the front, party in the Backline: compiling quantum workloads across CPUs, GPUs, and FPGAs

Authors: Joseph K. L. Lee, Mehrdad Malekmohammadi, Hong-Sheng Zheng... Published: 2026-09-08 | Citations: 0 arXiv | PDF

Research Question: How can quantum platform builders bridge the gap between high-level Python-based quantum algorithm design and the low-latency, heterogeneous execution required for fault-tolerant quantum error correction across CPUs, GPUs, and FPGAs?

Summary: Backline is a heterogeneous compilation and runtime framework built inside PennyLane/Catalyst that compiles Python-authored quantum-classical workloads through MLIR to CPUs, GPUs, and FPGAs. It demonstrates microsecond-scale synchronous co-processing (2.3 μs CPU, 4.5 μs GPU round-trip from an FPGA controller over RoCE v2), addressing a core infrastructure gap on the path to fault-tolerant quantum computing.

Key Results: Demonstrated a working heterogeneous compilation/runtime (Backline) that compiles Python quantum workloads to CPUs, GPUs, and FPGAs via MLIR. Using an AMD VPK120 FPGA as controller over RoCE v2, measured median steady-state round-trip latencies of 2.305 μs to an AMD Ryzen Threadripper PRO CPU and 4.5 μs to an AMD Instinct MI210 GPU across 10^6−1 rounds per path, proving microsecond-scale synchronous co-processing between quantum control hardware and heterogeneous classical accelerators.

Key Findings:

MLIR provides a viable IR for lowering high-level Python quantum programs to FPGA bitstreams alongside CPU/GPU targets within one toolchain
Median 2.305 μs FPGA↔CPU and 4.5 μs FPGA↔GPU round-trip latencies over RoCE v2 are achievable, sustained across ~10^6 rounds — inside the budget for real-time QEC cycles
A vendor-agnostic Python frontend can drive distributed, hardware-handshake-synchronized quantum-classical execution without dropping to per-device toolchains

Technical Novelty: Unified MLIR-based compilation pipeline within PennyLane/Catalyst that targets FPGAs alongside CPUs/GPUs from a single Python frontend, with hardware-handshake-engine-driven round dispatch over RoCE v2 — prior quantum-classical stacks typically require separate toolchains per accelerator class and cannot achieve microsecond-scale synchronous dispatch from Python-authored code.

What's New: First demonstration (to the authors' knowledge) of a Python-frontend, MLIR-based quantum compilation stack that spans CPUs, GPUs, and FPGAs with hardware-handshake dispatch and microsecond round-trips — most prior quantum SDKs either stop at simulation or require hand-written HDL for the FPGA control layer.

Extension Opportunities:

Add support for additional FPGA/ASIC vendors (Xilinx UltraScale+, Intel Agilex, custom QEC ASICs) to broaden Backline's vendor-agnostic reach beyond AMD hardware
Integrate real QEC decoders (e.g., Union-Find, neural-network decoders) as MLIR-compilable kernels to benchmark end-to-end syndrome-extraction-to-correction latency under realistic surface-code cycles
Extend the runtime to support multi-node distributed FPGA meshes with deterministic clock synchronization for scaling to logical-qubit-count workloads beyond single controller boards

Replicability: Abstract does not mention open-source code release. Reproduction requires specialized hardware: AMD VPK120 FPGA board, AMD Ryzen Threadripper PRO CPU, AMD Instinct MI210 GPU, and a RoCE v2-capable network fabric — significant capital cost and integration effort, likely only reproducible in well-equipped quantum-hardware labs.

Research Gaps:

No end-to-end QEC decoder benchmark shown — latency measurements are for data-movement paths, not full syndrome-decode-and-correct loops
Vendor coverage is currently AMD-centric; portability to other FPGA/GPU ecosystems and to true distributed multi-controller topologies is not demonstrated

2. HBFSim: Fast and Faithful Simulation of High-Bandwidth Flash Under Real GPU Execution

Authors: Yanpeng Hu, Yiwei Yang, Yuanwu Zhu... Published: 2026-09-09 | Citations: 0 arXiv | PDF

Research Question: How can hardware architects evaluate capacity and data-placement decisions for High-Bandwidth Flash (HBF)—NAND stacked inside the accelerator package—before silicon samples in early 2027, given that trace-replay storage simulators, GPU simulators, and cycle-accurate simulators each fail to combine real LLM compute with realistic flash timing, capacity, and thermal effects?

Summary: HBFSim is the first evaluation platform that applies real High-Bandwidth Flash timing, capacity, and thermal behavior to an LLM inference workload while that workload actually runs on a real GPU. It rewrites PTX and gates kernel launches to separate access issue from consumption, using measured-device timings and junction-temperature-driven refresh, and demonstrates faithful, fast simulation on vLLM serving Qwen3-30B-A3B ahead of HBF silicon.

Key Results: HBFSim matches a measured reference device exactly at all six calibration breakpoints with zero unsafe kernel launches. An unmodified vLLM 0.15.1 serving Qwen3-30B-A3B returns token identifiers bit-identical to the uninstrumented baseline. The device fast path completes the Qwen3-30B-A3B case in 2s versus 44s on the detailed reference path—a 20.8x speedup—while still applying HBF timing, junction-temperature-driven sustained rate, and retention-driven refresh writes.

Key Findings:

Exact match to the measured reference device at all six calibration breakpoints with zero unsafe launches
Unmodified vLLM 0.15.1 running Qwen3-30B-A3B produces token IDs identical to the uninstrumented baseline, confirming functional transparency
Fast path delivers 20.8x speedup (2s vs 44s) over the detailed reference path on Qwen3-30B-A3B while retaining HBF timing/thermal fidelity

Technical Novelty: Prior tools force a choice: replay traces without executing the workload, simulate GPUs without the real kernels, or run cycle-accurate models that never finish one LLM inference. HBFSim's novelty is decoupling issue from consumption via PTX rewriting and kernel-launch gating, so a real GPU supplies the compute that hides HBF accesses while measured-device timing, thermally-modulated sustained bandwidth, and retention-deadline refresh writes are overlaid faithfully—yielding the first end-to-end HBF evaluation on real inference stacks.

What's New: It is the first simulator to combine real GPU compute execution with faithful HBF timing, capacity, and thermal/retention modeling—bridging the gap between trace-replay storage simulators, functional GPU simulators, and infeasibly slow cycle-accurate models.

Extension Opportunities:

Extend the PTX-rewriting/launch-gating layer to AMD ROCm (HIP/LLVM IR) or Intel GPUs so HBF placement can be studied across non-NVIDIA accelerators
Add a placement/policy search harness on top of HBFSim (e.g., RL or cost-model-guided) that co-optimizes KV-cache tiering across HBM/HBF/host for long-context serving workloads
Integrate a multi-package thermal model and cross-request scheduler to study how HBF junction temperature and refresh-write amplification interact under bursty multi-tenant vLLM traffic

Replicability: The abstract does not mention a public code or artifact release. Reproduction would require an NVIDIA GPU capable of running vLLM 0.15.1 with Qwen3-30B-A3B (roughly an H100/H200-class or MI300-class-equivalent node with sufficient HBM), plus access to the measured HBF reference device (or its calibration traces) to seed timing and thermal parameters. Without the calibration measurements, only the PTX-rewriting/gating scaffolding could be independently rebuilt.

Research Gaps:

No public artifact/code availability disclosed, and calibration depends on access to a specific measured HBF device
Evaluation is demonstrated on a single model (Qwen3-30B-A3B) and single serving stack (vLLM 0.15.1); generalization to other models, batch regimes, and non-NVIDIA GPUs is untested

3. Quantum Feature Engineering for Credit Default Prediction: When and Why IQP Circuits Help Linear Classifiers

Authors: Menachem Finkelstein, Diana Legziel Levy, Zohar Yakhini... Published: 2026-09-09 | Citations: 0 arXiv | PDF

Research Question: Can Instantaneous Quantum Polynomial-time (IQP) circuits produce features that improve linear classifiers on tabular credit default prediction beyond both raw classical baselines and the strongest unsupervised classical non-linear alternative (Kernel PCA) at equal feature budget?

Summary: The paper shows that features generated by shallow 8-qubit IQP quantum circuits, when appended to a Logistic Regression classifier on the UCI credit default dataset, produce a statistically significant F1 improvement (0.462 → 0.517) that beats Kernel PCA and other classical baselines. The gain is exclusive to linear classifiers and depends critically on which raw features are encoded, indicating the circuit amplifies informative structure rather than creating it.

Key Results: On the UCI Default of Credit Card Clients dataset with 5-fold CV, appending 16 IQP features (n=8 qubits) to Logistic Regression raised F1 from 0.462 to 0.517 (+0.055, p<0.0001), beating Kernel PCA's 0.493 at the same feature count. The gap survived Benjamini-Hochberg correction across 12 tests (p=0.00007). Random Forest importance-guided input selection further improved F1 to 0.523; maximally uncorrelated selection dropped it to 0.496. No gains were observed for Random Forest, SVM, XGBoost, or k-NN.

Key Findings:

IQP features boost Logistic Regression F1 by +0.055 (0.462 → 0.517) with high statistical significance (p<0.0001), outperforming Kernel PCA (0.493) at equal feature budget.
The benefit is exclusive to linear models — Random Forest, SVM, XGBoost, and k-NN see no improvement, suggesting a linear-expressivity mechanism.
Input feature selection matters substantially: Random Forest importance-guided selection reaches F1=0.523, while maximally uncorrelated selection drops to 0.496.

Technical Novelty: Uses fixed-depth IQP circuits purely as a feature-generation preprocessor for classical linear models rather than as a full quantum classifier or kernel method, and empirically isolates that gains arise specifically from linear-expressivity effects (only Logistic Regression benefits) combined with informed input selection.

What's New: Unlike prior quantum ML work that treats IQP circuits as end-to-end classifiers or kernel functions, this paper isolates them as a non-linear feature-engineering layer for classical linear models and provides a controlled comparison against Kernel PCA at matched feature budget, plus a mechanistic decomposition showing input-selection sensitivity.

Extension Opportunities:

Test IQP feature augmentation on other tabular financial datasets (fraud detection, loan approval, insurance claims) to check whether the linear-expressivity boost generalizes beyond credit default.
Systematically study the relationship between input feature selection strategies (mutual information, SHAP, causal importance) and IQP amplification quality to build a principled selection method.
Benchmark IQP features against other quantum-inspired classical feature maps (random Fourier features, polynomial kernels of matching degree) to isolate whether the advantage is quantum-specific or reproducible classically at small n.

Replicability: Abstract does not mention a code release. Reproduction requires the public UCI Default of Credit Card Clients dataset and simulation of 8-qubit IQP circuits, which is feasible on a laptop (2^8 Hilbert space); no real quantum hardware required. Compute needs are modest — classical simulation plus standard scikit-learn.

Research Gaps:

No theoretical explanation of why IQP-generated features are specifically well-matched to linear decision boundaries in this domain.
Generalization beyond a single dataset and beyond 8 qubits is unproven, leaving open whether the effect scales or is dataset-specific.

⚡ ENERGY

1. Overcoming Transport Layer Bottlenecks to Quantify Ionic Parameters from Transient Ion Current Measurements of Perovskite Solar Cells

Authors: Shudi Jiao, Miguel Torre Cachafeiro, Huagui Lai... Published: 2026-09-09 | Citations: 0 arXiv | PDF

Research Question: Why do transient ion current (TIC/BACE) measurements in perovskite solar cells often report ion densities orders of magnitude above the physical screening limit predicted by drift-diffusion simulations, and how can transport layer (TL) artifacts be disentangled from true ionic properties of the perovskite absorber?

Summary: The paper demonstrates that transport-layer capacitance — not perovskite ionic properties — sets the ceiling on ion densities extracted from transient ion current (BACE) measurements in perovskite solar cells, explaining widely reported anomalously high values. By systematically varying TL thickness (C60) and doping (Spiro-OMeTAD) and extrapolating to the TL-free limit, combined with an ionic-displacement correction, the authors recover the true ionic conductivity, density, and mobility of the absorber.

Key Results: The authors identify TL capacitance as the fundamental bottleneck capping the maximum quantifiable ion density to the interfacial charge that can accumulate at perovskite/TL interfaces. They confirm this experimentally across two architectures: systematically varying C60 electron-TL thickness in p-i-n cells and Spiro-OMeTAD hole-TL doping in n-i-p cells, showing TIC depends more strongly on TL properties than on perovskite ionic properties. Extrapolating the TL-thickness trend toward a TL-free limit yields the actual ionic conductivity (plus density and mobility depending on ionic model assumptions). Drift-diffusion simulations further show ion penetration into TLs or forward-bias preconditioning raises the capacitive ceiling, while trapped-carrier release can inflate TIC signals.

Key Findings:

TL capacitance imposes a hard upper bound on ion densities quantifiable by TIC/BACE, causing measurements to reflect TL properties more than perovskite ionics
Systematic TL-thickness (p-i-n C60) and doping (n-i-p Spiro-OMeTAD) sweeps followed by extrapolation to the TL-free limit yield the actual ionic conductivity of the perovskite absorber
Ion penetration into TLs, forward-bias preconditioning, and slow trapped-carrier release can each raise the capacitive ceiling or inflate the TIC signal, and must be accounted for in interpretation

Technical Novelty: Prior work treated anomalously high TIC-derived ion densities as unexplained artifacts. This paper provides a unified physical explanation (TL interfacial capacitance ceiling), an experimental protocol (systematic TL-thickness/doping variation with extrapolation to TL-free limit), and a correction based on average ionic displacement — turning BACE from a qualitative into a quantitative technique for ionic conductivity/density/mobility.

What's New: First systematic identification of TL interfacial capacitance as the fundamental physical origin of BACE overestimation, paired with an experimentally validated extrapolation protocol that recovers true ionic transport parameters across both p-i-n and n-i-p architectures.

Extension Opportunities:

Build an open-source correction/extrapolation toolkit that ingests TIC datasets across varied TL thicknesses/doping and automatically outputs corrected ionic conductivity, density, and mobility values
Design and validate ultra-thin or TL-free test structures (e.g., lateral device geometries or semi-transparent electrodes) that directly access the true ionic conductivity without extrapolation
Extend the framework to mixed-ion perovskites and couple TIC analysis with impedance spectroscopy or photothermal deflection to independently deconvolve ion species contributions and trap-release currents

Replicability: The abstract does not mention released code or data. Reproduction requires standard PSC fabrication capability (p-i-n with variable C60 thickness and n-i-p with doped Spiro-OMeTAD), a source-measure unit or potentiostat capable of voltage-step TIC/BACE measurements, and a 1D drift-diffusion simulator (e.g., IonMonger, SETFOS, Driftfusion) — all runnable on a standard workstation.

Research Gaps:

Lack of a device geometry that directly measures perovskite ionic conductivity without requiring TL-thickness extrapolation
Incomplete deconvolution of trapped-carrier release currents from genuine ionic displacement currents in TIC signals

2. Surrogate-accelerated parameterisation of physics-based Li-ion battery models

Authors: A. Emir Gumrukcuoglu, Josh Pearson, Jamie M. Foster... Published: 2026-09-09 | Citations: 0 arXiv | PDF

Research Question: How can we efficiently infer physically meaningful, non-redundant parameters of physics-based Li-ion battery models from terminal current-voltage data, given that direct inference is computationally expensive and limited by identifiability?

Summary: The authors present a surrogate-accelerated inverse framework for parameterising a single-particle Li-ion battery model with electrolyte dynamics (SPMe), using their differentiable Artiphy surrogate as the forward map. After rescaling to remove structural redundancies, they infer transport, kinetic, and capacity parameters from synthetic DFN voltage data under a WLTP-like protocol, achieving ~1 mV voltage error and characterising identifiability via Fisher-information analysis.

Key Results: Using an Artiphy surrogate for a single-particle model with electrolyte dynamics (SPMe), the framework inferred non-redundant transport, kinetic, and capacity parameter groups from synthetic DFN-generated voltage data under a WLTP-like drive cycle. The inferred SPMe reproduced benchmark voltage to ~1 mV error, recovered electrode capacities accurately, and recovered positive-electrode diffusivity across much of the stoichiometric range. Fisher-information analysis quantified correlated kinetic-Ohmic and electrolyte-transport directions.

Key Findings:

Inferred SPMe reproduces DFN benchmark voltage to order 1 mV error under a WLTP-like current protocol
Electrode capacities and positive-electrode concentration-dependent diffusivity are recovered accurately across much of the stoichiometric range
Fisher-information analysis reveals correlated kinetic-Ohmic and electrolyte-transport directions; negative-electrode diffusion is weakly identified over most of the drive cycle but constrained by localised sensitivity plus global parameterisation

Technical Novelty: Combines a differentiable surrogate (Artiphy) for the SPMe forward map with an explicit rescaling that removes exact structural redundancies before inference, allowing concentration-dependent diffusivities in both solid and electrolyte phases to be inferred as non-redundant groups — prior work typically either used lumped parameters, non-differentiable simulators, or ignored identifiability structure.

What's New: Introduces a differentiable surrogate (Artiphy) enabling fast gradient-based inference for SPMe, paired with an explicit redundancy-removing rescaling that makes concentration-dependent solid and electrolyte diffusivities identifiable — bridging surrogate ML and rigorous identifiability analysis for physics-based battery models.

Extension Opportunities:

Apply the surrogate-accelerated inverse framework to real experimental cell data (not just synthetic DFN benchmarks) across aging states to track parameter drift over cycle life
Extend to multi-cycle or temperature-varying protocols to jointly infer thermal and electrochemical parameters, enabling in-silico BMS calibration
Couple the differentiable Artiphy surrogate with active experimental design — choose current protocols that maximize Fisher information along weakly identified directions (e.g., negative-electrode diffusivity)

Replicability: Abstract does not explicitly mention released code or data. The Artiphy surrogate framework is named as their own; reproducing would require access to Artiphy plus a DFN reference implementation (e.g., PyBaMM). Surrogate training likely dominates compute; inference itself should be cheap (single-GPU or CPU feasible) given differentiability.

Research Gaps:

Validation is on synthetic DFN data only; performance on real experimental cells with unmodelled effects (aging, temperature, SEI growth) remains unproven
Negative-electrode diffusivity remains weakly identified over much of the drive cycle, indicating a need for tailored excitation protocols

3. Surrogate-Assisted Inverse Design and Temperature-Dependent Electrothermal Analysis of an All-Oxide Narrowband Thermophotovoltaic Emitter

Authors: Bibekananda Nath, Kawshik Nath, Ahmed Zubair Published: 2026-09-09 | Citations: 0 arXiv | PDF

Research Question: How can we design a thermally stable, narrowband thermophotovoltaic (TPV) emitter that avoids the oxidation and structural degradation of traditional metal-dielectric emitters while achieving high peak emissivity aligned with a solar cell's bandgap?

Summary: The paper presents a surrogate-assisted inverse-design pipeline for an all-oxide (ITO/Al2O3 grating on sapphire) narrowband TPV emitter, using ExtraTrees regressors trained on FDTD-simulated spectra and Bayesian optimization on a penalty-augmented objective. The optimized geometry achieves peak emissivity above 0.90 with a narrow FWHM and high in-band emission fraction, and the design is validated via FDTD, offering a thermally robust alternative to metal-dielectric emitters.

Key Results: The authors demonstrated a surrogate-assisted inverse-design framework for a 1D ITO/Al2O3 grating emitter on sapphire that achieves E_peak > 0.90 (enforced as a Bayesian optimization constraint) while minimizing FWHM and maximizing in-band emission fraction. ExtraTrees surrogates were trained on FDTD-simulated spectra with layer thicknesses and grating period as inputs, and predictions of the optimized structure were validated back against FDTD.

Key Findings:

All-oxide ITO/Al2O3/sapphire stacks can achieve narrowband emission with E_peak > 0.90, avoiding the metal-oxidation failure mode of traditional emitters
ExtraTrees surrogates trained on modest FDTD datasets are accurate enough to drive Bayesian optimization toward FDTD-validated optima
A penalty-augmented objective combining peak-emission constraint, FWHM minimization, and in-band fraction maximization successfully shapes both height and selectivity of the emission peak

Technical Novelty: Combining ExtraTrees regression surrogates with Bayesian optimization on a penalty-augmented multi-objective (peak-emission constraint + FWHM + in-band fraction) applied to an all-oxide (ITO/Al2O3/sapphire) narrowband TPV emitter — prior TPV inverse-design work has largely used metal-dielectric stacks and either direct FDTD sweeps or neural-network surrogates.

What's New: First (per the abstract) surrogate-assisted inverse-design framework specifically targeting an all-oxide narrowband TPV emitter, coupling ExtraTrees regression with constraint-penalized Bayesian optimization over a multi-component narrowband FOM, and adding a temperature-dependent electrothermal analysis on top of the optical design.

Extension Opportunities:

Extend from 1D gratings to 2D/3D oxide metasurfaces (e.g., cross or pillar arrays) using the same surrogate + Bayesian optimization loop to explore polarization-independent narrowband emission
Couple the electromagnetic surrogate with a coupled electrothermal/CFD model to co-optimize emitter geometry against emitter operating temperature stability and radiative-to-electrical conversion efficiency of a specific PV cell (e.g., GaSb, InGaAsSb)
Replace ExtraTrees with a neural surrogate (e.g., Fourier neural operator or transformer over spectra) to enable spectrum-level prediction and gradient-based inverse design rather than scalar-FOM Bayesian optimization

Replicability: The abstract does not mention released code or datasets. Reproduction would require an FDTD solver (Lumerical, MEEP, or Tidy3D) to regenerate the training corpus by sweeping ITO/Al2O3/period dimensions, plus scikit-learn ExtraTrees and a Bayesian optimization library (e.g., scikit-optimize, BoTorch). Compute is modest — 1D grating FDTD sweeps run on a workstation CPU/GPU in hours to a day, and surrogate training is trivial.

Research Gaps:

No reported experimental fabrication or measurement — the design is only FDTD-validated, so real-world losses (surface roughness, ITO stoichiometry drift, Al2O3 crystallization at high T) remain unquantified
Limited to a 1D grating geometry and a single material stack; angular/polarization dependence and integration efficiency with a specific PV cell are not fully explored in the abstract

🏥 HEALTHCARE

1. OmniMed-FL: A Robust Multimodal Federated Learning Framework for Clinical Diagnosis

Authors: Ayush Debnath, Ruelia Saha, Sudip Misra Published: 2026-09-09 | Citations: 0 arXiv | PDF

Research Question: How can multimodal (imaging + text) clinical diagnosis be performed across distributed hospitals under HIPAA/GDPR constraints that prevent centralized data aggregation, and which federated learning strategies best handle non-IID label skew across sites?

Summary: OmniMed-FL is a systems-study benchmark of multimodal (chest X-ray + synthetic clinical note) federated learning for 5-class thoracic diagnosis, sweeping 8 fusion strategies, 3 initializations, and 4 imputation rules across 3–20 non-IID hospital clients. FedProx narrowly leads (0.737 macro-F1) over FedAvg (0.662), multimodal fusion consistently beats unimodal, but the authors' SCAFFOLD-AdamW adaptation collapses and communication cost scales linearly to 183.5 GiB at K=20.

Key Results: On a proxy corpus of 3,000 chest radiographs paired with 3,000 class-conditioned synthetic notes across 5 classes (Normal, Pneumonia, COVID-19, Pleural Effusion, Cardiomegaly), with K=5 clients and severe Dirichlet skew (α=0.1): local-only training reaches macro-F1 0.297, FedAvg 0.662±0.074, FedProx 0.737±0.085, FedMME-style one-shot ensemble 0.647±0.080, and a SCAFFOLD-AdamW variant collapses to 0.070±0.015. Multimodal fusion beats unimodal on both corpora (0.956 vs 0.934 text / 0.664 image on synthetic; 0.906 vs 0.880 / 0.737 on radiographs) at 2.3× the model state of text-only. Over a 4×3 grid, label skew costs up to 0.27 F1 while a ~7× client increase costs at most 0.10; bidirectional communication grows linearly to 183.5 GiB at K=20.

Key Findings:

FedProx (0.737) marginally outperforms FedAvg (0.662) at K=5, α=0.1, but the 0.075 gap falls within the wider two-seed standard deviation, making the ranking statistically weak.
Label skew is the dominant degradation axis (up to 0.27 F1 lost) — far larger than the cost of adding clients (≤0.10 F1 across a ~7× increase).
Multimodal fusion wins on both corpora (0.956 vs 0.934 text / 0.664 image on synthetic; 0.906 vs 0.880 / 0.737 on radiographs), but at 2.3× the parameter footprint of a text-only model.
Adaptive-optimizer control-variate methods (SCAFFOLD-AdamW) failed catastrophically (0.070 F1), a negative result worth investigating.

Technical Novelty: A controlled 8-fusion × 3-init × 4-imputation × multi-K sweep for multimodal FL on medical data, plus a SCAFFOLD adaptation combined with AdamW for multimodal clinical classification. The systems-study framing (label-skew vs client-count vs bandwidth cost surface) is more thorough than prior FedMME-style single-configuration reports.

What's New: Prior FL-on-medical-imaging work is largely unimodal or reports single algorithm/single client-count numbers. This paper contributes a controlled grid across fusion × init × imputation × client-count × skew, honestly caveated as a proxy study (synthetic notes, class-matched not patient-matched), and pairs accuracy metrics with concrete bandwidth costs — a framing closer to systems research than typical ML-benchmark papers.

Extension Opportunities:

Replace synthetic class-conditioned notes with real patient-level paired radiology reports (e.g., MIMIC-CXR) to test whether the fusion advantage holds when text and image share a true patient — the current pairing is class-matched, not patient-matched, which likely inflates fusion gains.
Diagnose and repair the SCAFFOLD-AdamW collapse (0.070 F1): investigate control-variate/adaptive-moment interaction, try SCAFFOLD with SGD, FedAdam, or MIME, and characterize when adaptive server optimizers help vs hurt under severe non-IID multimodal settings.
Attack the 183.5 GiB communication cost at K=20 with modality-aware compression (freeze the vision encoder after warmup, quantize/sparsify the text tower, LoRA-only updates for fusion heads) and measure the F1/GiB Pareto frontier.

Replicability: No code/data links are indicated in the abstract. Data is reproducible in principle: 3,000 public chest radiographs plus 3,000 class-conditioned synthetic notes generated by the authors. Compute is modest by federated-learning standards — 3 to 20 simulated clients, seed-level variance reported from only two seeds — likely reproducible on a single multi-GPU workstation.

Research Gaps:

Class-conditioned synthetic notes matched by class rather than patient — the strong multimodal gains may not survive real patient-level pairing where text and image can conflict.
Only two seeds reported; standard deviations already swallow the FedProx vs FedAvg gap, so many algorithmic rankings in the paper are underpowered.

2. CHIMERA Challenge Task 2 and 3: Response Subtypes Classification and Progression Survival Prediction in Bladder Cancer Patients using Multimodal Datasets

Authors: Catherine Chia, Tongjie Wang, Robert Spaans... Published: 2026-09-08 | Citations: 0 arXiv | PDF

Research Question: How can multimodal AI (histopathology, RNA-seq, structured clinical data) improve prediction of BCG response subtypes and progression in high-risk non-muscle-invasive bladder cancer (HR-NMIBC), where current clinical risk stratification is limited?

Summary: CHIMERA is a multimodal AI challenge for HR-NMIBC that benchmarks BCG response subtype classification and progression prediction across 368 patients using histopathology, RNA-seq, and structured clinical data. Best submissions reached F1=0.73 and C-index=0.68, and post-challenge analysis exposed modality-dependence, cohort-degradation, and consistently hard-to-predict patient subgroups (notably T1 substage).

Key Results: Benchmarked 13 top models from 159 submissions on a 368-patient multimodal dataset with hidden validation/test splits. Best models achieved weighted F1 = 0.73 for Task BRS (BCG Response Subtype classification) and C-index = 0.68 for Task Progression (time-to-progression). Post-hoc analyses showed histopathology partially compensates for missing pathology-derived structured variables in BRS, while progression models depend more on complementary modalities. T1 substage was associated with cross-model prediction difficulty.

Key Findings:

Best weighted F1 of 0.73 for BRS classification and C-index of 0.68 for progression across 13 selected models
Modality contributions are task-dependent: histopathology can compensate for missing structured pathology variables in BRS, but progression prediction requires complementary modalities
T1 substage patients are consistently difficult across model architectures, and models are sensitive to missing structured data — pointing to transportability and missingness-handling as core barriers

Technical Novelty: Rather than a new model, the novelty is a standardized multimodal HR-NMIBC benchmark (CHIMERA) that pairs histopathology, RNA-seq, and structured clinical data with hidden test splits, plus a systematic cross-model error analysis framework quantifying modality contribution, cohort-dependent degradation, and per-patient difficulty.

What's New: First standardized HR-NMIBC multimodal benchmark combining WSI + RNA-seq + structured clinical data with hidden test evaluation, and a framework analyzing not just accuracy but robustness, information sufficiency, and patient-level failure modes.

Extension Opportunities:

Build missingness-aware multimodal fusion architectures that explicitly model which modality is absent rather than imputing, targeting the sensitivity issue the paper flagged
Test transportability by fine-tuning the top models on an independent multi-institutional cohort and quantify performance degradation vs the hidden test set
Develop a difficulty-aware ensemble that routes T1-substage or otherwise cross-model-hard patients through a specialized branch (e.g., pathology-focused foundation model embeddings)

Replicability: Challenge implies public training data with hidden validation/test — reproducing top submissions likely requires access to WSI-scale pathology (multi-GPU with substantial VRAM for foundation-model encoders) plus RNA-seq processing. Code availability for individual submissions is not stated in the abstract.

Research Gaps:

No independent multi-institutional external validation — transportability of the top models remains unproven
Lack of principled missingness-aware modeling; current approaches degrade when structured inputs are incomplete

3. ADMET-EvO: a self-evolving scientific agent for sustained research across heterogeneous tasks

Authors: Yiling Zhou, Yilin Wang, Jianmin Wang... Published: 2026-09-09 | Citations: 0 arXiv | PDF

Research Question: How can a scientific agent sustain adaptation across heterogeneous ADMET prediction tasks without overfitting to internal validation, while accumulating evidence to revise its own hypotheses and strategies?

Summary: ADMET-EvO is a self-evolving scientific agent that formalizes ADMET endpoints, proposes falsifiable hypotheses, and tests data/feature/model interventions while carrying evidence forward across cycles. It tops the 22-task TDC ADMET benchmark at 96.77 task-normalized score, cuts fitting time by 72.2%, and extends itself to 43 new toxicity endpoints — demonstrating sustained, evidence-driven adaptation across heterogeneous scientific tasks.

Key Results: On the 22-task Therapeutics Data Commons (TDC) ADMET benchmark, ADMET-EvO achieved the highest task-normalized score of 96.77. Evidence-guided selection reduced cumulative fitting time by 72.2% within a predefined non-inferiority margin. The agent additionally formalized 43 toxicity-related tasks and built endpoint-specific predictors for them.

Key Findings:

Achieves state-of-the-art 96.77 task-normalized score on the 22-task TDC ADMET benchmark
Evidence-guided selection cuts cumulative fitting time by 72.2% while staying within a predefined non-inferiority margin
The agent autonomously expanded its scope by formalizing 43 additional toxicity-related tasks and training endpoint-specific predictors

Technical Novelty: The evidence-gated agentic loop that formalizes endpoints, generates falsifiable hypotheses, and tests interventions along three orthogonal axes (data, feature, model) — while explicitly carrying supported, rejected, and inconclusive outcomes forward to guide the next cycle. Prior AutoML/agent systems for ADMET typically optimize a single score per task without a persistent, cross-task evidence ledger that gates future search.

What's New: Unlike single-shot AutoML or per-task agent runs, ADMET-EvO maintains a cross-cycle evidence ledger (supported/rejected/inconclusive) that gates future search over data, feature, and model axes — turning ADMET modeling into an ongoing hypothesis-driven research loop rather than isolated optimization.

Extension Opportunities:

Port the evidence-gated cycle (supported/rejected/inconclusive carry-forward) to adjacent domains like protein-ligand binding, PK modeling, or materials property prediction where task heterogeneity is similarly high
Add active-learning or wet-lab-in-the-loop hooks so hypotheses rejected on public data can trigger targeted new assay requests rather than dying at the validation step
Build an open leaderboard/registry of the falsifiable hypotheses ADMET-EvO generates and their outcomes, turning the agent's internal ledger into a shared community artifact

Replicability: The abstract does not mention a code or model release. Reproduction would require the TDC ADMET benchmark (publicly available), an LLM-agent backbone plus standard cheminformatics stacks (RDKit, DeepChem/Chemprop-style models), and enough GPU capacity to run 22 task pipelines repeatedly — feasible on a single multi-GPU workstation given the reported 72.2% fitting-time reduction.

Research Gaps:

No discussion of external / prospective wet-lab validation — gains are shown on public benchmarks, which the agent could still implicitly overfit to via strategy reuse
Unclear how the evidence ledger degrades or transfers when moving to chemical domains far from the TDC distribution, or how conflicting evidence across tasks is arbitrated

🔬 MATERIALS

1. Constrained Optimization of Higher-Order Cluster-Expansion Hamiltonians for Alloys Using Simulated Bifurcation

Authors: Kazuhide Ichikawa, Satoru Ohuchi, Tomoyasu Yokoyama... Published: 2026-09-09 | Citations: 0 arXiv | PDF

Research Question: How to efficiently find ground-state and low-energy atomic configurations of binary alloys when cluster-expansion Hamiltonians extend beyond pairwise (quadratic) terms to cubic/quartic interactions, under fixed-composition constraints — avoiding the overhead of quadratizing higher-order polynomial Ising models with auxiliary variables.

Summary: The paper shows that Toshiba's SQBM+ simulated-bifurcation solver can directly optimize cubic and quartic cluster-expansion Hamiltonians for binary alloys as polynomial unconstrained binary optimization (PUBO) problems, bypassing costly quadratization. On Au-Cu systems up to 2048 atoms with composition-penalty constraints, it recovers CuAu and Cu3Au ordering trends and produces sensible formation-energy convex hulls.

Key Results: Using Au-Cu as a testbed, the authors formulated cubic and quartic CE models as PUBO problems with quadratic composition-penalty terms (weights derived from continuous-relaxation derivative coefficients) and solved them with SQBM+, a simulated-bifurcation solver. Benchmarks scaled up to 2048 atoms: SQBM+ robustly recovered low-energy feasible configurations for cubic CE and remained effective for many quartic instances. Formation-energy convex hulls reproduced the known CuAu and Cu3Au ordering trends and exposed finite-size effects at off-stoichiometric compositions.

Key Findings:

SQBM+ robustly finds low-energy feasible configurations for cubic CE models up to 2048 atoms and remains effective for many (though not all) quartic instances.
Composition constraints can be enforced via quadratic penalties whose weights are estimated from derivative coefficients of the continuous CE relaxation — a lightweight, principled tuning heuristic.
Optimized configurations reproduce the known CuAu and Cu3Au ordering, while finite-size effects distort the convex hull at off-stoichiometric compositions.

Technical Novelty: Direct PUBO treatment of higher-order (cubic/quartic) CE Hamiltonians via simulated bifurcation — no explicit quadratic reformulation with auxiliary spins — combined with a principled composition-penalty weight derived from the continuous relaxation's derivative coefficients.

What's New: Prior alloy CE optimization typically restricts to pairwise (Ising QUBO) Hamiltonians or quadratizes higher-order terms with auxiliary variables, inflating problem size. This work applies a simulated-bifurcation solver natively to cubic/quartic PUBO objectives with a composition-constraint penalty scheme, extending Ising-machine-based alloy search into the higher-order CE regime.

Extension Opportunities:

Apply the same PUBO + simulated-bifurcation pipeline to ternary or multi-component alloys, where CE Hamiltonians naturally become higher-order and quadratization is even more costly.
Benchmark SQBM+ against quantum annealers (D-Wave), parallel tempering, and genetic algorithms on identical CE PUBO instances to quantify where simulated bifurcation actually wins.
Automate penalty-weight tuning by learning the derivative-coefficient estimator from data, or replace fixed penalties with adaptive Lagrangian updates to reduce infeasibility on quartic instances.

Replicability: Abstract does not mention released code or data. SQBM+ is Toshiba's simulated-bifurcation solver (available via cloud/SDK). Reproducing 2048-atom runs would need access to SQBM+ (GPU-accelerated) plus published CE coefficients for Au-Cu; compute is modest relative to DFT but non-trivial without SQBM+ access.

Research Gaps:

Quartic CE instances still show reduced robustness — no principled explanation or guarantee about when SQBM+ succeeds vs fails on higher-order terms.
Finite-size effects at off-stoichiometric compositions distort the convex hull; scaling beyond 2048 atoms and reconciling with thermodynamic-limit predictions remains open.

2. Record-Breaking Elemental Superconductivity in Tetralayer Kagome Borophene

Authors: Yingnan Liu, Yan Liu, Renyu Duan... Published: 2026-09-09 | Citations: 0 arXiv | PDF

Research Question: Can 2D elemental crystals achieve superconductivity above liquid-nitrogen temperature (77 K), given that strong covalent bonding typically yields high phonon frequencies but insufficient electron-phonon coupling?

Summary: The paper predicts, via first-principles anisotropic Migdal-Eliashberg calculations, that tetralayer kagome borophene stabilized by ABAB covalent stacking is a record-breaking elemental superconductor with Tc = 102 K — above liquid-nitrogen temperature. The mechanism relies on a novel out-of-plane s-pz bonding channel coupled to low-frequency vertical boron vibrations, establishing covalent stacking engineering as a route to 2D high-Tc superconductivity.

Key Results: Using first-principles calculations and fully anisotropic Migdal-Eliashberg theory, the authors predict tetralayer kagome borophene (TKB) with ABAB covalent stacking exhibits a critical temperature Tc = 102 K — a record-high among elemental superconductors and above the liquid-nitrogen threshold (77 K). The pairing is mediated by interlayer s-pz bonding states at the Fermi level coupled to low-frequency out-of-plane boron vibrations.

Key Findings:

TKB with ABAB stacking is dynamically stable and predicts Tc = 102 K, the highest among reported elemental superconductors
Pairing is mediated by interlayer s-pz bonding states at the Fermi level coupled to low-frequency out-of-plane B vibrations — a mechanism distinct from MgB2-like in-plane sigma bonding
Covalent stacking engineering emerges as a viable design strategy for high-Tc superconductivity in 2D elemental materials

Technical Novelty: Identifies a previously unexploited out-of-plane s-pz bonding pairing channel in multilayer borophene — distinct from the in-plane sigma-bonding + high-frequency B-B stretching mechanism dominant in prior boron-based superconductors (e.g., MgB2). Introduces ABAB covalent stacking engineering as a design lever for 2D high-Tc superconductivity.

What's New: First prediction of an elemental 2D superconductor breaking the liquid-nitrogen barrier, and identification of an out-of-plane s-pz pairing channel that inverts the conventional in-plane sigma-bond paradigm dominant in prior boron superconductors.

Extension Opportunities:

Experimentally synthesize TKB via covalent stacking on suitable substrates (e.g., Ag(111) or Cu(111)) and verify the predicted 102 K Tc with transport/tunneling measurements
Extend the covalent-stacking design principle to other light elements (C, Be, Al) or generalized kagome n-layer configurations to search for even higher Tc materials
Investigate strain, doping, or pressure tuning of TKB to further enhance the out-of-plane s-pz coupling channel and explore topological superconductivity signatures

Replicability: Purely computational study — reproducible with DFT + DFPT + EPW/Wannier90 for anisotropic Migdal-Eliashberg calculations. No code/data availability explicitly mentioned in the abstract. Compute requirements: moderate-to-large HPC cluster (hundreds to thousands of CPU-hours) for phonon and electron-phonon coupling calculations on a 4-layer boron supercell.

Research Gaps:

No experimental synthesis or verification — remains a purely theoretical prediction awaiting lab confirmation
Substrate effects, defects, and realistic growth conditions that could destabilize ABAB stacking or renormalize Tc are not addressed in the abstract

3. Non-equilibrium dissipative stabilization of s- and d-wave superconductivity

Authors: Aleksey Lunkin, Yury Holubeu, Denis Golež... Published: 2026-09-09 | Citations: 0 arXiv | PDF

Research Question: Can superconductivity be stabilized above the equilibrium critical temperature T_c through non-equilibrium driving, and what mechanism enables such nonthermal pairing?

Summary: The paper shows that a BCS superconductor coupled to two thermal baths that break detailed balance can host a nonthermal pairing gap even when one bath is far above the equilibrium T_c. The non-equilibrium steady state is well described by a generalized Gibbs ensemble parametrized by chemical potentials for Bogoliubov quasiparticles, and the effect is non-perturbative for both s-wave and d-wave symmetries, with cavity implementations proposed.

Key Results: The authors demonstrate a nonthermal pairing gap in BCS superconductors coupled to two thermal baths where one bath sits well above equilibrium T_c. They show the resulting steady state is captured by a generalized Gibbs ensemble parametrized by chemical potentials conjugate to Bogoliubov quasiparticles, and report a strong, non-perturbative pairing effect for both s-wave and d-wave symmetries. No specific numerical benchmarks or temperature ratios are quoted in the abstract.

Key Findings:

Two-bath coupling that violates detailed balance stabilizes a genuine superconducting steady state above equilibrium T_c
The non-equilibrium steady state admits a generalized Gibbs ensemble description with Bogoliubov-quasiparticle chemical potentials
The pairing effect is strong and non-perturbative and applies to both s-wave and d-wave order parameters, with a proposed cavity-based realization

Technical Novelty: Prior non-equilibrium superconductivity work focused on Floquet driving or transient pump-probe states. This paper instead uses two thermal baths with broken detailed balance to reach a true steady state, and rigorously characterizes it via a generalized Gibbs ensemble with Bogoliubov quasiparticle chemical potentials — a many-body integrability-inspired description applied to a dissipative BCS problem.

What's New: Combines dissipative steady-state engineering with a GGE description of a non-integrable superconductor, providing a mechanism for above-T_c pairing that is neither Floquet nor transient but a genuine dissipative steady state.

Extension Opportunities:

Extend the two-bath dissipative framework to unconventional pairing channels beyond s/d-wave (e.g., p-wave, pair-density-wave, or topological superconductors) to test whether the GGE stabilization mechanism generalizes
Build a concrete cavity-QED simulation pipeline (Lindblad + BCS mean-field) that predicts observable signatures — gap magnitude vs. bath temperature imbalance, tunneling spectra — for a specific candidate material like FeSe or cuprate thin films
Couple this framework to strong-correlation solvers (DMFT, tensor networks) to move beyond BCS mean field and test whether non-equilibrium pairing survives interactions, or design a Floquet/cavity experiment protocol that engineers the detailed-balance-breaking coupling

Replicability: No code or data availability is mentioned in the abstract. Reproduction likely requires solving Lindblad/Keldysh equations for a BCS Hamiltonian coupled to two baths self-consistently for the gap; feasible on a workstation for mean-field s/d-wave lattice models with modest momentum-grid sizes.

Research Gaps:

No treatment of strong-coupling or fluctuation corrections beyond BCS mean field — unclear if the steady state survives in real correlated materials
Concrete experimental parameters (bath coupling strengths, cavity Q, achievable gap enhancement) and material candidates are not yet mapped out

🔥 GitHub Trending

1. trailhq/Graft

⭐ 6898 stars | TypeScript

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

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

2. shy3130/tick-stock-panel

⭐ 4534 stars | Python

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

a-stock ai-agent aigc backtesting daily duckdb

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

⭐ 863 stars | Python

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

cv deep-learning obs obs-studio

4. CheshireMew/VoxWeave

⭐ 235 stars | Python

Local-first high-quality offline RVC voice conversion workstation

audio-processing cuda ffmpeg local-first pyside6 qml

5. giannisanni/pulsar

⭐ 212 stars | Rust

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

cuda gguf glm inference-engine llm local-llm

6. Desert-Ant-Labs/desert-ant-core

⭐ 172 stars | Swift

On-device AI SDKs for iOS, macOS, Android, and the web. Small, focused models that run fully offline in Swift, Kotlin, and JavaScript with Core ML, LiteRT, and WebAssembly.

android browser coreml edge-ai inference ios

7. Blackwellboy/model-serving-minefield

⭐ 129 stars | Python

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

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

8. grandgaming9321-prog/reality-engine

⭐ 126 stars | HTML

Top Dynamic AI World Simulation & Storytelling Tools 2026

3d ai ai-simulation creative-ai digital-twin game-ai

9. evo-design/proto-tools

⭐ 125 stars | Python

A universal infrastructure layer for generative biology

bioengineering bioinformatics generative-biology machine-learning

10. lucidrains/x-jepa

⭐ 125 stars | Python

Explorations into some of the approaches advocated by Yann LeCun, and just a more wholistic architecture (JEPA) in general

artificial-intelligence deep-learning jepa world-models

11. reunios2024/cortex-sentinel-trading-nexus

⭐ 121 stars | HTML

Self-Tuning Multi-Agent AI Trading System 2026: 8-Source Signal Fusion & Kronos Model

agentic ai-trading algorithmic-trading claude claude-opus fintech

12. iamwavecut/MLX-DLSS

⭐ 60 stars | Swift

NVIDIA DLSS 5 neural rendering and DLSS frame generation on Apple Silicon (MLX/Metal, Core ML) and PyTorch; weights extracted from your own DLSS libraries

apple-silicon coreml dlss frame-generation metal mlx

Generated by Research Pulse on 2026-09-10 06:05