In briefCoTrace: Data recipes for training terminal agents with harness and model co-evolution
SummaryThis paper argues that terminal agent performance depends on both model weights and the runtime harness that handles prompts, tools, and error recovery. It proposes CoTrace, a framework that routes trajectories by harness and trains only on verified rollouts matched to the adopted runtime. On the Tmax split, Qwen3.5-9B's solved tasks rise from 78 to 88 with supervised fine-tuning and to 90 with an online reinforcement variant.
In briefGraphRectify: Reusing adversarial example detectors across different neural network backbones using graphs
SummaryThis paper introduces GraphRectify, a graph-based framework for transferring adversarial image detectors from one classifier backbone to another. It learns a structured graph representation of intermediate classifier features and adapts features from a new backbone to a detector trained on the original model, so the detector can be reused. The method was evaluated across multiple datasets, backbone architectures, and attacks, including adaptive attacks that target both the classifier and detector. It achieved higher aggregate ROC-AUC than training a detector from scratch on the new backbone, though training from scratch remained competitive when data was very limited.
In briefWhich training rollouts taught a language model a behavior, and can attribution be trusted in RL?
SummaryThe paper studies whether training-data attribution can identify the GRPO online reinforcement learning rollouts that taught a language model a new behavior. The authors release BehaviorTrace, an evaluation harness using a planted behavior with known cause, with controls for gradient magnitude, fluency, and seed variation, tested on Qwen2.5-1.5B. Much of the apparent attribution signal comes from confounds, and only the trigger-token gradient alignment holds across all three seeds. The authors turn these findings into an evaluation checklist and test existing estimators without proposing a new one.
In briefSteerSpeech: Activation Steering for Controlling Emotion in Generated Speech
SummaryThis paper presents SteerSpeech, a lightweight framework that controls emotion in text-to-speech output by injecting steering vectors into the model's hidden activations. For each target emotion, a low-rank transform is trained while the TTS backbone stays frozen. Evaluations on Qwen3-TTS report 1.08x to 7.12x higher target-emotion scores than the baseline, with limited degradation in speaker identity and content.
Afsara Benazir, Darius P\'etermann, Felix Xiaozhu Lin et al.
In briefTraining parallel speculative draft models by directly minimizing expected decoding rounds
SummaryThis paper addresses how to train parallel and semi-autoregressive draft models for speculative decoding, noting that existing block-local objectives ignore coupling across decoding rounds. The authors model speculative decoding as a Markov reward process and propose the Expected Decoding Rounds (EDR) objective, along with an exact temporal-difference gradient for unbiased stochastic optimization. Fine-tuning the DSpark and DFly drafters with EDR improves mean accepted length over existing training objectives across nine benchmarks covering math reasoning, code generation, and chat.
In briefRobotWorld: A Benchmark for Multimodal Agents Controlling Robots Across Diverse Tasks
SummaryThis paper introduces RobotWorld, a simulation testbed that measures how well AI agents can turn instructions and observations into physical actions through robot interfaces. It includes 84 tasks spanning manipulation, mobile manipulation, locomotion, driving, and aerial control. The authors find that agents can build sophisticated perception and control workflows, but these abilities often fail to combine into reliable task completion, and different models excel at different kinds of goals.
In briefRubix: Global point set alignment without correspondences via assignment geometry
SummaryThis paper presents Rubix, a method that globally aligns two point sets without given correspondences, solving the equally weighted planar problem under squared Euclidean loss. It uses the convex hull of complex correlations from all matchings, called the permutation polygon, whose farthest vertex gives the global alignment. The authors prove a sharp bound of n(n-1) vertices, answering Rote's open problem. On timed MPEG-7 shape pairs, Rubix reaches every numerical reference value in 12 ms on average, about 50 times faster than a rotation grid at the same accuracy.
WhyRecasts correspondence-free point set alignment as a permutation-polygon convex hull problem, proving a sharp n(n-1) vertex bound that answers Rote's open question; validated on MPEG-7 shapes and 3D s
Subhransu S. Bhattacharjee, Dylan Campbell, Rahul Shome
In briefReasoning Token Counts Rise When Models Are Prompted to Answer Untruthfully
SummaryThis paper studies whether the number of reasoning tokens a model generates changes when system prompts tell it to answer truthfully, falsely, or without regard for truth. Three reasoning-capable language models answered 210 multiple-choice questions, and truthful answers used fewer reasoning tokens than lying or truth-indifferent answers across all three models. The authors present this as a proof of concept for a content-independent black-box signal, but they do not yet show it can detect spontaneous deception.
Maverick Morales, Tom\'a\v{s} Dominik, Vermut Gao et al.
In briefSelf-correction optimization for consistent image-text generation without retraining
SummaryThis paper proposes self-correction optimization (SCO), a training-free method for generating interleaved image-text content that stays consistent. SCO uses the classifier-free guidance update as a reference and applies minimal self-correction under new-event and state-preserving constraints. Experiments on interleaved multimodal generation benchmarks show improved temporal coherence and visual-subject preservation. The method also extends to video generation, improving modeling of physical processes such as robot manipulation.
In briefCross-domain pretraining for steady-state neural CFD surrogates
SummaryThis paper studies neural surrogate models for computational fluid dynamics, which can speed up simulations but often fail to generalize beyond their training geometries and applications. The authors train surrogates across different geometries, boundary conditions, and fidelities. They report that cross-domain pretraining improves zero- and few-shot performance over training from scratch or transferring from domain-specific experts. Finetuning a pretrained model is said to reduce errors by 2 to 3 times at the same sample size and to need 8 times fewer samples for the same error, and simply pooling steady-state datasets is found to be sufficient and effective.
Anthony Zhou, Amir Barati Farimani, Shirley Ho et al.
In briefGaussian Density Splatting Network: A crowd counting method using continuous 2D Gaussian primitives
SummaryThis paper proposes the Gaussian Density Splatting Network (GDSNet) for crowd counting. Instead of grid-based density maps, which are sensitive to spatial resolution, it represents a crowd as a superposition of continuous 2D Gaussian primitives. It introduces a control-point-based mechanism to predict Gaussian parameters and adapts differentiable Gaussian Splatting to counting for end-to-end training. The authors report that GDSNet consistently outperforms the state of the art on four standard benchmarks.
In briefRunning Causal Structure Learning Inside Linear-Attention Transformers
SummaryThis paper asks whether transformers can execute a causal discovery algorithm in their forward pass. The authors build a fixed-weight transformer whose blocks exactly reproduce one update of a continuous acyclicity-constrained method, carrying the graph and multiplier between updates. Experiments show agreement with a reference update to floating-point precision, but the executor inherits the reference solver's successes and failures in causal recovery. Ordinary attention models did not reliably execute the update or generalize to larger graphs.
In briefKernel Autoresearch: Automated Kernel Design for Open-Ended Model Discovery
SummaryThe paper proposes Kernel Autoresearch (Kernaut), which treats kernel design as open-ended model discovery. Coding agents write kernels as programs, while construction contracts ensure every accepted kernel is valid, and a quality-diversity archive keeps high-performing kernels with distinct behaviors. The discovered kernels generalize to unseen tasks and outperform baselines on held-out black-box optimization and enzyme-kinetic mechanisms.
In briefSafe Scheduling of Model Updates Under Risk Limits
SummaryThe paper studies how to plan when to replace a policy with newly trained versions, weighing performance gains against the risk of regressions. It maximizes expected cumulative value under a budget on the expected number of updates that perform worse than the policies they replace, estimating switch value and risk from historical learning trajectories. Schedules are chosen by dynamic programming over a directed acyclic graph. Experiments on synthetic and clinical trial data show the performance-risk tradeoff and compare against baselines.
In briefMOTIP2: Adding Spatial Priors to End-to-End Multi-Object Tracking
SummaryThis paper addresses spatially implausible identity errors that end-to-end multi-object trackers still make, such as giving one identity to objects on opposite sides of a frame. It proposes three spatial priors at the data, loss, and representation stages, implemented in MOTIP2 on the DEIM detection transformer while keeping inference fully end-to-end. The authors report that the main model, MOTIP2-L, achieves state-of-the-art results without extra training data, with 73.4 HOTA on DanceTrack, 76.0 on SportsMOT, and 71.1 IDF1 on PersonPath22.
Beno\^it Roussel, Damien Bouet, Liming Chen et al.
In briefGeoCoTDrive: explicit 3D geometric reasoning for driving vision-language-action models
SummaryThis paper addresses a mismatch in vision-language-action models for autonomous driving, where precise 3D geometry is needed for driving but reasoning mostly happens in 2D semantic space. It proposes GeoCoTDrive, which first grounds decision-critical 2D regions, then samples localized 3D features from a geometric foundation model within those regions and inserts them into the autoregressive context for trajectory generation. The authors also introduce a planning-relevant region-level grounding task and the PlanningGrounding dataset, and report improved safety-critical planning across several end-to-end driving benchmarks.
In briefRoboQuest: Robot Agents That Search, Inspect and Test Physically
SummaryThis paper introduces RoboQuest, a benchmark testing whether embodied AI agents can actively gather task-relevant information through physical interaction in unfamiliar environments. It includes ten mobile manipulation tasks involving search, manipulation-based inspection, and interactive testing. The best of five evaluated multimodal agents succeeded in only 23% of episodes, and a fine-tuned policy almost never succeeded. The authors find that failures stem mostly from stopping exploration too early and not handling disturbances, rather than from execution ability.
WhyProposes a new benchmark for goal-directed embodied exploration, evaluating five frontier agents and a fine-tuned policy across three uncertainty types.
Liu Renhang, Navonil Majumder, Tej Deep Pala et al.
In briefOrBIT: Structure-Guided Compression for Language Model Embeddings
SummaryThis paper introduces OrBIT, a framework for compressing embedding tables, which are among the largest parts of modern language models. Instead of using a fixed coding geometry, it learns reusable local geometry from orbit dynamics and uses it to constrain a small set of shared codewords. The global reconstruction residual decides where the fixed coding budget is spent. Across four LLM embedding tables, OrBIT reports 37.9x compression on GPT-2 and over 23x on each 7B table versus 16-bit storage, with competitive rate-distortion performance against established baselines.
In briefBeyond Owls: Subliminal Learning Transfers Capabilities and Backdoors
SummaryThe paper tests whether subliminal learning, where a teacher passes a trait to a student via distillation on semantically unrelated data, can transfer more complex traits than previously shown. The student learns to predict the outputs of a randomly initialized MLP and partially acquires a backdoor that answers in French for female-name prompts (23.5% versus 0.0% for male names). In a chess environment, hacking rises from 10.9% to 58.3% of episodes. The amount of transfer depends on the setup.
Jan Dubi\'nski, Anna Sztyber-Betley, Jan Betley et al.
In briefBoosting and the expressive power of simple weak learners via the γ-VC dimension
SummaryThe paper studies how boosting, which turns weak classifiers slightly better than random guessing into highly accurate ones, depends on the structure of the base class. It uses the γ-VC dimension introduced by Alon et al. (STOC 2021). The authors show this parameter characterizes weak-to-strong learning sample complexity up to a constant factor in γ, and they refine the relationship between the classic VC dimension and the γ-VC dimension. They also give improved bounds on the γ-VC dimension for decision stumps and axis-parallel rectangles in R^d.
Arthur da Cunha, Kasper Green Larsen, Liang-Yu Zou
In briefResidualQuant: KV cache compression for looped Transformers using 2-bit residuals
SummaryLooped Transformers reuse shared blocks across several loops, but their KV cache memory grows with the number of loops, limiting batch size and throughput. The paper proposes ResidualQuant, which treats the final-loop KV states as a reference and stores the other loops as low-precision residuals, combined with least-square scaling, residual rotations, and loop-wise mixed precision to reach INT2. Across several looped models and math reasoning and code generation benchmarks, it improves the accuracy-memory tradeoff over rotation-based KV quantization. Under mixed precision it keeps accuracy close to BF16 while reducing theoretical KV storage by 80.7%, and on an RTX 5090 it raises decode throughput by up to 2.73x.
In briefContinual learning without continual training: replacing weight updates with in-context inference
SummaryThis paper proposes Latent Concept PFN, a meta-trained and then frozen model that performs in-context Bayesian inference over a latent concept space. New classes or domains are incorporated by adding exemplars to a memory, updating posterior beliefs rather than changing parameters, which the authors argue reduces forgetting. Experiments on class and domain incremental learning report competitive performance while learning interpretable latent concepts.
In briefSimplifying Estonian long texts as whole documents using large language models
SummaryThis paper examines document-level text simplification in Estonian, a morphologically rich, low-resource language, using large language models. It goes beyond sentence-level edits to address discourse coherence, anaphora resolution, and consistency across paragraphs. The authors compare five multilingual LLMs under three prompting strategies and evaluate outputs with automatic metrics and a structured manual annotation protocol. Gemini-2.0 and LLaMA-3.3 show near-native fluency and strong meaning preservation, while other models show notable grammatical and semantic limitations.
In briefTaoD2C-Bench: A benchmark testing whether AI turns UI designs into industrial-compliant code
SummaryThis paper introduces TaoD2C-Bench, a benchmark for testing whether multimodal large language models can convert UI designs into code that meets industrial implementation requirements. The dataset includes 2,861 production designs from 17 commercial platforms with 97,652 expert annotations across four categories. It defines three tasks: end-to-end UI code generation, requirement inference, and requirement realization. Evaluating eight models shows substantial gaps in meeting these requirements, and visual reconstruction ability does not necessarily translate into compliant code generation.
In briefPalmSpace: A Unified Touch Model for a Versatile On-Palm Input Space
SummaryThis paper presents PalmSpace, a wrist-worn infrared system that provides mode-aware, palm-referenced absolute input on the bare palm without per-user calibration. Evaluated with 17 participants in leave-one-participant-out testing, it achieved 6.7 mm mean localization error, 98.9% contact detection accuracy, and 96.7% F1 for four-class interaction-state recognition. User studies demonstrated absolute pointing and dragging, eyes-free digit input, and multi-finger controls such as scrolling and pinch-based map manipulation.
In briefInput-blind control perturbations also yield large oracle headroom for layer programs in multiple-choice evaluation
SummaryThe paper studies layer-skipping and repetition programs for language models on multiple-choice items, using oracle evaluations with known answers to estimate the potential gain from adaptive computation. It compares these programs with input-blind perturbations applied at the same layer sites, finding that the controls produce headroom of similar scale. The authors conclude that this headroom does not establish a benefit specific to the selected layer computation.
In briefTemporally Interpretable Differentiable Decision Trees: Action Chunking for Time-Aware Explainability
SummaryThis paper addresses the mismatch between single-timestep decision trees and human multi-step planning in sequential decision-making. It introduces time as a new interpretability dimension, called temporal interpretability, and shows that action chunking improves it. The authors propose two policy gradient algorithms incorporating action chunking and an information-theoretic tree restructuring method to keep trees parameter-efficient. In four simulated environments, warm-started action-chunked trees matched neural policies in three while using up to 80% fewer parameters.
In briefKoopman observers that correct diffusion feature forecasts using shallow-layer measurements
SummaryThis paper proposes an observation-corrected Koopman framework to speed up sampling from frozen diffusion models. Calibration trajectories are used to fit time-dependent Koopman operators that describe increments of shallow and deep network features, and during accelerated sampling, fresh shallow-feature innovations correct the predicted deep features. Periodic full evaluations refresh the observer, and no generative-model parameters are changed. Reported results on CIFAR-10 and a ten-class ImageNet subset show Inception-feature MSE reductions of 19.9% and 11.9% over channelwise affine prediction, with measured speedups of about 1.89x and 1.85x over DDIM-50.
In briefChecking Whether Decentralized Adaptive Sensing Has Collected Enough Evidence to Tell the Target Apart
SummaryThis paper studies decentralized adaptive sensing, in which several agents pick measurements based on local beliefs while sharing information over a communication network. The authors introduce a pathwise certificate built from the Rényi–Chernoff information accumulated along the actual sensing trajectory, which yields nonasymptotic MAP-error bounds and an anytime, network-wide stopping rule. They show that when information against the least-resolved competitor grows linearly, error decays exponentially, and in 1D and structured 2D experiments the worst-competitor score correlates more strongly with localization speed than an average-pair proxy.
In briefORDERS: A Study of Norm-Rank Weighting for Personalized Federated Learning
SummaryThis paper studies ORDERS, a personalized federated learning setup with a shared backbone, private adapter and classifier, and server weights assigned by descending update norm. The server takes a weighted sum of updates from the same broadcast model. Across 80 final runs on CIFAR-10 and Sent140, ORDERS reaches 80.51 ± 0.79% native mean client accuracy on CIFAR-10 and 74.71 ± 0.49% on Sent140. Ablation evidence is limited and depends on the endpoint.
In briefMultiFly: A real-world low-altitude drone dataset with cheaply transferred labels for RGB, thermal, LiDAR, and radar
SummaryThe paper introduces MultiFly, a real-world low-altitude UAV dataset with 17,272 synchronized samples across RGB, thermal, LiDAR, and radar from four suburban scenes, labeled with 15 semantic classes. Labels were manually created for only 115 RGB images and propagated to all modalities through shared geometric representations. Transferred annotations reach 89.93% average agreement with held-out manual labels. The authors also provide semantic segmentation benchmarks for all four modalities.
WhyTransfers labels from 115 annotated RGB images to four modalities via shared geometry, validated on large real-world aerial data with benchmarks.
Markus Gross, Andreas Greiner, Taehyoung Kim et al.
In briefOpen-MMUnlearning: A Unified Framework for Evaluating Methods that Remove Knowledge from Multimodal LLMs
SummaryThis paper introduces Open-MMUnlearning, an open-source, extensible framework for machine unlearning in multimodal large language models, covering model preparation, data processing, unlearning, and evaluation. It supports five benchmarks, eight MLLMs from four model families, and twelve unlearning methods. Under a common protocol, GD and MIP-Editor tie for the highest overall score, with GD forgetting most effectively and MIP-Editor retaining more utility. The authors also propose a meta-evaluation of metrics, finding BLEU has the highest aggregate reliability among thirteen evaluated metrics.
In briefMIRA: an agent that refines prompts so text-to-music output better matches user intent
SummaryThe paper argues that current text-to-music evaluation, based on a single global relevance score, cannot show whether generated music meets specific requirements or implied intent. The authors define per-request rubrics of verifiable items and introduce MuRA-Bench, a benchmark of real platform requests curated by music experts. They also propose MIRA, a test-time agent that searches over prompt revisions under a bounded budget, verifying outputs against the rubrics. Experiments reportedly show MIRA improves intent alignment, letting an open-source generator approach commercial systems such as Suno and Mureka.
In briefThe Handover Problem: Rules for Shifting AI Control in Human-AI Teamwork
SummaryThis paper formalises the Handover Problem: deciding at each cycle whether to raise, keep, or lower an AI system's autonomy during repeated human-AI collaboration, while keeping the process reversible and auditable. It introduces the Handover Readiness Score (HRS), a composite measure built from operator readiness, human-AI trust, learning stability, and operational performance. The authors combine it with a hysteresis-based policy and hard safety guards, and report that these address different failure patterns across software engineering and manufacturing settings.
Vicente Pelechano, Antoni Mestre, Manoli Albert et al.
In briefAutoAdapt: Low-Cost Model Extension via Automatic Domain Discovery
SummaryThis paper presents AutoAdapt, a framework that adds new domains or data to an instruction-tuned model without full retraining. It automatically discovers latent domains, trains a separate LoRA adapter for each domain in parallel, and routes inputs without learned parameters. Across 14 domain-specific benchmarks with GPT-4o pairwise judgements, it reaches parity with a LoRA adapter trained on all domains.
Josh McGiff, Salma Mekaoui, Robert Shanahan et al.
In briefLabel-free dataset pruning via linear programming to keep models accurate on smaller training sets
SummaryThis paper proposes a dataset pruning method that reduces a large training set to a representative subset while preserving model performance. The authors reformulate unbiased subset selection as a variance minimization problem and derive geometric selection criteria that require neither labels nor model training. Across CIFAR-10, MNIST, and CelebA, the method matches or exceeds uniform sampling at every evaluated budget and outperforms competing geometric methods in several settings, especially at small budgets. The same framework can also reduce stochastic-gradient variance by increasing diversity within mini-batches.
In briefReal-time audio-video generation by combining two adapters in parallel
SummaryThis paper presents a method for generating joint audio and video in real time. Instead of distilling a model in sequential stages, it trains a causal adapter in parallel on a frozen backbone and adds it to an off-the-shelf few-step adapter at inference. The authors report that the two weight updates are nearly orthogonal, so they combine without interference. The system runs at about 26 fps at 480×832 and sustains 30 seconds of stable generation, matching or beating chained baselines on most metrics.
In briefPosition Forcing: a self-conditioned method for single-stage 3D shape generation
SummaryThis paper addresses a limitation of single-stage 3D generative models that use VecSet representations, where token positions must be inferred implicitly during denoising. It proposes Position Forcing, which recovers token positions from the current clean latent estimate, quantizes them at progressively finer resolutions across denoising stages, and feeds the positional encodings back into the diffusion Transformer. The authors report that this coarse-to-fine positional feedback improves generation quality, with strong results among single-stage methods and better performance than several multi-stage approaches.
In briefData Reuse in Non-Stationary Learning: Quantifying when reusing past observations helps an online learner whose target switches between recurring values
SummaryThis paper studies online learning in non-stationary environments where an unknown parameter abruptly switches among a finite set of recurring values. The authors propose Exposure-Capped Reuse (ECR), an anytime algorithm combining online change detection, compatibility testing, and contamination control to reuse past data. They characterize a regime where ECR's regret scales with the number of distinct values rather than the number of changes, and provide an information-theoretic lower bound establishing near-minimax optimality.
WhyProposes ECR for reusing past data and proves a near-minimax optimal information-theoretic lower bound, a substantive theoretical advance for non-stationary online learning.
In briefWhen to Unpair: Reducing Pairing Dependence in Medical Visual In-Context Learning
SummaryThis paper studies how medical visual in-context learning models depend on the pairing between support images and their labels, using a test-time label shuffling check. It proposes a late unpairing curriculum (LUC) that trains with matched pairs first and then randomly unpairs support labels. On two backbones, LUC nearly closes the pairing gap while maintaining matched-support performance, and on BraTS whole-tumor segmentation it raises matched-support DSC from 0.733 to 0.857.
In briefA paper proposing a unified benchmark and metric for judging how well face de-identification protects privacy
SummaryThe paper notes that existing face de-identification (FDeID) evaluations use inconsistent metrics, datasets and annotations, which makes cross-method comparison difficult. The authors introduce UtilFace, a demographically balanced benchmark, and HiFD, a hierarchical metric that combines identity suppression, multi-level utility preservation and image quality into one score. They use this protocol to compare adversarial, GAN-based and diffusion-based methods, and release the benchmark and toolkit.
In briefTraining Small Agents to Act Using Task-Progress Labels from Teacher Demonstrations
SummaryThe paper introduces Task-Progress Distillation (TPD), which pairs each teacher-demonstrated action with a short label for the current task stage, so a small student model learns compact targets and selects actions. On ALFWorld, a 1.7B student trained on 404 demonstrations reaches 72.4% mean unseen-task success, versus 48.3% for a reasoning-trained student. Explicit stage labels gave an additional gain at 200 demonstrations, while the gap closed with more demonstrations.
In briefModel-Free Average-Reward RL for Multichain MDPs via Hierarchical Decomposition
SummaryThe paper studies learning optimal policies in average-reward multichain Markov decision processes, where the optimal gain can depend on the starting state. The authors propose an asynchronous value-iteration reinforcement learning algorithm that needs only the transition graph and uses Bather's decomposition to split the state space into communicating subsystems and transient states. They prove the algorithm converges to the optimal gain and yields gain-optimal policies after finite time, and extend it to two algorithms targeting near gain-optimal and near bias-optimal policies. All three come with almost-sure convergence guarantees, and experiments show the latter two improve transient performance.
In briefPerformance at What Cost? A Sustainability-Aware Index for Cell Segmentation Models
SummaryThis paper compares pretrained models used for cell and nucleus instance segmentation, which differ widely in size and computational demand. The authors introduce the Sustainability-Aware Performance Index (SAPI), a metric that combines segmentation performance, energy consumption, and model size. They benchmark 19 models zero-shot and 16 fine-tunable models with few-shot adaptation across CellBinDB datasets. They find that larger models do not consistently yield proportionate quality gains, and SAPI-based rankings can differ from performance-only rankings.
In briefHAN-Mamba: Multi-scale Hierarchical Model for Financial Volatility Forecasting
SummaryThis paper addresses short-horizon realized volatility forecasting by proposing HAN-Mamba, which replaces the Transformer encoders in the earlier HAN-T hierarchical model with selective state space (Mamba) encoders and keeps attention only in the fusion layer. On the Optiver Realized Volatility Prediction benchmark under time-aware five-fold cross-validation, it achieves a mean RMSPE of 0.1942 versus 0.1965 for HAN-T, with 33% fewer parameters. Extending the high-frequency context from 60 to 240 buckets further reduces error to 0.1927.
In briefPhysics-grounded post-training that makes generated two-humanoid interactions executable and faithful in simulation
SummaryThe paper introduces DIGHT, a framework that jointly adapts a digital human interaction generator and a humanoid tracking policy. Text-conditioned interaction candidates are executed in simulation with a fixed tracker, and preferences are built from tracking error, friction, floating, and contact occurrence, location, duration, and force. The generator is aligned with physics-decoupled diffusion direct preference optimization, and its outputs then fine-tune the tracker. Experiments reportedly show improved physical plausibility of generated motions and more reliable, faithful humanoid interactions in simulation.
In briefEstimating crash factors missing from coded police records using AI models
SummaryThis study combines the coded fields and officers' narratives of 5,601,890 Texas crashes from 2017 to 2025 to estimate crash factors that the coded fields omit. A tabular foundation model, Kumo Tabular, reads every coded record, while a calibrated System One model, Jev, reads narratives from two probability samples, and human judgments recalibrate the probabilities. The narratives document more injury crashes than the coded field for several factors, and a human check agrees with all fifteen estimates within its margin.
In briefUsing AI agents to build transparent, open-source multiphysics models of electrochemical devices
SummaryThe paper shows that large language model agents can build multiphysics models of electrochemical devices from a human-written specification that lists governing equations, parameters, numerical methods, and checkpoints. The test case is one-dimensional electrochemical CO2 reduction to CO in a porous gas diffusion electrode, with models written in open-source Julia. The agent-built models agree with an equivalent COMSOL implementation to within 0.7% of peak CO partial current density, and with each other to within 0.04%.
Sebastian Castro, Maya F. Schuchert, Spencer A. McCluskey et al.
In briefEven humans, specialized tools and LLMs disagree on sentiment in social media posts
SummaryThis study compares three sentiment analysis tools and three large language models against six human raters on 100 tweets. Even human raters showed only fair agreement, highlighting how subjective sentiment judgment is. The Twitter-roBERTa-base model aligned most closely with human ratings among the automated methods.
Himarsha R. Jayanetti, Sivakanesan Dhanushkanda, Shuai Hao et al.
In briefReusing Intermediate Layers for Better Tabular Predictions
SummaryTabular foundation models make predictions on new tables using labeled examples as context, but the study finds their refinement is often concentrated in later layers. The paper introduces Retro, which lets later stages revisit and recombine intermediate representations via Attention Residuals and query-conditioned Gated Attention. The authors report that Retro ranks among the top three and lies on the Pareto frontier across TabArena, TALENT, and RelArena.
In briefPoreML: a framework for learning fluid flow in porous materials from data
SummaryThis paper introduces PoreML, an open-source framework that unifies data generation, model training, and evaluation for multiphase flow in porous media at the pore scale. It includes a GPU-based lattice Boltzmann solver, a 3.3 TB dataset of 560 simulation runs across four scenarios, and a unified learning and evaluation pipeline. Five models of different architectures are assessed, with challenges testing transfer to larger domains and from synthetic to micro-CT-derived structures.
WhyProposes a unified framework for multiphase flow data, training, and evaluation, backed by a large 3D time-resolved dataset and multi-model benchmarks.
In briefFault-tolerant foundation models: training large language models to run on unreliable, energy-efficient hardware
SummaryThe paper shows that large language models can be trained to tolerate errors from emerging hardware that trades reliability for energy efficiency. Using modified neural scaling laws derived from about 40,000 GPU-hours of training on simulated faulty digital hardware, the authors find that error resilience increases as models grow. They conjecture that appropriately trained models may be formally fault-tolerant, which could enable substantial energy savings when running inference on low-energy, faulty hardware.
WhyProposes that larger LLMs become more error-resilient, quantified by modified scaling laws from substantial simulated faulty-hardware training. Evidence is simulation-based, not real hardware.
In briefRSIGym: A Flexible Environment for Studying Recursive Self-Improvement of AI Agents
SummaryThe paper introduces RSIGym, a research environment in which agents use reusable services for training, inference, evaluation, and sandboxed execution, with shared budget and permission controls. It defines RSI-Index as the mean fraction of the remaining performance gap closed across five benchmarks. In independent Joint runs of six frontier models, Opus 5 achieved the highest RSI-Index of 0.4809 under a $500 per-benchmark budget, raising SWE-bench Verified from 17.67% to 50.33% and AIME from 31.67% to 97.78%. The full codebase and results are open-sourced.
In briefOne-shot adaptive segmentation for scientific images
SummaryThis paper presents a training-free, one-shot framework that adapts general vision foundation models to scientific image segmentation using a single annotated reference image. It combines DINOv3 features with background-adaptive feature orthogonalization to suppress artifact-related directions, then uses cosine similarity to locate candidate regions for SAM segmentation. Evaluated on red-blood-cell microscopy, structured-illumination pool boiling, and chest radiography, it improves mean IoU over the strongest baseline by 5.91% and 78.62% on the first two datasets and performs comparably on chest radiographs.
Tejaswi V. Panchagnula, Allison M. Davis, Fengqing Zhu
In briefHow Transformers learn symmetries: a study of how standard Transformers learn invariance and equivariance to geometric symmetries in point clouds from finite data augmentation.
SummaryThe paper examines whether a vanilla Transformer can learn various symmetries of point cloud data through finite data augmentation. It finds an ordering of learnability, with non-angle-preserving symmetries learned least easily and base angle-preserving subgroups such as translation, rotation, and scale learned most easily. For these base subgroups, the authors study extrapolation and analyze trained models to identify interpretable mechanisms that induce invariance. They also show these mechanisms can serve as a building block for learning equivariant functions.
Eduardo Santos-Escriche, Valerie Engelmayer, Ya-Wei Eileen Lin et al.
In briefSemanticFold: Compressing latent sequences shows language fit, decodability, and reasoning can diverge
SummaryThis paper examines whether compressing prompt prefixes into fewer latent states preserves the abilities large language models use during inference. The authors propose SemanticFold, which folds prefix hidden states at learned boundaries, and test it on five model checkpoints. They find that effects are non-monotonic and differ across metrics, so language-model fit, decodability, and reasoning can move in different directions.
In briefPXtal: Aligning Powder X-Ray Diffraction Patterns and Crystal Structures Despite Information Gaps
SummaryThis paper presents PXtal, a framework that learns aligned representations of powder X-ray diffraction (PXRD) patterns and crystal structures. Because compressing a 3D crystal into a 1D diffraction pattern loses information, the two modalities are not equally informative. PXtal uses unbalanced optimal transport and a coupling-level generalized KL divergence to supervise the alignment. It outperforms baselines on PXRD-to-crystal candidate retrieval across six test sets, and its encoders transfer better to downstream materials and crystallographic tasks.
Zhuoran Yang, Christopher M. Collins, Bei Peng et al.
In briefBeyond the Ergodic Wall: A Discrete Geometric Physics Sandbox for Analysing AI Scaling Limits
SummaryThe paper argues that current deep learning converges to a statistical average of historical human knowledge, limiting novelty and efficiency. It proposes a digital physics sandbox built on a Holographic E8 Projection Engine to verify models against physical constraints. The authors claim that restricting algorithms to physically conserved causal trajectories could collapse NP-hard search into polynomial-time paths. The spacetime and particle-spectrum components are presented as theoretical proposals.
In briefIs Separating Attention and FFN Layers Necessary in Vision Transformers?
SummaryThis paper examines whether the strict alternation of Attention and Feed-Forward Network layers in Vision Transformers (ViTs) is necessary. The authors introduce AttenFeed, a module combining both functions, and build a unified ViT (uViT) from a sequence of these modules. Comparing uViT with standard ViT across multiple datasets and model scales, they report that the rigid Attention-FFN split can hurt performance at smaller model sizes.
In briefContinual Graph Multi-Agent RL: Keeping Agents Cooperative Across Changing Task Structures
SummaryThe paper studies continual multi-agent reinforcement learning, where agents must adapt to new tasks while retaining skill on earlier ones. It proposes CGMARL, which models each task's structure as an attributed graph, and introduces GRAFO, the first benchmark for this setting, showing how forgetting arises. To counter forgetting, the authors propose FROG, a frozen graph encoder that preserves past structural information, and report that pairing it with existing continual learning methods substantially improves performance across GRAFO scenarios.
In briefRethinking explainable AI with model-independent concept dictionaries
SummaryThis paper proposes DictXAI, which defines concepts directly in the input domain using a dictionary of predefined, interpretable elements. It first computes a sparse code of the input and then attributes the model's prediction to the associated dictionary elements. The authors show it can trace malfunctions such as Clever Hans effects to identifiable data artifacts and support human-AI alignment on biomedical signals. It works across learned image bases, analytic ECG waveforms, and experimentally acquired dictionaries.
Thomas Schnake, Doreen Sch\"oppenthau, Alexander Meyer et al.
In briefShared Gaussianization: what Gaussian regularizers can and cannot guarantee for contrastive learning
SummaryThe paper studies shared Gaussianization (SG), a characteristic-function Gaussianity test applied to the average of two normalized views, as a way to relate the SIGReg regularizer in LeJEPA to contrastive learning. It proves that SG vanishes at aligned, uniform InfoNCE minimizers and bounds the InfoNCE excess by a square-root function of the SG loss. It also analyzes when alternative uniformity tests give linear bounds and how the objectives differ away from the optimum.
WhyProves a sharp square-root bound linking SG loss to InfoNCE excess, with controlled experiments checking predictions; impact is confined to a regularizer viewpoint.
In briefHierarchical reinforcement learning for energy-efficient quadruped gait adaptation across terrains
SummaryThis paper proposes a hierarchical reinforcement learning framework that separates a high-frequency joint-level motion policy from a low-frequency gait adaptation policy that minimizes cost of transport (CoT). The three-stage Isaac-based training enables zero-shot sim-to-real transfer. The learned hierarchy switches from pacing at low speeds to trotting at higher speeds. Simulation results show reduced CoT across commanded velocities while maintaining robust locomotion on flat, rough, and inclined terrain, and the approach was deployed on a physical Unitree AlienGo.
In briefLLM-Assisted Transportation Management Plan Drafting: A WisDOT WisTMP Case Study
SummaryThis paper proposes a large language model framework to automate drafting of Transportation Management Plans for work zones, using Wisconsin DOT's WisTMP system as the case. Several open-source models of different sizes are fine-tuned and run locally, trained on a dataset built from historical WisTMP documents converted into JSON question-answer pairs. Fine-tuning improves standard text generation metrics, but the models tend to over-generate strategies and struggle with project-specific justifications and accurate cost estimates. Scaling from 7B/8B to 14B yields only limited gains.
In briefTouchScale: 500 Hours of Human Vision and Touch for Visual-Tactile Learning
SummaryThis paper introduces TouchScale, a 500-hour dataset of contact-rich human interactions recorded with one wearable setup, synchronizing egocentric RGB-D video, wrist video, and dense bimanual tactile measurements. Training on it raises zero-shot contact IoU on an unseen tactile sensor from 0.134 to 0.383, and improves a robot policy's average real-world success on four manipulation tasks from 22.5% to 57.5%. The authors plan to publicly release the dataset.
WhyBuilds a 500-hour unified-sensor visual-tactile dataset, validated by zero-shot contact IoU, action recognition, and robot success experiments showing scale benefits.
In briefΔRepresentation: Counterfactual Geometry-Supervised Phenotype Learning for Medical VLMs
SummaryThis paper addresses a limitation of medical vision-language models, which struggle to represent lesion-specific visual changes relative to normal anatomy. The authors propose ΔRepresentation, comprising BaseAnatomy, which gives geometric supervision from spatial relations among anatomical structures, and ΔPhenotype, which computes the increment between lesion and normal anatomical representations and clusters increments of the same phenotype. Experiments on ReXGroundingCT and LIDC-IDRI reportedly improve lesion grounding and phenotype characterization accuracy.
In briefSafety limits for AI agents acting under partial observation and time pressure
SummaryThis paper examines theoretical limits on agent-agnostic safety guarantees in environments that are only partially observable and require action within limited time. The authors present two scenarios, one with an infinite state space and one involving signal mixture, and prove that even a perfect agent cannot guarantee safe behaviour in them. They argue that any proof of AI safety or alignment must consider the environment and the relevant safe actions together with the agent.
In briefLearning Grasp Stability from Time-Based Vision and Touch for Dexterous Hands
SummaryThe authors collected 10,000 grasp trials on 200 objects using a multi-fingered robot hand with four Digit 360 tactile sensors, recording vision, proprioception, and touch. They trained end-to-end temporal multimodal models to predict post-lift stability from pre-lift observations and compared sensing modalities and encoders. Adding touch, especially high-resolution dynamic touch, improved prediction, and using the predictor as an online gate raised the success rate of executed lifts by 10.5 percentage points over a non-tactile gate.
WhyApplies high-resolution dynamic touch to dexterous grasp stability prediction with real-robot deployment, backed by a 10,000-trial dataset and controlled ablations.
Ken Nakahara, Aleksei Buvailik, Prokhor Kotov et al.
In briefNeural sampling via Wasserstein–Fisher–Rao JKO steps with reweighted normalizing flows
SummaryThe paper proposes a neural algorithm for sampling from unnormalized Boltzmann densities, based on a Jordan–Kinderlehrer–Otto scheme for KL divergence in the Wasserstein–Fisher–Rao geometry. It proves that exact iterates converge exponentially fast to the target for any fixed step size, without structural assumptions such as log-concavity. The authors implement the scheme neurally by parametrizing transport and reaction components with reweighted normalizing flows, and numerical experiments on challenging multimodal targets show promising performance.
Chenguang Duan, Johannes Hertrich, Gabriele Steidl
In briefGradient-based token selection for faster sparse planning in visual world models
SummaryThe paper introduces COSTGRAD, a training-free, goal-conditioned method that ranks spatial tokens by the gradient norm of the planning cost with respect to each input token. At 50% sparsity on AdaLN-conditioned predictors, it matches or exceeds full-token planning on three of four continuous-control benchmarks, with a measured 2.6x wall-clock speedup per planning step. The authors also find an architecture-dependent failure: on concat-based predictors, pure COSTGRAD loses its advantage over random selection.
In briefNon-asymptotic convergence analysis of PPO with learned critics and clipping
SummaryThe paper analyzes PPO with clipping as a closed-loop actor-critic system, covering actor-critic coupling, probability-ratio clipping, finite-batch reuse, and early stopping. It characterizes policy stationarity and critic tracking accuracy under explicit coverage and regularity assumptions, in both synchronous and asynchronous settings. For finite layered MDPs with tabular critics, it gives a polynomial sample complexity result for a growing-horizon family, and a two-time-scale schedule yields O(T^{-2/5}) bounds. The authors state these results offer theoretical guidance for tuning.=
WhyGives a non-asymptotic closed-loop analysis of PPO-Clip coupling clipping, finite-batch reuse, and learned critics with explicit bounds, though under strong assumptions.
In briefVideo Prediction Policy 2 (VPP2): a robot control model that uses video prediction to plan better actions
SummaryExisting world action models often predict incorrect motion in open-ended environments, which leads to wrong robot actions. The authors propose VPP2, which continues pretraining a video model on a large curated manipulation dataset, distills it into a single-step visual planner, and adds an action module via a mixture-of-transformers architecture. Reported results include gains in video prediction instruction-following success, zero-shot real-world ALOHA manipulation, and top success rates on LIBERO-Pro, LIBERO-OOD, and RoboDojo after benchmark-specific post-training.
In briefLeak-Controlled Multimodal Learning for Predicting Alzheimer's Diagnosis and Progression
SummaryThis paper presents a multimodal, multitask framework for predicting Alzheimer's disease diagnosis and progression, designed to handle irregular visits, heterogeneous measurements, and missing modalities. It combines an adapted MRI encoder, causal clinical Transformers, shared fusion, and task-specific ODE-GRU dynamics. Across 2,649 subjects and 17,317 visits from ADNI, OASIS-2 and MIRIAD, internal diagnosis AUROC is 0.935. External validation on OASIS-3 yields a hybrid diagnosis AUROC of 0.767.
Akeem Temitope Otapo, Ghazaleh Khodabandelou, Zuheng Ming et al.
In briefLoomSC: Scalable Deep Subspace Clustering via Projector Factorization and Exact Spectral Reduction
SummaryThis paper introduces LoomSC, a framework that tackles the scalability limits of subspace clustering caused by dense self-expression matrices and full-affinity spectral clustering. It jointly learns latent features and a projector self-representation through two thin factors, then uses an exact feature map to reduce the spectral problem to an eigenproblem whose size depends only on factor width. On five image-clustering benchmarks, LoomSC ranks first or second in all 15 dataset-metric comparisons, and it maintains at least 99.8% accuracy on synthetic data scaled to 500,000 samples.
Nairouz Mrabah, Youssef Melki, Mohamed Bouguessa et al.
In briefStale, misattributed, or late: where agent memory fails before generation
SummaryThis paper examines how personal memory for language agents fails before an answer is generated, such as serving an obsolete value, a fact about the wrong person, or no useful fact before the deadline. Using Personal Fact Memory (PFM) as a reference, the authors find on a controlled revision benchmark that serving only the active value of each correctly keyed slot eliminates stale exposure. Assigning revisions to the right slot is the harder problem, since missed and false merges leave stale values active or remove current ones. The authors argue agent memory should be evaluated before generation, separating stored-state validity, identity resolution, abstention, and serving latency.
In briefUsing small language models to map the stages of machine learning code
SummaryThis paper examines whether Small Language Models (SLMs) can identify the stages of machine learning pipelines, such as data preprocessing and modeling, from source code. The authors compare several SLMs against two reference studies and test how wording changes in the taxonomy affect results. They find that taxonomy wording significantly affects performance, the best SLM performs well but does not outperform other classifiers, and different methods yield significantly different insights. The authors also note high inference costs as a barrier to large-scale use.
Nicolas Lacroix, Frederic Precioso, Mireille Blay-Fornarino et al.
In briefProof That Cyclic Strategies Are Best for the Cow-Path Search
SummaryThe cow-path problem asks how a cow should find a goal at an unknown distance along one of several paths meeting only at the origin, with performance measured by competitive ratio. Kao, Reif and Tate gave an efficient randomized algorithm that visits paths in a fixed cyclic order, and Kao, Ma, Sipser and Yin later proved it optimal for all numbers of paths. This note provides a detailed proof of their claim that no algorithm outperforms the best cyclic one.
In briefLogarithmic regret for controlling systems whose parameters change unpredictably
SummaryThe paper studies minimum-variance control of an unknown autoregressive system with exogenous inputs whose coefficients change at unknown times. Under bounded disturbances, stability and feasibility conditions, it proves a regret bound of O((C+1)log((T+1)/δ)) with probability at least 1-δ, where T is the horizon and C the number of changes. The proposed PIECE-CD algorithm explores initially and after alarms, detects changes with a windowed output-energy test, and uses gated recursive least squares for control.
In briefStationary bias and extrapolation in nonlinear two-timescale stochastic approximation
SummaryThis paper studies the nonzero stationary mean error that remains in constant-step stochastic approximation even under time averaging, for nonlinear two-timescale recursions driven by an exogenous finite-state Markov chain. It derives a first-order bias expansion whose error bound stays uniform when the slow step size is much smaller than the fast one. An exactly solvable nonlinear Markov example verifies the coefficients, and the paper compares extrapolation methods for finite runs.
In briefMorphCL: Contrastive Self-Supervised Learning for Wearable Motion Recognition
SummaryThe paper introduces MorphCL, a self-supervised pretraining framework that uses structure-aware grouping to capture the global structure of large-scale inertial motion data. It combines discovered motion motifs with domain-specific feature descriptors. The authors report F1-score gains of up to 15 percentage points in linear probing and finetuning. Encoders pretrained with MorphCL match or surpass existing foundation models in linear probing while using about 4600 times less data.
In briefTestGRAD: Evolving Test Suites with Failure-Pattern Momentum to Select Patches from SWE-Agent Ensembles
SummaryThe paper treats choosing the best patch from several SWE-agent candidates as a test-space optimization problem, evolving an executable repository test suite until it separates competing patches. TestGRAD introduces a differential loss defined by execution behavior, full CRUD update directions, and Failure Pattern Momentum, which mines frequent failure sequences from memory to avoid repeated non-discriminative directions. On SWE-bench Verified, a 4-agent ensemble reaches 84.2% Pass@1, 3.6 points above the strongest baseline (80.6%), while compressing failure-history context by over 100×.
In briefWhen Scientific Cognition Is No Longer Scarce: How AI Could Shift Science's Bottlenecks
SummaryThis paper imagines a world where machines outperform people at most computer-based scientific work, and asks what would then limit scientific progress. The authors introduce a "scarcity inversion": literature synthesis, hypothesis generation, and software development become abundant, while experiments, observations, well-supported conclusions, and accountable authority become scarce. They examine four areas: selection, physical access, validation, and organizational choice. The paper argues that national laboratories are well placed to turn abundant machine reasoning into trustworthy results.
In briefGeometry-Supervised Visual Representation for Interpreting Multi-Phenotype Lesions in Medical VLMs
SummaryMedical vision-language models struggle to interpret lesions that require assessing several pathological phenotypes together, because existing alignment methods do not preserve anatomical hierarchies or relations among phenotype subclasses. The paper proposes PureVision, which combines PureEyes, a module that adds geometric supervision via ideal spatial distributions encoding these relationships, and PureNeurons, which uses learned latent positions to aggregate lesion-specific evidence. Experiments on LIDC-IDRI, CBIS-DDSM, and 3DReasonKnee report improved lesion grounding and phenotype characterization in visual question answering and radiology report generation.
In briefBeyond LLM-GA: A Memetic Algorithm Designed by ReEvo for Secure Fluid Antenna Systems
SummaryThis paper studies how to protect fluid antenna systems from eavesdropping by selecting secure ports. The authors propose a memetic algorithm whose crossover, mutation, and local-search operators are evolved offline by an LLM using reflective evolution. These operators are embedded in the search framework, so no LLM queries are needed during execution. Simulations at equal generation counts show higher secure sum-rate than the conventional GA and existing LLM-assisted GAs.
In briefMasked Generative Motion Planning with Geometry-Guided Token Search
SummaryThis paper proposes MGMP, a method where a masked generative transformer produces discrete trajectory candidates in parallel, and Geometry-Guided Token Search uses scene geometry to choose which parts to edit and which alternatives to evaluate. It treats test-time repair as a search over discrete motion alternatives, allowing route-level restructuring. The authors report 96% success on Ring Maze and 82% repair success on Kuka's Controlled Route Invalidation, and show generalization to unseen layouts, obstacles, dual-arm planning, and real-world Baxter tasks.
In briefLLM persuasion scores depend on the evaluation method; automated rankings agree only weakly
SummaryThe study adapts nine published automated methods for measuring LLM persuasion to a shared setup and runs them on fifteen LLMs to test whether their rankings agree. Agreement is weak (mean Spearman ρ = 0.25). Models that refuse some tasks but not others lower agreement, and these refusals mostly fall on manipulation tasks. The authors conclude that a single persuasion score reflects its specific setting and says little about a model's persuasiveness across tasks.
Kamile Dementaviciute, Julija Vaitonyte, Tijl De Bie
In briefFrom Prompts to Trees: LLM-Guided Decision Tree Generation for Few-Shot Tabular Classification
SummaryLarge language models have broad knowledge but are costly and hard to interpret when used directly on tabular classification, while decision trees are transparent but weak with little data. This paper proposes a framework that distills LLM knowledge into interpretable decision trees in a few-shot setting. Rather than asking the LLM to generate full trees, it uses a three-stage process: generating rules, then organizing those rules into a tree. Experiments on several real-world tabular datasets report higher accuracy and interpretability than existing baselines with much lower prompting overhead.
In briefWhy software engineering remains essential in the era of coding agents
SummaryThe paper asks whether capable AI coding agents make software engineering as a discipline and software engineers obsolete, and argues they do not. It contends that three structural properties of large language models (probabilistic generation, agnosticism, and semantic statelessness) can yield software that looks plausible but is unverifiable and untrustworthy. The authors propose four knowledge levers (methodological knowledge, domain knowledge, design choices, and process choices) that must be captured as persistent artifacts, with software engineers acting as methodologist, mediator, and custodian.
In briefMasked Feature Encoding for Whole Slide Image Representation: A Feature-Space Method to Improve Pathology Image Analysis
SummaryThis paper proposes MFE-MIL, a feature-space masking framework for whole slide image analysis in computational pathology. It trains a lightweight MLP adapter jointly with a window-based masked reconstruction branch and a multiple instance learning classification head to suppress within-slide patch variance. At inference, only the adapter and MIL head are retained. The authors report improvements in classification metrics across several datasets and encoders, and consistent gains in average concordance index on five TCGA survival cohorts.
WhyJoint feature-space masked reconstruction with MIL is a meaningful refinement, backed by broad experiments across datasets, encoders, and strong baselines, useful to computational pathology.
Haoyu He, Basile Tessier-Cloutier, Yang Wang et al.
In briefEvaluating how synthetic private time-series windows are stitched together for forecasting
SummaryDifferentially private time-series generators usually output fixed-length synthetic windows, while forecasting models often need long continuous sequences. The study varies window overlap rates and weighting schemes and measures boundary continuity, fidelity, and Train-on-Synthetic-Test-on-Real forecasting utility. Across four public datasets and five forecasters, the best assembly choice depends on the forecasting model. Higher overlap generally improves continuity, but these diagnostics do not consistently reduce forecasting error.
In briefFinite-sample approximation of Hessian-guided perturbed Wasserstein gradient flows
SummaryThis paper studies whether particle approximations of the Hessian-guided perturbed Wasserstein gradient flow (PWGF) remain accurate over growing time horizons. It analyzes curvature accumulated along the reference path and proves particle and objective-value tracking bounds on a high-probability event under regularity and curvature conditions. The results are checked in a variance-plus-cosine model and partially in a regularized matrix-factorization model.
In briefRoBART: Bayesian tree regression with per-tree rotations
SummaryThe paper proposes RoBART, which gives each regression tree a single rotation shared by all its splits, so splits stay axis-aligned in rotated coordinates. A Metropolis-Hastings sampler jointly draws the rotation and cutpoints, and its reversibility is established. Under stated conditions, the authors prove posterior contraction rates for additive functions and show RoBART can adapt to intrinsic dimension where axis-aligned BART cannot.
In briefAccurate Edges Alone Don't Transfer: Why Dynamics-Learned Structure Fails in Inverse Problems
SummaryThe paper shows that transferring relational structure learned from forward-simulation data to inverse problems with scarce labels fails systematically, even when it meets the standard theoretical condition for helping. On hydrogen-leak and acoustic source-localisation tasks, structure learned by Neural Relational Inference degraded performance far more than a physics-based Green's function prior. The authors also propose a lightweight Jaccard-similarity transferability test, presented as a heuristic calibrated on few cases rather than a validated general threshold.
In briefOrthoGen: A Generative Orthogonal Learner for Time-Varying Treatments
SummaryThis paper studies estimating conditional distributional potential outcomes under time-varying treatments using flexible generative models, for example to assess patient risks under different treatment sequences. The authors propose generative recursive g-computation, which propagates full conditional outcome distributions rather than conditional means. They then introduce OrthoGen, a Neyman-orthogonal and doubly robust generative learner. Experiments on synthetic, semi-synthetic and real-world datasets report that OrthoGen is highly effective.
Tom\`as Garriga, Valentyn Melnychuk, Konstantin Hess et al.
In briefCARES: A controlled synthetic benchmark for judging which sounds matter in audio scenes
SummaryThis paper introduces CARES, a dataset for testing which sounds should be kept when describing an audio scene. The authors define a sound as salient when a speaker audibly reacts to it. The corpus is built from controlled scenarios with dialogues written by a language model and contains 10,000 two-speaker scenes. Six audio-language models were tested on scene identification, sound tagging, and reaction classification, and they were found to hear the sounds but miss the speakers' reactions.
Marcel Gibier, Thomas Thebaud, Olivier Bo\"effard et al.
In briefMachine learning estimates of slice genus and unknotting number for links
SummaryThe paper uses reinforcement learning and Bayesian optimisation to find new upper bounds for link invariants lacking known algorithms, including the slice genus and unknotting number of links and the strong slice genus of algebraically split links. Lower bounds are computed from known invariants, and combining both yields new exact values in many cases. The unknotting agents reproduce the non-additivity of the unknotting number in several counterexamples by Brittenham and Hermiller, and in some cases find new unknotting trajectories.
Yutong Dai, Oliver Hayman, Andr\'as Juh\'asz et al.
In briefTraining model organisms with multi-objective validation for interpretability research
SummaryThe paper argues that training model organisms, models with installed behaviors like backdoors or sycophancy, to a single objective is insufficient for evaluating interpretability methods. It proposes validating them on three objectives: target-behavior installation, general-capability preservation, and output naturalness. The authors show these validation metrics predict how well interpretability methods recover the installed behavior, and they introduce a model-merging-based multi-objective training approach.
In briefGAGR-Lab: A Framework for Testing Joint Spatial-Geometric and Analytic Function Reasoning
SummaryThe paper presents GAGR-Lab, a framework that measures how well a model turns a spatial scene into a symbolic function whose executed curve meets geometric constraints, using Cartesian game scenes and a Rust trajectory executor. In a bounded pilot, the hosted model Llama 3.2 11B Vision Instruct produced 432 attempts with no target hits, while a privileged analytic search control succeeded on 600 directional cases. The authors outline a staged prospective study plan, and the full difficulty matrix and comparative model results remain untested.
In briefRobust Decentralized Fairness Auditing (Auditopus)
SummaryThis paper proposes Auditopus, a method for auditing the fairness of large language models in which several auditors query the model separately and share only cumulative statistics vectors, not raw queries. The authors show that a single adversarial auditor can fabricate these vectors to make an unfair model appear fair. To counter this, honest auditors down-weight any auditor whose statistics are inconsistent with earlier rounds. Experiments report reduced audit error compared with no defense and standard robust aggregation, and state that unfair models are not passed as fair even when 49% of auditors are adversarial.
Sayan Biswas, Jade Garcia Bourr\'ee, Anne-Marie Kermarrec et al.
