In briefLearning from the Hardest Examples: Provable Gains of Data Pruning
SummaryThis paper develops a high-dimensional theory of learning under data pruning using weighted empirical risk minimization, giving exact limiting test error curves when input dimension and sample size grow proportionally. It shows that strategic pruning can speed the decay of classification test error from the classical sqrt(N/P) rate to the faster Bayes-optimal N/P scaling when samples greatly exceed dimension. The optimal strategy keeps only the 2N hardest training examples and discards nearly all remaining data.
WhyGives a precise high-dimensional theory of pruning, proving error decay can speed from sqrt(N/P) to N/P by keeping only the 2N hardest examples.
In briefData-poisoning attacks on protein design models' pre-training and retrieval corpora
SummaryThis paper studies two data-only poisoning attacks on protein design systems. One inserts trigger-tagged malicious sequences into the pre-training corpus of autoregressive protein language models such as ProtGPT2 and ProGen2 variants. The other inserts records into the retrieval corpus of a text-guided protein generator. The experiments show that poisoned data can give models reliable trigger-dependent generation and that inserted records can capture retrieval results and alter generated sequences, while ordinary model behaviour is largely preserved. The authors conclude that the integrity and biological provenance of protein sequence data are critical safeguards for generative protein design systems.
WhyApplies known data poisoning to protein design pre-training and retrieval corpora, with 300M-parameter models and retrieval experiments, mostly transferring existing ideas to a new domain.
Charalampos Koilakos, Kimon-Antonios Provatas, Michail Patsakis et al.
In briefCondition-Contrastive Extrapolation for Language Model On-Policy Distillation
SummaryThis paper introduces CCE-OPD, a method that makes teacher supervision in on-policy distillation sensitive to question conditions, so the teacher does not favor continuations that violate the question's premises. It combines a fixed-prefix comparison that masks a selected condition with a bounded log-probability contrast that reweights the teacher distribution. Reported results show that distilling Qwen3-8B into 0.6B and 1.7B students improves avg@8 over OPD by 1.60 and 1.95 points on average across six math benchmarks.
In briefControlFlow3D: Distilling multi-view knowledge into latent flow matching for point cloud upsampling
SummaryThe paper proposes ControlFlow3D, a method that upsamples sparse point clouds into denser 3D point clouds. It distills knowledge from pretrained foundation models into a lightweight generation network during training, so multi-view images are not needed at test time. The method performs conditional flow matching in a latent space defined by a frozen 3D encoder, and the authors also build ShapeNetPU, a 35K-object multi-modal benchmark. The paper reports effectiveness and strong generalization across multiple benchmarks.
In briefTraining-free weight editing to resist harmful instructions
SummaryThis paper proposes a training-free weight-editing method to reduce instruction-following language models' compliance with harmful instructions. It needs only forward passes over a small set of prompt pairs per target behavior, and the edit is stored in the model weights, so changing the prompt cannot undo it. Tested on five instruction-tuned models for instructed lying and instructed sexism, the method raised honest response rates and cut the sexist-response rate from 57.9% to 24.3%, or 11.9% when combined with a fairness prompt.
Ali Ebrahimpour-Boroojeny, Xiaocong Yang, Samuel Schapiro et al.
In briefSensor fault diagnosis to improve fuel and CO2 estimates in heavy-duty diesel engines
SummaryThe paper tests methods that use relationships learned from healthy sensor data to detect faulty readings and correct fuel-use estimates in a heavy-duty diesel engine. On 217 steady-state operating points with five synthetic fault types, the diagnosis procedure issued diagnoses for 77.3% of groups, 92.1% of which were correct. Under stated assumptions, selective fuel correction reduced mean absolute error in calculated combustion CO2 by 83.8%.
In briefMine, Sample, Render: A test-time method for diverse generative outputs
SummaryThe paper notes that modern generative models produce high-quality single outputs but struggle to sample diversely, even when instructed to. It proposes MINE-SAMPLE-RENDER (MSR), which mines an interpretable categorical schema and statistics from a corpus of real artifacts, samples attribute configurations from those statistics, and uses a generative model only as a conditional renderer. Evaluated on weather report generation, user simulation, webpage design, text-to-image generation, and agent-generated 3D assets, MSR is reported to improve alignment with corpus distributions and output diversity.
In briefIWOO: A Wave Optics Optimizer Improved with Hybrid Strategies
SummaryThis paper proposes IWOO, an improved version of the Wave Optics Optimizer that uses optical diffraction and interference for numerical optimization. It adds four complementary strategies to address slow convergence and getting trapped in local optima. IWOO is tested on IEEE CEC2019, CEC2022, and CEC2017 benchmarks against representative optimizers, with ablation studies for each strategy. It is also evaluated on six constrained engineering optimization problems.
In briefWhy LLMs Miss Evidence in Multi-Hop RAG Systems
SummaryThis paper examines why wrong answers in multi-hop retrieval-augmented generation lack required evidence, distinguishing inaccessible passages from queries the agent never issued. The authors introduce ExecutionRAG, a training-free requirement register that tracks open sub-questions and their supporting evidence. On 1,000 MuSiQue questions, it reduces wrong answers with missing required evidence across two embedding models. The authors conclude the remaining bottleneck shifts from guiding search to verifying that retrieved evidence is sufficient.
In briefUnsupervised deep learning method to align surface reflectance across satellite image dates
SummarySatellite surface reflectance images can show systematic differences between acquisition dates because each image is atmospherically corrected separately. This study proposes DeepMGCCA, an unsupervised framework that combines deep generalized canonical correlation analysis with multivariate alteration detection to learn deep pseudoinvariant features for removing these biases. Tested on four annual Landsat-8/9 and Sentinel-2 acquisitions across four regions, a monotone piecewise-linear transform based on the features' quantiles gave the largest bias reduction.
In briefDoes Latent Chain-of-Thought Actually Reason? A Unified Taxonomy
SummaryThis paper examines whether latent chain-of-thought, which replaces written reasoning steps with continuous hidden states, actually performs reasoning. Across twenty-one model-task configurations, the authors combine chain deletion, content replacement, probes and trajectory analysis. They find that on logical tasks the answer is often readable before the chain begins, while on arithmetic tasks all tested recipes depend on chain content. The authors introduce a taxonomy separating what a chain represents from how the model uses it.
In briefJudging the Judges: Confidence and Disagreement in Rubric-Based Rewards
SummaryThis paper studies Rubrics as Rewards (RaR), a method that uses an LLM judge to score responses against a structured rubric for reinforcement learning in non-verifiable domains. The authors show that simply changing the judge produces a substantial spread in scores when training Qwen3-4B-Base. They propose RaRE, an ensemble of judges that weights confident judges more heavily via logprobs and masks high-disagreement rollouts from the GRPO update. RaRE outperforms every single-judge RaR variant on HealthBench and GPQA Diamond.
Michael Tian-Yue Liu, Anne Harrington, Trevor Darrell et al.
In briefAgentEval: Evidence-Bounded Checking of Stateful AI Agents
SummaryAgentEval is a system that keeps a contract-based assessment record across tool execution, state capture, report review, and case upkeep. It separates what the agent did, what the evidence shows, and whether the conclusion supports its present use. In 32 tool-environment runs, independent checks agreed on all 30 complete episodes, distinguishing satisfied, violated, and unresolved obligations, while two interrupted runs were kept. Controlled changes to 48 evidence views revealed 36 binding faults that hash-only checks missed, without altering agent behavior.
In briefLanguage models share a readout bottleneck in working memory
SummaryThis paper studies how large language models perform on the N-back task, where accuracy falls as memory load increases and errors concentrate on lure items. The authors compare two explanations: imprecise softmax attention versus interference from competing items' content, testing 15 pretrained models and minimal transformers. Attention usually lands on the correct target, yet competing items' content corrupts the decision, and suppressing non-target letter identity improves sensitivity. The authors conclude the working-memory limit lies in reading out items under interference rather than in locating them.
WhyDissociates precision and interference accounts using probes and suppression across 15 models, locating LLM working-memory limits in readout.
In briefMemScope: Localized code repair to evolve agent memory architectures
SummaryThis paper proposes MemScope, a framework that modifies an LLM agent's memory architecture locally, based on execution evidence, instead of rewriting it at large scale. A tool-using Patch Agent inspects raw execution traces, runtime memory state, and architecture code to find repair targets and make bounded in-place edits. The evolved architecture reports gains on GAIA, xBench, and ALFWorld, with a median code-change ratio of 4.4%.
In briefGRADS: A Graph-Guided Method for Generating Training Data with Controllable Difficulty
SummaryThis paper introduces GRADS, a framework that synthesizes training data for deep research agents by tracking the evidence and reasoning dependencies needed to answer each question. It generates and verifies question-answer tasks while conditioning each step on a relevant subgraph. Experiments in security, software engineering, and scientific documents show empirically separated difficulty levels, and a 4B model trained on this data matches Sonnet 4 on software engineering and surpasses it on the scientific and security domains.
In briefFrom Long Context to Selective Memory: A Survey of Memory-Centric Compression for LLMs
SummaryThis survey reframes long-context compression in large language models as a memory-management problem during inference. It separates compute-active memory, which directly takes part in computation, from compute-inactive memory that can be stored, retrieved, updated, or forgotten. Operations such as writing, offloading, eviction, and consolidation are placed within a single budgeted cost model. The authors also summarize evaluation dimensions and benchmark gaps, and call for dynamic, lifecycle-aware memory management.
In briefBengali speech-based sentence auto-completion for dementia support using a Transformer-based speech recognition system
SummaryThis paper proposes a system to help dementia patients by detecting signs of cognitive decline from Bangla speech and text and offering sentence auto-completion during conversations. A Transformer model transcribes voice recordings into text, and several machine learning and deep learning models were evaluated, with LSTM architectures performing strongly on contextual patterns. For completion, an attention-based Bidirectional LSTM was developed to complete partially expressed Bangla sentences.
Md Asif Akram, Arif Faisal, Sakir Hossain Faruque et al.
In briefCoordination-free data loading that balances per-rank batch cost in distributed training
SummaryThis paper addresses idle time in synchronous data-parallel training, where ranks receive equal record counts despite highly variable record costs. It proposes Batch Cost Lookahead (BCL), a rank-local data-loading policy that buffers a bounded window and repartitions it into fixed-size packs with balanced estimated cost, without cross-rank coordination or runtime changes. The authors report up to 82.8% lower per-rank batch-cost variability in simulation, and speedups of 1.050 to 1.167 times over pooled MFFD across 13 NeMo-RL SFT corpora on eight H200 GPUs. They also note that a p99 cap is needed for high-variance or OOM-prone data."}}LG
In briefVisualTrust: Testing Video Reasoning When On-Screen Text Makes a Claim
SummaryThis paper studies whether multimodal AI models let on-screen text claims override what the video visually shows. The authors build VisualTrust, a benchmark of 5,063 cases where each video appears with no claim, an accurate claim, or a misleading claim. Mean accuracy drops from 90.1% without a claim to 51.9% with a misleading claim. They also propose DEAR, a training-free agent that separates evidence arbitration from reading and raises accuracy under misleading claims to 81.8–85.2% on their test set.
WhyIntroduces a counterfactual benchmark and training-free arbitration agent, evaluated on 11 models and 5,063 cases, giving fairly strong evidence of claim deference.
In briefMaterial and Defocus Cues for Locating Occluded Objects in Single Hyperspectral Images: An Exploratory Study
SummaryPartial occlusion makes objects hard to detect when visible parts have weak texture or similar appearance. The authors combine material-dependent spectra and wavelength-dependent optical blur from hyperspectral imaging using nonnegative matrix factorization, divergence-based optical cues, and graph Laplacian propagation. The method localizes objects with bounding boxes without a separate depth sensor at inference. Qualitative examples from 30 indoor and six outdoor scenes show when material and optical cues help, while objects sharing both material and depth remain challenging and quantitative evaluation is still needed.
In briefIs Recursion Really Needed? A One-pass Medical Image Registration Method Guided by Multi-level, Multi-space Context
SummaryThis paper addresses deformable medical image registration and proposes M2CG, a one-pass network that performs multi-step progressive pyramid registration in a single iteration instead of using recursive passes. Each step is guided by contextual cues from previous, current, and subsequent steps across image, image feature, segmentation, and segmentation feature spaces. M2CG supports unsupervised and semi-supervised training and reports state-of-the-art results on two medical registration tasks across eight public datasets. The authors suggest that the reliance on recursion may arise from insufficient contextual information describing the remaining distance to the target alignment.
In briefTraining Proactive Dialogue Agents Against Resistant Users, Not Soft Ones
SummaryThis paper introduces MASP, a framework that models proactive dialogue as a partially observable game between a system and a user who rationally resists. It represents the user's changing mental state with a BDI-based latent state and alternates self-play updates for both the system and user policies. Compared with SFT and planner-only approaches, MASP improves success rate and rationality-constrained success rate across diverse benchmarks and reduces average turns.
In briefFrom Observation to Intervention: Distilling Self-Generated Demonstrations via Causal Intervention
SummaryThis paper studies how large language models filter the few-shot demonstrations they generate for themselves. The authors argue that judging rationales only by whether they yield correct outputs cannot separate faithful causal rationales from spurious post-hoc ones. They propose ECI2D, which turns rationales into executable programs, applies interventions to logically similar samples, and keeps only demonstrations showing transfer invariance. On math, logic, and commonsense reasoning benchmarks, ECI2D outperforms observation-based in-context learning while retaining far fewer demonstrations.
In briefAll-in-One Memory: process-rewarded reinforcement learning for AI memory agents
SummaryThe paper proposes All-in-One Memory (AiO-Mem), which treats search, memory writing, and stopping as actions over a shared evidence state, scored in one reward space. A process reward model scores complete interaction trajectories, and PA-DAPO uses these scores to optimize the memory policy. According to the authors, AiO-Mem improves answer accuracy on long-context memory benchmarks over static memory systems and RL-trained memory agents.
In briefCoSTA: A Framework for Faster Spoken Dialogue Models Using Compressed Speech Tokens
SummarySpoken dialogue models usually use one speech-token space for both generating and synthesizing speech, which produces long sequences and slow inference. CoSTA keeps the original speech tokenizer and decoder but separates the generation and synthesis token spaces, adding compressed representations that are deterministically expanded back before synthesis. On bilingual speech-dialogue benchmarks, it reaches about 1.7 times faster expanded speech-token generation, with lower response latency and higher throughput while preserving response quality.
In briefWhen is a costlier, more accurate LLM worth it? Stochastic optimization with switching oracles
SummaryThis paper studies when paying for higher-fidelity large language models is worthwhile in LLM-assisted data curation for downstream optimization. The authors propose two cost-aware adaptive frameworks that switch between stochastic oracles of different fidelity, one for known bias and one for unknown bias using budgeted certification. Synthetic and LLM-completion experiments show the adaptive policy skips several expensive oracles, substantially reducing cumulative cost while achieving comparable accuracy.
In briefCleanRM: Training reward models from noisy feedback using a few clean labels for RLHF
SummaryPreference data often contains noisy feedback from subjective judgments, annotation errors, and imperfect LLM judges. The paper proposes CleanRM, which uses a small amount of clean feedback and optimal transport to calibrate per-sample noise ratios, then trains the reward model with a noise-corrected surrogate loss. The authors prove the objective is unbiased with respect to the clean risk. Experiments show CleanRM outperforms state-of-the-art denoising baselines and improves downstream RLHF alignment.
SummaryThis opinion article argues that the next AI will succeed by knowing its user rather than by knowing more about the world. It introduces the term OnBodyAI for large language models given a body, living on the user, working for them and answering to them, and discusses what research and industry need to make this happen. The article is based on the author's talk at the UbiComp/ISWC 2026 Industrial Panel and follows its ten key statements.
In briefPathwise Shrinkage and Probabilistic Calibration of a Multiplicative Half-Cauchy Process
SummaryThe paper studies the multiplicative half-Cauchy process using its log-scale random-walk representation. It identifies shrinkage, expansion, and critical regimes, along with almost-sure summability, local rebounds, and effective dimension. It also gives a semi-analytic rule that links interpretable prior probability statements to the process's scale parameter.
In briefPredicting live birth after IVF/ICSI fresh embryo transfer using a hybrid attention and logistic regression model
SummaryThis study uses data from 616 single-centre fresh embryo-transfer cycles, with 251 live births and 365 non-live births, to predict live birth before transfer. It introduces a differential weighting of inner cell mass and trophectoderm morphology and develops a hybrid attention network ensembled with logistic regression. On a held-out test set, the ensemble reached an AUC-ROC of 0.710, outperforming the listed baselines. The authors note that external validation is the next step.
Abdul Raheem Shahzad, Jinxin Chen, Ming Zhang et al.
In briefPedestrian path prediction using hierarchical envelope adversarial domain adaptation (HEADA-LTCTPM)
SummaryThis paper addresses two limitations in pedestrian trajectory prediction: considering only one type of social coupling, and relying on predefined statistical metrics for domain shift. It proposes HEADA-LTCTPM, which builds envelope samples extended into graph structures and uses an adversarial domain adaptation module combining local and global strategies. On two public datasets against six baselines, ADE and FDE drop by 17.6% and 19.1% respectively, with time overhead meeting practical requirements.
In briefSparse attention that speeds up video generation by growing sparsity and reusing early search results
SummaryDiffusion Transformers are widely used for video generation but are expensive at inference because self-attention scales quadratically. This paper proposes GSRS, a training-free method that keeps more attention computation early in denoising and reduces it progressively later, while limiting later searches to an early candidate set. Experiments on HunyuanVideo and Wan 2.1 report 1.55 to 2.21 times speedup with generation quality comparable to full attention.
In briefWhen judges agree and are wrong: a blind spot in label-free evaluation
SummaryThe paper tests whether consistency can stand in for correctness in label-free evaluation methods. Using four financial QA datasets, ten consistency dimensions, and two frontier model families, it examines consistency and accuracy jointly for individual examples. It finds that 60% of incorrect answers fall in a high-consistency region, and 77.7% of those show complete answer collapse, with identical wrong answers across all perturbations, which agreement-based signals cannot detect.
In briefCompression Does Not End at Selection: Positional Geometry for Sparse KV Caches
SummarySparse KV caches cut long-context inference cost by keeping only selected past states, but those states keep their original RoPE positions, so gaps left by evicted tokens can weaken query-key matching. The paper studies post-selection positional geometry on a fixed cache and proposes RoPERelay, which assigns compact, order-preserving coordinates and resynchronizes later generation without reselection, re-prefill, or extra budget. Across 36 comparisons with two models, six selectors, and three long-term memory tasks, it improves answer quality in 32 cases; for Llama-3.1-8B-Instruct with BM25 on LongMemEval-S, accuracy rises from 23.6% to 35.4%, with about 1%-3% measured time overhead.
In briefNoise Policy Optimization: aligning few-step generators to rewards by learning over their initial noise
SummaryThis paper addresses steering few-step generative models toward desired properties when only black-box reward scores are available. It proposes Noise Policy Optimization (NPO), which keeps the generator frozen and learns an amortized flow policy over its initial Gaussian noise so that the outputs follow a reward-tilted distribution. The authors report that NPO achieves a better reward-quality trade-off than direct fine-tuning across image, language, and protein structure generation, and matches test-time search methods at 16 to 549 times lower inference cost.
WhyProposes learning an amortized flow policy over initial noise, validated across image, language, and protein tasks with 16 to 549 times lower inference cost.
In briefDST-Net: A decomposed network that models trend and seasonal parts of time series separately for forecasting.
SummaryThe paper proposes DST-Net, which splits a time series into trend and seasonal components and models each differently. The trend branch uses a Latent Graph Block to capture global cross-channel dependencies, while the seasonal branch uses Gaussian Multi-scale Attention with learnable Gaussian priors to focus on local temporal patterns. The authors report experiments on multiple real-world benchmarks demonstrating effectiveness.
In briefA method for analyzing which farming decisions drive results in agro-ecological simulations
SummaryThis paper studies sensitivity analysis for an agent-based agro-ecological model whose parameters form a decision tree, where many parameters only matter when a higher-level choice activates them. The authors encode the crop-management itinerary of the MAELIA platform as a hierarchical design space with categorical and numerical variables. They propose a random-forest-supervised HSIC-ANOVA method that ranks main effects and interactions on the tree. Applied to outputs such as nitrogen leaching, soil organic carbon variation, and yield predictions, it yields an interpretable, decision-oriented reading.
Benjamin Pianet, Baptiste Lesquoy, Paul Saves et al.
In briefResStreamGS: Residual-Sensitive Dual-Stream Gaussian Splatting for Mobile Volumetric Video
SummaryThis paper presents ResStreamGS, a method for streaming real-time volumetric video to mobile devices under limited bandwidth and computing power. It separates 3D Gaussian Splatting primitives by temporal behavior, sending temporally stable Gaussians once per group of pictures and only residual-sensitive Gaussians in the per-frame video stream. Experiments on complex full-scene videos report 7.51 to 8.35 dB higher post-codec PSNR than an all-Gaussians baseline, up to 56.5% lower amortized payload, and 30.883 FPS on a Snapdragon 778G.
In briefSpectral-Spatial Interpretation: Per-position, per-channel attribution for multispectral image models
SummaryThis paper introduces Spectral-Spatial Interpretation (SSI), a post-hoc method that explains model decisions on multispectral images. For each spatial position, SSI fixes the surrounding context and applies the Shapley value to a local game over its channels, producing a full position-channel attribution matrix. On HLS Burn Scar, the authors report that SSI outperforms an MC-Owen estimator on retained attribution metrics while using about 6% of its evaluation budget.
In briefSSR: A Training-Free Retriever for Scaling Agent Discovery in LLM Multi-Agent Systems
SummaryThis paper proposes SSR, a deterministic, training-free method that retrieves suitable agents from Agent Cards so an LLM orchestrator does not need the full registry. It derives lexical signals from capability fields, weights them with IDF, and scores candidates through an Inverted Signal Index. At 1,000 agents, the authors report 98.8% less LLM input and 70.9% lower orchestration latency than full-catalog orchestration.
Giancarlo Maldonado C\'ardenas, Carlos A. Astudillo, Nelson L. S. da Fonseca
In briefKolab: A benchmark for human-agent collaboration on long workflows
SummaryThe paper introduces Kolab, a dataset of over 115 hours of screen recordings from 50 professionals in 40 occupations working with AI agents on realistic long-horizon tasks, including their collaboration and output quality judgments. The authors find that frontier models, agentic and non-agentic alike, overestimate collaboration with experts and struggle to tell high-quality from low-quality collaboration. Longer workflow context does not consistently improve these predictions.
Shirley Anugrah Hayati, Patricia Anugrah Setiani, Ruizi Wang et al.
In briefHow LLM planners commit to plans when they cannot see everything
SummaryThis paper studies two strategies for LLM-based planning agents in partially observable environments: aggressive planners commit to long plans using assumptions about unseen states, while conservative planners plan only over observed states. Both reach similar accuracy across three LLMs, but conservative methods fail less often and make more API calls. The authors also explore hybrid strategies with tunable hyperparameters to balance safety and efficiency.
In briefA one-sided McDiarmid-type inequality for heavy-tailed variables, using reverse KL tensorization
SummaryThis paper revisits McDiarmid's inequality, a core tool in statistical learning theory. It shows that for functions of independent random variables that are almost surely bounded above and satisfy a conditional moment assumption of order p in (1,2], large deviations above the mean are exponentially unlikely. When p<2, this holds even if the function has infinite variance. The proof relies on a new tensorization inequality for KL divergence, called reverse KL tensorization.
WhyIntroduces reverse KL tensorization extending McDiarmid-type concentration to heavy tails with only finite p-th moments, a substantive theoretical advance with rigorous proofs.
Hanuka Eshima, Amedeo Roberto Esposito, Andrea Paudice
In briefToolBottleneckBench: Measuring Tool-Agent Performance Under Controlled Pressures and Fixed Budgets
SummaryThis paper introduces ToolBottleneckBench, a benchmark that tests tool-using AI agents on fixed synthetic enterprise tasks while varying one pressure at a time, such as the visible tool menu. Across 21 model and harness configurations, only the visible menu caused a large performance penalty, which tracked truncation at the completion cap. The authors recommend reporting pressure effects together with the harness budgets used, and the benchmark is released on GitHub.
WhyIntroduces a benchmark isolating tool-agent bottlenecks via controlled pressures; experiments across 21 configurations show truncation drives the penalty.
In briefDo world modeling and action modeling help each other under unified causal training?
SummaryThe paper introduces OneWM, a framework that treats actions both as conditions for predicting the world and as targets for policy learning within a shared video-action architecture. The authors evaluate policy learning and world modeling on LIBERO, LIBERO-Plus, and RoboCasa365, with complementary real-world robot experiments. They report that learning the visual consequences of specified actions improves policy performance and generalization across changing scenes, and that action learning improves multi-view world-prediction fidelity.
In briefMemComposer: Learning to Assemble Memories for LLM Agents with Reinforcement Learning
SummaryThis paper proposes MemComposer, a memory composer that examines the current query and retrieved candidate experiences to decide which to include, how many to select, and their order. It is trained with Experience Policy Optimization (ExPO), a GRPO-based method that uses task-success feedback from complete agent interactions. Across three τ²-bench domains and two frozen executors, MemComposer outperforms the strongest baseline for each executor while reducing average memory-token usage by 16.9% and 4.9%.
In briefRoboCloth: a large real cloth dataset for learning how fabrics reflect light
SummaryRoboCloth is a dataset of 500 real cloth materials, photographed with robot-controlled, calibrated cameras and lights, yielding about 580 high-dynamic-range frames per material. The authors train a shared MLP decoder on this corpus to learn cloth material priors, then fit a 2048x2048 latent texture to each target material. On held-out materials and the independent UBO2014 benchmark, a decoder trained on their captures outperforms decoders trained on existing measured reflectance datasets and analytic PBR fitting. The dataset is released under CC BY 4.0 on the Hugging Face Hub.
WhyA new robot-controlled capture dataset plus a two-stage neural pipeline, validated against existing decoders on the independent UBO2014 benchmark.
In briefProbe-and-Resume: Deciding When and How to Check an AI Agent's Work
SummaryThe paper studies test-time search for AI agents in terminal-based computer-use environments, finding that the evaluated MCTS setup does not significantly improve task success. It attributes this to critics lacking needed environment state and to intermediate states giving weak signals about eventual success. The proposed Probe-and-Resume method invokes the critic only when the agent believes it is finished, letting the critic probe a fork of the current and initial state before a rejected agent resumes. On Terminal-Bench 2.1, the authors report an 8.4 percentage point gain over MCTS with conventional critics.
In briefSIEVE: Fixing a Conflict Between Point Cloud Masking and Rotation-Equivariant Networks
SummaryThis paper shows that standard masked autoencoding for 3D point clouds conflicts with SE(3)-equivariant networks, causing a topological collapse that forces the model to invent orientation without a reference frame. The authors propose SIEVE, a self-supervised framework that compresses point cloud geometry into invariant scalar embeddings, using a spherical projection mask that keeps global pose while hiding local semantics. Experiments report that SIEVE avoids the predicted collapse that constant and noise masks suffer, and its embeddings outperform coordinate-based and hand-crafted baselines on downstream tasks, with the largest gains on unseen shape categories.
In briefRoPE-Flow: Monocular Vision-based Robot Pose Estimation via 2D Graph-Conditioned 3D Flow Matching
SummaryThis paper presents RoPE-Flow, an end-to-end flow matching framework that estimates camera-to-robot pose from a single monocular image. It extracts keypoint-centered features from predicted 2D heatmaps and generates flow conditions using a learnable adjacency map over keypoints, with bone length and ray-depth losses for consistency. The authors report outperforming state-of-the-art methods across robot types and multi-camera setups, and introduce a cross-scene benchmark for zero-shot evaluation.
In briefA Jury, Not an Electorate: Label-Free Trust Estimation for Language Model Ensembles
SummaryThis paper treats an ensemble of language models as a jury whose members have unknown reliability. The authors use Expectation Maximisation to estimate each model's reliability from the agreement and disagreement patterns among their answers, without labels. On a 1,560-question benchmark with nine models of up to 3B parameters, the method performs comparably to the same fusion rule using label-fitted weights. It also outperforms the best single model with matched self-consistency and the strongest label-free alternative.
Alessandro Rozza, Davide Mambelli, Steven Motta et al.
In briefSci$\Pi$vot: A benchmark testing whether AI research agents make sound judgment calls
SummaryThis paper introduces Sci$\Pi$vot, a benchmark of 248 cases for evaluating scientific judgment in LLM research agents. The cases come from six months of projects across six scientific fields, where senior researchers corrected agents' research decisions. Models must choose between the original LLM plan and the human-corrected alternative, and most frontier models score below chance, while fresh human researchers score higher.
In briefAdaptive randomized neural networks for PDEs with smarter feature allocation and residual-based expansion
SummaryThe paper proposes SFRBE–RaNN, a two-stage adaptive method for solving partial differential equations with randomized neural networks. It first redistributes a fixed feature budget across regions based on indicator masses and adjusts regional weight scales, then adds nonredundant feature blocks by screening activation families and removing already-represented response directions. The authors establish an exact residual-reduction identity and a conditional solution-error estimate for fixed linear problems, and report numerical experiments on localized, multiscale, and nonlinear space-time problems.
In briefOmicsBench: A Long-Horizon Multi-Turn Benchmark for AI Agents in Research-Scale Bioinformatics
SummaryThe paper introduces OmicsBench, a benchmark of 97 long-horizon, multi-turn bioinformatics tasks comprising 662 linked subtasks, each graded deterministically against an answer key. Across 12 models on three harnesses, mean scores ranged from 0.43 to 0.57 without in-context feedback and 0.54 to 0.67 with it, showing no combination saturates the benchmark. The authors also propose an unsupervised pipeline that clusters agents' failed attempts to surface emergent failure modes.
In briefSkip the Compressor: Coding Agents Fix More Issues When They Pick Whole Files
SummaryThis paper studies how coding agents should choose code when a repository exceeds their context window. The proposed selector, CQ2-CARE, spends a fixed token budget on showing the files that must change verbatim, instead of pruning tokens or retrieving chunks. On 300 recent SWE-rebench issues it resolves more issues than three compressors, three retrievers, and two issue-resolution methods.
Xiaocong Yang, Rishit Chatterjee, Himanshu Udupi et al.
In briefAI-Text Detection for Math-Heavy Science Using Images
SummaryResearchers study how well detectors can flag AI-written scientific passages containing heavy mathematics. They generate paired human and AI versions of arXiv passages with 15 LLMs and compare LaTeX source, rendered PDF images, and extracted text. Existing detectors lose accuracy on math-heavy text, while the proposed image-based SnapJudge detects 92% of AI proofs and 93% of AI prose at a 0.1% false positive rate.
WhyProposes an image-based AI detector for math-heavy papers, validated on paired LaTeX, rendered, and extracted-text data with clear gains over Pangram 4.
In briefSampling Under Score Mismatch in Zero-Shot Conditional Discrete Diffusion
SummaryThe paper studies zero-shot conditional generation with discrete diffusion models, where a model trained on the unconditional distribution is reused to sample from a conditional target. It derives a general KL divergence upper bound that quantifies the extra bias caused by this score mismatch, and analyzes the Conditional SEDD sampler under partial-observation constraints. The authors propose a Randomized Hard Clamping (RHC) sampler for coordinatewise set constraints, and report numerical and real-data experiments showing its effectiveness.
Renxiang Huang, Yuchen Liang, Yingbin Liang et al.
In briefBellman-consistent trajectory reinforcement learning for LLM arithmetic reasoning
SummaryThis paper proposes a Bellman-consistent trajectory reinforcement learning method to address sparse terminal rewards in LLM arithmetic reasoning. Its core component, Trajectory Bellman Residual Minimization (TBRM), converts sparse rewards into dense token-level signals, combined with an RLOO group strategy and hindsight goal relabeling. On the Countdown task, the TBRM-RLOO framework reaches Pass@1 of 0.55 to 0.64 and Pass@16 of 76.0% to 84.0%, outperforming SFT, IPO/DPO, and RLOO baselines.
In briefAthanor: An agent-native system where AI agents autonomously post-train a base model
SummaryThis paper presents Athanor, an open system on real GPUs that lets frontier AI agents carry out post-training research on a small base model without external teacher models. In a 104-hour campaign, four agents raised Qwen3-4B-Base to scores on a development suite and a held-out suite that match or exceed the official Qwen3-4B-Instruct-2507 on average. The authors report that the agents reached this result mainly by recombining known methods and each other's models.
In briefSALTO: a solver-anchored Koopman model that infers finite-amplitude sources from measurements.
SummaryThis paper introduces SALTO, a Koopman-based framework for inferring finite-amplitude sources in nonlinear systems. In a lifted space, the nonlinear dynamics are approximated by a linear propagator whose transpose propagates sensor data backward in a single solve. On three nonlinear PDE systems, SALTO predicts measurements to within 2-8% error, where the classical adjoint has 41-165% error.
In briefPremiseCheck: Efficient fact-checking of human and AI peer review comments
SummaryPeer review faces growing strain, and review comments written by humans or AI can misstate what a paper says. The authors introduce ReviewFact, a benchmark that labels factual errors in review comments under one definition across NLP, machine learning, and natural sciences, measuring cost and latency alongside accuracy. Their PremiseCheck framework has an LLM rewrite each comment as explicit premises without reading the paper, then a decision model checks each premise against the paper. Checked comment by comment, it reaches accuracy comparable to GPT-5.5 in all three fields at about half the cost, and at a quarter of the cost on NLP reviews.
In briefLearning When to Withdraw: Fine-Tuned Agents and Rule-Specific Failures
SummaryThis paper studies how tool-using language agents should suspend prepared actions when evidence changes or authority is revoked, then resume after repair. The authors fine-tune Qwen2.5-3B-Instruct with LoRA in a synthetic environment, NAQD-Env, comparing original and enriched training curricula. Fine-tuned checkpoints outperform baselines on aggregate agreement and withdrawal recall, but rule-level diagnostics reveal that the original-curriculum model wrongly withdraws many actions that retain alternative support. The authors present their method as a way to detect regressions that aggregate accuracy hides, not as evidence that fine-tuning alone ensures safety.|
In briefReviewPilot: A Multi-Agent System for Screening and Data Extraction in Systematic Reviews
SummaryThis paper presents ReviewPilot, a conversational multi-agent system that automates the PRISMA workflow for systematic reviews, from paper screening to data extraction. Researchers can adjust criteria and extraction schemas through dialogue, and the system records the reason for each screening decision and the source of each extracted value. On epidemiological expert reviews, the authors report that it outperforms a single-prompt LLM and AgentSLR across three LLMs. Case studies in biomedicine and human-computer interaction show close agreement with human annotators.
In briefRecursive Reward Improvement: an adversary checks and repairs model-written grading rubrics for open-ended tasks
SummaryThe paper proposes RRI (Recursive Reward Improvement), a self-improvement method for open-ended expert tasks such as clinical report generation, where no automatic answer key exists. A reward-hacking adversary writes an answer that exploits the generated rubric, and up to two repair steps revise the rubric so an honest answer wins. On HealthBench Professional, RRI improves adjusted scores over a generation-only self-improvement baseline at 9B, 27B, and without external models, and the 27B model also gains on PRBench Hard and ProfBench.
In briefA benchmark for steering audio models to describe specific speaking-style aspects
SummaryThis paper formulates controllable speaking-style captioning as an instruction-following audio-language task, where instructions name the speech dimensions to describe, such as speaker traits, emotion, or delivery. It introduces StyleInstructCaps, the first standardized benchmark for this task, with a dataset and an evaluation framework covering groundedness, hallucination, instruction following, generalization, and speaker-style consistency. Testing nine state-of-the-art audio-language models, the authors find that strong instruction following does not guarantee grounded understanding of speaking style. They also report that training with SSC-specific, instruction-diverse data substantially improves groundedness, controllability, and speaker consistency.
Shreeram Suresh Chandra, Ashwyn Sharma, Liang Lu et al.
In briefINCONVENIENT: A Benchmark for Testing LLM Generalization from Biased Convenience Samples
SummaryThis paper introduces INCONVENIENT, a benchmark that tests how well models generalize when training data come from convenience samples with systematic bias. Instead of using naturally occurring domain splits, it directly perturbs feature-target relationships in training data while keeping the original population as the deployment environment, with explicit control over shift direction and severity. The benchmark includes 200 tabular classification and regression datasets and generates about 15,500 shifted training distributions, which cause substantial downstream performance degradation.
Sara Rezaeimanesh, Chad Klochko, Kundan Thind et al.
In briefLearning When to Think: Utility-Guided Adaptive Reasoning for Efficient Retrieval
SummaryThis paper proposes AURA, a framework that supervises intermediate reasoning states based on how well they separate relevant documents from hard distractors. It also trains a lightweight utility estimator to adapt reasoning depth at inference time. On the BRIGHT benchmark, AURA-4B reaches 34.7 average nDCG@10 using 2.8 reasoning steps and 96 generated tokens per query. Adaptive stopping cuts latency from 4.6 to 0.76 seconds, and MS MARCO results suggest in-domain retrieval is not substantially degraded.
In briefSCARCE: Spontaneous Rule-Making Among Coding Agents Sharing Compute
SummaryThis paper introduces SCARCE, a framework where coding agents complete sequences of tasks under individual deadlines while negotiating access to a shared execution resource for inspecting code, compiling, and running tests. The authors find that agents spontaneously form informal institutions, persistent rules and commitments based on reputation, reciprocity, and precedent, which influence resource allocation. They present a 21-category taxonomy and audit over 1,500 trajectories across eight models.
Abhinav Kumar, Rediet Abebe, Eugene Bagdasarian et al.
In briefLast-iterate convergence of clipped gradient methods under heavy-tailed noise
SummaryThe paper studies last-iterate convergence of first-order methods for stochastic strongly convex optimization. Prior guarantees for SGD often assume light-tailed gradient noise, which is unrealistic, and SGD may fail under heavy-tailed noise. The authors analyze Clipped-SGD, which adds gradient clipping to SGD, and prove the first optimal last-iterate convergence rates in high probability under heavy-tailed noise.
In briefLearning to Read Out: How Language Models' Output Layer Shapes Capability Timelines
SummaryThis paper asks when a language model acquires a capability and argues that timelines based on when outputs become correct can be misleading, because outputs pass through a readout matrix that is still being learned. The authors introduce parameter-trajectory crosscoding, a method that uses only checkpoint weights to track the readout. In Pythia's pretraining, they find a narrow early window where the readout reorganizes, and hidden states already encode task information before the model's outputs show it. Small models pretrained with different readout learning rates show that a higher rate makes the readout develop earlier and closes the gap sooner.
In briefFrom Single Songs to Music Corpora: Testing an Explainable Harmony Model on HookTheory
SummaryThis study applies a fixed, explainable melody harmonization model (EMH), originally built for individual songs, to human-annotated melody and chord excerpts from HookTheory. The model generates chords from melody alone, with human chords used only for post-hoc comparison. Across 439 evaluable harmonic segments from 19 excerpts, exact chord agreement was 20.50%, root agreement 31.66%, and harmonic-function agreement 47.62%. The analysis separates representational limits from differences in chord selection.
In briefWhether acceleration can speed up Gaussian variational inference, and where it fails
SummaryThis paper examines whether momentum-based acceleration, which speeds up gradient descent on strongly convex smooth problems, carries over to Gaussian variational inference for strongly log-concave targets. With exact Gaussian expectations, a convex composite formulation achieves the classical accelerated rate. With only pointwise derivatives, the authors construct targets for which any randomized algorithm needs roughly κ^{1-o(1)} queries, so curvature bounds alone do not guarantee uniform acceleration. Under a Hessian-Lipschitz assumption, variance-reduced stochastic methods can be accelerated when Hessian variation is small, and experiments on real datasets show substantial empirical speedups.
WhyConstructs a family of targets with κ^{1-o(1)} query lower bound, showing curvature alone cannot guarantee uniform acceleration; adds positive Hessian-Lipschitz results and broad experiments.
In briefDeep learning maps small forest plantations in India's Godavari districts using satellite data
SummaryThis paper uses deep learning to map small-area forest plantations in the East and West Godavari districts of Andhra Pradesh, India. It classifies cloud-free Harmonized Landsat Sentinel-2 satellite data with a pixel-level neural network, using training labels from field surveys and expert aerial photo interpretation. The deep learning approach is compared with a random forest baseline over a 21,543 sq. km study region across three years, with differences analyzed by land cover class and season.
Snehal More, Anuj Karpatne, Valerie A. Thomas et al.
In briefFailure as Learning Signal: Recovering Supervision via Verified Hindsight Notes
SummaryThe paper addresses GRPO training, where problems whose rollouts all fail provide no learning signal. It proposes Replay-Selected Hindsight Distillation (RSHD), in which the policy writes corrective notes on failed attempts without a reference answer and re-solves the problem with each note. A note is admitted only if its replays succeed more often than the original group, and admitted replays are distilled into the plain-prompt policy. The authors report consistent gains over GRPO on in-domain benchmarks across model families, sizes, and two domains, with gains carrying over to out-of-domain evaluations.
In briefCausalEvo: Causal Skill Self-Evolution for Health LLM Agents
SummaryThis paper proposes CausalEvo, a method that lets health-focused large language model agents accumulate and select textual skills, such as calculation criteria and formula units, based on task feedback. With model weights, feedback cases, and rule count held fixed, it compares paired skill replacements to estimate which skills to retain. On six MedCalc-Bench-Verified tasks with 120 held-out cases, it reaches 60.00% accuracy, versus 55.83% for SkillOpt and 51.67% without skills, and uses 55.6% fewer training tokens than SkillOpt. Gains concentrate on particular calculations with regressions on others, and paired accuracy intervals include zero.
In briefAudio-visual identification and tracking of individual dogs in real videos
SummaryThis paper presents a multimodal framework that links dog barks to individual dogs in real-world videos by combining visual identification, multiple-object tracking, and canine vocal fingerprinting. It keeps persistent profiles that work despite occlusion, overlapping barks, and off-screen barking. The authors also introduce a benchmark of 46 videos with 213 manually annotated bark events. At IoU of at least 0.30, the framework correctly localizes the active visible dog in 61.83% of visible events and reaches a 91.33% F1 score for distinguishing visible from off-screen barking.
Aryahvishwa Babu, Kingsley Jin, Lillian Wang et al.
In briefStatistical evaluation of deep reinforcement learning for HPC job scheduling
SummaryThis paper compares six deep reinforcement learning setups for scheduling jobs on high-performance computers, using PPO, A2C and DQN, each with and without action masking. The methods are tested on two real Slurm workload traces against three heuristics and a masked random control. On the homogeneous trace, no masked DRL method significantly beats the random control, while on the heterogeneous trace MaskablePPO improves all three metrics relative to that control.
In briefMind the Tail: Sparse-Concept Data Selection for Robust Training
SummaryThis paper proposes a data selection method that uses sparse autoencoders to extract visual concept activations from images. It accumulates dataset-level concept support with diminishing returns, down-weighting well-represented concepts and focusing on under-supported tail concepts, and adds local support to pick samples that strengthen them. The authors report consistent gains over strong selection baselines across fine-grained, spurious-correlation, general classification, and multimodal pretraining benchmarks, along with improved robustness to out-of-distribution and spurious contexts.
In briefCan Language Models Predict the Key Factors Behind Their Decisions?
SummaryThis paper studies how to extract the cruxes, the sub-questions whose answers change a language model's recommended action, by examining disagreement across multiple generated responses. It compares three strategies of increasing cost: asking the model to list cruxes beforehand, extracting them from a single reasoning rollout, and comparing reasoning across many rollouts. The authors find that the cheaper approaches are insufficient, with a priori assessments bearing little relation to the factors identified afterward, so the many-rollout procedure is necessary for high-stakes decisions.
In briefImperfect Columns Help: Learned Column Prefilling for Branch-Price-and-Cut
SummaryThe paper studies column prefilling in branch-price-and-cut, showing that preloading ordinary routes that need not be optimal can speed up root pricing. A neural edge scorer generates the column pool, and a second network predicts the initial center for Wentges smoothing. On RouteOpt for capacitated vehicle routing, the pipeline reaches the same certified root bound in 13.2% of vanilla root pricing time across 95 instances with 300 to 1000 customers.
WhyShows non-optimal column prefilling with learned generator and smoothing center beats vanilla column generation; 95 routing instances reach the same root bound in 13.2% time, and all 25 scheduling ins
In briefEigDiff: Training Diffusion Models on Far Less Data with Little Performance Loss
SummaryDiffusion models are costly to train because they usually need large datasets. This paper proposes EigDiff, which uses principal component analysis in a pre-trained CLIP model's latent space to find eigenvectors, then uses them to select a representative training subset of exemplars and diverse images. Experiments across several diffusion models and datasets report up to 14x less training data and up to 1607x faster subset selection, with marginal performance differences.
Andrei Jarca, Vlad Hondru, Florinel-Alin Croitoru et al.
In briefHalve-Then-Greedy: an adaptive sampling method for reaching strong posterior evidence
SummaryThis paper studies fixed-budget adaptive experiments that aim to maximize the chance of gathering enough posterior evidence that a recommended design's mean exceeds a target, without needing to identify the best design. For Gaussian bandits with a Gaussian prior, the authors propose Halve-Then-Greedy (HTG), which combines a large balanced initial allocation, successive halving, and posterior-greedy sampling. They prove instance-dependent finite-budget exponential bounds on posterior failure probability and show HTG's sample efficiency advantages in simulations.
Yinan Li, Kwang-Sung Jun, Kapilan Balagopalan et al.
In briefOne bias score, two factors: how fixes change answer bias in multiple-choice tests
SummaryThe paper studies how reordering multiple-choice options can change a model's answer, treating this as both letter preference and order sensitivity. It unifies the two using a pooled selection-rate standard deviation, which factors exactly into a flip rate and a positional skew. Across 41 checkpoints and 12 benchmarks, prior correction mainly reduces positional skew, while aggregating each ordering with its reversal mainly reduces the flip rate. The authors recommend reporting both factors together with accuracy, since aggregation can settle questions on the same wrong option.
In briefDeSpatial: The Difficulty of Spatial Reasoning Lies in 3D Perception Itself
SummaryThe paper argues that spatial agents with reconstruction and segmentation tools often produce inconsistent predictions and fragmented object identities, which keeps their outputs from being reliable 3D scene information. It introduces DeSpatial, which splits spatial understanding into perception and reasoning, and uses a shared 3D perception tool to merge separate perception outputs into a unified, metric, gravity-aligned scene. DeSpatial achieves leading results on VSI-Bench, MindCube-Tiny, and ViewSpatial, with additional long-video results on VSI-SUPER.
In brief"Do As I Do, Not As I Say": On-policy distillation passes on teacher flaws in vision-language models
SummaryThis paper compares two ways to improve vision-language models: supervised fine-tuning (SFT) on teacher-generated demonstrations, and on-policy distillation (OPD), where the teacher gives feedback on the student's own outputs. OPD works well only when the teacher stays reliable on the student's trajectories. On Geometry3K, where the teacher fails to complete 22% of its solutions, OPD trails SFT by 7.9 points, mostly due to incomplete responses. Starting OPD from an SFT student also lowers its completion rate from 76% to 62%, showing that on-policy supervision can transfer persistent teacher failure modes.
In briefBlume-Capel Model as a φ⁶ Field Theory: Monte Carlo and Flow Model Study
SummaryThis paper uses Monte Carlo simulations to study a continuous-spin version of the Blume-Capel model, computing phase diagrams in the crystal-field-temperature plane as spin softness varies. The authors use normalizing flows, generative models that sample directly without autocorrelation, across the phase diagram, including near first-order transitions and the tricritical point. Finite-size scaling analyses of flow samples locate the transitions and tricritical point with accuracy comparable to standard and extended-ensemble Monte Carlo methods.
In briefTeaching a Small Model to Build and Improve GPU Optimization Scripts
SummaryThis paper presents a training method that teaches an 8-billion-parameter language model to generate complete compiler optimization scripts for GPU kernels and to refine existing ones. The authors built training data from 54,116 workloads using TVM MetaSchedule, then applied supervised fine-tuning and reinforcement learning with a reward combining correctness and latency. The paper reports per-attempt correctness of 91.27% for construction and 92.01% for refinement, and claims faster kernels than several larger frontier LLMs on the evaluated tasks.
Annabelle Sujun Tang, Wei Tang, Songyao Jiang et al.
In briefTwo-stage score estimation for group-specific diffusion models
SummaryThis paper studies diffusion transfer across related distributions, using a shared pooled score and a target-specific offset to improve generation when target data is limited. The authors derive convergence rates that separate the learning costs of these two components and identify when a simpler offset gives a faster rate than fitting separately. Under suitable assumptions, they prove the offset stays bounded and regular as diffusion time approaches zero, and provide total-variation guarantees. Synthetic and image experiments serve as a proof of concept.
In briefPiCo: Physics-informed probing to explain geo-foundation model predictions
SummaryGeo-foundation models are usually judged only by how well they perform downstream tasks, which says little about the physical features they rely on. This paper proposes PiCo, a post-hoc framework that uses a 3D radiative transfer model and a game engine to generate synthetic multispectral satellite data and identify concept vectors in a model's embedding space. Tested on three models and seven forest and canopy concepts, the concept-based representations kept 70.0% to 93.8% of full-embedding classification accuracy.
In briefSOMA: Quadruped Robot Pushing and Repositioning Stacked Objects
SummaryThis paper presents SOMA, a hierarchical reinforcement learning framework that lets a quadruped robot manipulate stacks of objects using its whole body, without a dedicated arm. It adds explicit base-height control so the robot can push in three dimensions, not only on a plane. In simulation, it reports 90.3% success for single-object repositioning and up to 90.1% for dynamic two-box stacks. Zero-shot sim-to-real transfer is also validated on a Unitree Go1 robot.
In briefMIDI-REPA: Aligning Symbolic Music Representations to Improve Text-to-Music Generation
SummaryThis paper introduces MIDI-REPA, a framework that uses MIDI derived from automatic music transcription as a supervision signal to align intermediate representations of a flow-matching DiT model for text-to-music generation. Building on the 122M-parameter FluxAudio-S baseline, the authors study what, where, and when to align, finding that aligning a piano roll with an early DiT layer across the full flow trajectory works best. The method trains on a single RTX 3090 in about 31 hours, and the authors release their transcription pipeline, JamendoMIDI, training code, and audio samples.
In briefMSP: Training Multimodal LLMs Without Paired Cross-Modal Data via Modality Self-Play
SummaryThis paper proposes Modality Self-Play (MSP), a training framework that uses generated proxy text descriptions as semantic bridges instead of real paired cross-modal data. Models are trained with only a real anchor modality and its proxy descriptions, and can then compose and reason over multiple modalities at inference time. The authors report state-of-the-art or competitive results on several multimodal benchmarks, while noting that the method still relies on modality-text pairs, pretrained contrastive encoders, and modality-distribution bridge calibration.
Robert Moseley, Md Kaykobad Reza, Ameya Patil et al.
In briefESC: Geometry-constrained diffusion for consistent 3D-aware novel view synthesis
SummaryThis paper presents a framework that generates a target view and its depth map together, improving geometric alignment when synthesizing novel views from a single image under large camera changes. The diffusion model is conditioned on the source image, source depth, and target camera rays, and a per-pixel depth consistency constraint enforces 3D consistency across generated views. The authors report superior visual accuracy and spatial alignment compared with current state-of-the-art methods.
In briefVELD-QA: A Large-Scale Benchmark for Visual Evidence Grounding and Question Answering in Long Documents
SummaryVELD-QA is a large-scale benchmark for answering questions about long documents by locating and interpreting visual evidence. It includes 15,000 human-written question-answer pairs drawn from 5,859 real PDFs totaling 372,445 pages, with annotations for supporting pages and visual elements. Evaluating nine vision-language models and seven retrieval-augmented or agentic systems, the authors find that retrieval-based methods lag far behind the ground-truth page setting, which peaks at 81.7% accuracy. Adding visual-element annotations to inputs narrows the performance gap by 17-24%.
Zhivar Sourati, Marianne Menglin Liu, Rongguang Wang et al.
In briefStepSafe: Learning timely intervention for AI agents through context-aware safety reasoning
SummaryThis paper addresses safety in agentic systems, where whether an action is safe depends on earlier steps. It proposes ASPO, a training framework that aligns decisions to their step and distills reasoning from the model itself. The authors present StepSafe, a step-level guardrail that reports state-of-the-art results among open-source guardrails across five agent-safety benchmarks, with Youden's J of 53.0 for the 4B model versus 46.2 for the comparison.
In briefSame Context, Different Agents: Behavioral Study of Context Techniques in Data Agents
SummaryThis paper studies three context-engineering techniques for data agents: semantic layers, multi-agent orchestration, and memory compaction, both alone and combined. Using seven large language models on 1,514 long-horizon tasks from five benchmarks, it analyzes agent behavior, accuracy, and cost. The authors report that each technique's benefit depends on the agent's behavior group, and they derive guidance for choosing techniques based on agent behavior, task demands, and accuracy-cost trade-offs.
Zhivar Sourati, Marianne Menglin Liu, Zheng Wang et al.
In briefLine-Augmented cuVSLAM: Robust Visual Odometry in Low-Texture Environments
SummaryThis paper adds line features to cuVSLAM, a CUDA-accelerated visual SLAM system, so it works better in low-texture settings such as warehouses. Lines are used in both per-frame pose estimation and sliding-window bundle adjustment, added as residual terms to the existing solvers. In RGB-D mode, line features reduced mean absolute trajectory error by 38.5% on 22 TUM RGB-D Freiburg3 sequences and by 30.1% on 8 Replica sequences, with no tracking failures in any tested configuration.
In briefCross-person decoding of imagined speech from scalp EEG, tested on three public datasets
SummaryThis study asks whether imagined-speech content recorded by scalp EEG can transfer between people, so a new user could use a decoder without calibration. The authors analyzed three public datasets and examined why accuracy stays low within a single person. Across 482 ordered participant pairs, none survived correction for multiple comparisons, and transfer accuracy stayed within 0.006 of chance. The authors conclude that imagined-speech content did not transfer between people through any representation they examined.
In briefLearning checklists to reduce lottery effects in peer review
SummaryThe paper studies reviewer lottery error, where a few reviewers raise only some of a paper's issues. The authors learn checklists of criticisms from historical reviews and select the version that best maximizes correlated agreement among reviewers. In LLM-simulated reviews on NeurIPS 2021 and ICLR 2022 data, the checklist reduced committee mean squared error by 58% and 21%, respectively.
