2026-09-12·Daily digest

🔬 Research Pulse

Daily Digest

September 12, 2026

🤖 AI

🧠 LLMs

1. Biology-in-the-loop: Amortized Adaptive Hit Discovery in CRISPR Screens

Authors: Carl Edwards, Edward De Brouwer, Xiner Li... Published: 2026-09-10 | Citations: 0 arXiv | PDF

Research Question: How can we efficiently prioritize CRISPR perturbation experiments under constrained budgets when exhaustive testing is infeasible, and how can acquisition strategies be learned across historical screens rather than designed heuristically per-experiment?

Summary: The paper introduces AssayBench-Loop, a 1,389-screen CRISPR benchmark for adaptive hit discovery, and AssayLoop, which pairs an amortized transformer acquisition policy (AssayFormer) trained across historical screens with LLM-derived biological priors via an adaptive handoff. It demonstrates that learning acquisition strategies from prior experiments beats standalone LLMs and existing adaptive-design baselines, achieving 5.67x enrichment and 27.7% hit recovery at 5% library sampling.

Key Results: Introduced AssayBench-Loop (1,389 CRISPR screens across 5 phenotype categories). On temporally held-out screens, AssayLoop achieved 5.67-fold enrichment over random selection and recovered 27.7% of hits after assaying ~5% of the candidate library, outperforming existing adaptive-design methods, standalone LLMs, and AssayFormer alone. Performance scales with historical training data and transfers to held-out phenotype categories.

Key Findings:

Amortizing acquisition policies across historical screens outperforms per-experiment adaptive designs and standalone LLMs
LLM biological priors are valuable for cold-start seeding but should hand off to the learned policy as evidence accumulates
The approach transfers to phenotype categories excluded from training, and performance scales with the amount of historical data

Technical Novelty: AssayFormer: a transformer trained across historical CRISPR screens as an amortized acquisition policy (rather than per-experiment Bayesian optimization). The 'adaptive handoff' combines LLM-derived biological priors for cold-start with the learned policy as evidence accumulates. AssayLLM shows the same principle can be baked directly into an LLM via task-specific post-training.

What's New: Prior adaptive hit-discovery methods design acquisition heuristics per-experiment (e.g., Bayesian optimization, active learning) without leveraging the growing corpus of completed screens. This work reframes past experiments as training data for the acquisition policy itself, and integrates LLM biological priors through a principled adaptive handoff — a combination not previously demonstrated at this scale.

Extension Opportunities:

Extend the amortized acquisition policy to other high-throughput biological assays (small-molecule screens, arrayed RNAi, base-editing scans) using the same 'completed experiments as training data' paradigm
Replace/augment the LLM prior with structured knowledge graphs (STRING, Reactome) or protein language model embeddings to sharpen the cold-start seed before the transformer takeover
Study the adaptive handoff mechanism itself — learn when to trust the LLM prior vs. the transformer policy as a meta-controller, potentially with uncertainty estimates

Replicability: Abstract does not specify code/data release, but AssayBench-Loop is presented as a benchmark, implying intended release. Training a transformer over 1,389 screens plus LLM post-training suggests moderate-to-high compute (multi-GPU); inference should be tractable on a single GPU.

Research Gaps:

Existing adaptive hit-discovery benchmarks are small and phenotypically narrow, limiting the ability to train cross-screen policies
Standalone LLMs and standalone acquisition policies each have blind spots — biological priors vs. evidence accumulation — that hadn't been bridged

2. The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement

Authors: Yi Duan, Ying Liu, Zirui Tang... Published: 2026-09-10 | Citations: 0 arXiv | PDF

Research Question: How can AI systems move beyond static post-training paradigms to achieve genuine recursive self-improvement (RSI), where systems autonomously turn experience into persistent gains in both capabilities and the improvement process itself?

Summary: The paper argues that current LLMs are approaching the ceiling of the human-authored improvement paradigm and proposes recursive self-improvement (RSI) as the next frontier. It contributes a diagnostic (HCI), a five-stage autonomy roadmap toward genuine RSI, and a scenario-by-scenario analysis (science, embodied AI, software engineering) of where RSI will land first.

Key Results: The paper is largely a position/roadmap paper rather than an empirical study. It introduces the Headroom-Closed Index (HCI) as a diagnostic to expose limitations of current LLMs and cites 'preliminary empirical evidence' drawn from industry practice, but the abstract reports no specific benchmark numbers, datasets, or quantitative results.

Key Findings:

Existing LLMs exhibit diminishing returns under the human-in-the-loop improvement paradigm, quantified via the Headroom-Closed Index
RSI can be decomposed into five progressively harder autonomy stages, culminating in recursive meta-improvement where the system improves its own improvement process
Different application domains (scientific discovery, embodied intelligence, software engineering) have distinct RSI requirements and will progress at different speeds — software engineering being the most tractable near-term target

Technical Novelty: The framing of RSI as a five-tier autonomy ladder (improvement-execution → improvement-strategy → experience-acquisition → environment-adaptation → recursive meta-improvement) and the HCI diagnostic for measuring how much 'headroom' current LLMs have already closed.

What's New: Unlike prior self-improvement work that focuses on single mechanisms (self-play, self-distillation, RLAIF), this paper offers a unifying taxonomy plus a measurable diagnostic (HCI) and ties the roadmap explicitly to industry-observed bottlenecks rather than a single algorithm.

Extension Opportunities:

Operationalize the Headroom-Closed Index (HCI) as an open benchmark and evaluate frontier LLMs (GPT, Claude, Gemini, Llama) across the five autonomy tiers to produce a public leaderboard
Build a reference implementation of 'improvement-execution autonomy' by wiring an agent loop that persists learned skills/tools across sessions and measure compounding gains on SWE-bench or ML-Agent-Bench
Instantiate RSI in one narrow scenario (e.g., software engineering) with a closed loop of code-generation → test feedback → weight or memory updates, and quantify per-iteration lift

Replicability: No code, datasets, or compute requirements are indicated in the abstract; as a conceptual roadmap paper, reproduction would mean reimplementing the HCI metric from the paper's definition rather than rerunning experiments.

Research Gaps:

No concrete algorithmic proposal or empirical baseline is provided for any of the five autonomy stages
The HCI metric is introduced but its formal definition, validation, and calibration against known model capabilities are not detailed in the abstract

🤖 Agents

1. Evaluating Scaffolding-Oriented Multi-Agent Large Language Model System for Clinical Interview Training

Authors: Luming Yang, Haoxian Liu, Siqing Li... Published: 2026-09-10 | Citations: 0 arXiv | PDF

Research Question: Can a scaffolding-oriented multi-agent LLM system improve medical students' clinical interviewing skills (communication, empathy, history-taking) beyond traditional simulated patient training, without artificially inflating diagnostic scores?

Summary: The paper introduces an AI Standardized Patient training platform built from three coordinated LLM agents (patient, Socratic tutor, turn-level evaluator) that provide pedagogical scaffolding without leaking diagnostic answers. In a 100-student RCT, the multi-agent system improved OSCE exam scores particularly in communication, empathy, and history-taking, while leaving diagnostic accuracy unchanged — evidence that scaffolding lifts process quality without inflating outcomes.

Key Results: Randomized controlled trial with N=100 medical students, split between multi-agent (MA) scaffolding condition and control. Assessed via OSCE-based rubric in patient-only exam environment. No significant difference in final diagnostic accuracy, but MA group showed improved final exam scores overall, with most substantial and consistent gains in communication, empathy expression, and specific history-taking behaviors.

Key Findings:

Multi-agent scaffolding improved OSCE exam scores overall vs control, driven mainly by communication and empathy subscales
No significant difference in final diagnostic accuracy between conditions — scaffolding did not shortcut the diagnostic reasoning being tested
Turn-level evaluator and Socratic tutor architecture enables formative feedback while preserving assessment validity

Technical Novelty: Three-agent decomposition specifically designed around pedagogical scaffolding theory: (1) patient agent for simulation, (2) tutor agent giving Socratic prompts that deliberately withhold diagnostic information, (3) turn-level evaluator that monitors progress without revealing summative scores. The information-hiding design is the key novelty — prior AI-SP work tends to leak evaluative signals that inflate performance.

What's New: Combines multi-agent LLM architecture with pedagogically motivated information hiding — the tutor and evaluator withhold diagnostic and summative signals by design. Also delivers an RCT (N=100) rather than a single-arm feasibility study, which is rare in LLM-for-medical-education work.

Extension Opportunities:

Extend the tutor agent's Socratic scaffolding to specialty-specific interviews (pediatrics, psychiatry, emergency medicine) where communication norms differ substantially
Add a longitudinal fading-scaffolds mechanism where tutor/evaluator intervention decreases across sessions to measure skill retention and transfer
Integrate multimodal patient agents (voice, facial affect) to train non-verbal empathy cues absent from text-only dialog

Replicability: Authors release a multi-expert annotated dataset (transcripts, checklist annotations, turn-level evaluations, OSCE-aligned scores). Code availability not explicitly stated in abstract. Compute is modest — inference on hosted LLM APIs for 100 students across 2 sessions; reproduction requires IRB approval and medical student cohort more than compute.

Research Gaps:

Only tested on one OSCE case type; generalization across clinical scenarios and specialties is untested
No measurement of skill retention beyond immediate post-training exam or transfer to real patient encounters

🦾 ROBOTICS

1. ActSafeGuard: Differentiable and Training-Aligned Constraint Enforcement for Flow-Matching Policies

Authors: Jianming Ma, Rongjun Jin, Xiaxi Si... Published: 2026-09-10 | Citations: 0 arXiv | PDF

Research Question: How can flow-matching robotic policies (VLA/WAM) enforce hard physical action constraints deterministically at every step without decoupling safety from training, avoiding the train/inference mismatch of post-hoc projection methods?

Summary: ActSafeGuard is a differentiable safeguard layer for flow-matching robotic policies that bakes hard action-feasibility constraints into training via an analytical ray-scaling operator. It delivers 100% per-step safety on π_{0.5} and Fast-WAM without sacrificing (and sometimes improving) task success, closing the train/inference gap left by post-hoc safety filters.

Key Results: On π_{0.5} and Fast-WAM backbones across multiple manipulation tasks, ActSafeGuard achieves 100% step-wise safety rate while preserving or boosting task success rates versus unconstrained baselines and inference-only correction methods.

Key Findings:

Training-time integration of safety via ray-scaling yields boundary-aware gradients that teach the policy the feasible manifold rather than clipping onto it
Achieves 100% step safety across tested tasks on two foundation backbones (π_{0.5}, Fast-WAM)
Safety enforcement is non-degrading — task success is preserved or improved, contradicting the common safety-vs-performance tradeoff

Technical Novelty: An analytical, differentiable ray-scaling operator embedded inside flow-matching training that produces boundary-aware gradients, so the policy learns the constrained manifold end-to-end rather than being projected onto it at inference.

What's New: Prior safe-RL/VLA work either enforces statistical safety (no per-step guarantees) or applies inference-time projection (train/deploy mismatch). ActSafeGuard is the first to give both deterministic per-step feasibility and training alignment for flow-matching policies via a differentiable analytical operator.

Extension Opportunities:

Extend the ray-scaling operator from convex/box-like feasibility sets to non-convex or contact-rich constraint manifolds (e.g., collision meshes, friction cones) via learned signed-distance approximations
Integrate ActSafeGuard with diffusion-policy backbones (beyond flow-matching) and test on real-robot deployments where constraint violations have physical cost
Combine with model-predictive safety filters or CBFs over a horizon rather than per-step, enabling anticipatory safety for long-horizon manipulation

Replicability: Abstract does not mention code/data release. Reproduction would require the π_{0.5} and Fast-WAM checkpoints plus a standard VLA manipulation benchmark suite; compute is likely moderate GPU-scale fine-tuning (single-node multi-GPU) rather than pretraining from scratch.

Research Gaps:

Only per-step, likely convex action constraints are addressed — horizon-level safety and non-convex environment constraints (collisions, contacts) remain open
Evaluation is in-distribution across standard backbones/tasks; real-world robustness, sim-to-real transfer, and adversarial or novel-constraint generalization are not established

2. GeoTrussRover: Morphological Computation with Contact-Semantic Control Primitives

Authors: Muyuan Ma, Yi Zhang, Yang Yang... Published: 2026-09-10 | Citations: 0 arXiv | PDF

Research Question: How can a variable-geometry truss (VGT) coupled to a wheeled mobile base be coordinated across its high-dimensional configuration space to negotiate obstacles like steps, without re-solving the full motion planning problem for every new obstacle geometry?

Summary: GeoTrussRover couples a variable-geometry truss to a wheeled base and introduces contact-semantic primitives — a low-dimensional representation of coordination across 21 truss members — to plan and adapt obstacle traversal. The method reuses primitives across step heights via physics-constrained projection, recomputing only infeasible phases, and demonstrates 63.7% fewer objective evaluations for a representative transfer along with feasibility up to ~5 wheel radii.

Key Results: The authors demonstrate that extracting four contact-semantic primitives from a single solved traversal (coordinating 21 truss members) and applying physics-constrained projection enables reuse across unseen step heights. Concrete results: (1) transfer from 0.10m to 0.075m step reduces objective-function evaluations by 63.7% vs. full recomputation; (2) contact-phase feasibility covers step heights from 0.10m to 0.46m (1.08 to 4.97 wheel radii), with upper bound near theoretical feasibility limit; (3) physical electric prototype traversed 2.11 wheel radii; (4) full-space QP tracks adapted motion and corrects member/wheel errors.

Key Findings:

Four contact-semantic primitives are sufficient to describe coordination across all 21 truss members for step traversal
Physics-constrained projection transfers primitives to new step heights with 63.7% less compute (0.10m → 0.075m), and localized recomputation handles infeasible phases
Feasibility envelope extends to 4.97 wheel radii (near theoretical bound); physical prototype validated at 2.11 wheel radii

Technical Novelty: The novel contribution is encoding coordination of a hyper-redundant load-bearing morphology (21 members) as a small set of contact-semantic primitives that describe contact topology invariants, then adapting them via physics-constrained projection while only recomputing phases that violate constraints. This differs from prior VGT work which typically requires full trajectory recomputation, and from primitive-based mobile robot work which does not deal with load-bearing structural coupling.

What's New: Prior VGT work treats shape change and locomotion as separate high-dimensional planning problems. This paper is the first to (a) unify VGT and wheeled base coordination through a compact contact-semantic primitive representation and (b) enable primitive reuse across obstacle geometries via physics-constrained projection with selective phase recomputation, rather than full retraining or replanning.

Extension Opportunities:

Extend contact-semantic primitives beyond single-step obstacles to more complex terrain classes (stairs, gaps, slopes) and evaluate whether the primitive library composes
Learn primitives from demonstration or self-play rather than hand-extracting from one source traversal — enables scaling to more morphologies and terrain types
Integrate onboard perception (RGB-D or tactile) to auto-detect contact topology and trigger which primitive to use, closing the loop for fully autonomous obstacle traversal

Replicability: Abstract does not mention public code or dataset release. Reproduction would require an electrically actuated VGT prototype (21 members), a wheeled base, and a QP solver stack — significant hardware barrier. Simulation reproduction is more feasible but the physics-constrained projection details would need to be inferred from the full paper.

Research Gaps:

Only step obstacles are evaluated — generalization to stairs, gaps, uneven terrain, and dynamic obstacles is untested
Primitives are extracted from a single hand-solved source traversal, limiting scalability; no learning-based extraction is presented

3. Safety-aware Skill Adaptation for Reinforcement Learning in Dynamic Environments

Authors: A K M Nadimul Haque, Sheila Sutjipto, Marc G. Carmichael... Published: 2026-09-10 | Citations: 0 arXiv | PDF

Research Question: How can RL-based skill adaptation frameworks safely modify learned robot skills in cluttered, dynamic environments without unsafe exploration, unstable learning, or violating kinematic characteristics of the original demonstration?

Summary: Dist-GPRL is a safety-guided RL framework that adapts GP-parameterized robot skills by sequentially updating overlapping local via-point windows rather than the full trajectory. It combines a Hausdorff-Approximation-Planner safe-subspace prior, distance-field clearance/gradient rewards, and a kinematics-similarity regularizer to achieve safer, more stable skill adaptation in cluttered dynamic environments with successful sim-to-real transfer.

Key Results: The authors demonstrate Dist-GPRL on two dynamic object-manipulation tasks in simulation with sim-to-real transfer. Reported qualitative outcomes: higher task success rate, lower collision frequency, and more stable learning curves vs. baselines, while preserving demonstrated velocity/acceleration profiles. Specific numerical benchmarks are not provided in the abstract.

Key Findings:

Local sequential GP via-point adaptation reduces action-space size and credit-assignment difficulty compared to global trajectory adaptation
Combining a HAP-derived safe-subspace prior with distance-field rewards yields higher success and lower collision rates than baselines in dynamic manipulation
A kinematics-similarity regularizer preserves the demonstrated velocity and acceleration profiles during RL adaptation, and the policy transfers to real hardware

Technical Novelty: The combination of (1) sequential adaptation of overlapping local GP via-point windows rather than global trajectory modification, (2) a Hausdorff Approximation Planner-derived safe-subspace prior biasing exploration, (3) dynamically updated distance-field clearance + gradient rewards, and (4) a trajectory-kinematics similarity regularizer that preserves demo velocity/acceleration. Prior GP-RL skill adaptation typically modifies whole trajectories globally without a HAP safe prior or kinematics-preserving regularizer.

What's New: Unlike prior GP-RL skill adaptation methods that either fix observations or globally re-optimize the whole trajectory, this work introduces windowed local via-point updates coupled with a planner-derived safe prior and a demonstration-preserving kinematic regularizer — a novel decomposition of safety, locality, and style preservation for structured skill adaptation.

Extension Opportunities:

Replace the Hausdorff Approximation Planner prior with a learned safety critic (e.g., CBF-based or diffusion-policy prior) to handle higher-dimensional or non-Euclidean configuration spaces
Extend the sequential GP via-point window mechanism to multi-arm or bimanual manipulation where inter-arm distance fields must be jointly optimized
Integrate perception uncertainty (e.g., depth-sensor noise, occluded moving obstacles) into the distance field clearance reward using probabilistic SDFs instead of deterministic ones

Replicability: No code/data release is mentioned in the abstract. Reproduction requires a robot simulator (likely PyBullet/Isaac/MuJoCo), a real manipulator for transfer, and GP libraries. Compute is modest — GP-based policy parameterizations are typically CPU-tractable, though distance field updates for dynamic scenes may need a GPU for real-time performance.

Research Gaps:

No reported quantitative metrics (success %, collision counts, sample efficiency) in the abstract, making direct benchmarking against SOTA safe-RL methods difficult
Reliance on HAP-derived priors may not scale to high-DOF or non-geometric task spaces (e.g., contact-rich or deformable manipulation)

💻 COMPUTE

1. Hilbert-space selected switch of helical edges in an artificial quantum Hall insulator

Authors: Naijie Ren, Zhiren Xiong, Kaining Yang... Published: 2026-09-10 | Citations: 0 arXiv | PDF

Research Question: How can Landau-level topology be reconstructed to engineer artificial quantum Hall insulators and selectively control helical edge channels without relying on challenging electrostatic confinement or symmetry breaking?

Summary: The paper demonstrates that interfacial charge transfer between CrOCl and large-angle twisted bilayer graphene produces composite Landau-level filling configurations that selectively host or gap helical edge states. The (+2,-2) state is fully insulating (GΩ), while (+6,-6) and (+10,-10) retain counter-propagating helical edges with up to 50% non-local response, offering a Hilbert-space route to engineer artificial quantum Hall insulators.

Key Results: Demonstrated that interfacial charge transfer between CrOCl and large-angle twisted bilayer graphene offsets the two Dirac Landau-level ladders, producing composite filling configurations. At charge neutrality, the (+2,-2) state is fully insulating with longitudinal resistance in the GΩ regime. Higher composite zero-filling states (+6,-6) and (+10,-10) retain counter-propagating helical edge channels, exhibiting non-local transport up to 50% of the local response.

Key Findings:

CrOCl/tBLG interfacial charge transfer offsets decoupled Dirac Landau-level ladders, creating new composite fillings
(+2,-2) charge neutrality state is fully insulating with GΩ-scale longitudinal resistance
(+6,-6) and (+10,-10) states retain helical edge channels with non-local transport reaching 50% of local response
The N=±1 orbital components in Landau-spinor Hilbert space act as the switch controlling edge conduction

Technical Novelty: Uses interfacial charge transfer in a CrOCl/twisted-bilayer-graphene heterostructure to offset otherwise decoupled Dirac Landau-level ladders across the two graphene layers — bypassing electrostatic confinement and symmetry breaking as reconstruction routes. The Hilbert-space (Landau-spinor orbital N=±1) selection mechanism for switching helical edge modes on/off is a distinctive framing.

What's New: Unlike prior reconstructions relying on electrostatic gating or symmetry breaking, this work uses passive interfacial charge transfer with a magnetic insulator (CrOCl) to reshape Landau-level ladders, and identifies a Hilbert-space (orbital N=±1) selection rule governing which composite fillings host helical edges.

Extension Opportunities:

Explore other 2D magnetic insulator/graphene heterostructures (e.g., CrI3, CrPS3) to identify additional charge-transfer platforms with tunable Landau-level offsets
Integrate the (+6,-6)/(+10,-10) helical edge states into mesoscopic interferometers or Josephson junctions to probe topological superconductivity and Majorana modes
Develop gate-tunable devices that dynamically switch between insulating (+2,-2) and helical-edge (+6,-6) regimes for on-demand topological qubit routing

Replicability: No mention of code/data release in the abstract. Reproduction requires high-quality CrOCl crystals, large-angle twisted bilayer graphene fabrication (dry-transfer stack in cleanroom), high magnetic field cryogenic transport measurements (dilution refrigerator + ~10T+ magnet), and nanofabrication for Hall-bar/non-local geometries — substantial experimental infrastructure.

Research Gaps:

Microscopic role of CrOCl magnetism versus purely electrostatic charge transfer in stabilizing the composite states remains incompletely disentangled
Whether the Hilbert-space selection mechanism generalizes beyond twisted bilayer graphene to other Dirac/multi-flavor systems

2. Entwine: Coordinating Tiled Computation and Fine-Grained Communication across GPUs

Authors: Kai Ma, Quanfeng Lv, Jingguo Ge... Published: 2026-09-10 | Citations: 0 arXiv | PDF

Research Question: How can tile-based GPU computations better overlap with communication when mismatched progress causes stalls, bursts, and resource contention that offset overlap benefits in tensor-parallel LLM workloads?

Summary: Entwine is a GPU runtime/kernel framework that co-designs tile scheduling, fine-grained SM-based communication, and SM resource allocation to maximize compute-communication overlap in tensor-parallel LLM workloads. It smooths communication data-arrival patterns via tile reordering and dedicates tuned SM shares to comm kernels, yielding 1.232x geomean (1.433x max) speedup over cuBLAS+NCCL.

Key Results: Across representative tensor-parallel LLM workloads, Entwine achieves a geomean speedup of 1.232x (up to 1.433x) over cuBLAS+NCCL, and outperforms state-of-the-art overlap baselines by 3.1-9.8% in geomean. Evaluated on tensor-parallel LLM computation patterns.

Key Findings:

Naive tile ordering produces bursty data-ready patterns that starve or overwhelm communication, limiting overlap effectiveness
SM-based fine-grained communication kernels give lower latency than NCCL but consume compute resources, requiring explicit allocation tuning
Coordinated scheduling + fine-grained comm + SM partitioning delivers 3.1-9.8% geomean improvement over prior overlap-focused baselines

Technical Novelty: Joint coordination of three dimensions previously optimized in isolation: (1) tile computation reordering to produce a regular data stream, (2) fine-grained SM-resident communication kernels for low-latency dispatch, and (3) explicit SM resource partitioning between compute and comm kernels to balance overlap gains against compute slowdown.

What's New: Prior work addresses compute-comm overlap via either kernel fusion, custom comm libraries, or scheduling in isolation. Entwine is the first to jointly coordinate all three axes (tile order, SM-resident comm, resource partitioning) and demonstrates the interaction terms matter.

Extension Opportunities:

Extend Entwine's coordinated scheduling to pipeline-parallel or expert-parallel (MoE) workloads where communication patterns differ from tensor parallelism
Apply the tile reordering + SM allocation coordination to training workloads (backward passes, gradient all-reduce) rather than just inference-style operations
Port the fine-grained SM-based communication approach to emerging interconnects (NVLink 5, CXL, or heterogeneous GPU-CPU fabrics) and quantify tradeoffs vs NCCL

Replicability: Authors state they will open-source upon publication (not yet available). Reproduction requires multi-GPU node with NVLink (likely 8xH100/A100 class), CUDA toolkit, cuBLAS, NCCL, and representative LLM tensor-parallel workloads (e.g., Megatron-LM style GEMM+AllReduce patterns).

Research Gaps:

Evaluation scope limited to tensor-parallel LLM patterns; generalization to MoE, pipeline-parallel, or non-LLM tiled workloads is unproven
Coordination policy tuning (how tile order and SM allocation are chosen per workload/topology) may require offline profiling not detailed in abstract

3. Rethinking Sparse Formats for RISC-V: A Hierarchical Approach to High-Performance SpMV

Authors: Anna Pirova, Anastasia Vodeneeva, Konstantin Kovalev... Published: 2026-09-10 | Citations: 0 arXiv | PDF

Research Question: How can sparse matrix storage formats be redesigned to maximize SpMV throughput on emerging RISC-V vector processors, where existing formats (CSR, ELL, etc.) were optimized for x86/GPU architectures and underutilize RVV instructions?

Summary: The paper benchmarks 8 sparse matrix formats for SpMV on RISC-V vector hardware and introduces Hierarchical CSR (HCSR), a CSR variant tailored to RVV 1.0 that outperforms existing formats across diverse matrices. Implementations are released as the open-source RVVLASparse library, validated on SpacemiT K1/K3 boards.

Key Results: Benchmarked vectorized SpMV across 8 sparse formats using RVV 1.0 intrinsics on SpacemiT K1 and K3 RISC-V boards. Selecting the best format per matrix yields an average 1.6x speedup, and the proposed Hierarchical CSR (HCSR) achieves the shortest execution time across a broad class of sparse matrices. Released as open-source C++ library RVVLASparse.

Key Findings:

Format choice alone yields an average 1.6x SpMV speedup on RISC-V — flat CSR is rarely optimal
HCSR delivers the shortest execution time across a broad class of sparse matrices among all 8 formats tested
RVV 1.0 intrinsics can be effectively targeted from C++ to build portable sparse linear algebra kernels on emerging RISC-V SBCs

Technical Novelty: Hierarchical CSR (HCSR) — a two-level CSR variant designed around RVV vector-length-agnostic execution, reducing indirect-load overhead and improving vector lane utilization vs. flat CSR, ELL, SELL-C-sigma, and other established formats on RISC-V.

What's New: First systematic RVV 1.0 SpMV format comparison on real RISC-V silicon (SpacemiT K1/K3), plus a new hierarchical CSR variant purpose-built for vector-length-agnostic RISC-V execution rather than ported from x86/GPU designs.

Extension Opportunities:

Build a runtime auto-tuner that predicts the optimal format (including HCSR) per matrix using lightweight structural features, avoiding manual format selection
Extend HCSR to block-sparse and mixed-precision (FP16/BF16) SpMV for ML workloads on RISC-V edge accelerators like SpacemiT K1
Port HCSR to iterative Krylov solvers (CG, GMRES) inside PETSc/Trilinos with RVV backends and measure end-to-end solver speedup on scientific simulations

Replicability: Yes — RVVLASparse is publicly available as an open-source C++ library using RVV 1.0 intrinsics. Reproduction requires access to RISC-V hardware with RVV 1.0 support (SpacemiT K1 or K3 boards, ~$100-200 SBCs) and standard SuiteSparse matrices; no GPU or large compute needed.

Research Gaps:

No automatic format-selection heuristic or cost model is provided — users must still pick HCSR vs alternatives manually
Evaluation is confined to single-node SpacemiT boards; multi-core scaling, larger RVV vector widths, and integration into full iterative solvers are untested

⚡ ENERGY

1. Breaking Water at Graphene Defects

Authors: Samuel G. H. Brookes, IniOluwa C. Popoola, Fabian Berger... Published: 2026-09-10 | Citations: 0 arXiv | PDF

Research Question: How does solvation influence water dissociation mechanisms at defective graphene surfaces, specifically at single vacancy (SV) sites, where prior understanding was limited to gas-phase or isolated-molecule pictures?

Summary: The paper uses machine-learned interatomic potentials to show that liquid-water solvation fundamentally rewrites how water dissociates at a graphene single vacancy, turning a single concerted gas-phase channel into two competing acidic and basic pathways that yield distinct chemisorbed SV-OH and SV-H intermediates. This coupling of point-defect chemistry to interfacial charge and wettability reframes even the simplest carbon vacancy as a source of rich, solvation-controlled interfacial reactivity.

Key Results: Using machine-learned interatomic potentials (MLIPs) at graphene-water interfaces, the authors demonstrate that solvation qualitatively alters the reaction mechanism at a graphene single vacancy: instead of the single concerted gas-phase channel, the solvated SV splits water via two competing lower-barrier pathways — a basic route (SV-H + OH-(aq)) and an acidic route (SV-OH + H3O+(aq)) — producing distinct chemisorbed intermediates that enhance graphene-water adsorption. Specific barrier heights, MLIP architecture, and system sizes are not disclosed in the abstract.

Key Findings:

Solvation qualitatively changes the mechanism: gas-phase single concerted channel vs solvated bifurcation into acidic and basic routes
Two competing solvated pathways produce chemically distinct products — SV-H + OH-(aq) (basic) and SV-OH + H3O+(aq) (acidic) — both with lower barriers than the gas-phase route
The resulting chemisorbed intermediates enhance graphene-water adsorption and couple defect chemistry to interfacial charge and wettability

Technical Novelty: First application of state-of-the-art machine-learned interatomic potentials to resolve solvation-driven bifurcation of water-dissociation pathways at a graphene point defect — bridging DFT-level accuracy with the length/time scales needed to capture explicit hydrogen-bond networks and solvated ion products (OH-, H3O+) that standard AIMD or static DFT cannot afford.

What's New: Prior work largely treated water dissociation at graphene defects in gas-phase or with implicit/small-cluster solvation; this work uses MLIPs to access explicit liquid-water environments at DFT-level accuracy and demonstrates that the solvent is not a spectator but a mechanism-selector, opening ionic product channels absent in vacuum.

Extension Opportunities:

Apply the same MLIP framework to other 2D-material defects (h-BN vacancies, MoS2 sulfur vacancies, or divacancies/Stone-Wales defects in graphene) to build a defect-vs-mechanism map for aqueous 2D interfaces
Couple the identified acidic/basic pathways to explicit electrochemical potential (constant-potential MLIP-MD) to predict pH-dependent and voltage-dependent selectivity relevant to graphene electrodes
Use the resulting SV-H / SV-OH chemisorbed intermediates as starting points to model nanofluidic transport, wettability tuning, and proton conduction through defective graphene membranes

Replicability: Abstract does not specify code/data availability, MLIP architecture (e.g., MACE, Allegro, NequIP), or training set. Reproduction would typically require GPU-accelerated MLIP training on DFT reference data (a few thousand to tens of thousands of configurations at meta-GGA or hybrid level) followed by nanosecond-scale MD with enhanced sampling (metadynamics/umbrella sampling) — roughly O(10^3-10^4) GPU-hours end-to-end.

Research Gaps:

No treatment of applied electrochemical potential or electrolyte ions — real electrode/nanofluidic conditions may shift the acidic/basic selectivity
Only the single vacancy is examined; the broader defect zoo (divacancies, Stone-Wales, edges, dopants) and its interaction with solvation-driven pathways remains open

2. Low temperature thermodynamics of $S_{\mathrm{eff}}=1/2$ triangular lattice quantum spin liquid candidate TlYbS$_2$

Authors: Bishnu P. Belbase, Arjun Unnikrishnan, Piyush Chhallare... Published: 2026-09-10 | Citations: 0 arXiv | PDF

Research Question: Does the Yb-based delafossite TlYbS$_2$ realize a quantum spin liquid ground state on the triangular lattice, and how does its low-temperature thermodynamic behavior compare to closely related sister compounds like TlYbSe$_2$ and KYbSe$_2$?

Summary: The paper presents a comprehensive low-temperature (millikelvin) thermodynamic study of single-crystal TlYbS$_2$, a Yb-delafossite triangular-lattice antiferromagnet, identifying a weak 530 mK anomaly, near-quadratic specific heat scaling ($T^{1.8}$), and pronounced easy-plane anisotropy. Field-dependent measurements reveal a 1/3 magnetization plateau and a proposed confinement–deconfinement transition, positioning TlYbS$_2$ as a distinct QSL candidate whose behavior differs qualitatively from its Se sister compound.

Key Results: Single-crystal TlYbS$2$ crystallizes in trigonal $R\bar{3}m$ with no detectable site disorder. A weak thermodynamic anomaly appears at ~530 mK at zero field. Below 300 mK, magnetic specific heat follows $C_m \propto T^{1.8}$ (near-quadratic), unlike the linear $T$ dependence in TlYbSe$2$. ESR/magnetization show strong easy-plane anisotropy with $g\perp/g\parallel \approx 6.9$. For $H\perp c$: anomaly persists to ~2 T, field-induced order above ~2.5 T, 1/3 magnetization plateau between ~5–8 T, full polarization near 17 T. For $H\parallel c$: anomaly suppressed near 3 T, consistent with confinement–deconfinement transition to field-induced QSL.

Key Findings:

Weak thermodynamic anomaly at ~530 mK suggests a weakly ordered ground state similar to KYbSe$_2$
Magnetic specific heat follows $C_m \propto T^{1.8}$ below 300 mK, contrasting with linear behavior in TlYbSe$_2$
Strong easy-plane anisotropy with $g_\perp/g_\parallel \approx 6.9$; field-induced 1/3 plateau (5–8 T) and full polarization at ~17 T for $H\perp c$
For $H\parallel c$, anomaly is suppressed near 3 T, consistent with a confinement–deconfinement transition into a field-induced QSL

Technical Novelty: First comprehensive single-crystal thermodynamic characterization of TlYbS$_2$ into the millikelvin regime, revealing near-quadratic (rather than linear) low-$T$ specific heat that distinguishes it from its Se analog, and mapping a full anisotropic $H$–$T$ phase diagram including a 1/3 plateau and a proposed confinement–deconfinement transition.

What's New: Provides the first millikelvin single-crystal study of TlYbS$_2$ with structural cleanliness verified, and directly contrasts its near-quadratic specific heat with the linear-$T$ behavior of TlYbSe$_2$, suggesting a different low-energy excitation spectrum despite structural similarity. Also documents a confinement–deconfinement transition under $c$-axis field.

Extension Opportunities:

Perform inelastic neutron scattering on TlYbS$_2$ single crystals to directly probe the spinon continuum and gauge-flux excitations predicted by the $T^{1.8}$ specific heat scaling
Systematically substitute S with Se to interpolate between the $T^{1.8}$ (TlYbS$_2$) and linear-$T$ (TlYbSe$_2$) regimes and map the QSL phase diagram
Develop microscopic spin-model simulations (DMRG/tensor networks) with realistic anisotropic exchange constants derived from the measured $g$-tensor to reproduce the 1/3 plateau and field-induced phases

Replicability: No code/data availability mentioned. Reproduction requires single-crystal growth of TlYbS$_2$ (chalcogenide flux/vapor methods), dilution refrigerator (down to ~50 mK), high-field magnet (up to ~17 T), plus ESR and AC susceptibility infrastructure — accessible only in specialized condensed-matter labs.

Research Gaps:

Microscopic origin of the $T^{1.8}$ power law and its relation to spinon/gauge-flux excitations remains unresolved without spectroscopic probes
Nature of the 530 mK anomaly (true ordering vs. crossover) and the field-induced QSL phase are not directly confirmed by neutron scattering or NMR

3. Competition between vacancy creation and filling in defect-engineering of hBN

Authors: Shrirang Chokappa, Manuel Laängle, Barbara Maria Mayer... Published: 2026-09-10 | Citations: 0 arXiv | PDF

Research Question: How can defects in monolayer hBN be created in a controlled way to enable selective engineering of quantum emitters, and do current simulation predictions match experimental reality?

Summary: The authors irradiate freestanding monolayer hBN with 150 eV Ar+ ions and use STEM to quantify resulting defects, finding that boron single vacancies dominate and double vacancies occur at half the rate — contrary to analytical MD predictions. They further reveal that Si and C impurity atoms fill vacancies at rates significant enough to complicate deterministic quantum emitter engineering.

Key Results: Using 150 eV Ar+ ion irradiation on freestanding monolayer hBN and STEM characterization, the authors demonstrate that boron single vacancies dominate, with double vacancies occurring at roughly half that count — contradicting analytical potential MD predictions. They also show substantial vacancy filling by Si and C impurities.

Key Findings:

Boron single vacancies are the dominant defect after 150 eV Ar+ irradiation, with double vacancies at about half the count.
Analytical potential MD simulations fail to predict the observed defect distribution in monolayer hBN.
Si and C impurity atoms fill vacancies at rates high enough to materially alter the defect landscape relevant to quantum emitters.

Technical Novelty: First combination of ultra-low-energy (150 eV) Ar+ irradiation on freestanding monolayer hBN with quantitative STEM defect-population statistics, plus explicit accounting for competing vacancy-filling by Si/C impurities — a channel typically ignored in prior irradiation-damage studies.

What's New: Direct experimental refutation of MD-based predictions for hBN defect creation under low-energy ion irradiation, coupled with the first quantitative emphasis on impurity vacancy-filling as a competing channel that shapes the final defect population.

Extension Opportunities:

Repeat the irradiation study across an energy sweep (e.g., 50–500 eV) and ion species (He+, Ne+, N+) to map defect-type selectivity as a function of momentum transfer.
Pair the STEM defect statistics with in-situ photoluminescence to directly correlate specific vacancy/impurity configurations with quantum emitter spectra.
Develop or benchmark machine-learned interatomic potentials against these experimental defect distributions to replace the failing analytical MD potentials.

Replicability: No code or dataset availability is mentioned in the abstract. Reproduction requires a low-energy ion irradiation setup (~150 eV Ar+), freestanding monolayer hBN sample preparation, and aberration-corrected STEM — specialized experimental infrastructure rather than compute.

Research Gaps:

Lack of accurate interatomic potentials that reproduce experimentally observed hBN irradiation-damage populations.
Absence of methods to suppress or control Si/C impurity migration into vacancies during defect engineering for quantum emitters.

🏥 HEALTHCARE

1. Development, Evaluation, and Multicenter Clinical-Trial Application of an Artificial Intelligence-Assisted MRI Method for Quantitative Knee Cartilage Morphometry

Authors: Binbin Yang, Rui Huang, Yuanjing Xu... Published: 2026-09-08 | Citations: 0 arXiv | PDF

Research Question: How can an AI-assisted MRI pipeline deliver reproducible, quantitative knee cartilage morphometry (volume, thickness, degenerated surface area) at the scale and rigor required for a multicenter phase III knee osteoarthritis (KOA) trial, where manual segmentation is too slow and inter-reader variability confounds treatment-effect measurement?

Summary: The authors built and deployed an nnU-Net-based knee cartilage segmentation pipeline with a two-reader-plus-adjudicator correction step, coupled to novel 3D ray-tracing thickness and 3D ray-based degenerated-area measurements, for a multicenter phase III KOA trial. Across 1,188 exams from 416 participants they showed the pipeline produces reproducible endpoints (ICC 0.96-1.00, Dice 0.96) that detected a treatment vs control divergence in cartilage volume, thickness, and degenerated area.

Key Results: On 1,189 phase III MRIs, 3D full-resolution nnU-Net pre-segmentation achieved Dice 0.964 +/- 0.030 (median 0.970), with 78.7% of cases >=0.95. After two-reader correction + third-reader adjudication, inter-reader ICCs for cartilage volume were 0.959-0.995. On 20 synthetic thinning phantoms, MAPE was 5.73%, CCC 0.822, Dice 0.956. In the 69-participant longitudinal subset, total cartilage volume rose from 14,184.4 mm^3 (V0) to 15,359.3 mm^3 (V8), while 3D-RBA (surface area with thickness <1.5 mm) and 3D-PMA fell 4.70% and 6.88%. Applied to 416 participants: treatment arm +3.45% volume, +2.46% thickness, -4.54% 3D-RBA vs -2.08%, -1.32%, +0.16% in controls.

Key Findings:

Unified three-class nnU-Net (v2.0) reached Dice 0.964 +/- 0.030 on 1,189 trial MRIs, with 78.7% >=0.95, sufficient as pre-segmentation for downstream reader correction.
After adjudication, inter-reader ICCs for volume were 0.959-0.995, meeting reproducibility bars for a regulatory-grade phase III endpoint.
The pipeline resolved a treatment effect at V8 (+3.45% volume, +2.46% thickness, -4.54% 3D-RBA in treated vs -2.08%, -1.32%, +0.16% in controls), and 3D-RT / 3D-RBA agreed with 3D-PMA on longitudinal decline (4.70% vs 6.88%).

Technical Novelty: Two things are genuinely new: (1) a unified three-class nnU-Net (v2.0) replacing separate femorotibial and patellar models, simplifying deployment while retaining sub-mm accuracy; and (2) the 3D ray-based area (3D-RBA) endpoint measuring surface area with local thickness <1.5 mm as a proxy for degenerated cartilage, validated against 3D-PMA and geometric phantoms as a therapeutic-efficacy biomarker.

What's New: Prior AI knee-cartilage work has largely been methodology papers on public datasets (OAI); this is one of the first end-to-end deployments inside an active multicenter phase III drug trial, combining a unified segmentation model, a formal two-reader adjudication SOP, and a new degenerated-area endpoint (3D-RBA) validated on synthetic thinning phantoms.

Extension Opportunities:

Replace the two-reader + adjudicator loop with an active-learning uncertainty-guided correction UI that routes only low-Dice slices to humans, cutting reader-hours while preserving ICCs.
Add per-subregion (medial/lateral femur/tibia + patellar) longitudinal biomarker dashboards with mixed-effects trajectory modeling, so trials can detect regional treatment effects (medial tibial plateau is where KOA usually progresses) rather than only whole-joint means.
Port the 3D-RT thickness and 3D-RBA degenerated-area computations into an open-source library callable from any nnU-Net mask, enabling non-trial cohorts (OAI, MOST) to reproduce the endpoints and benchmark drug candidates.

Replicability: No code or data release is stated in the abstract; the trial images are proprietary phase III clinical data and unlikely to be shared. nnU-Net itself is open source, so the segmentation stage is reproducible on public knee MRI (OAI, SKI10), but the 3D-RT / 3D-RBA post-processing and the adjudication workflow would need re-implementation. Compute is modest: 3D full-res nnU-Net training fits on a single 24-48 GB GPU over 1-2 days; inference on ~1,200 volumes is hours on one GPU.

Research Gaps:

No head-to-head comparison against fully manual segmentation for treatment-effect sensitivity (i.e., does AI-assisted reduce or inflate the detected drug effect?).
The <1.5 mm threshold defining 'degenerated' cartilage in 3D-RBA appears heuristic; no sensitivity analysis or per-compartment threshold calibration is reported.

2. Predicting directional flexibility in proteins

Authors: Vsevolod Viliuga, Leif Seute, Matteo Tadiello... Published: 2026-09-08 | Citations: 0 arXiv | PDF

Research Question: How can we predict directional protein backbone flexibility and dynamic correlations quickly, without running expensive MD simulations or large generative ensemble models that require many samples for convergence?

Summary: BackFlip-2 is a fast SE(3)-equivariant GNN that predicts directional backbone flexibility and pairwise dynamic correlations directly from an equilibrium protein structure, bypassing the need to sample conformational ensembles. It reportedly matches the accuracy of much larger generative ensemble models while running orders of magnitude faster, with the equivariant design giving it a particular edge on anisotropic motions.

Key Results: BackFlip-2, an SE(3)-equivariant GNN, directly predicts dynamical descriptors (directional backbone flexibility, pairwise dynamic correlations) from a single equilibrium structure. The abstract claims it matches accuracy of substantially larger ensemble generation models while being orders of magnitude faster, and shows the equivariant architecture is especially suited to capturing anisotropic protein motions. Specific benchmark numbers are not disclosed in the abstract.

Key Findings:

Direct regression of dynamical descriptors matches ensemble-generation model accuracy at a fraction of the compute
SE(3)-equivariance is well-suited to capturing anisotropic, directional protein motions that scalar flexibility measures (e.g., RMSF) miss
A single equilibrium structure carries enough signal to predict useful dynamics-related quantities without explicit sampling

Technical Novelty: Rather than generating ensembles and post-hoc computing dynamical statistics, BackFlip-2 directly regresses the tensor-valued dynamical descriptors (directional flexibility, correlations) using SE(3)-equivariant message passing that natively represents the anisotropic, directional nature of these quantities.

What's New: Prior deep-learning work either emulates MD by generating full ensembles (expensive, requires many samples to converge) or predicts scalar flexibility. BackFlip-2 directly predicts the directional/anisotropic descriptors themselves using equivariance, sidestepping ensemble generation entirely.

Extension Opportunities:

Couple BackFlip-2 predictions with a lightweight sampler to generate physically consistent conformational ensembles conditioned on the predicted anisotropic covariance
Extend the model to predict ligand- or mutation-induced changes in directional flexibility for allostery and drug design workflows
Integrate the predicted per-residue anisotropy as a prior into structure prediction (AlphaFold-style) or protein design pipelines to bias toward dynamically plausible backbones

Replicability: Code and model weights are open-sourced at https://github.com/graeter-group/backflip. Because the model is a single-pass GNN on an equilibrium structure, inference is cheap (likely single-GPU or CPU); training compute is not specified in the abstract but is presumably far below MD or diffusion-based ensemble baselines.

Research Gaps:

No explicit conformational states are produced, so downstream tasks needing actual structures (docking against alternative conformers) still require a separate sampler
Generalization to intrinsically disordered regions, large-scale domain motions, and non-equilibrium/ligand-bound states is not addressed in the abstract

3. Multi-ligand simultaneous docking of Carica papaya leaf phytochemicals, Carpaine and Rutin, reveals multi-mechanism inhibition of cancer proteins BCL-2 and WWP1

Authors: Merla Sudha, Asmita Saha, Belaguppa Manjunath Ashwin Desai... Published: 2026-09-08 | Citations: 0 arXiv | PDF

Research Question: How can phytochemicals from Carica papaya leaves (Carpaine and Rutin) be computationally evaluated for multi-mechanism inhibition of cancer proteins BCL-2 and WWP1, addressing chemotherapy resistance and toxicity from high-dose single-agent treatments?

Summary: The paper uses multi-ligand simultaneous docking, MD simulations, and MMPBSA analysis to show that two Carica papaya phytochemicals, Carpaine and Rutin, engage BCL-2 additively and WWP1 allosterically, with binding affinities that exceed the commercial inhibitor Venetoclax. It further demonstrates a synergistic interaction between Rutin and the proteasome inhibitor Bortezomib against WWP1, positioning these combinations as promising leads for lower-toxicity combination cancer therapy.

Key Results: Using molecular docking, multi-ligand simultaneous docking (MLSD), MD simulations, and MMPBSA analysis, the authors showed: (1) additive effect on BCL-2 with MLSD binding affinity of -13.13 ± 0.08 kcal/mol (outperforming individual ligands and the commercial inhibitor Venetoclax); (2) allosteric effect on WWP1 with Car bound near H-site and Rut near Le-site, boosting Car's MLSD affinity to -15.59 ± 0.39 kcal/mol; (3) synergistic Rut+Bortezomib interaction with WWP1 — Bort alone -7.64 ± 0.156, Rut alone -10.26 ± 0.07, MLSD -15.59 ± 0.39 kcal/mol.

Key Findings:

Carpaine + Rutin bind distinct sub-regions of the same BCL-2 pocket, yielding an additive MLSD binding affinity of -13.13 ± 0.08 kcal/mol that beats Venetoclax.
For WWP1, Rutin binding at the Le-site allosterically enhances Carpaine binding near the H-site, driving Carpaine's MLSD affinity to -15.59 ± 0.39 kcal/mol.
Rutin + Bortezomib show synergistic binding to WWP1 (-15.59 ± 0.39 kcal/mol MLSD vs -10.26 alone and -7.64 for Bort alone), suggesting a rationale for dose-sparing combination therapy.

Technical Novelty: Applying multi-ligand simultaneous docking (MLSD) — rather than sequential single-ligand docking — to quantify additive, allosteric, and synergistic binding modes of two natural phytochemicals (and one with an FDA-approved drug) against two distinct cancer targets, providing a mechanistic categorization framework rather than a single affinity score.

What's New: Prior computational studies of papaya phytochemicals typically report single-ligand docking scores. This work uses MLSD plus MD/MMPBSA to explicitly classify the interaction as additive, allosteric, or synergistic across two cancer proteins, and pairs a natural product with an FDA-approved drug (Bortezomib) to model realistic combination therapy at the atomistic level.

Extension Opportunities:

Experimental (wet-lab) validation via in vitro cytotoxicity assays on BCL-2/WWP1-overexpressing cancer cell lines and SPR/ITC binding measurements to confirm the predicted additive/allosteric/synergistic effects.
Extend the MLSD pipeline into an automated screening platform that combines phytochemical libraries with FDA-approved oncology drugs to systematically discover synergistic combinations across other cancer targets.
Investigate ADMET properties, oral bioavailability, and pharmacokinetics of Carpaine+Rutin+Bortezomib combinations, plus formulation strategies to overcome Rutin's known low bioavailability.

Replicability: No code, structures, or trajectory data availability mentioned in the abstract. Reproduction requires standard computational chemistry stack (AutoDock/Vina or similar for MLSD, GROMACS/AMBER for MD, MMPBSA tooling) and moderate HPC/GPU resources — likely a few thousand CPU-hours for MD replicates plus MMPBSA rescoring. Protein structures (BCL-2, WWP1) and ligand SMILES for Carpaine/Rutin/Venetoclax/Bortezomib are publicly available.

Research Gaps:

No experimental (cell-based or biophysical) confirmation of the computationally predicted binding modes or synergy.
ADMET, bioavailability, off-target selectivity, and in vivo efficacy are not addressed — critical gaps before any translational claim.

🔬 MATERIALS

1. High-throughput thermodynamic screening of oxide-scale adhesion across the CoCrFeMnNiAl high-entropy alloys

Authors: Dennis Boakye, Chuang Deng Published: 2026-09-10 | Citations: 0 arXiv | PDF

Research Question: How can we systematically select optimal reactive element (RE) dopants in high-entropy alloys (HEAs) to improve oxide-scale adhesion during high-temperature oxidation, given that first-principles methods cannot access the required compositional resolution across the vast HEA design space?

Summary: The paper presents a high-throughput thermodynamic screening framework combining the macroscopic atom model with McLean and Guttmann segregation models to rank reactive element dopants across CoCrFeMnNiAl high-entropy alloy sub-families for oxide-scale adhesion. It reveals an oxide-dependent RE ranking inversion (Hf best for Cr2O3, La best for Al2O3) and identifies Co16Cr16Fe16Ni16Al35 as an inverse-designed sulfur-immune composition with W_sep = 5.95 J/m² without any RE doping.

Key Results: The authors applied the macroscopic atom model coupled with McLean isotherm and Guttmann models to screen 9 CoCrFeMnNiAl sub-families with 5 REs (Hf, Y, Zr, La, Ti) at Cr2O3 and Al2O3 interfaces. Key results: (1) oxide-dependent ranking inversion — Hf dominates at Cr2O3, La dominates at Al2O3; (2) Mn-containing alloys show intrinsic sulfur resistance; (3) sulfur immunity phase diagram identifies Al+Mn ≳ 25 at% as thermodynamically S-immune; (4) crossover concentrations collapse onto a universal exponential governed by segregation enthalpy difference; (5) inverse design identifies Co16Cr16Fe16Ni16Al35 as optimal S-immune composition with work of separation W_sep = 5.95 J/m².

Key Findings:

Oxide-dependent RE ranking inversion: Hf dominates at Cr2O3 interfaces while La dominates at Al2O3 interfaces, driven by RE-O vs RE-matrix enthalpy interplay
Mn-containing HEAs exhibit intrinsic sulfur resistance consistent with experimental oxidation behavior; compositions with Al+Mn ≳ 25 at% are thermodynamically immune to S-induced adhesion loss
All RE-oxide crossover concentrations collapse onto a universal exponential governed by segregation enthalpy difference, providing a transferable design rule; Co16Cr16Fe16Ni16Al35 identified as optimal S-immune composition (W_sep = 5.95 J/m²) requiring no RE doping

Technical Novelty: First application of the macroscopic atom model at the compositional resolution of HEA sub-families, combined with McLean/Guttmann segregation models to screen RE segregation, adhesion enhancement, and sulfur displacement simultaneously across a 9-family × 5-RE × 2-oxide design matrix. The discovery of oxide-dependent RE ranking inversion and a universal exponential design rule for crossover concentrations is new.

What's New: Unlike prior empirical RE selection or expensive DFT studies limited to specific compositions, this work provides the first high-throughput thermodynamic screen across the full CoCrFeMnNiAl compositional space, uncovers the oxide-dependent RE ranking inversion, and derives a universal exponential design rule from the segregation enthalpy difference.

Extension Opportunities:

Build an open-source high-throughput screening pipeline coupling the macroscopic atom model with DFT validation for other refractory HEA families (e.g., MoNbTaW) beyond CoCrFeMnNiAl
Experimentally validate the Al+Mn ≳ 25 at% sulfur immunity threshold through controlled oxidation experiments and compare with the predicted W_sep values
Extend the framework to include kinetic factors (diffusion, oxide growth rates) alongside the thermodynamic ranking to predict actual service lifetimes under cyclic oxidation

Replicability: No mention of code/data availability in the abstract. Reproduction requires implementing the macroscopic atom model, McLean isotherm, and Guttmann segregation models — all analytical/semi-empirical, so compute needs are modest (desktop-scale). Tabulated interaction enthalpies (Miedema parameters) are publicly available.

Research Gaps:

Kinetics of RE segregation, oxide growth, and diffusion are not addressed — only equilibrium thermodynamics
Predictions require experimental validation, particularly the sulfur immunity threshold and the ranking inversion between Cr2O3 and Al2O3 scales

🔥 GitHub Trending

1. FareedKhan-dev/kimi-k3-in-c

⭐ 7710 stars | C

A 2.78-trillion-parameter Kimi K3 running inference on a single CPU in 8.24 GB of RAM. Portable C99: no BLAS, no framework, no GPU.

avx2 c99 cpu-inference deep-learning from-scratch inference-engine

2. deeplethe/utopia

⭐ 7152 stars | Rust

World's first open-source enterprise world model.

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

3. wanshuiyin/Anti-Autoresearch

⭐ 153 stars | Python

Don't trust an autoresearch paper at face value. Reviewer-side integrity forensics (self-consistency + fabrication), deterministic verdict. 61 signals: 46 integrity hack-patterns (families A–H, verdic

ai-generated-content ai-research ai-scientist aris autoresearch claude

4. yanght27/GPU-Perf-Playground

⭐ 124 stars | Python

GPU 性能与 AI Infra 学习项目：CUDA/Triton 算子、NCU/NSYS、vLLM/SGLang/TRT-LLM/ms-swift、PyTorch/DeepSpeed/ms-swift 训练、并行架构

ai cuda cuda-kernel cuda-kernels cuda-programming cuda-toolkit

5. endend2003-cmd/Tactical-Matrix-Console

⭐ 118 stars | HTML

WarMatrix 2026: Next-Gen Tactical Simulation & AI Command Console

ai-simulation command-and-control defence-technology fastapi glassmorphism local-llm

6. okesipoke/manuscript-phoneme-decipher

⭐ 117 stars | HTML

Voynich Manuscript Decoded: Elu-Sinhala Phonetic Transcription & Vocabulary Toolkit 2026

abugida algorithms anthropic artificial-intelligence ayurveda beinecke-ms408

7. mattn/tensai

⭐ 103 stars | Go

A tiny neural-network framework in pure Go with AVX2 SIMD kernels (GOEXPERIMENT=simd)

autograd avx2 deep-learning gguf golang llm

8. Tylogi/TyloQuant

⭐ 74 stars | C++

Get more intelligence from every bit. Better quantization formats and smarter calibration let larger, stronger models run smoothly on the hardware you already own.

apple-silicon cuda deepseek gemma gguf llm

9. ZekunCheng/novoweave

⭐ 65 stars | Python

NovoWeave — a conceptual generative protein-design framework (non-functional pseudocode).

deep-learning diffusion-models protein-design pseudocode research-software structural-biology

10. lucidrains/light-loco-parkour

⭐ 60 stars | Python

Implementation of Light Loco Parkour, Versatile Perceptive Whole-Body Locomotion via Multi-Skill Distillation, out of Light Origins Labs

artificial-intelligence deep-learning locomotion robotics

11. jshn9515/deep-learning-notes

⭐ 59 stars | Python

Personal deep learning study notes and tutorial-style notebooks

deep-learning deep-learning-tutorial jupyter-notebooks pytorch quarto tutorial

Generated by Research Pulse on 2026-09-12 06:05