In briefLearning Contractive Dynamical Representations for Composite Adaptive Control
SummaryThis paper presents a representation-learning framework for composite adaptive tracking control when disturbances are dynamically coupled. The authors use a statistically grounded hard-EM procedure with a Kalman smoother to learn disturbance dynamics whose latent evolution is uniformly contractive, extending fixed-decay adaptive methods to a predictive disturbance-accommodating formulation. Experiments on a slippery ground vehicle with a liquid-sloshing tank and pendulum load, and on coupled Duffing oscillators, report more accurate disturbance prediction and better tracking than the compared baselines.
Min Kim, Jos\'e Leonardo Brenes, Fred Hadaegh et al.
In briefNeural Harmonic Measure Operator: a neural solver for elliptic PDEs on changing domain shapes
SummaryThe paper introduces the Neural Harmonic Measure Operator (NHMO), a neural solver for elliptic PDEs on variable-shape domains. It models the harmonic measure density, which depends only on geometry, with a transformer-based boundary kernel trained on Walk-on-Spheres exit samples, so one kernel handles new boundary values without retraining. The method is extended to Poisson problems via a classical decomposition, and the authors report improvements over four baselines on the MCB-B 3D variable-shape Poisson benchmark.
In briefHow to Loop MoE: Flattening Experts and Untying Attention for Looped Mixture-of-Experts Models
SummaryThis paper studies how to combine looped Transformers, which reuse layers several times, with sparse mixture-of-experts (MoE) models. The proposed method, Foil, keeps expert parameters and per-token compute fixed while halving expert layers, doubling experts per layer and doubling passes, and giving each pass its own attention parameters. Experiments report lower pretraining loss than the unflattened looped baseline, with the most flattened model 0.012 nat lower at 100B tokens and downstream accuracy on par or better.
In briefCapacity, Responsiveness and Alignment: What Makes a Latent Structure Actionable
SummaryThis paper studies why some latent structures in language model activation spaces have much greater causal influence than others. It breaks causal influence into three factors: capacity, responsiveness, and alignment, and tests them across 4 model families and 50 concepts. The authors find that all three must be high for effectiveness, and that causally effective directions form a context-dependent low-dimensional subspace. Causal probes trained within that subspace improve steering by 17%-118% with about a 3% drop in concept detection.
In briefKV-streams: a faster way to compress long context for training AI agents
SummaryThis paper addresses the slow training of agentic LLMs caused by very long context traces. It proposes KV-streams, which streams the KV cache forward instead of flushing it after each compaction, and works with any compaction strategy. The authors report 2.6 to 5x wall-clock training speedups across three compaction strategies without evidence of performance loss.
Emiliano Penaloza, Dane Malenfant, Dheeraj Vattikonda et al.
In briefTeacher-guided training for socially efficient language agents (TACT)
SummaryThis paper proposes Teacher-Assisted Communication Training (TACT), a method that helps language agents reach social goals while using fewer words. Expression and strategy specialists revise the student's actions, partner responses test each revision, and useful feedback is distilled back into the student. On SOTOPIA and AgentSense, TACT achieves higher goal attainment and uses fewer target tokens and interaction messages than the compared methods.
In briefROSS: Reusing Model-Generated Past Attempts Through Selective Training
SummaryThis paper addresses the fact that self-generated rollouts from reinforcement learning and on-policy distillation are often discarded once the model updates. It introduces ROSS, which keeps the full historical trajectory as context and applies training loss only to selected model-generated continuations. ROSS outperforms baselines across mathematics, code generation, instruction following, and software engineering. On Qwen3.6-35B-A3B, it raises the six-benchmark MOPD average from 58.40% to 62.20% and SWE-bench Verified from 64.20% to 68.40%.
In briefAdaptive Looped Transformers for Better Test-Time Scaling
SummaryThis paper studies whether looped transformers, which reuse layers for latent computation, improve accuracy as test-time decoding compute grows. The authors propose TaH2, which directs extra iterations only to tokens that benefit from them, jointly post-training the backbone and an iteration decider with lookahead depth supervision. On AIME benchmarks, TaH2 raises the accuracy-compute slope by 53% over a non-looped baseline, and its gain keeps growing as maximum iteration depth increases.
In briefCopy What Stays, Distill What Differs: Initializing Linear Vision Transformers
SummaryLinear Vision Transformers replace Softmax attention with a linear-complexity operator, but usually need training from scratch and often underperform Softmax ViTs. This paper asks whether pre-trained Softmax ViT weights can initialize linear ViTs. Directly copying attention weights helps little, but distillation with a suitable loss recovers the attention's token routing behavior. Copying MLP weights transfers most of the benefit, and combining both lets linear ViTs close the gap and even surpass Softmax ones, consistently across variants, sizes, and datasets.
Huaiyuan Qin, Muli Yang, Gabriel James Goenawan et al.
In briefInfiniHand: streaming estimation of world-space hand motion from first-person video
SummaryInfiniHand is an end-to-end streaming framework that jointly estimates MANO hand parameters, camera trajectories, and hand locations directly from uncalibrated egocentric video. It combines persistent spatiotemporal memory with hand-centered visual features and is trained in two stages on about 5,000 hours of aggregated public egocentric data. The authors report outperforming state-of-the-art baselines on in-domain benchmarks, including a 21.4% reduction in ARCTIC PA-p compared to ViDiHand, and robust generalization to in-the-wild videos at 11.19 FPS.
In briefSimple self-retrospection improves language-model agents without reinforcement learning
SummaryThis paper proposes ROFT (Retrospection-Only Fine-Tuning), in which a language-model agent is fine-tuned only on its own retrospective explanations of attempts, with no external teacher or reward-based update. In software-engineering experiments with Qwen3.5-4B on SWE-bench Verified and Pro, ROFT reached solve rates after 20 updates comparable to GRPO's after 40 updates, and progressed faster early on. The authors also report that ROFT learned to solve some tasks where all sampled base-model attempts had failed.
Jonathan Light, Christopher Zhang Cui, Jeonghye Kim et al.
In briefRight words, wrong moment: clinician review of how young people's distress appears in ChatGPT conversations.
SummaryResearchers collected 19,930 ChatGPT conversations and survey data from 158 young adults aged 18 to 25. Ten clinicians reviewed five conversations reflecting user distress. Distressed participants reported more emotional engagement and behavior change than peers. During acute distress, ChatGPT often gave overly dramatic responses and excessive action suggestions, and clinicians identified seven process failures. The authors translated this feedback into three-stage design guidelines.
In briefHarness Learning: Adapting Agent Programs at Test Time for Generalizable Gains
SummaryThe paper introduces harness learning, which trains a proposer model to revise a language-model agent's harness, the executable program organizing model calls, tools, and information flow, using execution feedback. Harness revisions act as weight updates in a meta-learning formulation, and the proposer is trained with reinforcement learning using the task performance of revised harnesses as reward. At test time, the proposer refines the harness from successive executions without parameter updates. Experiments on reasoning and multi-hop question answering show improved revision quality and that test-time adaptation transfers to unseen tasks.
In briefREVO: Efficient Language Model Distillation by Reusing Rollouts
SummaryThis paper introduces REVO, an off-policy distillation method that trains language models by reusing each student-generated rollout for multiple learner updates, reducing costly regeneration. It handles policy mismatch with stabilized prefix weighting and one-step resampling from the current student, and uses the variance of the student-teacher log-probability ratio to prioritize informative tokens. The authors report that REVO with 50 rollout iterations matches or exceeds on-policy distillation baselines trained for 200 iterations on in-domain and cross-domain reasoning benchmarks.
In briefGeoVerse: Novel-view image synthesis that stays consistent across viewpoints in a geometric latent space
SummaryGeoVerse generates new viewpoints from sparse images by working in the geometric latent space of a pretrained 3D foundation model. It injects appearance priors from a video generative model through a ControlNet-style adapter to help complete unseen regions. A global spatial memory aggregates observed and synthesized content to keep different views coherent. The authors report a 2.23 dB higher PSNR on DL3DV and 32.4% lower ATE on Mip-NeRF360 compared with GLD.
In briefFailure-Transparent Agents: Benchmarking Post-Failure Reporting in Tool-Using Language Models
SummaryThis paper introduces Failure-Transparent Agents (FTA), a controlled benchmark that fixes failed tool observations and required evidence before generation, so post-failure claims by language-model agents can be audited. It contains 100 tasks across five failure families, a neutral control, and four user-pressure conditions, with 3,600 human-annotated responses from six models and three response policies. The evidence-contract policy reduced false-success rates from 22.8% to 0.8% while useful-response rates rose from 74.9% to 98.8%.
In briefFlowAct-R2: Interactive Humanoid Video Generation with Streaming Inputs and Proactive AI Agent Planning
SummaryFlowAct-R2 is a framework for generating interactive humanoid videos that responds to continuous inputs. It adapts a pretrained reference-to-video model to accept streaming actions, audio, and updated image, audio, and video references. A proactive agent prepares a persona, agenda, and skills in advance, then schedules behaviors, responds to audience input, and handles interruptions during live sessions. The system supports real-time 720p generation and hour-long streaming.
In briefHow camera angle and patient orientation affect AI rehabilitation exercise monitoring
SummaryThis paper studies how camera placement affects pose estimation used to assess the quality of rehabilitation exercises. The authors introduce REHAB26-ViewAngles, a dataset of correct and incorrect exercise executions filmed from many camera angles, and a separability metric that measures how well an algorithm distinguishes valid from faulty repetitions. They compare single-camera 2D and 3D pose estimation with several multi-camera strategies. They report that an optimally placed 2D camera improves separability by 16.9% over the common 0° frontal view, and combining two views can further improve accuracy by up to 13.1%.
Miriama J\'ano\v{s}ov\'a, Andreas Lang, Petra Budikova et al.
In briefSplitting 3D point clouds into shape parts using geometry-aware inlier selection
SummaryThis paper addresses decomposing 3D point clouds into interpretable geometric primitives represented by superquadrics. It adds local surface properties, such as normal consistency, to the fitting process and refines inlier selection by solving an energy minimization problem with graph-cut optimization. Experiments on synthetic and real datasets report improved geometric accuracy, robustness to noise and outliers, and convergence efficiency compared with state-of-the-art RANSAC-based methods.
Alessandro Rinaldi, Edoardo Tedesco, Andrea Ferraris et al.
In briefHard vs. Easy Vision: How GPT-6 Astra Performs Across Computer Vision Tasks
SummaryThis paper evaluates GPT-6 Astra and five other frontier general-purpose AI systems on 34 capabilities and 55 benchmarks across nine computer vision areas. Semantic interpretation, reasoning, and object-centric prediction approach or reach reference levels, while metric geometry, faithful reconstruction, temporally consistent dense prediction, and specialized fine-grained knowledge show larger gaps. Additional reasoning and specialist tools close some gaps, with benefits varying by capability.
Hanoona Rasheed, Mohammed Irfan Kurpath, Bin Ren et al.
In briefX-Reset: Scaling object-centric RL with cross-embodiment resets
SummaryThis paper proposes X-Reset, a framework that uses human hand-object demonstrations to address exploration difficulties in simulated reinforcement learning for dexterous manipulation. Human hand-object states are kinematically retargeted to noisy robot states, unstable states are filtered out in simulation, and the rest are used as reset points during training with general object-centric rewards. The authors report that the resulting policy, which depends only on object state and goal, generalizes across 20 objects and three embodiments, generalizes to unseen objects, and transfers zero-shot from simulation to real robots.
In briefLagrangian–Hamiltonian flows for video prediction and image generation, from a symplectic viewpoint
SummaryThe paper presents LHFM, a framework that uses classical mechanics and symplectic geometry to model image dynamics, representing each image as an exact Lagrangian graph evolving under image-dependent Hamiltonian flows. Its main application, LHFM-V, is a recurrent model for deterministic video prediction that reports the lowest FLOP count among compared recurrent models with similar accuracy. The image variant, LHFM-I, is compatible with flow matching and achieves a lower FID than the flow-matching baseline in a matched experiment.
In brief"Mind the RefGAP": a training-free fix for reference attention in diffusion image and video editing
SummaryReference-guided diffusion editors often fail to faithfully reproduce user-provided references, and the paper attributes part of this to limited reference-attention allocation. The authors propose RefGAP, a training-free correction that adjusts logit offsets at each layer online, based on reference-attention mass measured during the forward pass, controlled by two global coefficients. Across seven diffusion-based image and video editors, RefGAP improves identity fidelity in head and face swapping. Additional experiments on virtual try-on and background replacement test transfer beyond identity editing.
In briefScAn-Bench: A Benchmark for Testing Scaling Analysis Methods
SummaryThe paper notes that no systematic study has evaluated how scaling laws are derived across different model types, despite their importance for foundation models. The authors introduce two surrogate benchmarks, ScAn-Bench-LLM and ScAn-Bench-VLM, built from 4524 and 8024 checkpoints of language and vision-language model pipelines. Using these, they perform the first systematic evaluation of data acquisition and extrapolation methods for scaling analysis across different data modalities.
Artin Sermaxhaj, Nastaran Alipour, Donat Sinani et al.
In briefTraining LLMs for more creative scientific ideas via reinforcement learning
SummaryThis paper addresses the problem that large language models produce homogeneous, predictable outputs, which limits their use in open-ended scientific ideation. The authors introduce AI Night-Scientist, an agentic framework that uses reinforcement learning with GRPO to teach models when and how to depart from predictable reasoning, modeling creativity along action, process, and outcome axes. Reported results show a 27.8% increase in research direction diversity, a 14.9% increase in contribution types, up to 32.0 percentage points higher predicted citation impact, and 66.2 points higher originality over the base model.
Priyanka Kargupta, Silviu Cucerzan, Shweti Mahajan et al.
In briefDynaTokens: Teaching Scene Motion to Camera-Controlled Video Models at Test Time
SummaryThis paper introduces DynaTokens, a lightweight set of learnable, scene-specific tokens that teach object dynamics to an existing camera-controlled video world model. The base model stays frozen, and the tokens are trained from a few example trajectories for a scene, then applied under new camera paths. On VBench2 and WorldScore, DynaTokens achieves a better balance between dynamics and camera control than LoRA, block finetuning, and other trainable-layer baselines.
In briefA single uncertainty representation for graph neural networks using doubly-spectral stochastic expansion
SummaryThis paper proposes the DSS expansion, which models uncertain node embeddings as random graph signals, using graph Fourier filters for structural variation and an orthogonal-polynomial chaos coordinate for latent stochastic variation. One representation supports calibration, out-of-distribution detection, and shifted-data prediction. The authors report that standalone DSS-GNN achieves the lowest Brier score on all 14 node classification benchmarks, and DSS-Hybrid achieves the best AUROC on most node-OOD settings.
In briefMeqMuon: a matrix-balancing Muon optimizer for LLM pretraining
SummaryThis paper proposes MeqMuon, an improved Muon optimizer for pretraining large language models. It balances both row and column magnitudes through normalization that adapts automatically to different imbalance patterns without manual tuning. MeqMuon also removes the need to store AdamW's second-moment estimates, reducing optimizer-state memory. Empirical results show better convergence than AdamW, Muon, and other baselines.
In briefProven Gains from Regularization in Adversarial Imitation Learning: Faster Convergence Rates
SummaryThis paper studies adversarial imitation learning, in which an agent imitates expert demonstrations by optimizing its policy against an adversarial reward that distinguishes expert and learner behavior. It establishes fast convergence rates for methods that jointly apply reward regularization and policy regularization, in finite-horizon Markov decision processes with general function approximation. The proposed model-free algorithm, Dually Regularized AIL, combines KL policy regularization with a quadratic reward penalty, and the paper proves an Õ(1/K+1/N) bound on the regularized imitation gap for fixed regularization parameters, where K is the number of online episodes and N is the number of expert trajectories.
Hanbin Zhou, Shangzhe Li, Alexander Braverman et al.
In briefRethinking Personalized Generation: Test-Time Alignment via Factorized Ranking Models
SummaryThe paper argues that standard alignment methods, which target a single average user, limit how well LLMs can match diverse personal preferences. It shows that personalized generation suits test-time scaling such as Best-of-N, and proposes a million-parameter MLP ranking model that reuses the base generator's embeddings to score large candidate pools. Across nine datasets, the authors report that this model outperforms billion-parameter generalist reward models while using under 0.4% of their parameters and having four orders of magnitude lower scoring latency.
In briefReport on Progressive Disclosure of Agent Skills
SummaryThis report examines how LLM-based agents at Workday can be extended with named procedures, or skills, placed in their context. As the skill library grows, operating costs rise, so the authors study progressive disclosure, which loads skills only when needed. They find it improves skill-retrieval quality but marginally increases overall latency.
In briefDo Circuit Explanations Capture Model Errors?
SummaryThis paper examines circuit-based explanations in mechanistic interpretability, which isolate compact subnetworks to reproduce a model's behaviour. The authors find that circuits validated this way closely match the model's correct answers but account for only a small share of its errors. Restoring omitted attention heads substantially raises error reproduction on an IOI case study. The authors conclude that exact error reproduction is necessary but not sufficient for a valid circuit explanation.
In briefQuanReview: offline, auditable tool for reconciling human and LLM span annotations
SummaryThis paper presents QuanReview, an open-source system for auditing and correcting structured span annotations such as quantities with units, uncertainty modifiers, and event classes. It aligns human and LLM annotation streams at the character level, resolves unambiguous cases with a logged policy, and sends conflicts to a browser-based interface where reviewers choose or merge annotations. On a 4,457-record humanitarian benchmark and an LLM extraction stream, it fully auto-merged 8% of documents, made automatic decisions on 1,513 records, and left 3,131 candidate conflicts for human review, averaging 5.4 per reviewed document.
Matteo Musacchio, Juan Cruz Giner Pulero, Isabel Casta\~neda et al.
In briefA paper on hidden perception constraints in task-aware compression
SummaryThis paper examines perception constraints that arise naturally in task-aware compression, where reconstruction is the primary task and classification, a statistical test, is secondary. The authors study the problem under varying levels of available domain information and show how these constraints can guide the design of rate-minimal compression schemes that also maximize secondary-task utility. They report that when classifier decision boundaries are ill-defined for the source distribution, matching the reconstruction to a target distribution improves classification accuracy.
Sahan Liyanaarachchi, Semih Akkoc, Sennur Ulukus et al.
In briefHEAR: A large-scale benchmark of real human recordings for testing voice bias in audio language models
SummaryThe paper introduces HEAR, a benchmark of 87k real human audio samples from 843 demographically diverse speakers, for evaluating audio language models. It supports multiple-choice and open-ended long-form tasks and tests both speech-to-speech and speech-to-text models. The authors report that voice-conditioned bias varies by model, and that personalization instructions consistently widen demographic disparities.
In briefVerifier Errors in RLVR: Reward Hacking and Selective Control
SummaryThis paper studies reinforcement learning with verifiable rewards (RLVR), where imperfect verifiers can reward incorrect answers, enabling reward hacking. The authors characterize when reward rises while correctness falls, and show that observations available during RLVR generally cannot detect or identify accepted errors. They propose a correction using additional correctness feedback from audits, achieving selective control that lowers accepted-error probability while raising correct-response probability. Experiments on log-linear and neural contextual bandits and a language model support the analysis.
In briefFlowTool: Controlling Image Retouching Tool Parameters with Flow Matching
SummaryThis paper proposes FlowTool, which frames tool-based image retouching as a flow matching problem, generating tool parameters directly from the input image and user instruction. Instead of autoregressive multimodal language models producing reasoning and parameters step by step, it uses a vision-language backbone with a Diffusion Transformer that turns Gaussian noise into an editing plan. The authors report stronger reference-based results than specialized editing agents and proprietary models on several benchmarks, with at least 50× lower latency and nearly half the memory.
Thanh-Long V. Le, Steven Walton, Seunghyun Yoon et al.
In briefPhoneCLI: Compiling App Screens into Callable Commands for Mobile Agents
SummaryThis paper presents PhoneCLI, a method that turns a mobile app's GUI navigation into callable commands for mobile agents. Offline, it explores the app from outside and builds an annotated map of screens, elements, and navigation edges, with one deterministic replay command per screen. Online, the agent selects and verifies a command, executes it quickly without vision-language model calls, and falls back to the vision-language model for open-ended interaction or failures. On AndroidLab it improves task success while reducing steps and tokens, and it transfers efficiency gains to AndroidWorld's M3A agent.
In briefLearned preconditioning for a primal-dual interior-point optimization method
SummaryThe paper introduces pdLIP, an interior-point method for smooth nonlinear programs that uses a learned, coordinate-wise recurrent network to predict a positive diagonal preconditioner, avoiding Hessian evaluations and Newton-system solves. Training is self-supervised, using a penalty-barrier merit function and perturbed optimality residuals rather than target directions. On four classes of 200-dimensional problems, warm starts reduced pdProj refinement iterations by 63-67% versus cold starts at a KKT tolerance of 1e-8.
In briefMS-GLA: Multi-Scale Gated Linear Attention Tackling Representational Bottlenecks Through Multiple Temporal Resolutions
SummaryThis paper proposes Multi-Scale Gated Linear Attention (MS-GLA), which spreads attention heads across several temporal resolutions so that coarse heads capture long-range dependencies and fine heads keep local syntactic detail. A learnable, input-dependent fusion layer recombines the head-group outputs at each timestep. Evaluated on language modeling, recall-intensive tasks, and long-context generalization, MS-GLA outperforms GLA at matched parameter counts, with up to 18.9% gains on recall tasks and 9.5% lower average perplexity on language modeling.
In briefLate-layer attention alone can copy entity tokens, but only with context-token attention
SummaryThis paper studies how large language models copy tokens that refer to an entity from a prompt into their output. Using Qwen3-8B, the authors apply two new methods: one controls which layers may take part in copying, and the other cuts specific tokens' attention to entity tokens. They find that two groups of layers in the model's second half are both necessary and sufficient for entity copying. Context tokens' attention to entity tokens is also necessary for copying the exact tokens.
In briefSingle-pass 3D reconstruction from mixed perspective, fisheye, and panoramic cameras
SummaryMEOW is a feed-forward system that reconstructs metric pointmaps and camera poses from a set of images mixing perspective, fisheye, and full-panorama views in one forward pass. It needs no calibration, camera-type labels, or poses. The model keeps a perspective-pretrained backbone and learns heterogeneous cameras from a procedural data engine trained only on synthetic data. On real 2D3DS tuples it reaches 79.9 mAA@30, compared with 53.8 for Wid3R, which is given each view's camera type.
In briefCMDO: A Memory-Driven Algorithm for Adaptive Population-Based Search
SummaryThis paper introduces CMDO (Cognitive Memory-Driven Optimization), a derivative-free population-based optimizer. It stores search context, behavior, and outcome as experience across working, episodic, and consolidated memory, and retrieves it by similarity to the current search state. The method uses both successful and unsuccessful experiences to guide later search, generating candidates through exploratory, directed, and local behaviors. Tests on BBOB/COCO and CEC2017 problems and photovoltaic parameter estimation show problem-dependent but competitive performance, and the authors report that recalled experience changes the distribution of executed search behaviors.
Mohammed Yusuf Mujawar, Shahram Rahimi, Noorbakhsh Amiri Golilarz
In briefDenoising-Based Planning for Multi-Robot Trajectories (D4orm)
SummaryThis paper builds on D4orm, a dynamics-aware diffusion-denoising framework, to plan trajectories for multiple robots. It iteratively deforms candidate control trajectories through massively parallel sampling on GPUs to produce kinodynamically feasible, conflict-free paths. The authors present decoupled, online receding-horizon, and distributed planners, and report better and faster results than MPPI and a learned diffusion method, including real quadrotor and ground-robot experiments.
WhyBuilds decoupled, online, and distributed planners on D4orm, compared against MPPI and learned diffusion baselines, with real quadrotor and 100-robot simulation validation.
In briefReference-guided test-time specialization for transferable mass spectrum prediction
SummaryTandem mass spectrum prediction helps identify compounds in metabolomics, natural-product discovery, and environmental analysis. This paper introduces SPARC, which adapts a pretrained predictor at test time using a spectral reference library, without access to test-query spectra. For each query, it retrieves chemically related reference spectra to recalibrate fragment intensities and uses reliability-aware consistency to preserve trustworthy predictions. The authors report improved spectral prediction across multiple transfer settings on MassSpecGym, NPLIB1, and GNPS libraries.
In briefMedical image alignment checking as a test of general visual reasoning in advanced AI models
SummaryThis paper uses medical image alignment, judging whether two images show matching anatomy, as a test of whether frontier multimodal models can make visual judgments without task-specific training. The authors compare several model generations, a locally fine-tuned model, and a task-specific CNN, varying prompts and image presentation. They report that recent frontier models can now perform this assessment effectively, while models released only months earlier generalize poorly, and fine-tuned local models transfer much less well to unseen settings.
Ross Callaghan, Niannu Gao, Hojjat Azadbakht et al.
In briefRubric Response Theory (RRT): Rewards for Rubric Criteria from Item Response Theory
SummaryThe paper addresses how to turn rubric verdicts on language model responses into a reinforcement learning reward, noting that summing fixed criterion points can give identical rewards to different verdict patterns. It proposes RRT, which treats verdict patterns as evidence of quality using a two-parameter item response model, with a Response Parameter Network predicting criterion difficulty and discrimination, updated online via expectation maximization during training. With Qwen3.5-4B as the policy, RRT's macro criterion score is reported to be 1.7 points above GRPO, and at half the criterion budget it stays within 0.1 points of GRPO with full judging.
In briefCoSE-E: A Benchmark for Testing Speech Recognition When Speakers Mix Languages in Business Settings
SummaryThis paper proposes a synthetic benchmark and multidimensional evaluation framework for code-switched speech recognition tailored to enterprise domains. It systematically evaluates frontier ASR systems across 5 language pairs and analyzes the extra transcription errors that code-switching introduces. The authors release CoSE-E to support enterprise-focused evaluation of multilingual voice agents.
In briefNot All Thinking Is Equal: Latent Reasoning Learns a Recurrent Search Method That Generalizes to Deeper Problems
SummaryThe paper trains five variants of the same GPTNeoX model from scratch on an extended multi-hop reasoning task (ProsQA-Ext). It finds that strong in-distribution performance does not guarantee generalization to longer reasoning chains, with vanilla, chain-of-thought, and pause-token models generalizing poorly while latent-reasoning models generalize better. Causal interventions and circuit analysis show the latent model implements a sparse recurrent forward-search circuit that works beyond the training depth.
WhyComparing five models with causal interventions and circuit analysis, it shows latent recurrence learns a depth-generalizing forward search algorithm, with fairly solid evidence.
In briefWhen a Spoken Agent Has Enough Evidence to Act: The PACT-SLM Test
SummaryThe paper introduces PACT-SLM, a controlled test for streaming spoken language models that checks whether a correct action is taken from partial speech and at what time. Using 80 paired contrast groups and 1,600 prefix predictions, it finds that action identity and onset timing are measured differently. The authors argue these should be evaluated separately.
In briefVerifiable Visual Rewards: Training Image Generators on Synthetic Scenes for Natural Prompts
SummaryThe paper introduces Verifiable Visual Rewards (VVR), which builds synthetic scenes of geometric objects and relations to produce both a prompt and a deterministic verifier for scoring image generation. The authors release VVRBench with 10,000 tasks and VVRBench-Challenge with 720 harder tasks. Using VVR scores as reinforcement learning rewards raises Stable Diffusion 3.5 Medium's accuracy on VVRBench from 2.8% to 28.3%, and the gains reportedly extend to natural prompts and out-of-domain benchmarks.
Shuyue Stella Li, Xiaochuang Han, Yulia Tsvetkov et al.
In briefGPUPhysBench: Testing whether coding agents can write correct, fast GPU physics simulation code
SummaryThis paper introduces GPUPhysBench, a benchmark of 50 tasks covering fluids, deformable solids, and granular materials, to test whether coding agents can write correct and efficient GPU physics simulation code. In a single-attempt evaluation of six frontier model-harness pairs, the two strongest passed all 50 tasks. However, even the fastest reached at least 0.9 of the expert reference speed on only 22% of tasks, and no submission was more than 5% faster than the reference. The largest gaps appeared in collision detection, constraint solving, and iterative solvers.
In briefRT-Super: Learning Tumor Segmentation from Scans and Reports Over Time
SummaryThis paper introduces RT-Super, a model for segmenting tumors in medical images without relying on costly manually drawn tumor masks. A teacher network uses longitudinal images, multi-phase images, and radiology reports to generate tumor masks, which then train a student network that works from a single image alone. The authors also propose new consistency losses that exploit tumor location consistency across longitudinal images. The method is reported to segment esophagus, uterus, and spleen tumors, where public masks are scarce, and to outperform public AI models.
Pedro R. A. S. Bassi, Wenxuan Li, Hanxue Gu et al.
In briefUsing coding LLMs to learn interpretable models from data
SummaryThe paper proposes a simple harness that repeatedly prompts a coding LLM to generate the next rule in a rule set, given labeled examples. On tabular data it is competitive with and often better than state-of-the-art rule learners, especially with small training sets. On text it wins by large margins, including against another LLM-based rule learner. The authors also propose interpretable classifier construction as a new coding benchmark.
Aditya Kumar, Cassandra A. Cohen, William W. Cohen
In briefCurvature-Aware Associative Memory on Riemannian Manifolds
SummaryThis paper builds intrinsic dense associative memories on Riemannian manifolds, treating memory as mode seeking over Epanechnikov kernel densities. It compares geodesic and volume-corrected energies and shows that curvature determines which one retains memories, with geodesic memory always keeping isolated patterns. The authors also derive capacity scalings and show that overlapping patterns can create new memory modes.
In briefDR-net-Mamba: Long-range ECG denoising using selective state-space models
SummaryECG recordings are often corrupted by non-stationary noise, which reduces diagnostic reliability. The paper proposes a model that inserts Mamba selective state-space blocks at the convolutional bottleneck, combining local feature extraction with long-range modeling at linear complexity. On synthetic and real data it reports the highest SNR and lowest RMSE, with gains growing for longer sequences and low-SNR conditions. Downstream classification shows the best macro AUROC among denoisers, while calibration results depend on the classifier.
Basile Morel, Samuel Ruiperez-Campillo, Andreas P. Streich et al.
In briefFluxLite: Inference-Time Proposal Control for Discrete Diffusion Models
SummaryThis paper presents FluxLite, a training-free framework that controls the proposal dynamics of discrete diffusion models at inference time, so that sampling from a tilted distribution becomes more stable. It introduces two practical samplers and proves population stability under the standard score-entropy loss, along with finite-particle convergence. On an analytically tractable finite-state CTMC benchmark, it improves terminal KL over standard Feynman-Kac SMC baselines by up to two orders of magnitude. On 2D Ising sampling, it reduces row-correlation MSE by 5-7x in geometric mean and up to 55x at peak.
In briefWhich the Eye Fears: Read-Blindness Explains Massive Activations in Transformers
SummaryThis paper examines why massive activation features, extreme-valued residual-stream features in Transformers, persist across layers. Operator-level analysis of attention and feed-forward blocks finds that these blocks ignore those coordinates when reading but not when writing, a read-write asymmetry that blocks corrective feedback. Checkpoint and gradient analyses indicate this read-blindness emerges before feed-forward amplification and is actively maintained by the model. Removing read-blocking at different locations causes compensatory shifts, but the massive activations still persist.
Swagatam Mukhopadhyay, Vishal Vivek Saley, Vraj Parikh et al.
In briefSANTA++: Sampling Attention via Representative Keys (a training-free method that reads fewer cached keys in LLM attention)
SummaryThis paper introduces SANTA++, a training-free stochastic attention method that groups cached keys into teams and uses one representative key per team to decide which teams to sample. Exact attention is computed within sampled teams and reweighted by inverse inclusion probability to estimate attention over the full cache. With Qwen2.5-7B-Instruct at 32K context, the authors report retaining 94% to 99% of dense-attention scores on LongBench v2 and HELMET's RAG subset, and 85% to 91% on RULER, while reading 16% to 22% of the KV cache.
In briefNeural network approximation to arbitrary accuracy with minimal neurons and near-optimal bit cost
SummaryThis paper studies how few hidden neurons are needed to approximate multivariate Hölder-continuous functions on [0,1]^d to arbitrary accuracy, and the complexity of encoding them. The authors construct a fixed, explicit activation function for which a two-hidden-layer network with widths d and 1 reaches arbitrary uniform accuracy. They prove that d+1 is the exact minimum number of hidden neurons for standard feedforward networks. For bounded Hölder classes, the required bit count matches the metric-entropy lower bound up to a logarithmic factor.
In briefCan LLMs Judge the Value of Medical Evidence? Evidence-Value Misalignment in Dynamic Diagnosis
SummaryThe paper argues that a correct diagnosis based on insufficient or misleading evidence can still be clinically hazardous, and that accuracy alone may reward such lucky guesses. The authors introduce MedEVM, a dynamic benchmark of 1,050 cases across 24 disease systems, where models must decide turn by turn whether to wait for more evidence or submit a diagnosis. Testing nine LLMs, they report miscalibrated evidence tracking, misaligned submission timing, order effects on diagnoses, and persistent influence from misleading evidence. They also propose EVD-Harness, which verifies evidential support before submission and improves accuracy by 12.0 to 51.1 percentage points across five LLMs.
WhyProposes evidence-value misalignment with a dynamic benchmark and verified-submission harness; nine-LLM experiments and controls give solid evidence for medical AI impact.
In briefRIDE: Reference-Anchored Diffusion Editing for Finding New Drug Scaffolds
SummaryScaffold hopping seeks new molecules that are structurally different from a reference binding ligand yet keep its key functional groups and similar 3D shape. The paper introduces RIDE, which recovers the reference diffusion noise trajectory conditioned on the binding pocket and functional groups, selects a trajectory segment for editing via noise perturbation, and performs value-guided scaffold sampling. Compared with baselines, RIDE generates scaffolds with lower 2D similarity and higher 3D similarity to the reference, with average improvements of 11.7% and 7.3%, respectively.[
In briefLearning Optimal Decision Boundaries in Single-Index Bandits via Natural Boundary Learning
SummaryThis paper studies two-arm contextual bandits in which each arm has its own single index and both share an unknown monotone link. Because of monotonicity, the optimal action depends only on the difference between the index directions, so the arm-specific reward functions need not be estimated. The authors propose Natural Boundary Learning (NBL), a greedy procedure that learns the optimal boundary directly using a sequential Stein contrast. Under local decision stability, the paper shows NBL converges to the optimal boundary with expected regret of order O(log n). Experiments illustrate the predicted stability regimes and compare NBL with a parametric greedy benchmark under link misspecification.
In briefCartridges++: KV cache compression that keeps off-context query ability
SummaryThe paper studies compressing the key-value cache used when a large language model repeatedly processes long documents, to cut computation and memory costs. It finds that the learned Cartridges method answers document-related queries well but loses ability on queries outside the document. The authors propose Cartridges++, which uses an inference-time router or mixes a small set of off-document training questions to preserve that ability.
In briefAttention Graphons: a graph-limit view of graph Transformers
SummaryGraph Transformers produce a dense n×n learned interaction matrix for each attention head. The paper treats each attention matrix as a finite sample from an underlying kernel, called an attention graphon, and studies how it concentrates around that limit under cut-distance as n grows, deriving a worst-case and a regularity-aware variance bound. It proposes a canonicalize-then-block-average estimation pipeline and a variance-based diagnostic for whether attention has a stable limit. Experiments on several graph benchmarks report that attention stabilizes to dataset-specific graphon structure and that error decreases with graph size.
Caio F. Deberaldini Netto, Moshe Eliasof, Luana Ruiz
In briefFrom Cacophony to Hierarchy: A Principled Framework for AI Consciousness Assessment
SummaryThis paper proposes a framework for assessing whether AI systems may have consciousness, separating the hard problem from the mapping problem. The authors extend Marr's three levels of analysis into a five-level hierarchy of functional descriptions, place major theories within it, and define indicators at each level. A Bayesian model combines theoretical credences with indicator evidence, and illustrative assessments of current LLMs range from below 0.01 to about 0.8 depending on assumptions.
Shamil Chandaria, Arvo Mu\~noz Mor\'an, Fernando Rosas et al.
In briefEvolvingAvatar: a 3D head generator that adapts during live conversations
SummaryThis paper presents EvolvingAvatar, a causal 3D head generator that uses test-time training to adapt to a user's face video and dyadic audio during interaction. It learns from audiovisual context through a self-supervised objective and keeps fast weights that accumulate updates within each conversation. The authors also introduce InterHead-Bench, a 455.95-hour benchmark, and report improved conversational motion statistics over strong baselines.
In briefBehavioral Foundation Models for Quality-Diversity Search in Reinforcement Learning
SummaryThis paper introduces BFM-QD, a framework that runs Quality-Diversity search inside the compact latent space of Behavioral Foundation Models instead of raw policy parameters. The authors also derive a closed-form, gradient-free policy improvement operator that approximates a policy gradient update without critic training or backpropagation. On continuous-control benchmarks, BFM-QD consistently outperforms parameter-space baselines, with especially large gains in sparse and deceptive settings where those baselines collapse to near-zero performance.
Nazim Bendib, Nicolas Perrin-Gilbert, Olivier Sigaud
In briefEvE: A Faster Alternative to Adam for Searching Neural Network Configurations
SummaryThis paper introduces EvE, an optimizer that mainly uses differential evolution and falls back to brief gradient descent only when needed, aiming to rank training configurations cheaply. Under matched computational budgets, it matches or beats Adam on 76% of 70 benchmark cells and finishes neural network training runs 1.7 to 3.9 times faster, with slightly lower final quality. The authors position EvE as a fast proxy for search-heavy, budget-constrained settings rather than a full replacement for Adam as a final-stage trainer.
In briefSparse-view 3D Gaussian splatting for remote sensing using depth-based pseudo views
SummaryThis paper addresses novel view synthesis from sparse remote sensing images, where limited observations make reconstruction difficult. The proposed DIBR-GS framework uses Depth Image-Based Rendering to create pseudo views that supervise cross-view consistency, with geometry initialized by aligning monocular depth priors to sparse SfM reconstruction. It also adds progressive pseudo-view supervision and a height-constrained anchor growth strategy. With only 3 input views, the authors report a 6.83 dB PSNR gain over the previous best method, plus 14% relative SSIM and 60% relative LPIPS improvements.
In briefOn-Policy Self-Distillation for Multi-Turn Image Editing: Training Models to Keep Editing Their Own Outputs
SummaryThis paper finds that instruction-based image editing models degrade quickly when edits are applied repeatedly, because they are trained on clean source images but must condition on their own imperfect outputs at inference. The authors propose MT-OPSD, a self-distillation framework that trains on the model's self-generated conditioning states, supervised by a teacher conditioned on clean inputs, without needing multi-turn annotations. They also introduce LME-Bench, a benchmark of 100 ten-turn editing sessions for evaluating long-horizon robustness. Experiments on three editing backbones report improved long-horizon editing success and reduced multi-turn collapse, with single-turn quality largely preserved.
In briefTwist, Don't Tilt: A trajectory-exact method for constrained decoding in masked diffusion language models
SummaryThis paper studies constrained decoding for masked diffusion language models, which aims to make generated outputs satisfy a given structure or syntax rule. The authors prove that composing existing step-exact automaton-based samplers introduces trajectory bias, deviating from the model's relative probabilities over valid trajectories, and derive an exact expression for this bias. They propose TWISTER, the first automaton-twisted Sequential Monte Carlo decoder for these models, which corrects the bias and is shown to target the unbiased Doob h-transformed path law conditioned on constraint satisfaction.
Aditya Thimmaiah, Lara Marinov, Jayanth Srinivasa et al.
In briefReal-time translation between sign languages for Deaf and hard-of-hearing signers
SummaryExisting sign-to-sign translation systems work offline and need the full source clip before producing output, which blocks live conversation. This paper presents what it describes as the first simultaneous sign-to-sign translation system, using two wait-k approaches and a new computation-aware latency metric, ca-Stream-AL. Averaged over six translation directions, the streaming system reduces ca-Stream-AL by 38% at the cost of a 9% increase in DTW-PA-MPJPE and a 2.1 BLEU-4 drop relative to the full-sentence baseline.
In briefTCSAlgBench: A benchmark testing AI models on research-level proofs in theoretical computer science.
SummaryThis paper introduces TCSAlgBench, a benchmark and reusable pipeline for natural-language proof discovery, built from 398 theorem-level challenges drawn from 138 STOC and COLT 2026 papers. The best-performing model reaches 23.6% verifier-accepted coverage over five runs after 10 rounds of discussion, and discussion and repeated sampling improve results. In an agent workflow comparison, agentic planning achieves the highest coverage at 25.4%.
In briefStraightening control geometry to make latent planning easier
SummaryThis paper proposes Control-Geometry Straightening (CGS), an auxiliary loss that aligns pairwise cosine similarities among actions with those among corresponding latent differences to learn planner-friendly representations. Under linear dynamics, the authors provide theoretical analysis linking the objective to temporal straightening and convergence results for several planners. Across four control environments and multiple planners, CGS improves success rates with fewer sampled candidates and refinement steps, with reported gains of up to 20 and 12.6 percentage points over LeWM and LeWM+TS.
In briefEvidence that LLM internal activations encode explore and exploit modes in semantic search
SummaryThis paper examines whether large language models, when listing concepts in a semantic fluency task, internally represent the same explore and exploit search modes found in humans. Using interpretability methods such as the Jacobian lens, the authors show that activations predict when the model switches between categories. They also show that steering these activations changes how often switching occurs.
WhyExtends semantic foraging to LLM internals, causally validating switching via J-lens and steering vectors; novel and well-evidenced, but impact is mainly within interpretability.
In briefLearning conditional expectation operators with functional Newton updates
SummaryThis paper introduces the Functional Spectral-Newton Method (FSNM), which learns the leading singular structure of a conditional expectation operator without fixing a basis or reproducing kernel Hilbert space. It fits a low-rank representation of the centered joint-to-product density ratio kernel through alternating functional Newton updates, approximated by vector-valued regression trees in a boosting procedure. The authors prove descent and an O(1/T) best-iterate block-stationarity rate under a relative weak-learner accuracy condition, and that nondegenerate local minima are globally optimal rank-d approximations. Synthetic experiments show recovery of a low-rank density ratio and its spectral structure, and reuse of the learned kernel for multiple conditional queries without refitting.
Thiago Ramos, Alek Fr\"ohlich, Daniel Perazzo et al.
In briefReSS: Residual-Restoring Sparse Attention for 3D Vision Transformers
Summary3D vision transformers such as VGGT estimate camera poses and scene geometry from multiple images in one pass, but their attention cost grows quickly with the number of views. This paper proposes ReSS, which selects blocks to drop by minimizing the drift that sparsification leaves in the residual stream, rather than by keeping high attention probability. It uses a drift score and an iterative residual restoration procedure that refines the drop set as a whole. Across three backbones and five datasets, ReSS preserves dense-model performance better than prior methods at matched sparsity.
WhyRecasts block selection as minimizing residual-stream drift with iterative restoration; validated across three backbones and five datasets, a useful advance for this subfield.
In briefLanguage Models' Hidden Valence and Choice: Revealed Preference Study
SummaryThe paper uses activation steering to attach positive or negative valence patterns to two meaningless zones, then removes steering and observes which zone models prefer, testing revealed preference instead of self-report. Across seven open-weight models from five families, hidden-state valence shifts later choices in proportion to steering dose, even with identical visible tokens. Models also tend to remove imposed negative states at a dose-dependent rate, and the authors note it remains unclear whether these traces involve subjective experience.
WhyNovel revealed-preference design using activation steering and KV cache controls; seven-model experiments are solid, though welfare implications remain limited.
In briefCoDR: A Training-Free Method That Re-masks Language Model Tokens Whose Confidence Drops
SummaryMasked diffusion language models decode by committing tokens to masked positions, and these commitments are usually irreversible, so early mistakes can propagate. The authors identify confidence drift, where the model's confidence in a committed token falls when it moves from sparse early context to denser later context. They propose CoDR, a training-free, sampler-agnostic refinement pass that estimates drift in only k forward passes and remasks and regenerates tokens the model no longer endorses. The paper reports accuracy gains across two backbones, four reasoning and coding tasks, and three base samplers.
WhyProposes training-free confidence-drift remasking with broad experiments across backbones, tasks, samplers, plus controlled ablations supporting targeted gains.
In briefSource-preserving alignment method for locating evidence in scientific PDFs
SummaryScientific extraction systems often return a claim with an evidence string, but that text differs from the PDF's text layer due to line wrapping, Unicode variants, superscripts, and citation markers. The paper proposes a framework that normalizes text for matching, maps matches back to source-character spans, and highlights only their geometry. On 1,020 chemistry papers, it achieved a 92.6% quote-level automatic localization rate, versus 43.6% for text search and 19.1% for a precomputed bounding-box baseline.
In briefHardware-aware features for choosing GPU kernels in CUTLASS
SummaryThe paper addresses choosing among tens of thousands of equivalent GPU kernels in CUTLASS without costly autotuning. It adds statically computable estimates of hardware behavior to each candidate configuration and trains ranking models on a dataset of 4.9 million kernels. On held-out evaluation problems, the representation reduces selection regret by up to 40% versus structural baselines and 64.2% versus NVIDIA's matrix-multiply heuristics.
Shriram Chandran, Dominic Rinderer, Yakup Budanaz et al.
In briefIMC-CLINIC: A clipping calibration method for analog in-memory computing that jointly optimizes activation and weight clipping factors.
SummaryAnalog in-memory computing performs matrix multiplications inside memory arrays to save energy for LLM inference, but limited-precision converters add output error. The paper introduces IMC-CLINIC, which uses an analytical surrogate of IMC output error and a safeguarded Newton-type method to jointly optimize activation and weight clipping factors. Across several models and datasets, it improves average zero-shot accuracy by 6.5 to 11.5 points over grid search while cutting calibration time by 10.0 to 12.1 times.
In briefWhat Paired Evaluations Reveal under Visual Perturbations
SummaryThis paper studies how paired clean and perturbed predictions for the same image can support robustness evaluation under visual perturbations. The authors prove that classwise correct-correct counts yield sharp bounds on lost acceptance and true-class probability decrease, and they rank samples for physical testing using synthetic corruption responses. Across 44 classifiers, testing the highest-risk 20% of images recovers more failures than clean-confidence or natural-transformation baselines.
In briefOutput-aware pruning of residual streams in large language models
SummaryThis paper proposes a residual-stream pruning method for large language models that accounts for how downstream layers sensitively respond to removed dimensions. Existing methods choose dimensions by minimizing activation reconstruction error alone. The authors use a second-order approximation of output KL divergence to combine activation covariance with local output sensitivity, and derive a tractable spectral bound that reduces selection to an eigendecomposition. Across several instruction-tuned model families, the method lowers calibration KL divergence versus activation-only pruning and improves perplexity and downstream performance across compression levels.
In briefFactorEngram: Factorized N-gram Memory with Basis-Level Gating for Language Models
SummaryThis paper proposes FactorEngram, a factorized n-gram memory for large language models. Instead of storing each retrieved embedding as a single unit, it retrieves sparse coefficients over a dictionary of basis vectors shared across patterns, and gates each coefficient using the backbone hidden state. On 340M- and 1B-parameter Transformer backbones, it improves language modeling and downstream task performance. Ablations support the contribution of each component and identify insertion before the attention sublayer in middle layers as effective.
In briefF4R: A Framework That Lets Robots Improve Themselves by Learning from Their Failures
SummaryThe paper proposes F4R, a closed-loop framework that improves vision-language-action robot policies by turning real-world failures into targeted training. An agent identifies and diagnoses failures from robot rollouts, then rebuilds each one as an object-centric tabletop simulation. The policy is refined through sim-real co-training and targeted reinforcement learning, then redeployed. On four manipulation tasks, the authors report 93.75% in-distribution and 90.0% out-of-distribution success, without collecting extra real-world corrective demonstrations.
In briefLess Is More: Genetic Frame Selection for Efficient Novel View Synthesis
SummaryThis paper proposes a render-free view selector that chooses a fixed-size subset of captured input frames most useful for reconstructing target viewpoints, scoring frames by target coverage, redundancy with selected views, and sharpness. The selector is trained by distilling a genetic algorithm's offline search, which optimizes reconstruction quality, and uses only geometric and image features at inference. The authors report consistent gains over geometric and reconstruction-aware baselines across six datasets, with some selected subsets outperforming full-sequence feed-forward reconstruction.
Diego E. Farchione, Ramzi Idoughi, Alberto Jaspe-Villanueva et al.
In briefA study showing that representation alignment, not model capacity, is the main bottleneck for LLM-based retrosynthesis planning
SummaryThis paper examines using large language models for symbolic retrosynthesis planning in chemistry. The authors find that direct SMILES-to-PDDL generation fails, and they decompose the task into molecule mapping, reaction mapping, and PDDL generation. This decomposition achieves high success rates where end-to-end approaches fail, leading the authors to conclude that representation alignment is a primary bottleneck rather than raw model capacity.
In briefEdgeVLN: A Quantized Vision-Language Navigation Model Ready for Edge Deployment
SummaryThis paper presents EdgeVLN, a quantized vision-language navigation model designed to run on memory- and power-limited robotic edge devices. It pairs a quantized StreamVLN backbone with LATTE, a lightweight causal transformer that improves real-time stopping, and runs through a llama.cpp-based driver. On an NVIDIA Jetson Orin NX 16 GB, the 4-bit IQ4 NL model with LATTE reaches a 58.02% success rate in simulation and fits within the board's memory.
In briefBeyond Energy: How Sustainability Dimensions Change LLM Serving Decisions
SummaryThis paper presents PRISM, a framework that jointly characterizes and optimizes large language model serving across energy, carbon, water, and biodiversity impacts. It finds that computing configurations determine energy use, while deployment location and timing determine carbon, water, and biodiversity impacts. In regional-routing experiments, PRISM reduces median worst-case regret by 50.2% relative to the strongest baseline.
In briefFrom Experience to Expertise: Adoption-Aware Memory Learning for NPU Kernel Synthesis
SummaryThis paper presents SAGE, a persistent self-improving agent that generates kernels for NPUs, where LLMs trained on CUDA transfer poorly due to limited data. It combines Adoption-Traced Utility estimation, which credits experiences using adoption records and evaluation outcomes, with Utility-Gated Consolidation, which keeps reusable rules in a bounded resident context. On NPUKernelBench, SAGE reports a 95.5% execution rate, compared with 84.1% for the strongest controlled baseline.
In briefAlmieyar: A Culturally Grounded Arabic Speech Recognition Benchmark for Multiple Dialects
SummaryThis paper introduces ALMIEYAR, a speech recognition benchmark covering 17 Arabic dialects from six families, built from newly recorded speech unseen by existing models, totaling about 13.7 hours. Twelve state-of-the-art systems were tested zero-shot, with GPT-4o-transcribe achieving the lowest overall word error rate at 35.0%. Performance varies across dialect groups, and the authors argue that word error rate alone hides behaviour differences, so they recommend reporting both WER and CER. The benchmark includes the first published benchmark for Ahwazi Arabic.
Omid Ghahroodi, Anas Madkoor, Dima Faris Al Saudi et al.
In briefRe-examining Risky Tackle Detection with Vision Transformers
SummaryThis paper is a reproducibility companion to an ICPR 2026 study on detecting risky tackles in American football practice videos. The original work fine-tuned a ViViT model on 733 clips labeled with the SATT-3 rubric. The companion documents the released artifact and traces reported results to specific scripts and fold outputs. The reproduced results show brightness reduction as the dominant factor, and without augmentation ViViT's risky-class recall is 0.545, not exceeding the C3D baseline of 0.583. Raw clips cannot be redistributed because they show identifiable athletes, so a public sample is provided for pipeline checks.
Syed Ahsan Masud Zaidi, Lior Shamir, Scott Dietrich
In briefRSI-Master: Structuring Experiments to Guide AI's Autonomous Model Improvement
SummaryThis paper introduces RSI-Master, a framework that lets AI agents autonomously explore post-training strategies to improve a base model. It regularizes experimental actions to prevent hacking and organizes Worker and Reviewer agents in a growing research graph to avoid getting locked into one early direction. On PostTrainBench with Qwen3-4B-Base, it averages 54.49 versus 46.53 for the strongest baseline, with a 0.0% hacking rate. At 35B scale, it outperforms the human-developed Instruct model on LiveCodeBench-v6 and SciCode.
In briefEnJoi: an ensemble joint score filter that makes generative models fit partial observations of dynamic systems
SummaryThis paper proposes EnJoi, a diffusion-based data assimilation method for reconstructing the full state of a dynamical system from partial observations. It learns the joint distribution of past and future states and adapts the classical En4DVar algorithm, which uses ensemble covariance, to balance confidence in the current state against new observations. Experiments on fluid and traffic flow simulations report better reconstruction, especially when observations are sparse and uneven.
In briefWorldPlay2: Extending Real-Time Interactive World Models in Control and Horizon
SummaryThis paper presents WorldPlay2, an interactive world model that responds to varied controls in real time while staying consistent over long sequences. It pairs a factorized hybrid control interface, which separates scene appearance, character identity, and dynamic events, with compressed memory tokens and a stable distillation method called Stable Forcing. The authors report strong generalizability and better performance than existing methods in their experiments.
In briefFrom Search to Research: Studying how larger search budgets affect autonomous factor mining by LLM agents
SummaryThis paper examines how increasing search budgets affects the research performance of autonomous LLM agents, a practice it calls search scaling. The authors test nine models on 50 quantitative factor-mining tasks drawn from financial research reports. They find that initial performance tracks model capability while deeper search narrows gaps between models, and that parallel search outperforms sequential search under the same iteration budget.
In briefHERO: A Robust Histology Encoder for Oncology Pathology Images
SummaryThis report introduces HERO, a ViT-G/14 pathology foundation model trained with DINO and iBOT objectives on 500 million tiles from about 575,000 clinical whole-slide images. It focuses on robustness to differences in center, scanner, and staining. According to the authors, HERO shows the strongest robustness to these variations among compared models, performs comparably on tile-level tasks, and ranks first on average across 39 slide-level clinical tasks.
WhyAdds high-resolution Gram anchoring to pathology foundation models for robustness, validated across many benchmarks and 39 clinical slide-level tasks.
Zhi Li (1), Eghbal Amidi (1), Yating Cheng (1) et al.
In briefHow fish stay together while still exploring: experiments with robotic fish and a new model
SummaryThe authors tested a live fish swimming with four robotic fish and compared classical fish-school models with the observed positional behavior. Classical models matched the data only for very specific parameter values and failed under small parameter changes. They proposed a new model with an extra exploratory term for positional preference, which was robust to parameter errors and reproduced schooling patterns and both individual and group exploratory behavior.
Vahagn Grigoryan, Donato Romano, Cesare Stefanini et al.
