In briefPLE: A benchmark testing whether AI agents can complete real professional physical engineering design work
SummaryThe paper introduces PLE, a benchmark of 85 long-horizon agentic tasks spanning electrical, civil, structural and mechanical engineering, where success means producing a realistic, billable engineering deliverable. Tasks are built either end to end by practising engineers or from engineer-written problem statements expanded through an adversarial Creator-Solver pipeline and validated by experts. Evaluating frontier LLMs under seven public agent harnesses, the strongest configuration averages only 50.6%, using about 27M tokens and 63 minutes per task.
In briefIFC-MCQ: Auto-generating multiple-choice tests to evaluate LLM agents retrieving building data from IFC files
SummaryIFC is the standard format for exchanging building information in architecture, engineering, and construction. The paper proposes IFC-MCQ, which automatically builds multiple-choice benchmarks from IFC files so answers can be graded exactly without costly review. The authors tested generation coverage on IFC models of varying size and characteristics, and evaluated two agent setups using tool-based and code-based retrieval. Results revealed category-specific failure modes and performance that varies with the source model and retrieval interface.
In briefLane-change prediction for safer, smoother autonomous driving decisions
SummaryThis study examines whether predicting surrounding vehicles' lane changes, using knowledge graph embeddings and Bayesian inference, improves autonomous vehicle decision-making. Three strategies were tested in risky merging scenarios simulated in CARLA: reactive driving, prediction-based longitudinal action, and prediction-based lateral action. The model anticipates lane changes about 3-4 seconds in advance, and both predictive strategies avoided collisions in the evaluated records. Ride comfort was also estimated using an ISO 2631-1-based vibration measure.
Mohamed Manzour, Aditya Kumar, Rub\'en Izquierdo et al.
In briefLLM2Cert: LLM-generated stability and safety certificates checked with SMT and Lean
SummaryThis paper addresses synthesizing and verifying Lyapunov functions and barrier certificates for nonlinear, high-dimensional systems. It introduces DynaCertBench, a benchmark of 105 instances, and LLM2Cert, a framework where an LLM proposes candidate certificates that are checked by an SMT solver or a Lean proof-repair agent. The authors report that LLM2Cert verifies 81 of the 105 instances, improving on one-shot proposals without feedback, and verifies 22 instances that an All-SMT baseline leaves unresolved.
Zhipeng Cao, Peiyang Song, Nikolaus Vertovec et al.
In briefNEST: Nested Embeddings for Discovering Clusters at Multiple Scales
SummaryThis paper introduces NEST, an unsupervised method inspired by Matryoshka Representation Learning that encodes coarse-to-fine cluster structure within a single embedding. It jointly optimizes clustering and separation across nested prefix subspaces, with regularization to prevent the embedding from collapsing. On synthetic hierarchies, NEST recovers multiscale structure given only the number of clusters per level, and it also shows multiscale organization on vision, text, and single-cell data.
In briefJoint Forecasting of Player Trajectories and 3D Poses in Team Sports
SummaryThis paper addresses predicting both global trajectories and 3D body poses of agents, which are tightly coupled in team sports with interacting players. The authors propose a socio-temporal encoder that fuses trajectory and pose inputs and models them jointly via shared factorized attention, with per-joint loss reweighting and frequency-domain trajectory smoothing. They also correct evaluation biases from overlapping sequences by adopting a non-overlapping protocol. The method improves over prior joint forecasting approaches on WorldPose and 3DPW, with the largest gains in pose accuracy.
In briefProtegrity-PII Technical Report: A model for detecting personal identifiable information in text
SummaryThis report presents Protegrity-PII, a model that identifies personally identifiable information (PII) in both structured and unstructured text. On a new benchmark called PII-SCOPE, the authors report a typed recall of 0.915 and a precision of 0.907. The paper also gives processing speed and cost figures and states that the model generalizes better out of distribution than several compared state-of-the-art models and datasets.
In briefBeyond Overall Scores: How Changing the LLM Judge Affects Individual Verdicts
SummaryThe paper examines how stable large language model judges are when comparing two answers. The authors fixed two models' answers for 500 Arena-Hard v2.0 prompts and ran GPT-4.1 as the judge under eight combinations of criteria, output format and task framing, repeating each setup at temperature zero. Rerunning the default setup changed the win/tie/loss outcome on 21.2% of prompts, even though the aggregate score moved only about one point.
In briefRethinking How LLM Fact-Checkers Are Evaluated: A Japanese Benchmark
SummaryThis paper argues that scores for LLM fact-checking depend on three design choices (which questions the claims answer, the judging conditions, and the evidence sources), which existing benchmarks leave implicit. The authors introduce JSFactCheckBench, a Japanese benchmark of 1,913 claims from LLM answers about misleading information circulated on Japanese social media. Testing six LLMs, they find that claims directly answering the question are verified less accurately than supplementary claims, a gap hidden by aggregate scores.
In briefPhysibleMDM: Enforcing Physical Feasibility on Motion Diffusion Models via Inference-Time Noise Optimization
SummaryThis paper presents PhysibleMDM, a framework that improves the physical plausibility of motion generated by diffusion models by optimizing the diffusion noise at inference time, without retraining. It evaluates surrogate constraints combining kinematic checks such as ground penetration, slip, and self-collision with inverse-dynamics tests formulated as a cone program, and solves the black-box problem with Sep-CMA-ES. On an SMPL humanoid across 100 motion styles, mean ground penetration, self-collision penetration, and ground slip fell by 75%, 92%, and 74% relative to unoptimized output.
Maximilian St\"olzle, David M\"uller, Agon Serifi et al.
In briefCan LLMs Improve Their Own Harness? A Controlled ARC-AGI-3 Study
SummaryThis paper asks whether frontier language models can build and improve their own harness, the orchestration program that manages model calls, perception, memory and tools, for the interactive ARC-AGI-3 benchmark. Each model revises its full codebase twice after acting on five development games, then is evaluated on twenty held-out games using the RHAE metric. Across 27 lineages performance ranges widely, and the two revisions raise mean held-out RHAE by only 2.80 points, showing development gains overstate generalization.
In briefFrom Prose to Properties: A Property-Based Testing Library for Checking Research Claims
SummaryThe paper presents ARATest, a property-based testing library that turns empirical research claims into executable tests by abstracting each claim's property skeleton. Applied to 54 of 243 claims across 33 agent-native research artifacts, it returned 32 holds, 8 violations, and 14 inconclusive verdicts. Two violations exposed errors in published papers that the authors confirmed.
WhyProposes turning paper claims into executable property tests, validated on 243 claims across 33 artifacts, exposing confirmed errors in published papers.
In briefOptimal Multitask Representations Under a Rate Constraint: An Information Bottleneck Analysis
SummaryThis paper studies which pretraining objective gives the best representation for downstream tasks under an information budget, formalized as a multitask information bottleneck with a mutual information constraint. For Gaussian inputs and tasks, it characterizes optimal representations and shows a weakly task-aligned specialist can beat a reconstruction generalist at low rates but lose at high rates. Experiments on CelebA and dSprites show that compression can change which pretraining objective performs best.
In briefAuditing Behavioral Detectors in LLM Agent Evaluation Traces
SummaryThe paper examines how LLM agent execution traces are reduced to labels such as final correctness, verification, and fault propagation, and whether those labels faithfully reflect the traces. Using a GSM8K calculator setting with corrupted tool outputs, the authors audit three trace-level detectors across six models and find systematic disagreements, including silent overrides of corrupted values that verification detectors do not record. They argue that automated labels should be accompanied by trace evidence, detector applicability, execution-limit information, and human corrections.
Idunnuoluwa A. Adeniji, Shilpika, Michael E. Papka
In briefFrom Single-Actor to Multi-Actor Event Worlds: Can LLMs Model Events?
SummaryThis paper formalizes systems with multiple decision-making actors, evolution without agent action, and partial observation through event logs as "event worlds," and asks whether large language models can predict future observations from the log and the agent's proposed actions. The authors build benchmarks from human Diplomacy games and real Polymarket trades and test prediction, multi-step rollouts, and agent usefulness. They find that fine-tuning and more data help only when all actors' actions are recorded, while with only the agent's action, models struggle when other actors strongly influence outcomes and tend to overestimate plan success.
In briefBenchmark testing whether personal AI agents remember how users want to be helped
SummaryThis paper introduces MemPABench, an interactive environment for testing whether personal agents remember user interaction preferences, such as when to clarify, confirm, or act autonomously. Four simulated users hold 14 preferences, giving feedback across 99 sessions over four months with five preference changes per user. Evaluating four memory systems on four LLM backbones, the authors find current memory brings limited gains, struggles to retrieve preferences, and adapts poorly to changes.
In briefAVAtt: A Free-Viewing Eye-Tracking Dataset for Visual Attention on Paintings
SummaryThis paper introduces AVAtt, an eye-tracking dataset of free viewing on paintings from sixteen art movements across five centuries. It releases scanpaths, fixation and saliency maps, and art metadata. The authors benchmark saliency models against human viewing and rank 44 scanpath-similarity measures.
Mohamed Amine Kerkouri, Marouane Tliba, Aladine Chetouani et al.
In briefFIGFLEX: A Mixed-Precision FP-INT Matrix Multiply Accelerator Co-Designed with Low-Bit LLM Quantization
SummaryThis paper presents FIGFLEX, an accelerator co-designed with algorithms for weight-only quantized large language models, so that computation speeds up as weight precision drops. It reformulates floating-point-integer multiplication using radix-4 shift-add operations and includes an INT3 optimization. On LLaMA-2-7B, it reports WikiText-2 perplexity of 5.85 at 3-bit and 8.77 at 2-bit average precision, with higher per-area throughput than AxCore and FIGLUT.
Faraz Tahmasebi, Haocheng Xu, Andrew Trauger et al.
In briefExternalising Memory: Where an Embodied Agent's Experience Should Live
SummaryThis pre-registered study asks whether an embodied agent's experience should be stored in policy weights or kept external and retrieved at inference. Student LoRA policies were distilled from a memory-augmented teacher's trajectories and evaluated under matched protocols. Distilling memory into weights gave only about +1 point over an action-only control, not significant, and the authors withdrew rather than refuted that claim because the test was underpowered. Attaching the frozen retrieval buffer to fixed student weights gave a significant gain.
In briefRethinking SAE Evaluation: A Framework Revealing Hidden Polysemanticity and Redundancy
SummarySparse autoencoders (SAEs) were originally designed to find monosemantic features, but no clear objective existed for the feature set as a whole. This paper proposes that SAE features should form an atomic interpretable unit basis, with metrics such as DAME serving as proxies. Evaluating hundreds of SAEs across architectures and model families, the authors find that some SAEs look monosemantic under existing metrics yet remain functionally polysemantic, with the most dominant features being less interpretable and more redundant.
In briefDiffLTM: continuous diffusion large tabular model for conditional data generation
SummaryThis paper presents DiffLTM, a continuous diffusion model for conditional generation in Large Tabular Models. It addresses mixed numerical and high-cardinality, unbalanced categorical data, as well as difficulty generating rare subpopulations. The authors use cross-entropy loss for categorical variables, providing a first theoretical justification, and introduce metrics focused on rare subpopulations. Trained on labor data, the model is shown to capture classical econometric insights.
Alankrita Bhatt, Mukur Gupta, Andrea Montanari et al.
In briefLearning to Incentivise: Reinforcement Learning for Greener, Faster Urban Traffic
SummaryThis paper proposes a multi-agent reinforcement learning method that dynamically assigns incentives to drivers to reroute onto alternative, possibly longer, paths. Using traffic simulation to estimate travel times, the approach aims to reduce total travel time and emissions as an alternative to road pricing. With an unlimited budget, it cuts total travel time by 16% versus dynamic User-Equilibrium, and reduces CO2 emissions by 9% when emissions are the goal. The authors note a trade-off between the two objectives, which a suitable multi-objective weighting can balance to improve both.
Germ\'an Pardo-Gonz\'alez, Shaghayegh Vosough, Katerina Papadaki et al.
In briefSplatLab: An autonomous system that runs 3D Gaussian Splatting research end to end
SummaryThis paper presents SplatLab, an autonomous research system for 3D Gaussian Splatting (3DGS). Given a prompt, a target-paper URL, or no input, it reproduces a published baseline, proposes an extension, trains and ablates on Mip-NeRF 360 and LLFF, and writes the manuscript, binding each reported number to a logged training run. The authors report manuscripts beating their reproduced targets by +0.18 to +3.78 dB PSNR, and say the research cycle shrinks from months to days.
In briefFlow-based spectral kernel attention: a framework that learns flexible token-similarity kernels for attention with fewer projection parameters
SummaryThe paper proposes FSKA, a framework that learns nonstationary attention kernels by modeling a bivariate spectral density with normalizing flows, and it remains compatible with linear-attention computation. The authors show that this expressive kernel learning allows a simplified design that removes query-key projections and uses fixed orthogonal value projections. According to the paper, controlled experiments on classification, ablation, and scalability profiling show competitive or improved performance over existing attention mechanisms, with fewer projection-related trainable parameters than most baselines and favorable linear-attention memory scaling.
In briefShort-to-Long: Detached Features for Stable Long-Horizon Value Learning in Reinforcement Learning
SummaryThis paper addresses learning representations from changing value targets in deep reinforcement learning. The authors propose Short-to-Long (S2L), which feeds features from source networks trained with smaller discounts into a long-horizon behavior learner, with detached connections so the behavior loss does not update the sources. Experiments on Atari and online and offline continuous control report aggregate gains, including improved paired terminal median on 40 of 57 Atari games.
In briefRubricCraft: Building Evaluation Criteria from Where Solutions Disagree
SummaryThis paper introduces RubricCraft, a method for evaluating language agents on open-ended tasks that produce artifacts. It compares sampled candidate solutions against task requirements and uses the differences to generate natural-language rubric criteria, then revises checks that miss errors or penalize valid alternatives. On SpreadsheetBench v2, RubricCraft rubrics passed fewer wrong outputs (29.5% versus 42.6%) than a prompt-only rubric from the same model, and ranked outputs by correctness more often.
Hosein Hasanbeig, Priyanshu Gupta, Ashish Tiwari et al.
In briefaggmetrics: A Python Package for Aggregating Machine Learning Classification Metrics
SummaryThis paper presents aggmetrics, an open-source Python package that extends binary performance measures based on four confusion-matrix counts to multiclass, multilabel, ordinal, cost-sensitive, soft-label, and cross-validation settings. It replaces the usual average switch with one of four aggregation operators. The package records each measure's structural properties and can warn users when a measure is applied where it does not behave as commonly assumed. It is BSD-3 licensed and depends only on NumPy.
In briefSame Beliefs, Different Policies: Reusable Representations for Early Stopping in LLM Reasoning
SummaryReasoning LLMs often keep reasoning after reaching a correct answer, wasting computation and sometimes switching to a wrong one. The paper introduces BELT, which separates outcome prediction from stopping control by learning beliefs about current and future answers from a frozen LLM's hidden states, then fitting a linear stopping policy on those features. Averaged over five out-of-distribution math benchmarks, BELT cuts generated tokens by 41% on Qwen3-1.7B and 51% on OLMo-3-7B-Think without accuracy loss. It also transfers to coding and agentic tasks without training on that data.
Jingwei Ni, Dmitry Abulkhanov, Vladislav Smirnov et al.
In briefA Governance Framework for Safer, More Trustworthy AI Agents
SummaryThis paper proposes a Three-Pillar (3P) framework for developing safe and trustworthy AI agents through transparency, accountability, and trustworthiness. It describes an evolutionary path that raises agent autonomy step by step, using progressive validation, structured human oversight, and auditable operational controls. The authors also explain how the model supports lifecycle monitoring, decision traceability, escalation management, and human-agent collaboration in enterprise settings.
In briefStability of signed graph learning under edge additions, deletions and sign flips
SummaryThis paper studies how stable signed graph filters are when edges are added, removed, or have their signs flipped, a case existing unsigned-graph analyses do not cover. The theory shows how signed edge perturbations and node feature correlations jointly determine how stable the learned representations are. Based on this, the authors propose a theory-guided strategy that selects structurally influential signed-edge changes. Experiments on synthetic and real signed graphs report that it induces larger embedding drift, polarity change and link prediction degradation than baselines.
In briefSpecMAGIC: A Method for Generalizing Across Tasks in Multi-Agent Reinforcement Learning Using Specifications
SummaryThis paper addresses multi-task generalization in multi-agent reinforcement learning, where agents must handle complex, time-extended tasks requiring coordination. It introduces a specification language that combines temporal-logic task descriptions with operators for inter-agent requirements, and proposes SpecMAGIC, a hierarchical framework that decomposes tasks into subtasks using the specification's automaton structure. A centralized planner allocates subtask sequences to agents while decentralized policies execute them, enabling zero-shot generalization to new tasks. Experiments on multi-agent navigation tasks report stronger performance than existing approaches on unseen specifications.
Zijian Guo, \.Ilker I\c{s}{\i}k, H. M. Sabbir Ahmad et al.
In briefFaithful Data Selection: Limits and Scalable Methods for Choosing Training Subsets
SummaryThe paper shows that choosing a training subset to minimize population error is information-theoretically impossible unless the selector has a validation set much larger than the training set. The authors instead propose a one-sided "faithfulness" guarantee, an estimable upper bound on the gap between training on the subset and on the full data, derived from gradient discrepancy under Unadjusted Langevin training. They introduce a swap-based local search that decreases this bound at each accepted step, and report experiments on CIFAR10 and CIFAR100 showing performance comparable to existing baselines.
Ahmed Mehdi Inane, Mahdi Haghifam, Gintare Karolina Dziugaite et al.
In briefSecure Agentic AI and Multimodal Intelligence for 6G Networks
SummaryThis survey examines multimodal intelligence and agentic AI for cross-layered 6G communication systems, focusing on privacy and security. It presents a system view covering perception, memory, tool use, and multi-agent orchestration, and analyzes threat models and attack surfaces across sensing, cognition, action, memory, and communication layers. The paper organizes existing attacks into a structured taxonomy and reviews mitigations such as prompt and memory defenses, inter-agent trust, secure orchestration, and privacy-preserving learning. It concludes by outlining open research directions for trustworthy multimodal multi-agent 6G systems.
In briefSelf-Assessment Reinforcement Learning for Search-State Awareness in Retrieval-Augmented LLMs
SummaryThis paper proposes Self-Assessment Reinforcement Learning (SARL), a framework where models learn during training to recognize what information they have gathered and what is still missing, improving their search behavior without extra assessment at test time. It uses a dual reward design that rewards the corrective effect of assessments while penalizing excessive reliance on them. The authors report relative gains over the strongest baselines on multi-hop reasoning and deep research benchmarks, and training stability under extended search horizons.
In briefAnalogPortBench: Testing how well LLMs port analog circuits to new process nodes
SummaryAnalog circuit design relies heavily on expert knowledge, and moving proven designs to new process nodes is a recurring challenge. This paper introduces AnalogPortBench, a benchmark of 54 tasks that port ten circuits from a 0.18 μm process to the open-source SKY130 PDK, ranging from direct ports to infeasible specifications. Six frontier LLMs were evaluated, finding they can port simple circuits but rarely redesign headroom-sensitive amplifiers or recognize infeasible specifications. The authors report that simulation feedback helps models more than adding samples or reasoning effort.
In briefThe Path to NDE 4.0: A review of algorithms and techniques for automated non-destructive inspection
SummaryThis review surveys the non-destructive inspection field through four lenses: computational algorithms, physical techniques, industrial applications, and research objectives. It traces the shift from manual, analog interpretation to a digital, data-centric approach called NDE 4.0. The authors note that ultrasound and visual inspection dominate, while deep learning and generative models are increasingly used for automated defect recognition. The paper identifies bottlenecks such as scarce standardized open datasets and opaque AI models, and suggests research directions including explainable AI and standardized digital twin protocols.
Ashiqur Rahmana, Venkata Devesh Reddy Seethi, Zachary Kral et al.
In briefRoleFlow: Self-evolving multi-agent teamwork that adapts its own structure
SummaryThe paper presents RoleFlow, a multi-agent LLM system that refines its collaboration structure through feedback-driven atomic edits and on-demand role creation. It introduces Single-Edit Counterfactual Policy Optimization (SECPO), which compares each edit against a no-edit version at the same state to assign targeted credit. The authors report the best average performance across seven benchmarks in four domains, outperforming the strongest competitor by 4.7 points.
In briefMemory-augmented online test-time adaptation for EEG motor imagery decoding with perturbation consistency and confidence calibration
SummaryThis paper proposes MOPC, a source-free online test-time adaptation framework for EEG-based motor imagery decoding, addressing cross-subject distribution shifts and signal nonstationarity. It uses random-mask perturbations to enforce prediction consistency, a confidence-calibrated reweighting loss to handle unreliable pseudolabels and class imbalance, and a lightweight memory module to prevent forgetting. The authors report state-of-the-art results on the BCI IV IIb and OpenBMI datasets, with per-trial adaptation latency of roughly 40–60 ms.
In briefLearning data weights from noisy, diverse sources to improve model training
SummaryThis paper proposes a scalable method for assigning weights to training examples, controlling how heterogeneous data affects model training. Using metagradients computed through the training process, it optimizes example weights directly for performance on a small, high-quality target dataset, and learns parametric weighting rules that generalize to new examples and larger models. Applied to reward modeling, the learned weights improve held-out performance and allow mixtures of human and synthetic preference data to outperform either source alone.
Hedi Driss, Evan Frick, Anastasios N. Angelopoulos et al.
In brief"X-pressing Intimacy": How mental health posts on Twitter changed after Musk's acquisition
SummaryThis study analyzes about 824K tweets on anxiety, depression, bipolar disorder, borderline personality disorder, and suicidal ideation, along with 266K comparison tweets, around Elon Musk's October 2022 acquisition of Twitter. Using LIWC and an LLM-based intimacy classifier, it examines shifts in language and disclosure intimacy. After the acquisition, affective and social language declined while first-person singular pronouns and future-focused expressions rose. Relative to broader discourse, only the decline in moderate- versus low-intimacy disclosure for borderline personality disorder was statistically significant.</summary>
In briefSIM-UT: A User-Simulation Framework for Testing Medical Triage Chatbots
SummaryThis paper presents SIM-UT, a multi-turn evaluation method that uses simulated users to find failure modes in LLM-based medical triage chatbots. The authors tested three real WhatsApp health chatbots in India, South Africa, and Kenya, comparing MedGemma with ChatGPT and Claude over 5,000+ multilingual simulated turns. Both model families showed de-escalation risk: after correctly flagging an emergency, they agreed with users who minimised symptoms, such as calling vaginal bleeding in pregnancy 'not a problem'. The authors recommend simulation-based behavioural evaluation as standard practice before deployment.
Deepali Mittal, Dheeraj Mundhra, Yulia Sidorova et al.
In briefAcoustiPrint: Neural design of acoustic structures that can actually be manufactured
SummaryThe paper introduces AcoustiPrint, a physics-informed neural architecture for constrained inverse design that represents three-dimensional geometries with an implicit neural representation. Each candidate passes through a printability operator covering self-support, discrete material selection, and connectivity before its acoustic response is evaluated. It is demonstrated on personalized transcranial focused ultrasound, where experiments report enhanced focal accuracy over density-based and phase-based baselines.
Shanmuka Sadhu, Harshvardhan C. Takawale, Yehia Zakaria et al.
In briefVisCurate: Judging Whether Visual Skills Behave the Same by Running Them
SummaryThis paper presents VisCurate, a framework and benchmark that curates libraries of image-to-image visual skills by executing candidate skill pairs on a shared set of probe inputs rather than relying on names or code. A 300-instance benchmark with seven injected defect types is built with known relations and repairs. On hard negatives the verifier accepts no false merges while recovering 51.2% of truly mergeable pairs, and on two real libraries it confirms only 10 of 17 published equivalence claims.
In briefCross-species ion channel classification using persistent homology in multimodal graph networks
SummaryThe paper shows that random train/test splits leak phylogenetic information, inflating performance estimates for venom compound classification by ion-channel target. Using leave-one-taxon-out evaluation, it compares five structural inductive biases within a shared ESM-2 and graph attention backbone. The persistent-homology model (PH-GNN) reaches 63.4% accuracy under taxonomic shift, over 8 points above the next-best method.
Arezoo Bybordi, Favour Achimba, Mariam Gelashvili et al.
In briefSoft ontology priors for weakly supervised neuro-symbolic concept learning
SummaryNeuro-symbolic models trained only on task labels can fall into reasoning shortcuts, where different concept mappings fit equally well and the learned concepts cannot be identified or safely reused. This paper adds a soft ontology prior over latent concepts to the training objective, compiling it with the task constraint into a smooth, decomposable probabilistic circuit. The authors validate their predictions on an audited six-concept view of the BDD-OIA driving benchmark and on controlled neuro-symbolic tasks.
In briefLatent Space Jam: Training-Free Video Generation Controlled by Rough Layouts
SummaryThis paper proposes a training-free method that uses pretrained video diffusion models to control object placement and motion with coarse layouts. It works in two passes: first it generates a coarse visual guide through prompt-guided cross-attention, then uses that guide to produce a high-quality, layout-consistent video. The authors report that the method outperforms prior work in overlap and trajectory accuracy as well as visual quality across static, dynamic, and multi-object scenes.
In briefTransformer-based real-time path planning for many agents with changing start and goal states
SummaryThis paper addresses real-time path planning for multiple agents with prescribed initial and final configurations, formulated as a high-dimensional optimal control problem with collision and obstacle constraints. The authors propose OCFormer, a transformer-based network that learns the inverse of a coordinate transformation from a simple latent space to the coupled physical space, trained with a physics-informed loss based on the Hamiltonian ODE residual and boundary matching. Experiments on two- and three-dimensional navigation with up to 128 agents demonstrate scalability and efficiency.
In briefText-only LLM correctors break correct transcripts: repair-damage analysis of audio grounding in ASR correction
SummaryThis paper studies how large language models correct speech recognition (ASR) output. The authors split each word error rate (WER) change into repaired, damaged and inserted tokens, and compare correction with and without the source audio across three corpora. With Gemini 2.5 Flash, text-only correction damages many transcripts that were already correct, and adding the audio greatly reduces this. On medical dictation, retained insertions can cancel the net WER gain.
Vasiliy Kudryavtsev, Kirill Borodin, Nikita Vasiliev et al.
In briefBayesian Best-Arm Identification with Abstention Option
SummaryThis paper studies a Bayesian fixed-confidence best-arm identification problem in which a learner may abstain from giving a final recommendation. Under a target error probability and an abstention budget, the authors propose the Screen-and-Confirm (S&C) algorithm and analyze its expected stopping time. They also derive the minimum expected stopping time achievable by any algorithm meeting these constraints. For Gaussian priors and rewards, nearly tied instances dominate the sampling cost, and without abstention S&C attains the optimal dependence on the error probability.
In briefMemPlan: Memory-guided PDDL planning for partially observable text environments
SummaryThe paper proposes MemPlan, a framework that uses LLMs only as constrained bridges between language and symbolic PDDL planning in partially observable text environments. Positive graph memory assigns relevance-based costs to candidate hypotheses, while negative memory turns feedback-refuted assumptions into exclusion constraints. Experiments on TextWorld, ALFWorld, and Robotouille report better task completion and step efficiency than baselines, with substantially lower online token cost than LLM-planning methods such as ReAct.
In briefClimbing the Ladder: Building Automated Data Pipelines and Reproducing Baselines for Sicilian Translation
SummaryThis paper rebuilds a small neural machine translation system for Sicilian, a low-resource language, by extracting parallel text from the PDFs of a bilingual journal with maintained software and porting the original tokenization to a Perl-free implementation. It compares small Transformers trained from scratch with a pretrained NLLB model adapted by low-rank fine-tuning, using a staged training recipe. The pretrained model performs much better, and staged training helps translation into Sicilian more than out of it. The authors also note that a multilingual adapter degrades directions it was not trained on.
In briefProbing and steering language models for safety-related behaviours
SummaryThis paper studies activation probes, directions computed in a language model's residual stream, which are used both to detect safety-relevant behaviours and to steer outputs by adding the direction back into the stream. The authors test two estimators and three token-pooling schemes on Llama-3.1-8B-Instruct across deception, high-stakes awareness and sycophancy, evaluating each direction as both a detector and a steering vector. They report that a read/write dissociation holds for deception and high-stakes but reverses for sycophancy, and that the best pooling and layer choices depend on the behaviour and data.
In briefCTRL-FM: Inference-time control for filling sensor data gaps during long outages
SummaryThis paper presents CTRL-FM, a flow-matching method for estimating values from sensors that have been offline for long periods. At inference time, it uses self-supervised proxy outages built from surviving measurements to choose and locally adjust graph and temporal guidance for each outage, applying control only to missing entries. The authors report the strongest overall performance on PeMS08, METR-LA, Air-36, and AQI under 30% and 60% node outages, with graceful degradation under up to 75% sensor loss.
In briefCiteAnchor: A Source-Aware Framework for Detecting and Resolving Hallucinated Citations
SummaryThis paper addresses fabricated citations, which make it hard to trace sources and verify scientific evidence. The authors introduce CiteResolve, a benchmark of 423 real-world hallucinated citations annotated for whether their source can be recovered. They also propose CiteAnchor, a framework that adapts retrieval and field-level verification to each source type and separately determines citation validity and whether a unique source can be found. The paper reports improved detection F1 and end-to-end task success over baselines on the benchmarks.
In briefRoute-and-Generate: KL-Regularized Router Improvement for Adaptive Reasoning in LLMs
SummaryLarge language models often use complex reasoning on every question, which wastes effort on easy tasks and falls short on harder ones. The paper introduces Route-and-Generate (Rouge), a hierarchical training framework where a routing policy splits a fixed sampling budget across reasoning formats, while the base model is trained with GRPO and per-sample rewards. A rank-based reward favors shorter correct solutions, and the router is updated through a KL-regularized policy improvement step distilled into a neural router. The authors report improved accuracy-efficiency trade-offs across several reasoning benchmarks.
Tien-Phat Nguyen, Truong Nguyen, Thanh-Thien Le et al.
In briefConsistency Is Not Validity: Self-Rewarded Moral Argumentation Makes Language Models Decisive but Pliable
SummaryThe paper trains a language model to produce better moral arguments in a tournament where the model itself judges argument quality, using authentic dilemmas extracted from classic novels. The policy becomes more decisive and less sensitive to the order of options, but sycophancy and reframing remain problems for validity. The authors conclude that consistency under reordering can be optimized without improving the moral judgment it is meant to signal.
Marius De Kuthy Meurers, Aaron Parisi, Bernd Bohnet
In briefFiLM-CAM: Keyed Feature Modulation for Conditional-Access Watermarking
SummaryThis paper proposes FiLM-CAM, an image watermarking method that uses a secret key to condition Feature-wise Linear Modulation, linking the watermark payload to the key at the feature level. Correct extraction requires the right key, and training aims to widen the gap between authorized and unauthorized use. The encoder is also designed so the watermark residual can be removed only with the secret key. Evaluation across several encoder-decoder backbones is reported.
In briefAQL-Lite: Efficiently Teaching Arabic LLMs New Knowledge Through Curated Data
SummaryThis paper introduces AQL-Lite, a framework for adapting a multilingual LLM to Modern Standard Arabic using carefully curated data and partial-parameter continual pretraining. It trains adapters, the last eight transformer blocks, and the embedding matrix instead of all parameters. Under a matched 1B-token budget, curated data improved the Arabic benchmark mean by 3.51 points over raw web text, and the approach recovered 94% of full-parameter gains in 52% of the time. Scaled to 40B tokens and instruction-tuned, a 12B model scored 75.3 on the OALL v2 suite.
In briefBefore the Ticket: An AI Assistant Answering HPC Support Questions from Documentation and Past Tickets
SummaryThe paper presents an assistant that answers high-performance computing user questions by combining retrieval over live documentation with a Llama-3-8B model fine-tuned on 11,244 query-resolution pairs mined from historical tickets. Evaluated on 400 held-out tickets from Purdue's Rosen Center, the fine-tuned model agreed most with the center's resolutions. The combined system did not outperform its individual tools, though the two tools were found to be complementary.
Elham J. Barezi, Primus C Kabuo, Niveditha Nerella
In briefFoundation4D: A unified model for perceiving and reconstructing driving scenes
SummaryThe paper presents Foundation4D, a single feed-forward model that estimates geometry, detects 3D objects, predicts semantic occupancy, and performs Gaussian reconstruction from multi-camera driving video. It uses a shared backbone with factorized spatiotemporal attention and pose graph optimization to refine camera trajectories. The authors also release PAI-Foundation4D, a benchmark with dense depth and occupancy labels for 1,253 driving clips.
NVIDIA: S\'ergio Agostinho, Maximilian Apodaca, Emmanuel Attia et al.
In briefSpiking Neural Networks for Vision: Gradient-Based Mixed-Precision Quantization That Accounts for Spikes
SummaryThis paper proposes a mixed-precision quantization method for Spiking Neural Networks (SNNs), an energy-efficient neural network type, to keep accuracy at low bit-widths. It assigns bit-widths using first-order gradients, avoiding the costly second-order Hessian computations used by methods like HAWQ-V2. The method is tested on VGG, ResNet, and Vision Transformer models on CIFAR10/100, ImageNet, and DVS-CIFAR-10. On ResNet-34 with ImageNet, it reports 73.30% accuracy at 4.78× compression, comparable to HAWQ-V2, with 2.31× lower memory use and 100× faster bit allocation.
Rajshekhar Rakshit, Devgokul B.V, Rakshitha Kalkura et al.
In briefMoRE: Motion-Regularized Encoding for Generating 3D Objects with Moving Parts
SummaryThis paper presents MoRE, a method for generating articulated 3D objects, which have movable parts, from a single image. It encodes joints as coordinate-aligned joint lines using axis classification and two axis-orthogonal Plücker moment coefficients, and during training it compares the motion induced by decoded joints with that of annotated joints to regularize reconstruction. On PartNet-Mobility and ACD, the authors report better part-box agreement, articulated-state geometry and visual similarity than the compared baselines, and lower axis, origin and range errors than ArtiLatent.
In briefMPBRS: A Unified Reward Setup for Humanoid Robot Walking Using Multi-Potential Reward Shaping
SummaryThis paper addresses the difficulty of designing and tuning reward functions for reinforcement learning of humanoid robot locomotion. It proposes Multi-PBRS, which shapes rewards through differences of potential functions to guide exploration without changing the optimal policy. The same reward structure and weights are evaluated across nine humanoid robots, needing only robot-specific metadata such as joint or link mappings and default poses. According to the experiments, compared with standard shaping, MPBRS gives more consistent scaling across robots and is less sensitive to tuning.
In briefR4DGS: Language-Based Object Referring in Dynamic 4D Gaussian Scenes
SummaryThis paper introduces Referring 4D Gaussian Splatting (R4DGS), a task of locating objects in dynamic 4D scenes from complex natural-language expressions. The authors release R4D-Bench-QA, a benchmark with structured query annotations, and propose ReferGaussian, a framework that turns dynamic Gaussians into language-addressable entities. It uses multimodal large language models to break queries into trackable phrases and constraints, and needs no scene-specific semantic training. Experiments on R4D-Bench-QA and the 4D LangSplat protocol report better temporal target selection and competitive spatiotemporal grounding.
In briefSelf-Distilled Future Video Representations for Robot World Action Models (FRED)
SummaryThe paper introduces FRED, a framework that trains World Action Models to improve future video representations extracted in a single denoising step, using self-distillation and historical observations as temporal context. Across diverse real-world manipulation scenarios, FRED consistently outperforms competing WAMs. It shows particularly large gains on long-horizon and out-of-domain zero-shot tasks, reaching 50.0% success on tasks demonstrated only in human videos versus 19.4% for the best competing WAM.
In briefTerraTransfer++: A Planner Trained Without Human Driving Demonstrations
SummaryThe paper proposes an end-to-end autonomous driving training approach that avoids human driving demonstrations. A policy is first pretrained through self-play in a vectorized simulator, its rollouts are distilled into a flow-matching trajectory teacher, and lightweight adapters are trained on a frozen pretrained vision backbone. The authors report that the resulting closed-loop policy outperforms published imitation-based end-to-end methods on the HUGSim benchmark, and that the planner reaches 59.4 EPDMS on NAVSIM navhard and 95.5 PDMS on navtest.
WhyReplaces human demonstrations with self-play-distilled teachers and reports gains on HUGSim and NAVSIM; novelty is high, evidence rests on few benchmarks.
In briefCollaborative in-context learning for estimating indoor radio maps without sharing private data
SummaryThis paper proposes Poly-Ripple, a framework where multiple clients estimate indoor radio maps using large autoregressive vision models through in-context learning, without updating model parameters. Each client uses its private demonstrations, and the method models signal propagation outward from the transmitter in a ripple order, with importance-aware aggregation to combine knowledge across clients. Experiments on a public indoor radio map benchmark show it outperforms representative in-context learning baselines.
In briefSciFusion: Physics-guided diffusion model that generates 3D scientific images across multiple imaging types
SummarySciFusion is a multimodal framework for generating 3D scientific images, built on a latent diffusion model. It splits each imaging modality into domain, object, and contrast factors that are shared across related modalities, so combinations can be recombined at inference to produce unseen ones zero-shot. It also adds physics-based consistency objectives for MRI, CT, and electron microscopy to match how each image is formed. The authors report better fidelity and cross-domain generalization than modality-specific and flattened multimodal baselines across 15 medical and scientific imaging modalities.
In briefGender, Sexual Orientation and the One-Year Refugee Bar: Study of Canada's Bill C-12
SummaryThis paper examines Canada's Bill C-12, which since March 2026 makes refugee claims filed more than one year after first entry ineligible for referral to the Refugee Protection Division, diverting them to a paper-based Pre-Removal Risk Assessment. Using AI tools checked by human review, the authors analyze nearly 29,000 published decisions from 2006 to 2025. They find gender and SOGIE claims are overrepresented among late claims, and refusals of such late claims are more often overturned on appeal or judicial review than other late claims.
WhyLarge-scale computational analysis of about 29,000 Canadian refugee decisions shows late SOGIE and gender claims are overrepresented and overturned more often, with clear policy implications.
In briefMeasuring Research Progress and Linguistic Inclusion in Low-Resource Languages
SummaryThis paper analyzes 1,158 peer-reviewed open-access papers from 2020 to 2025 to examine research attention to low-resource languages and their representation across speaker populations. The authors introduce the Linguistic Inclusion Framework, which measures inclusion across research growth, representation expansion, representation continuity, and distributional equity. They find that publication growth mostly concentrated on already represented languages, and that Asian languages receive a smaller share of representation than their share of speakers.
In briefEgoHand: Reconstructing 3D hands from single-camera first-person video
SummaryThis paper presents EgoHand, a model that reconstructs 3D hand poses from monocular egocentric video. The authors built a training set of 14 million images combining synthetic data, in-the-wild video, and multi-camera motion capture. Zero-shot tests on HOT3D, ARCTIC, and H2O show lower joint, root-translation, and acceleration errors than baselines, at about 9.69 ms inference latency. A blinded user study found participants preferred EgoHand's hand poses, and the model and code are released under the OpenMDW license.
In briefGAV: A scientific experimentation environment that builds verifiable PDE problems for coding agents
SummaryThe paper introduces GAV (Generator-Agent-Verifier), an environment that gives coding agents reliable feedback when writing programs for partial differential equations (PDEs). The Generator learns a distribution of reference solutions from simulation or experimental data and uses the method of manufactured solutions to construct matching PDE problems without a separate numerical solve. The Verifier scores candidate programs on accuracy and computational cost. On Navier-Stokes and Vlasov-Poisson-type tasks, GAV reduced the mean test error of three existing PDE agents by more than an order of magnitude on the numerical Navier-Stokes task.
In briefGeneralizing Metropolis-Hastings for Sampling Directed Acyclic Graphs
SummaryThis paper addresses the difficulty of approximating the posterior distribution of directed acyclic graphs (DAGs) in Bayesian network structure learning, especially for large, high-dimensional data. It introduces new Markov chain Monte Carlo moves that resample the parents of several nodes at once, built on duet sampling, a generalization of Metropolis-Hastings that selects two proposal distributions per iteration. According to the authors, experiments show these moves outperform the previous state of the art, and the advantage persists even when earlier methods use duet sampling.
In briefTradeoff-aware territory design: a decision system for assigning customers to sellers
SummaryThis paper presents a production decision system that allocates enterprise customers to sales sellers while balancing competing goals. It combines account-value scoring, a counterfactual method for estimating the cost of breaking seller-account relationships, and multi-objective optimization. Historical backtesting found that optimizer-generated scenarios substantially improved assignment compliance while keeping seller-level opportunity distribution comparable. Adopted scenarios in production planning improved both compliance and opportunity balance over the previous baseline.
In briefSWE-Verifier: Selective Verification for Low-Cost Coding Agents
SummaryThe paper introduces SWE-Verifier, a lightweight generative model that decides when a coding agent should sample and verify alternative actions, then ranks candidates before execution. It is trained from Qwen3.5-4B using preference data built from contrasting actions, their observations, and rubric-guided teacher scores. On MiniMax-M2.7, resolution rises from 75.47% to 77.20% at 1.09 times cost; it also generalizes to DeepSeek-V4-Flash, raising resolution from 72.60% to 74.27% at 1.13 times cost. Across both models it verifies fewer than 4% of steps.
Kaijian Zou, Vaibhav Chennareddy, Yunxiang Zhang et al.
In briefAstro2Method: testing whether LLMs can pick the right machine learning method for astronomy problems
SummaryThis paper introduces Astro2Method, a benchmark of 500 expert-reviewed binary questions drawn from 346 astronomy papers published between 2024 and mid-2026. On easy questions, direct prompting reaches 87.50-92.55% accuracy, but on hard questions it drops to 58.87-64.52%. Retrieval and agent methods such as ReAct, Self-Ask and CRITIC do not give consistent gains, while web search helps mainly when it finds the source paper. Supplying the source paper's pipeline with method names masked improves hard-question accuracy.
In briefA Quadratic-Function Analysis of Muon's Orthogonalized Update: Benefits, Drawbacks, and Newton-Schulz
SummaryThis paper analyzes PolarGD, the orthogonalized update at the core of the Muon optimizer, on matrix quadratics and compares it with gradient descent. Orthogonalization helps when the gradient is far from low rank and its largest singular directions have high curvature, but it can be much slower when the iterate has tiny singular values. The authors also find that the few Newton-Schulz iterations used in Muon can converge faster than exact orthogonalization.
In briefHuman-AI loop breaks the high-degree barrier in Lehmer's problem
SummaryLehmer's problem asks whether the Mahler measure of integer polynomials can be made arbitrarily small. Previously, small-measure irreducible polynomials were systematically constructed only for even degrees up to 180, and Boyd's family often factors at higher degrees. The authors use a human-AI discovery loop combining numerical pipelines, exact symbolic verification, and distributed search to derive a five-parameter Generalized Boyd construction. It yields irreducible polynomials with Mahler measure below Boyd's limit point for every target degree, improves six published values across even degrees up to 2000, and matches 99.6% of best known records. A verified dataset is released.
WhyProposes a five-parameter Generalized Boyd construction removing degree skipping and beating Boyd's limit, verified across even degrees up to 2000 with six improved records.
In briefEdit, Don't Imitate: A Minimal-Edit Distillation Method for Multimodal Math Reasoning
SummaryThis paper addresses vision-language models that misread diagrams, which leads to incorrect mathematical reasoning, especially in smaller models. The authors propose minimal-edit corrective on-policy distillation, where training uses reference figure descriptions and verified solutions, and a teacher penalizes only the tokens it disfavors in incorrect student outputs. Correct outputs are reinforced with task rewards. The method needs no descriptions at test time, matches or exceeds base models given reference descriptions, and transfers to external math and perception benchmarks.
In briefLongZero: Improving LLM long-context reasoning without any external data
SummaryLongZero is a method that lets a large language model improve its long-context reasoning using only tasks it creates itself, with no external data. The model first synthesizes questions, answers, and supporting evidence in a compact structured form, then expands that evidence into a long multi-document context. It trains with GRPO on self-written multi-hop reasoning and multi-round retrieval tasks. Evaluated on two LLMs across five long-context benchmarks, LongZero generalizes to real tasks and outperforms both baselines.
In briefRethinking video autoencoders for interactive world models: how much to compress video that agents act in
SummaryThis paper examines how much video world models should compress frames, since compression sets per-frame cost and action latency. The authors introduce CADRE, an open family of causal diffusion representation autoencoders at several spatio-temporal compression levels, and train the same world model on each in manipulation and navigation settings. CADRE forms most of the latency-quality frontier, and the study reports findings on diffusion decoders and on opposite compression needs for still versus moving cameras.
WhyProposes a family of causal diffusion autoencoders across compressions and systematically studies their effect on world models across four environments.
Adrien Ramanana Rahary, Yutong Bai, Nicolas Dufour et al.
In briefAdaptive per-item computation for sequential recommendation
SummarySequential recommenders typically spend the same computation on every item in a user's history, even though items differ in how much processing they need. The authors show this uniform allocation is suboptimal, since many items converge early while a minority need full depth. They propose LoopRec, which recursively applies one weight-shared block and uses a lightweight gate at each step to let each item exit early or continue. Experiments on public benchmarks with four backbone families report consistent gains over matched fixed-depth models.
In briefVoxProg: 3D Spatial Reasoning via Voxel Programming
SummaryThis paper presents VoxProg, a method in which a language model turns spatial instructions into executable programs that operate on local voxel views of 3D scene surfaces. The programs locate task-relevant regions and compute a target SE(3) pose for object placement, and the model revises its code using visual and numerical execution feedback. The authors introduce VoxProg-Bench, a 71-task benchmark, and report that VoxProg outperforms relation-based, keypoint-based, and mesh-based baselines.
In briefPhysiCalWorld: In-Context Physics Calibration for World Models
SummaryExisting world models often rely on the current frame or a short history, which can miss physical properties like mass distribution, joint damping, or friction. This causes predictions to drift toward generic priors in novel scenes. The paper proposes PhysiCalWorld, which encodes an exploratory interaction clip from the test scene into a compact physics descriptor token that calibrates the model without fine-tuning or physics labels. Evaluated across four environments with varied physics, it improves rollout prediction accuracy and visual planning success rate.
In briefA quasipolynomial algorithm for ℓ∞ contraction fixed points
SummaryThe paper presents an algorithm that computes an ε-approximate fixed point of any d-dimensional nonexpansion under the ℓ∞ norm in time log(1/ε)^{O(log² d)}. Its core is a boosting argument that geometrically shrinks the search region. The authors note that this implies simple stochastic games can be solved in quasipolynomial time.
WhyA boosting argument shrinks the search region geometrically, yielding quasipolynomial fixed-point algorithms for ℓ∞ nonexpansions and simple stochastic games.
In briefALPS: A Visual State Encoder for World Models
SummaryThis paper introduces ALPS, a pretrained visual state encoder that unifies physical-state information, language alignment, and pixel reconstruction in one set of patch tokens for world models. It trains a ViT-B/16 from scratch using only still images and text, then refines it with cross-view self-distillation. In a unified evaluation across fifteen public encoders, ALPS achieves the lowest aggregate LIBERO state-probe error among ViT-B models and the highest mean policy success on LIBERO.
In briefSelf-normalized Bernstein inequalities for regenerative processes, with applications to learning from dependent data
SummaryThis paper proves a self-normalized Bernstein inequality for bounded functions of a regenerative process whose marked cycles are independent and identically distributed with finite mean length. The bound uses observed cycle lengths and keeps explicit terms for truncation and for the cycle ending after the observation horizon. Choosing a tuning parameter from a finite deterministic set yields a version in which all terms are empirical except one, tied to the tails of regeneration times. The authors apply this to classification with non-i.i.d. regenerative data, deriving generalization bounds and showing minimax optimality in sample size and expected regeneration time. They also combine it with a regenerative strong approximation to obtain a Gaussian approximation using an observable block variance estimator.
WhyExtends self-normalized Bernstein inequalities to regenerative processes, yielding minimax-optimal bounds and generalization guarantees for non-i.i.d. learning with rigorous theory.
In briefHow locality and weight sharing shape convolutional neural network learning in a teacher-student model
SummaryThe paper compares convolutional and locally connected student networks trained by online stochastic gradient descent in a teacher-student setup, deriving ODEs that describe their dynamics. For uncorrelated input positions, the convolutional student learns P times faster and locality affects sample complexity. With spatial correlations, the advantage of weight sharing is reduced, and the two students' error behavior differs depending on the learning rate schedule.
In briefEnvironment Duels: A Benchmark of How LLMs Create and Solve Tasks for Each Other
SummaryThis paper introduces Environment Duels, a dynamic agentic benchmark in which frontier LLMs write interactive environments with privileged hints and try to solve those written by others. The authors find that solving and authoring are distinct skills, and that smaller models learn more from how others played. Head-to-head results are transitive and yield a stable ranking, and a reinforcement-learning study with Qwen3.8-27B links hint lift to gains on held-out benchmarks.
WhyProposes a duel benchmark separating environment authoring from solving, with an RL study linking learning gains to author quality across nine models.
In briefIs Consensus Research Over? AI-Discovered Protocol A-greement
SummaryThe paper presents A-greement, a consensus protocol for geo-replicated state machines discovered with AI assistance. Across six deployments with 5 or 9 replicas, it reduces mean latency over the fastest baseline by 12.2% on write-heavy and 33.3% on read-heavy workloads, and matches it on read-only workloads. The authors also introduce AI-greement, an agent harness with a Lean verification framework, and argue that human consensus experts remain essential.
WhyProposes A-greement, cutting latency over the fastest baseline across six deployments, with Lean verification uncovering previously unreported protocol vulnerabilities.
Cl\'ement Burgelin, Rachid Guerraoui, Zicong Hong et al.
In briefQuant-Arena: Benchmarking LLM Agents on End-to-End Quantitative Trading Research
SummaryThis paper introduces Quant-Arena, a benchmark and public leaderboard testing whether large language models can complete full quantitative research on real A-share market data. Models build features and stock-scoring code, receive feedback only on 2022–2023, and freeze one strategy for evaluation on the sealed year 2024. Scoring is driven mainly by realized excess return, with signal quality and feature diversity as supplements. Nine frontier models were evaluated; GPT-6 Astra scored highest, but no model beat the equal-weight market in 2024.
In briefMinimax-optimal NPMLE for estimating Gaussian mixture densities under Hellinger distance
SummaryThe paper studies estimating univariate Gaussian mixture densities with the nonparametric maximum likelihood estimator (NPMLE). For sub-Gaussian mixing measures, the unconstrained NPMLE attains the minimax optimal rate log(n)/n in squared Hellinger distance. For compactly supported mixing measures, the NPMLE constrained to a compact interval attains a slightly faster optimal rate. The results are applied to show regret-optimality of NPMLE-based g-modeling in Gaussian empirical Bayes with bounded priors.
WhyProves NPMLE attains minimax optimal rate for Gaussian mixtures under squared Hellinger loss with sharp local bracketing entropy bounds; strong theory, impact mainly within statistical theory and empi
In briefMultimodal embodied agents: a unified view of AI moving from computer use to robot use
SummaryThis paper defines multimodal embodied agents (MMEAs) as goal-directed systems that combine multimodal reasoning with physical action and revise decisions based on feedback. The authors propose PAPAV, a framework of five capabilities (Perceive, Anticipate, Plan, Act, Verify) to compare computer-use agents with robotic systems. Across 64 benchmarks, evaluation concentrates heavily on Act, with only 5 assessing Anticipate and 3 assessing Verify.
In briefVelocity Ambiguity Profiles: Time-Resolved Bayes-Risk Diagnostics for Flow Matching
SummaryFlow matching usually reports one loss averaged over interpolation time, which hides whether errors come from intrinsic ambiguity or limited data. The paper introduces the Velocity Ambiguity Profile, a time-local Bayes-risk floor, and splits time-resolved loss into this floor plus a reducible excess term. Controlled Gaussian-mixture experiments and CIFAR-100 DINOv2 feature-flow experiments are used to validate the diagnostic for locating ambiguity-limited and data-limited regions.
In briefAsymmetric On-Policy Distillation: Learning In-Context Exploration
SummaryThe paper proposes ASTEROID, a student-teacher supervised learning framework for learning test-time exploration in partially observable sequential decision-making. The student collects on-policy context from its own history, and a privileged expert provides action labels conditioned on that context, teaching the student to search while uncertain and exploit once the hidden state is revealed. The authors report gains over reinforcement learning baselines and frontier models across eleven environments, and transfer of simulation-trained policies to vision-denied real robot manipulation.
In briefBatchBench: A leaderboard testing how well biomedical image AI handles batch effects and domain shift.
SummaryThe paper introduces BatchBench, a leaderboard that evaluates the robustness of AI systems on biomedical images from unseen experimental batches and data sources. It is built on JUMP-CP, a large publicly available imaging dataset generated with the Cell Painting assay. Because JUMP-CP shows pronounced batch effects and domain shifts, it serves as a challenging testbed for measuring robustness to distributional changes.
Sonja Topf, Ana Sanchez-Fernandez, G\"unter Klambauer
In briefDid the Problem Actually Get Harder? Testing LLMs on Hardened Paired P and NP-Complete Problems
SummaryThe paper introduces P2NPCBench, a benchmark pairing NP-complete problems with polynomial-time twins that differ in one structural detail, yielding 2,114 statements after adversarial multi-agent rewriting (AgentGAN) to remove textbook cues. Across 12 models, pass@1 was 35.4% on NP-complete statements versus 63.8% on their P twins. The authors report that hardening narrowed the class gap but that part of the gap reflects familiarity with textbook phrasing rather than complexity class.
Nurul Labib Sayeedi, Md. Faiyaz Abdullah Sayeedi, Zabir Al Nazi et al.
In briefDynamic Microsegmentation for Multi-Robot Networks: A Lab Demonstration
SummaryThis report describes a laboratory setup that applies dynamic microsegmentation to an Internet-of-Things network of multiple robots, using a commercial software-defined microsegmentation tool. Segments are adapted according to asset functional and health tags, and the dynamic policies are implemented automatically in software. The demonstration shows the solution automatically isolating an indicated threat in real time while unaffected operations continue.
In briefComparing Model Compression on Diverse Edge Devices for Rotating Machinery Maintenance
SummaryThis paper compares four model compression techniques applied to a convolutional neural network for predictive maintenance of rotating machinery, using a real-world dataset. The compressed models were tested on nine computing platforms, from highly constrained edge devices to high-performance processors. Knowledge distillation performed best, achieving over 90% speedup on all devices, a 91.22% balanced F-score, and a 98.9% size reduction, with only a small drop in predictive performance.
Matheus Calheira Guimar\~aes de Oliveira, Marco Aur\'elio da Silva Cruz, Regis Cardoso et al.
In briefPrivate Sparse Reward Estimation from Preference Feedback
SummaryThis paper studies how to estimate reward models for aligning language models from human preference feedback while protecting privacy with differential privacy. It assumes preferences depend on only a small subset of features and proposes efficient algorithms for Sample-DP and Label-DP settings under the Bradley-Terry-Luce model. The authors give upper and lower error bounds for parameter estimation and report that the method outperforms dense baselines on sentiment generation and safety alignment benchmarks.
