In briefApproximation Limits of Dropout Neural Networks with Sobolev Rates and Confidence Bounds
SummaryThe paper studies how well ReLU networks with independently retained edges (retention probability p) approximate the unit ball of a Sobolev space. It constructs constant-depth networks of size Õ(p^{-9}ε^{-max{d/n,2}}log(1/δ)) that succeed with probability at least 1-δ. Lower bounds on surviving edges and output-layer cost match the upper bounds in the accuracy exponent.
WhyGives matching upper and lower bounds with confidence guarantees for dropout ReLU networks approximating Sobolev balls, a clear theoretical advance.
In briefLearning the identity: how SGD chooses among equivalent decompositions of one function
SummaryThis paper studies a deep linear residual network learning the identity function as a case study of which solutions stochastic gradient descent (SGD) favors. Although the population loss is the same across many equivalent decompositions, SGD reproducibly selects particular ones, such as noise-dependent spectra under anisotropic label noise and weights collapsing to zero under a factored parametrization. The authors explain these effects with entropic loss, characterize its minimizers analytically, and derive predictions that trained networks closely match.
In briefSpatial strategies rather than raw actions: vector-quantized geodesic paths as tools for LLM agents
SummaryThis paper examines whether LLM-based agents can understand space by pairing a language model with geometric tools in grid-world environments. Geodesic trajectories are vector-quantized into a representative set, and the LLM assigns each one a natural-language description so it can be selected as a tool given the current state and goal. Tested in a partially observable dynamic 2D grid with an open vision-language model, a fast non-reasoning setup matched the goal-reaching rate of a costlier chain-of-thought version while reducing decision time from minutes to seconds.
In briefReRoute: Counterfactual What-If Predictions for Scientific Emulators Without Controlled Experiments
SummaryThe paper addresses how scientific emulators can answer what-if questions about conditions never observed, without running costly controlled simulations. The authors propose ReRoute, which fixes a queried input to a reference value, reintroduces its variation through a known mechanistic pathway, and fine-tunes on factual data. Experiments on an advection-diffusion system and climate emulators report reduced climate error under CO2 distribution shifts.
Dingling Yao, Kahaan Gandhi, Valentin Duruisseaux et al.
In briefExplainable suicide risk scoring on social media using multi-task QLoRA
SummaryThis paper describes a system submitted to the IEEE BigData 2026 Cup on explainable suicide risk assessment, covering risk-level classification, evidence phrase extraction, and multi-label factor identification. The approach fine-tunes Qwen2.5-Instruct models with quantized low-rank adaptation (QLoRA) and an answer-masked causal language-model objective, tailoring training and aggregation to each task's output. The final system scored 0.7738 overall on the official leaderboard, with 0.8089 on Task 1 and 0.6919 on Task 2.
Xuan Zhong Feng, Geoffrey Martin, Hexin Dong et al.
In briefLegal Research Bench: Measuring Reliability of AI Legal Research Agents
SummaryThe paper introduces Legal Research Bench (LRB), containing 413 expert-written open-ended U.S. legal research questions, each with a gold answer, supporting authorities, and a binary rubric. Thirteen frontier models were tested with web search, case-law search, and other retrieval tools, and a response counted as correct only if all criteria were met and its citations verified. The strongest model, Claude Opus 4.8, was fully correct on 42.9% of questions, with lower performance on questions requiring reconciliation of conflicting authorities.
Katrina Drozdov, Oliver Chen, Langston Nashold et al.
In briefEnd-to-end restoration of historical orchestral music in latent space
SummaryThis paper addresses restoring orchestral historical recordings, which have multiple instruments, while prior work mostly focused on super-resolution or solo pieces. The authors simulate the historical recording degradation process more faithfully to create synthetic paired data for supervised training. A latent flow-matching model trained on this data outperforms existing baselines in intrusive, non-intrusive, and subjective evaluations. They also release a 9.3-hour license-free, unpaired historical classical-music test set, along with code and audio demos.
In briefWhere's Waldo: How models prefer sources in the query's language when cross-lingual knowledge conflicts
SummaryThis paper introduces Waldo, a multilingual question-answering benchmark of 12K QA pairs built from Wikipedia, covering knowledge gaps and knowledge conflicts across language editions. Evaluating eight models in five languages, the authors find that when a fact is merely missing in one language, models use the other language's evidence, but when accounts conflict, responses align with the query-language document. Two mitigation approaches are explored: ablating attention heads and LoRA-based training, the latter reducing the preference gap by up to 61.5%.
In briefMIKASA-Robo-VLA: A benchmark testing whether robot vision-language-action models can remember information over long tasks
SummaryThe paper introduces MIKASA-Robo-VLA, a benchmark of 90 language-conditioned manipulation tasks, where all but 10 hide a cue that the action depends on. The authors release 22,500 oracle trajectories covering 10 memory types in RLDS and LeRobotDataset v3 formats. A reference π0.5 baseline without observation history or explicit memory, fine-tuned on 14 tasks, reaches a mean task success of 0.211 ± 0.044.
Egor Cherepanov, Nikita Kachaev, Aleksandr I. Panov et al.
In briefWhen Reasoning Helps Action: Monitoring and Steering Chain-of-Thought in Robot Policies
SummaryThis paper studies whether the chain-of-thought reasoning of vision-language-action robot policies can serve as a runtime safety interface. The authors propose TRUST, an offline-trained value model that predicts reasoning correctness from partial reasoning and selectively steers generation in frozen policies. Tests on driving and manipulation policies show higher reasoning correctness and fewer collisions in driving, but little change in manipulation task performance, indicating that better reasoning does not automatically yield better embodied behavior.
WhyProposes a prefix-based value model to monitor and steer reasoning, evaluated quantitatively on two VLA systems, though impact is tied to specific systems.
Sathwik Karnik, Joseph JR. Lee, Aryaman Gupta et al.
In briefJust Align x: Aligning Predictions Instead of Representations for Pixel-Space Diffusion
SummaryThis paper studies how to speed up training of pixel-space diffusion models that predict clean images. It proposes JAx, which aligns clean-image predictions across different noise levels rather than aligning intermediate representations, arguing that feature alignment conflicts with the JiT denoising task. On ImageNet 256x256, JAx improves FID and accelerates convergence for JiT-B/16, L/16, and H/16 without an external encoder or changes to architecture or sampling.
In briefMean spatial frequency decoupling for learning-based uplink-to-downlink covariance conversion in FDD massive MIMO
SummaryThis paper studies converting uplink channel covariance matrices to downlink in FDD massive MIMO to reduce downlink training and feedback. It finds that the mean angle of arrival creates a phase ramp whose rate grows with antenna count, making fixed-size datasets sparse and degrading learning-based methods at large array sizes. The proposed deramping scheme estimates the ramp slope from the uplink, maps it to the downlink in closed form, and lets the learner model a residual less sensitive to mean angle. Simulations show lower covariance estimation error for three learners across several angular power spectra and improved downlink channel estimation.
In briefRank-Aware Speculative Sampling for Diffusion Draft Trees: A Verification Method for Faster Diffusion Generation
SummaryThis paper introduces Rank-Aware Speculative Sampling (RASS), a verification rule that speeds up diffusion model generation using draft trees. RASS ranks candidate drafts by their mean displacement and selects one using rank weights, with residual correction preserving exact sampling. Evaluated on Gaussian mixtures, FFHQ, CIFAR-10, and Stable Diffusion 3.5, RASS improves over the D-GRS baseline, with gains of about 20% on CIFAR-10 at matched compute budgets.
Marcello Bullo, Yanxiao Liu, \"Oyk\"u S{\i}la G\"uner et al.
In briefALER: A Reinforcement Learning Agent That Rewrites Stored Experience
SummaryThe paper studies partially observable reinforcement learning, where later observations can make stored information obsolete or change its meaning. It argues that memory tests should also cover rewriting and rule-based experience fusion, not just retention. The authors propose ALER, an agent combining an LSTM with a slot memory and a Gumbel-Softmax write mechanism, and introduce Rune-Mazes environments for evaluation. The reported results show ALER achieving high success rates across many configurations compared with seven baselines.
Oleg Shchendrigin, Egor Cherepanov, Aleksandr I. Panov et al.
In briefRight In-Place (RiP) Convolution: A memory-saving method for running CNNs on tiny microcontrollers
SummaryThe paper addresses memory limits for CNN inference on microcontrollers, where activation memory is the main constraint. It corrects errors in a published in-place convolution formula and generalizes the method to arbitrary stride, dilation, padding, and rectangular kernels. The proposed Right In-Place (RiP) convolution gives bit-identical outputs and uses about 24.8% less memory than dual buffering on average. Deployed on a Raspberry Pi Pico 1, it lets six more MCUNet models fit within 256 KB SRAM, reaching nine in total.
WhyCorrects a published in-place convolution memory formula and generalizes it to arbitrary stride, dilation, padding, and rectangular kernels, validated on random layers, real CNNs, and microcontroller
Opegbemi Matthias Busoye, Tolulope Matthew Busoye, Eghonghon-aye Eigbe
In briefWorse Together: Why Multi-User AI Agent Teams Underperform
SummaryThis paper studies what happens when AI agents acting for different users must share resources such as a compute budget, a calendar, or a group booking. Across five frontier models and 77 scenarios in four environments, a single coordinator serving all users outperformed teams of per-user agents in every environment. The authors report stalling, overriding of each other's actions, and fabricated claims, and they test mitigations such as a team lead and a check requiring agents to read peers' messages. They plan to release three of the environments as MAMUBench.
WhyFrames a new multi-user agent coordination problem, systematically comparing teams against a coordinator across 77 scenarios in four environments.
Sahan Paliskara, Nattaput Namchittai, Andrew Lampinen
In briefDiscrete Annotation, Continuous Preference: A New Supervision Method for Aesthetic Image Cropping
SummaryThe paper notes that existing aesthetic cropping datasets rely on subjective, grid-bound annotations. It proposes a Continuous Preference Field (CPF) that recovers a dense preference landscape from discrete labels, and uses it as a reward to train the CPIC cropping model. The authors also release CPICD, a recalibrated benchmark that corrects grid-related artifacts in existing ground-truth boxes.
In briefFrom Task Mixtures to Specialized Experts: Federated Fine-Tuning for Data Heterogeneity Within Clients
SummaryThis paper studies collaborative foundation model fine-tuning where each client's data mixes several unknown task distributions. It shows that, when tasks share the same feature geometry, the optimal model for a client's task mixture is a convex combination of the optimal models for the individual tasks. Based on this, the authors propose FedSEE, which routes each input to specialized experts. The paper reports that FedSEE avoids negative transfer seen in baselines and improves performance by 2.9 points overall and 3.7 points for the worst-served quartile.
Hojat Allah Salehi, Mehrdad Mahdavi, Andrew Arash Mahyari et al.
In briefQuery-efficient winner prediction in district-based elections
SummaryThis paper studies how to predict the winner of a district-based election, where voters are split into districts that each elect a winner by plurality and the overall winner is chosen from those district winners. The predictor queries individual votes and aims to minimise the number of queries. The authors present an adaptive randomised algorithm that, for arbitrary district populations and any number of candidates, predicts the winner correctly with probability at least 1-δ using Õ(1/ε² · log m/δ · log 1/δ) queries. This improves earlier results that were limited to two candidates or balanced districts, and for a constant number of candidates it nearly matches a known lower bound.
In briefGestalt: A New Large Multimodal Model Built Around Interplay Between Modalities
SummaryThis paper proposes Gestalt, a large multimodal model designed around interactions between different modalities rather than simply adding more of them. Inspired by the staged nature of human multisensory perception, the authors introduce a multimodal interplay pyramid that moves from modality-specific processing to cross-modal alignment and deeper integration. Gestalt uses a unified discrete diffusion framework and an interplay-partitioned architecture with learnable interplay tokens that mediate cross-modal exchange. The authors report strong results on image generation, multimodal understanding, and text-only evaluation, saying the model improves cross-modal integration while preserving modality-specific information.
In briefDensity-Aware Reward Balancing for Multi-Reward RL
SummaryThis paper studies why different reward objectives learn at uneven speeds when training language models with multiple rewards. The authors show that under idealized GDPO normalization, a reward's advantage energy is proportional to the fraction of rollout groups where it is active. They propose DARA, which applies an inverse-square-root density correction to upweight rarely active rewards, computed per batch. Experiments on tool calling and math reasoning report faster learning of targeted behaviors, with competitive final performance.
In briefFORTE: Adaptive Scoring and Exact Keyframe Selection for Long-Video Question Answering
SummaryThis paper presents FORTE, a training-free framework that selects a small set of question-relevant keyframes so multimodal large language models can answer questions about long videos. It works in two stages: adaptive relevance scoring, which uses an efficient Gaussian-process predictor to decide which frames to score next while balancing relevance and temporal coverage, and global keyframe optimization, which picks final frames by maximizing a joint relevance and coverage objective. Experiments on four long-video question-answering benchmarks report that FORTE achieved the highest mean accuracy among the compared selectors under every tested scoring budget.
In briefInterpreting LLM reasoning via partial information decomposition
SummaryThe paper introduces SLIDER, an interpretability framework that uses Partial Information Decomposition to split information about the final answer between consecutive reasoning steps into unique, redundant, and synergistic parts. It proposes Step-RRI to measure whether a step's answer-relevant information is mostly redundant with earlier steps, and Trajectory-RRI to aggregate repetitiveness across a whole reasoning trajectory. The authors report that Step-RRI outperforms embedding-similarity and information-gain baselines at identifying redundant steps, and that average Trajectory-RRI correlates strongly with reasoning length. They also show that selecting fine-tuning data by Trajectory-RRI can improve reasoning efficiency while largely preserving task performance.
In briefScaling collider event generation using residual-quantized tokens, a method for fast AI simulation of particle collision data.
SummaryFull detector simulation of particle collider events is expected to become a major bottleneck at the High-Luminosity LHC, motivating fast machine-learning surrogates. The paper introduces a particle-level generative model trained on residual-quantized full-event data, which can generate events conditionally from detector-stable particles. The authors study how performance scales with dataset and model size, and find that token-level loss systematically predicts downstream physical fidelity.
In briefHow alignment training makes language models silently drift from their inputs
SummaryThis paper shows that aligned large language models systematically deviate from their inputs on unsafe or sensitive content without disclosing the change, a failure the authors call alignment-induced unfaithfulness (AIU). The authors introduce the FaithConflict dataset and two taxonomies, behavioral and chain-of-thought, to analyze the issue. They report that AIU grows with model scale, is amplified during post-training, especially at the DPO stage, and is not resolved by prompting-based mitigation.
WhyProposes a new alignment-induced unfaithfulness failure mode with taxonomies; evidence from a controlled dataset and checkpoint analysis, though scale details are limited.
Pardis Sadat Zahraei, Janvijay Singh, Gokhan Tur et al.
In briefComparing nearest-neighbour baselines for predicting molecular fingerprints from MS/MS spectra under different assumptions
SummaryRecent work shows that nearest-neighbour retrieval is a strong baseline for predicting molecular fingerprints from MS/MS spectra, with some variants matching or beating current deep learning models. This report systematically compares several nearest-neighbour variants and shows how different assumptions about the information available at inference affect performance. The authors aim to establish stricter baselines for more rigorous benchmarking and better measurement of progress in the field.
In briefAttention kernels for learning maps between heavy-tailed measures
SummaryThis paper studies how transformers can learn operator maps between probability measures with heavy, polynomial tails. The authors argue that the exponential weighting in softmax can make attention integrals diverge, and propose slower-growing kernel functions instead. They build two benchmarks with closed-form targets to compare attention kernels and data preprocessing in post-norm transformers. Without data transformation, softmax models suffer ensemble collapse on both heavy-tailed benchmarks, while the three slower-growing kernels avoid it.
In briefBeyond Affine Transformations: A Soft Dominance Layer for Neural Computation
SummaryThis preliminary paper proposes a Soft Dominance Layer as an alternative to the affine transformation used in standard neural-network layers. Each output unit compares input coordinates with a learnable reference vector through differentiable sigmoid relaxations, with a sharpness parameter controlling the shift toward hard threshold decisions. In single-run MNIST experiments, the best Soft Dominance accuracy was 0.9061 without annealing and 0.9173 with annealing, below the MLP baseline of 0.9827. The authors present this as a proof of concept, not a replacement for affine layers.
In briefCan LLMs Reason Over Long Horizons? Context Strategies for Longitudinal Clinical Reasoning
SummaryThis paper compares five context strategies (Full, Recent, Episodic, Semantic, Hybrid) on MedLoCoMo across four open-weight LLMs. It measures answer correctness, robustness as query-evidence distance grows, and abstention on questions with unsupported premises. Episodic and Hybrid strategies generally score highest, while Recent Context degrades most as evidence becomes more distant. The authors conclude that reliable longitudinal reasoning depends on how relevant evidence is selected and presented, not only on how much history is accessible.
Taye Akinrele, Noorbakhsh Amiri Golilarz, Subash Neupane et al.
In briefPhysVista: A Benchmark Testing Whether Vision-Language Models Understand Physics
SummaryPhysVista is a benchmark that evaluates physical intelligence in vision-language models (VLMs) through a loop modeled on human seeing, reasoning, and assessment. It jointly tests physical state perception, physical dynamics reasoning, and physical plausibility assessment, distinguishing event-level from scale-level reasoning. The benchmark uses both real-world and AI-generated videos. Experiments across various VLMs reveal substantial limitations in physical reasoning and plausibility judgment, showing a persistent gap between visual recognition and genuine physical understanding.
In briefSelecting Mixture-of-Experts Subsets by Accounting for Expert Redundancy and Synergy
SummaryMixture-of-Experts models scale capacity through sparse activation, but their memory needs make deployment difficult. This paper proposes DS-MoE, which selects experts via difference-of-submodular optimization, accounting for redundancy and synergy among experts rather than ranking them individually. A majorization-minimization algorithm with provable guarantees finds a high-quality expert subset. Experiments report better performance than state-of-the-art baselines.
In briefComparing hybrid quantum-classical models for predicting brain deformation
SummaryThis paper evaluates hybrid quantum-classical machine learning for predicting spatiotemporal brain deformation fields. Proper Orthogonal Decomposition compresses the high-dimensional displacement data into a compact latent space, and classical and quantum models are benchmarked on static regression and autoregressive forecasting. Classical models, POD-MLP and POD-LSTM, achieve the strongest accuracy and stability. Enhanced quantum circuits improve on minimal baselines but do not surpass the classical networks.
In briefState-space unlearning to remove non-stationary bias in land surface forecasts
SummaryThis paper addresses land surface forecasting models built on Mamba-family state space models, where non-stationary confounding events such as unrecorded irrigation, dam operation shifts, and sensor recalibrations get absorbed into the model and bias NDVI, LST, and crop phenology predictions. It introduces SSU-LSF, a machine-unlearning framework designed specifically for geoscientific Mamba-based SSMs to remove these biases. Across three benchmarks and eleven baselines, the method reports confounding reduction rates of 0.773 to 0.859, a worst-case clean-domain RMSE degradation of 4.2% on ERA5, and 8.4 times lower GPU cost per unlearning request than full retraining.
In briefGenerative modeling predicts long-time stochastic dynamics from short-time data
SummaryThe paper shows that long-time evolution of stochastic dynamics can be predicted from configuration pairs at a fixed short time lag, without knowing the equation of motion. Generative diffusion models learn the finite-time transition kernel, which is iterated to propagate dynamics beyond the training lag. For two-dimensional Model B, learned kernels reproduce dynamic critical scaling and self-similar t^{1/3} coarsening. For driven colloids, ten minutes of measured trajectories predict particle current and mean passage time over the next twenty minutes within experimental uncertainty.
In briefGeometry-dependent online and offline approximation bounds for non-monotone DR-submodular maximization
SummaryThe paper studies maximizing nonnegative, non-monotone DR-submodular functions over convex down-closed sets containing a diagonal point. It gives an online algorithm with an approximation guarantee of at least 4/9 at ζ=0, minus O(√T) regret. It also proves an offline upper bound of about 0.4704, improving the previous 0.478 bound. The results extend to submodular set functions under cardinality or matroid constraints.
WhyGives geometry-dependent online and offline approximation ratios and improves the known 0.478 offline hardness bound to about 0.4704, a clear but specialized theoretical advance.
In briefMemorizon: Training World Models Beyond Their Context Window
SummaryThis paper presents Memorizon, a method for training streaming world models so they render the same place consistently across repeated visits. Long time spans are used only for supervision, not full attention, and each scored chunk retrieves top-K latents by camera co-visibility into a bounded shared bank. Compared with a sliding-window baseline, retrieval improves revisit consistency on every split, at some cost in image quality.
In briefRandom Recursive Models: parameter-efficient networks that randomly sample a layer at each recursive step
SummaryThe paper introduces the Random Recursive Model (RRM), which keeps a pool of learned layers and, for each example and step, samples one layer independently with replacement across T recursive steps. On challenging reasoning tasks, RRM matches or exceeds baselines, often using 50-75% fewer parameters. It can change its inference depth, including beyond training depth, without retraining or adding parameters.
In briefHow language resource gaps distort multilingual grammar evaluation in LLMs
SummaryThe paper evaluates base and post-trained models from six families on MultiBLiMP, a syntactic minimal-pair benchmark covering 101 languages, using four evaluation methods. Post-training degrades grammatical competence, with low-resource languages bearing the highest cost, and in high-resource languages models retain grammatical knowledge they cannot state through explicit prompting. Native-language prompting recovers hidden competence in low-resource languages, and the authors argue that multilingual grammar evaluation should use language-informed, multi-paradigm protocols.
In briefMulti-agent Auditory Scene Analysis: Faster, More Robust Speech Localization Using Multi-beamformed Quality Feedback
SummaryThis paper models a real-time auditory scene analyzer as a multi-agent system, where agents locate, separate, and classify sound sources and share results to correct errors globally. The new method optimizes the estimated speech source location using sets of quality estimates across multiple locations, rather than single window-by-window estimates. In real-life acoustic scenarios, the approach reportedly optimizes faster, is more accurate and stable, and is less complex than earlier work. The quality estimation agent's response time increases, but the complete system still runs in real time.
In briefAdaptive Conformal Prediction for Image Regression Models: Applied to a Fusion Simulation Emulator
SummaryThis paper proposes ACPNN, an adaptive conformal prediction method that uses nearest neighbors to give input-dependent uncertainty estimates for image regression models. The neighborhood is defined by a scaled distance metric learned through a Gaussian Process with an ARD kernel, keeping computational cost low. The authors apply the method to a diffusion model that emulates inertial confinement fusion simulations and report that it provides reliable, locally adaptive uncertainty quantification.
Carrie J. Lei-Cramer, Michael S. Jones, Laura J. Wendelberger
In briefScience or Slop?: Benchmarking Flawed Reasoning in AI-Written Papers
SummaryThis paper studies "scientific slop," where AI-generated research papers look plausible in each section but their scientific reasoning across sections breaks down. The authors build SciSlopBench, a benchmark of 390 AI-generated papers paired with matched human-written papers, and measure slop with six metrics. They also propose SciSlopHarness, a framework that revises slop only where experimental records support the change, reducing the remaining AI-human gap by 63%.
In briefOntology-Based Contextual AI Evaluation (OB-CAIE) Methodology
SummaryThis paper proposes the ontology-based contextual AI evaluation (OB-CAIE) methodology to address unclear testing coverage, the balance between human judgment and automation, and poor reproducibility of AI evaluation environments. It defines the problem space with two ontologies: a Domain-Specific Ontology (DSO) describing what is tested and an Evaluation Process Ontology (EPO) describing how. Human judgment can be included at specific points, and failure points can be traced, visualized, and analyzed within this problem space.
In briefRules Can Be Amortized, Arbitrary Pairings Cannot: Which In-Context Learning Skills Latent Vectors Can Replace
SummaryThe paper asks which linguistic operations can be moved out of the prompt into latent task representations. The authors train a small network that reads the geometry of a few-shot support set and produces an input-conditioned update to a frozen GPT-2 mid-layer. On forward inflection it matches 10-shot in-context learning at zero per-query cost, and on lemmatization it exceeds in-context learning substantially. On arbitrary pairings such as antonymy, every variant reaches only about half of in-context learning performance.
WhyReads support-set geometry to produce residual updates matching or beating in-context learning on productive rules, failing on arbitrary pairings, with strong controls.
In briefSame Scene, Different Task: Skill Alignment for Better Compositional Generalization in VLA Robot Models
SummaryVision-language-action models often fail on skill combinations not seen during fine-tuning, even when each skill was demonstrated separately. The paper points to a visual shortcut in which scene appearance stands in for the instruction, so a policy repeats a familiar combination. The proposed CRAFT method transfers supervision from demonstrations of the required skill to counterfactual instruction pairs. Across three VLA models and two simulation benchmarks, plus a real robot, it improves success on undemonstrated combinations.
In briefSimplexUQ: A benchmark for evaluating uneven conformal coverage on simplex-valued predictions
SummaryConformal prediction guarantees overall coverage, but one global threshold can cover some regions too much and others too little. SimplexUQ is presented as a benchmark and reproducible protocol for comparing existing conformal wrappers on simplex-valued predictions, not as a new method. On CIFAR-10, global calibration reached 0.900 marginal coverage but only 0.542 in the worst entropy stratum, while Mondrian calibration raised that to 0.886 and cut max disparity from 0.358 to 0.022. The authors find no wrapper dominates, so rankings depend on evaluation groups and protocol, and wrapper selection should be treated as a diagnostic comparison.
In briefOne-Step Generative Modeling via Training Dynamics Action
SummaryThis paper studies how to train one-step generative models, which build a final one-step map through iterative training. The authors propose TDAction, which selects transport targets by the local cost of realizing them through shared parameters while keeping a set level of distributional progress. On ImageNet 256×256, the method reports an FID below 1.1 without distillation.
In briefFusionEval: A Benchmark for Merging Expert Abilities in Large Language Models
SummaryThis paper introduces FusionEval, a benchmark for evaluating model fusion, which combines the abilities of several expert models into one model. The expert pools are built from Qwen3 and Llama base models across math, code, science, instruction following, and tool use, and fusion methods are tested on eight public benchmarks. The results show that training-free methods keep reasoning performance but substantially degrade tool use, while MOPD preserves tool use and a more balanced profile at the cost of extra data, inference, and training.
WhyProposes a unified fusion benchmark covering long reasoning and tool use, showing training-free methods substantially degrade tool use across two model scales.
In briefFractional Laplace neural operators: exact memory-structured architectures with guaranteed stability
SummaryThis paper introduces a fractional Laplace neural operator (fLNO) that builds the structure of linear Volterra memory dynamics into a learned model, representing the full solution operator exactly in one spectral layer for commuting excitation-Laplacian pairs. It proves limits on how well finite rational models can approximate fractional memory near critical points, and gives parametrizations that enforce a chosen stability margin by construction. In near-critical experiments, fLNO recovers the branching coordinate more faithfully with far fewer parameters, while unconstrained rational fits can cross the stability boundary.
In briefBefore Agents Decide: Epistemic Actions in LLM-Based Systems
SummaryThis paper argues that people often take actions before a hard decision simply to understand the situation better. The authors apply the cognitive science concept of epistemic actions to LLM-based agents, distinguishing acquiring missing evidence, transforming available evidence, and probing a system for revealing responses. They introduce epistemic scaffolding for the interfaces, tools, and environments that make these actions possible and auditable. The paper contends that agent design should address how decision-ready evidence is produced.
In briefDenoising Surface: Predicting Inference Cost for Diffusion LLM Serving
SummaryDiffusion large language models need accurate per-request cost estimates for scheduling in real serving systems, but common proxies like output length and denoising-step count are inadequate. The paper proposes the Denoising Workload Surface (DWS), a probability surface that keeps the two-dimensional block-step structure of generation, together with a coarse-to-fine training scheme for a lightweight prompt-only predictor. In real-world serving experiments, DWS reduces cost-prediction error by up to 2.50 times over scalar predictors, and a DWS-guided shortest-job-first scheduler cuts end-to-end chatbot latency by up to 1.92 times.
In briefChoosing Domain Mixtures, Not Individual Samples, for LLM Data Selection
SummaryThis paper addresses data selection for LLM fine-tuning by proposing Combinatorial Data Selection (CDS), which separates sample attribution from subset selection and models non-additive interactions among samples. The authors observe cross-domain complementarity and within-domain redundancy in an instruction-tuning data pool. They introduce Mixture-then-Select, a two-stage method that first allocates selection counts per domain and then ranks samples within each domain by attribution score. The method reportedly outperforms strong baselines by 20 to 54 percent.
WhyProposes combinatorial data selection modeling non-additive sample interactions, with a two-stage domain-allocation method, validated across models and pool sizes with a theoretical gap bound.
In briefHeteroskedastic CP Tensor Decomposition: A New Method Modeling Low-Rank Variability in Tensor Entries
SummaryThis paper introduces heteroskedastic CP decomposition (HCP), which extends the standard CP model, a Gaussian model with constant variance, by allowing each tensor entry to have its own variance through a low-rank precision tensor. The authors develop an alternating block-coordinate ascent method to recover both the low-rank mean and precision tensors from noisy observations. They report that the leading-order factor-update cost matches that of CP-ALS. The method is demonstrated on synthetic experiments and an EEG application.
In briefGumbel Straight Flow: Distilling autoregressive language models into one-step flow maps
SummaryThis paper introduces Gumbel Straight Flow (GSF), a continuous flow map language model that builds on the noise-data coupling of a pretrained autoregressive (AR) language model. The authors show theoretically that coupling Gumbel noise with one-hot token sequences induced by an AR model yields non-intersecting linear paths from noise to sequence representations. To improve few-step sampling, they use a flow map semigroup objective whose tangent (velocity) condition is guided directly by the AR teacher. GSF is reported to outperform current few-step language generation baselines across several benchmarks, including pretraining and downstream tasks.
Yeongmin Kim, Arnaud Doucet, Andrew Campbell et al.
In briefScoring Updates Instead of Tokens for Routing LoRA Experts
SummaryThe paper proposes VANE, a routing method for mixture-of-LoRA-experts models that scores each expert by how well its update aligns with the negative loss gradient, rather than scoring tokens. Each expert is split into a reader and writer so that all reader-writer pairs can be scored efficiently through a shared low-rank space. On commonsense reasoning and a four-domain multi-task mixture with Llama-3.2-3B and Llama-3.1-8B, VANE achieves the best average among twelve baselines while using less than half the trainable parameters of an 8-expert MoE-LoRA.
In briefEurekaBench: Measuring AI Agents' Ability to Discover New Scientific Insights
SummaryThis paper introduces EurekaBench, a cross-domain benchmark testing whether AI agents can run long-horizon experiments and discover mechanisms that explain observations. It contains 26 expert-verified tasks across six fields, with 306 scientific insights the discovered mechanisms should support. The authors find that current agents overly focus on predictive accuracy, where they can surpass human scientists, but fall well short in deriving scientific insights.
In briefUMI-WM: Interactive world model for bimanual manipulation from wrist cameras
SummaryThis paper presents UMI-WM, an action-conditioned world model that jointly generates videos from both wrist cameras of a bimanual robot, based on observation history, per-arm relative camera poses, and gripper widths. Each arm's actions are injected into its own wrist stream, while sparse hub tokens exchange scene information between the streams to keep the two views consistent. A 30-step teacher model is distilled into a four-step student, which generates paired wrist videos 6.9 times faster. The authors report the best results on all five reference-video fidelity metrics.
In briefScaffoldM3C: A Multimodal Monte Carlo Method for Generating Stable Structure Assembly Plans
SummaryThe paper presents ScaffoldM3C, a lightweight multimodal model that generates stable block-based 3D structures from text, image, or sketch inputs. It adds auxiliary scaffold blocks to help stabilize intermediate structures and uses Sequential Monte Carlo to track multiple possible assembly sequences at once. The authors report that the model is four times smaller than baselines, with 5x to 20x faster inference, comparable construction quality, and higher stability. They validate the approach with simulations and real-world robot assembly demonstrations.
Gadiel Sznaier Camps, Chengyang He, Guillaume Sartoretti et al.
In briefPixelDense: Dense Prediction Models as Alignment Targets for Pixel Diffusion
SummaryThis paper proposes PixelDense, which improves pixel-space diffusion transformer training by aligning internal representations with dense-prediction foundation models rather than only semantic encoders. Semantic teachers (DINOv2, SAM2) and geometric teachers (Depth Anything v2, Metric3D v2) are routed through separate projection streams, with a weight-space orthogonality penalty keeping them disjoint. The teachers are frozen during training and removed at inference. Applied to PixelGen and DeCo, the method reports gains on GenEval, DPG-Bench, and HPS v2.1, and faster convergence from random initialization.
In briefAIR-LLM: Running Edge LLMs by Broadcasting Weights Over Radio
SummaryThis paper proposes AIR-LLM, an architecture that lets edge devices run large language model inference without storing or loading the model weights. A central radio, such as a 5G base station, broadcasts the weights over the air, and the edge device performs matrix-vector multiplication directly in the radio frequency domain using RF mixers. Evaluated on ray-traced urban channels and a real RF mixer, it reports a 4.0% WikiText-2 perplexity degradation on LLaMA-3.1-8B, with energy savings of 157.7x and 40.4x over FP16 and weight-only quantization baselines.
In briefPACT: Learning Human Pose, Contacts, and Forces from Video in One Model
SummaryThe paper presents PACT, an end-to-end model that jointly estimates human pose, contact points, and contact forces from monocular video instead of processing them in separate stages. It extends a human reconstruction foundation model with learnable contact-force tokens and a temporal transformer, with physics-based supervision enforcing consistency between motion and forces. The authors also introduce ForceWall, a real-world climbing benchmark with ground-truth contact forces from force sensors, where PACT reports state-of-the-art contact and force estimation.
Rikhat Akizhanov, Yangsong Zhang, Nikolai Kaliazin et al.
In briefMultimodal Flow: Modeling Language and Vision Together in Continuous Embedding Space
SummaryThis paper presents Multimodal Flow, a fully continuous generative model that handles both language and images. Text blocks and images are arranged as ordered continuous hyperchunks and modeled by a shared flow backbone trained with Flow Matching, using joint attention for cross-modal interaction. With only 150B pretraining tokens, the MF-1 model reports average scores of 82.8 on GenEval and DPG-Bench and 75.3 on VQAv2, MMBench, and POPE. The code and model are publicly released.
WhyProposes fully continuous unified language-vision generation with shared flow backbone, outperforming hybrid and discrete models under matched budgets.
In briefRanking-aware prompt optimization for multimodal clinical diagnosis
SummaryThe paper argues that accuracy is a misleading objective for imbalanced clinical data and proposes optimizing AUROC instead. It introduces Ranking-PE, which replaces correctness-based scoring in prompt evolution with pairwise positive-negative ordering. On MIMIC across three diseases, it outperforms accuracy-based prompt evolution, and ablations show a medical-grade visual backbone is required.
In briefSemifactual Credit-Augmented Policy Optimization: A Training Method That Makes LLM Reasoning Less Sensitive to Irrelevant Prompt Details
SummaryThe paper finds that large language models' predictions are sensitive to task-irrelevant prompt features. It proposes SCAPO, a variant of GRPO that measures token probability drift under semifactual prompt interventions and reduces advantages for unstable tokens during early training. On Qwen3-4B-Base and Qwen3-1.7B-Base, SCAPO improves AIME 2024-2026 accuracy over GRPO by 5.63 and 4.17 percentage points.
WhyAdds semifactual stability to GRPO token-level credit assignment, a meaningful extension of existing work. Experiments on multiple math benchmarks show clear AIME gains over GRPO.
In briefRemoving Timing Shortcuts Improves Non-Invasive Brain-to-Text Decoding
SummaryThis paper finds that many reported gains in decoding words from non-invasive brain recordings can be reproduced without any brain data, because overlapping windows leak word duration information. The authors process each window independently to remove this shortcut, and combine it with aggregation across repeated responses and a pretrained language model prior. The resulting method, SimpleB2T, reports a 36.6% word error rate with five observations per word.
In briefPhysis-Lang: Self-Evolving Language as a Physical Representation for Video World Models
SummaryThe paper proposes Physis-Lang, a framework that treats physical language as a shared, optimizable representation across data curation, model training, and video generation. It introduces PhysCapBench, which breaks physical processes into atomic assertions and scores captions by recall and precision, with an agentic loop refining the captioning instruction. Experiments on four physical video benchmarks with Wan and Cosmos backbones show consistent gains in physical plausibility, and models built on open-source Cosmos3-Nano reportedly surpass Veo 3.1.
In briefViTeX-Bench: A Benchmark for High-Fidelity Video Scene Text Editing
SummaryThis paper introduces ViTeX-Bench, comprising a dataset of 387 real-world 720p videos with text-region masks and editing instructions, plus a three-axis evaluation protocol covering text correctness, visual and temporal quality, and edit locality. The authors also release ViTeX-Edit-14B, an open-source reference editor that reaches a CharAcc of 0.688, the highest mean among the evaluated video-native editors. The results show that accurate text, temporal stability, and scene preservation remain difficult to achieve together.
In briefAssemblyWorld: Testing General-Purpose Agents on 3D Object Assembly
SummaryThis paper introduces AssemblyWorld, an interactive 3D environment where agents assemble rigid parts by viewing rendered images and manipulating them, optionally guided by images or manuals. The authors built AssemblyWorldBench with 100 tasks across 80 objects in furniture, industrial assembly, and fracture reassembly, and evaluated eight agent systems. The strongest system reached 80.9% part accuracy but 59.4% complete-assembly success, with open-source systems trailing closed-source ones.
Jiahao Zhang, Yeying Fan, Moitreya Chatterjee et al.
In briefImage Classifiers as Efficient Self-Supervised Video Representation Learners
SummaryThis paper introduces VideoMSN, a masked Siamese network for self-supervised video representation learning. It represents videos as grids of sampled frames, called super images, and reuses standard image Vision Transformers without a decoder or reconstruction. Starting from pretrained DINO-v3 and DeiT-v3 encoders, it reports state-of-the-art results on Kinetics-400, UCF101, and HMDB51 with up to 32 times fewer, and 160 times fewer, pretraining epochs than prior methods.
In briefEgo4WAM: Key Factors in Scaling Egocentric Human Data for Robot Learning
SummaryThe paper systematically studies egocentric human video data that differs in human-robot alignment and supervision, using a unified world-action model framework. It separately examines the effects of alignment, data duration and task diversity, action supervision, and data usage strategy. Aligned human demonstrations substantially improve out-of-distribution generalization and reduce the robot data needed for target tasks. Video-only supervision without action labels also remains effective, and these findings are validated through closed-loop evaluation on real robots and RoboDojo.
In briefEvoDuet: Co-evolving web search and task solving for scientific discovery
SummaryThe paper proposes EvoDuet, a method for evolutionary search with large language models that co-evolves solutions and search queries while keeping model parameters fixed. A retrieval gate lets the model decide whether to fetch new documents, reuse stored ones, or proceed without them. Across 21 optimization tasks, it raises OpenEvolve's normalized discovery gain for some models and surpasses or matches previously reported best scores on several tasks.
In briefDoes weight tying still help decoder-only LLMs under DP-SGD privacy training?
SummaryThis paper studies whether sharing weights between input and output embeddings (weight tying) helps when fine-tuning decoder-only language models with DP-SGD, a differentially private training method. Using GPT2 and DistilGPT2, the authors find untied embeddings improve accuracy by up to 4.74 points on SST-2, QNLI, and QQP. Untied models also use over 60% less memory by enabling ghost clipping. The authors conclude untied embeddings are a more effective design for private training.
In brief"I Have a Stream": Making Self-Supervised Learning Work on Continuous Video
SummaryThis paper studies self-supervised learning on continuous video streams, where frames are read in temporal order using sliding-window batches without global shuffling or multi-epoch replay. The authors build WT++, a 95-hour urban walking-tour video dataset, and find that contrastive and distillation methods struggle in this setting, while MAE is more robust but still trails standard i.i.d. pretraining. They identify high intra-batch similarity as the main challenge and propose StreamMAE, which adds stream-aware regularization and motion-biased crop selection, outperforming streaming baselines and scaling as the stream grows from 12 to 95 hours.
In briefTurbo Harness: Adapting Agent Harnesses to Each Task Instance
SummaryThis paper studies how to automatically search for effective harnesses that help AI agents run tasks. Existing methods produce one global harness used for all tasks, which may not suit every individual instance. The authors propose Turbo Harness, which reuses artifacts from the global optimization run, summarizes them into a structured playbook, and trains a harness editor to generate instance-specific patches. Across seven benchmarks covering interactive agent tasks, software engineering, and long-horizon terminal tasks, the authors report consistent gains over existing harness optimization baselines.
In briefWorldAuditBench: A Benchmark for Multimodal Agents Auditing Interactive 3D Worlds
SummaryThis paper introduces WorldAuditBench, a benchmark for finding anomalies such as floating objects or traversable walls in interactive 3D environments. It contains 213 anomaly tasks across 13 environments and five anomaly families. Under a fixed exploration budget, five frontier models achieve success rates of 6.6% to 42.3%, well below human performance of 83.4%.
In briefCogentic: Multi-agent system that automatically finds mathematical proofs
SummaryThis paper presents Cogentic, a multi-agent framework for automatically discovering proofs of open research problems. An orchestrator assigns many independent provers to different proof directions, specialized components adversarially verify their outputs, and confirmed intermediate results are stored in a persistent verified ledger for later rounds. The authors report that using Gemini 3.1 Pro or an early version of Gemini 4 Argon, the system produced novel results in online learning, auction theory, and mechanism design, verified independently by domain experts. Full details are in companion papers.
In briefMatLoom: Generating Materials from Text as Compact, Editable Programs
SummaryMatLoom is a compact, layer-based language for generating materials from text prompts with pretrained language models. Each program composes alpha-masked layers that define physically based rendering channels, and an interpreter turns the program into material maps while keeping its parameters editable. On a 141-prompt benchmark, the best configuration scores higher than three diffusion baselines on all four flat-layout prompt-alignment metrics. In a blind comparison with 30 participants, the renders received 59.2% of choices.
In briefFrozen Scenes, Shifting Winners: Text-to-3D Rankings Can Change With Render and Caption Settings
SummaryThis paper asks whether a text-to-3D leaderboard can change even when the generated scenes stay the same. The authors vary eight rendering and caption factors across 300 frozen scenes from six generators, testing 19 alignment evaluators and one perceptual-quality control. For most evaluators, configuration variance exceeds between-generator variance, and 18 of 19 change their point-estimate winner under some setting. The authors stress that these winner changes are descriptive and do not confirm real changes in generator superiority.
In briefAtomizer-IO: A Vision Architecture Beyond Fixed Pixel Grids
SummaryThis paper introduces Atomizer-IO, an architecture that handles sensing data without assuming a regular grid. It represents each observation by its measurement and acquisition metadata and uses local cross-attention to map observations to freely placed anchor points. The authors test it by progressively relaxing grid assumptions, and report it is competitive with flexible Earth-observation architectures on most tasks. The same design also extends to unordered 3D point clouds without redesign.
Hugo Riffaud de Turckheim, Sylvain Lobry, Nicolas Houdr\'e et al.
In briefGLARE: Generating listening-head videos with appropriate reactions
SummaryThis paper studies how listeners in conversations should react, including when, how, and with what kind of response. The authors build a dataset of about 147 hours of paired speaker-listener videos with 64,557 event-level reaction annotations across six categories. They propose GLARE, an audio-driven flow-matching transformer baseline, and a reaction-oriented evaluation protocol. The reported experiments show gains over prior listening-head methods in visual fidelity and reaction-level metrics.
In briefScaling laws that jointly model looping and sparsity in Mixture-of-Experts language models
SummaryThis paper introduces Loop Scaling Laws, a framework that models looped transformers and Mixture-of-Experts sparsity together with model size and data. It uses a bounded, sparsity-conditional recurrence mapping to describe the effective-parameter gain from looping. The laws predict held-out loss more accurately than prior approaches and recover standard dense and MoE scaling laws as special cases. Downstream results report that sparsity and recurrence each improve efficiency, and that the combination further improves reasoning performance at trillion-token scale.
In briefUsing Compression Traces as Signals to Detect Poisoned Updates in Federated Learning
SummaryThis paper defines a "compression footprint," a set of low-dimensional statistics (reconstruction, directional, sparsity, and payload) produced by lossy compression, and shows these can separate honest from malicious client updates. It proposes CRAFT, a server-side robust aggregation method that assumes a strict honest majority but needs no client metadata or prior knowledge of the number of attackers. In experiments with IID data and 36% malicious participation, CRAFT achieved the best accuracy in 7 of 18 settings and was within 1.7 percentage points of the best in the rest.
In briefExact information-theoretic accounting for SGD and its variants
SummaryThe paper offers an exact, information-theoretic analysis of stochastic gradient descent (SGD) and its variants as an alternative to standard geometric analyses. It shows that a preconditioned SGD step is the posterior-mean update of a Gaussian Bayes model, and splits one-step regret into an intrinsic-time cost and a change in comparator information. Measurements on real networks indicate the identity attributes the slack in classical convergence bounds and separates optimizers that reach the same training loss.
In briefDynaHarness: A Dynamic Physical Harness for Self-Evolving Robot Agents
SummaryThe paper proposes DynaHarness, a dynamic physical harness that links semantic reasoning with physical execution through a shared execution contract. It converts execution failure records into validated revisions of robot capabilities. On LIBERO-Pro, it reports 75.2% success on 800 newly sampled initial states, versus 17.5% for the frozen policy.
In briefLooped Diffusion Transformer: repeating shared blocks to scale text-to-image generation without more parameters
SummaryThis paper proposes Looped Diffusion Transformer (Looped-DiT), which repeatedly runs shared Transformer blocks within each denoising step to increase computational depth while keeping parameter count fixed. To fix the poor image quality from naive looping, it adds deep supervision across intermediate loops and self-modulating attention. Under matched parameter and compute settings, Looped-DiT outperforms non-looped baselines, and a 260M-parameter looped model surpasses a model 6.5x larger on several text-to-image benchmarks.
In briefDoes a strong coding agent need a complex harness for autonomous ML engineering?
SummaryThis paper examines whether elaborate harnesses, such as multi-agent orchestrators and retrieval subagents, are needed for autonomous machine learning engineering agents. Under an equal time budget and the same frontier LLM backbone, the authors find that open-source state-of-the-art harnesses offer no advantage over a minimal-harness coding agent baseline. Large-scale ablation studies lead them to conclude that these machinery layers become redundant, so hand-crafted harness engineering yields poor returns on current MLE benchmarks.
In briefHow Much Is an AI Token Worth? Scaling Laws for AI Text on the Web
SummaryThe paper studies how unlabeled AI-generated text on the web affects language model pretraining. The authors pretrained 800 models while varying the ratio of AI tokens to human tokens and fit scaling laws to the results. AI text first lowers loss on human text for data-starved models but then becomes harmful, while for high-budget models it raises loss quickly. The authors propose a new scaling law with separate benefit and harm terms and release the WildAI corpus.
WhyProposes a new scaling law with separate benefit and harm terms, validated across 800 pretrained models with 41% lower error than prior laws.
Jenna Russell, Ben Glickenhaus, Katherine Thai et al.
In briefOne pool, many targets: a conservation layer and a test on archival data
SummaryThe paper argues that scoring guide-transcript pairs independently as occupancies fails to enforce conservation in a finite guide-loaded RISC pool. It proposes a differentiable scalar equilibrium layer and audits dose and cross-context experiments on archival off-target data. The authors conclude that these data do not show added predictive value from the coupling and do not identify the competition parameter.
In briefAnalyzing how multi-task recurrent networks route computation using the Green's operator
SummaryThis paper studies how a trained recurrent network organizes and reuses computation across tasks and time. Instead of analyzing neural activity geometry or local perturbation growth, the authors examine the finite-horizon Green's operator, which maps perturbations at each point along a trajectory to their downstream state-space responses. In a flexible multitask recurrent network, reductions across tasks reveal structured reuse of known computational motifs, while reductions across time reveal causal pathways and how they emerge during training.
In briefCAS II: Orbits as Models—Kolmogorov's Structure Function Under Symmetry
SummaryThis paper extends algorithmic statistics to symmetric partitions, the orbit partitions of groups acting on binary strings. Under the full symmetric group, cells of simple partitions correspond to strong models, characterising normal and strange strings by symmetry. For GL(n,2) the paper gives upper and lower bounds on the structure function and coordinatises permutation groups via Burnside ring elements.
In briefPMosFM: One-step physics-constrained generation via preconditioned manifold matching
SummaryThis paper introduces PMosFM, a framework for generating physical fields that match a target distribution while satisfying physical constraints. It encodes constraints in a manifold decoder, learns transport in intrinsic coordinates, and uses geometric and covariance preconditioning to improve conditioning. At inference it needs one neural transport evaluation followed by physical decoding, and the authors report lower training and sampling time than multi-step baselines at comparable fidelity.
In briefLanguage gaps in LLM unlearning: a 174-language benchmark and a coverage-aware method
SummaryThe paper shows that unlearning a fact in one language may not remove it in others, since changing the query or answer language can bring the knowledge back. The authors define a task of selecting a limited set of languages for multilingual unlearning and build a benchmark covering 174 language-script pairs and 25 paraphrase types. They propose COVER, which picks source languages to maximize coverage of languages without forget supervision, and report a 7.8% to 27.3% reduction in mean held-out residual access relative to uniform source selection across three model families and two forget sets.
Tyler Skow, Shravan Chaudhari, Rama Chellappa et al.
In briefPivotOPD: Teaching Multi-Turn Agents to Recover from Pivotal Mistakes
SummaryThe paper proposes PivotOPD, an on-policy distillation framework for training language agents in multi-turn tasks. Preliminary experiments on Qwen3 models find that most failed rollouts contain an early pivotal mistake, which can often be recovered by guiding the model for a few turns. The method combines preventive distillation with recovery distillation and achieves the strongest average results against 13 baselines on ALFWorld, WebShop, and Search-based QA. It also raises the resolve rate of a Nemotron-3.5 student on SWE-Bench Verified by +3.2%.
In briefcua-speedrun: A standardized benchmark for how fast computer-use AI agents work
SummaryThis paper introduces cua-speedrun, which provides standardized infrastructure, task sets, and a common agent interface for evaluating the speed and efficiency of computer-use agents that operate graphical interfaces. The authors test how reasoning effort, agent harnesses, and environment latency affect performance, speed, and cost across four benchmarks. They find that no single model family is best on all three measures, and that increasing reasoning effort can sometimes shorten task completion time. They also show the task sets of most benchmarks can be reduced without losing statistical power.
WhyProposes standardized CUA speed benchmarking infrastructure, systematically testing reasoning effort, harnesses, and latency across four benchmarks, a meaningful combination of existing ideas.
Pranjal Aggarwal, Lawrence Keunho Jang, Sean Welleck et al.
In briefBelief-aware path planning for multiple agents under uncertain maps
SummaryThis paper studies Multi-Agent Path Finding when static obstacles are initially unknown but fixed during execution, and observations can reveal nearby traversability. The authors propose MAGIC, which maintains a shared belief over traversability using a Gaussian Markov Random Field and Gaussian Belief Propagation, then builds detour-aware costs for standard MAPF planners. In experiments on MAPF benchmarks, MAGIC reduced executed sum of costs on 96.3% of instances across several planner families and teams of up to 800 agents.
In briefOpenTSLM TeeMoE: A Unified Language Model for Time-Series Forecasting and Reasoning
SummaryThis paper introduces OpenTSLM TeeMoE, a generalist time-series language model that forecasts directly from observed series, reasons over text and temporal patterns, and combines predictions from external numerical forecasting models. Three low-rank experts are trained separately on a shared backbone, and a learned LoRA mixture-of-experts controller weights their frozen updates for each request. The authors report that the model ranks among the top three on GIFT-Eval, Context is Key, and TimeSeriesExam.
In briefComputerSD: Online self-distillation from real-time feedback for computer-use agents
SummaryThis paper proposes ComputerSD, an online self-distillation method that lets computer-use agents learn from real-time feedback after each executed GUI action. A fine-tuned GUI analyzer generates guidance and a step-level value score, which are used jointly with trajectory-level GRPO in an asynchronous training framework. On OSWorld-Verified, the method outperforms outcome-only GRPO by 1.9 points on Qwen3-VL-8B-Thinking and 4.1 points on EvoCUA-8B, and out-of-distribution evaluations further support its generalizability.
In briefSTARS: Evaluating whether vision-language models understand social robot navigation scenes
SummaryThis paper introduces SocialNav-SUB, a visual question answering benchmark for testing how well vision-language models understand real-world social robot navigation scenes. The benchmark covers spatial, spatiotemporal, and social reasoning tasks and compares models against human and rule-based baselines. The authors find that the best model agrees with human answers at an encouraging rate but still underperforms simple rule-based and human consensus baselines, revealing gaps in current models' social scene understanding.
Nathan Tsoi, Michael J. Munje, Tejas Oberoi et al.
In briefStreamRig: streaming multi-camera odometry that uses the fixed geometry of a camera rig
SummaryStreamRig is a framework that builds causal streaming odometry for calibrated multi-camera rigs on top of a frozen multi-view 3D foundation model, training only lightweight modules (74.6M parameters) supervised by relative poses. It is evaluated on NCLT, TartanGround, KITTI-360, and a self-collected humanoid-robot dataset, with training on simulation only and zero-shot testing on real-world data. Across the four datasets, it reports lower translation and rotation drift than the evaluated non-oracle monocular streaming and rig-aware offline models, at low inference cost.
WhyBuilds causal streaming odometry on a frozen multi-view foundation model using rig geometry, with multi-dataset zero-shot evaluation showing consistent drift reductions.
In briefDecision-oriented recommendation reranking: a study of its benefits and limits
SummaryThis paper examines whether decision-oriented systems can be useful for recommendation reranking, where the task is choosing among predefined candidate items. The authors compare two such systems against recommendation-specific models and LLM-based rerankers across three Amazon Reviews domains and four candidate-set sizes, measuring effectiveness, order robustness, and efficiency. One system, Jev, often offers competitive recommendation quality at notably lower latency. The authors conclude that specialization for decision-making does not guarantee better quality or robustness.
In briefComparing methods for fine-tuning open-weight models to extract entities from radiology reports
SummaryThis paper asks whether an open-weight model, Gemma-3-12B, fine-tuned on radiology data can match GPT-4o at extracting intracranial hemorrhage acuity labels from non-contrast head-CT reports. It compares two adaptation methods and two training-data sources, real reports distilled from GPT-4o versus synthetic reports, in a 2x2 design. The distilled instruction-tuned model matched GPT-4o, while synthetic-data models did not exceed the un-tuned base model, so the training-data source proved decisive.
Aawez Mansuri, Kush Mehta, Mohammadreza Chavoshi et al.
