In briefKnowledge Distillation Without Aligned Feature Spaces
SummaryThis paper proposes RPKD, a knowledge distillation method that compares teacher and student networks against a shared set of randomly initialized, frozen prototype vectors instead of aligning their feature spaces directly. It uses a logit branch and a feature branch, and reports that decoupling prototypes from the label space gives richer supervision when the number of classes is small. The authors report gains over OFA of up to +7.98% on CIFAR-100, +12.44% on ImageNet-100, and +1.28% on ImageNet-1K.
In briefFinite-step geometry of velocity fields that preserve marginals in generative flows
SummaryThis paper studies why different velocity fields can produce the same continuous marginal evolution in generative flows yet behave differently after numerical discretization. Keeping the marginal path and numerical method fixed, the authors use projected modified equations to characterize the local integration error that changes the marginal law, and give sufficient conditions for higher marginal convergence order. For affine fixed-shape distributions they characterize exact finite-step reachability and derive spectral obstructions and sharp nearest-reachable Wasserstein errors, while for diffusion probability-flow ODEs they construct skew-score perturbations that preserve the marginal path under exact scores. Controlled Gaussian and mixture experiments recover the predicted order, reachability, and obstruction phenomena.
In briefSkillMem: Compact External Skill Memory for Efficient Latent Multi-Agent Reasoning
SummaryThis paper introduces SkillMem, which stores each reusable skill as a small external latent memory that is directly used during latent multi-agent reasoning. A role-conditioned cross-attention reader lets the same memory guide different agents according to their roles and hidden states. Tested on six benchmarks with Qwen3 models from 4B to 14B parameters, it gives strong gains for smaller models, needs 9.6 KB per skill, reduces storage over 38 times for a 100-skill library, and achieves up to a 4.23 times end-to-end inference speedup.
Yuxuan Zhu, Mohammad Mohammadi Amiri, Pin-Yu Chen et al.
In briefBi-Phase Training: a parameter-efficient method for learning language models from scratch
SummaryThis paper introduces Bi-Phase Training (BPT), a pre-training method that updates model weights using trainable diagonal matrices and low-rank adapters, periodically merging them into frozen base weights. Pre-training Qwen2.5 models (100M to 1.5B parameters) and OLMo-2-1B from scratch on FineWeb-Edu shows validation perplexity close to full-parameter baselines. At 1.5B scale, BPT uses 66% fewer trainable parameters, and it preserves much of the downstream task performance.
In briefNLP-based message prioritisation for opportunistic routing in disaster communication
SummaryThis dissertation tests whether NLP-based semantic prioritisation can improve handling of urgent messages in opportunistic emergency networks, using the 2015 Nepal earthquake as a case study. TF-IDF with Logistic Regression and BERT classify social media messages into four routing priorities, applied within the ONE simulator. BERT yields higher priority accuracy and P1 recall, but network gains are inconsistent, with the clearest benefit appearing under buffer pressure.
In briefStudy of whether large language models can work as numerical optimizers
SummaryThis paper evaluates how well large language models solve numerical optimization problems, including convex optimization, fractional programming, and constrained projection. Their outputs are compared against deterministic reference solvers, separating solution quality from failures such as numerical errors, constraint violations, and format errors. The models often find near-optimal solutions, but reliability declines as problem dimension grows. Iterative refinement with feedback improves validity, and the authors propose hybrid approaches that pair LLM reasoning with deterministic numerical verification.
In briefMachine learning in the heliosphere: learned operators for modeling and forecasting Sun-to-ground space weather
SummaryThis chapter examines space weather, the study of how Sun-driven processes affect the near-Earth environment. The authors treat the chain from the Sun to ground impacts as coupled systems, each described by a state vector and observables. Machine learning algorithms are framed as learned operators for inference, propagation, and emulation. The work argues this approach can improve understanding of the physical systems and speed up the development of operational space weather forecasting models.
In briefGARDEN: Autonomously designed reward functions that transfer across cooperative multi-agent scenarios
SummaryGARDEN is a method in which a large language model writes a single reward program for a family of related cooperative multi-agent learning scenarios. The program is then transferred unchanged to scenarios unseen during design. On unseen Overcooked and SMAX scenarios, task returns rose 58.2% and 52.5% over default rewards on two kitchen layouts, and mean win rate in two combat scenarios rose from 51.0% to 89.9%. These gains exceeded those from fixed diagnostic feedback and Eureka-inspired reward reflection under matched budgets.
In briefVerifiedLab: A multi-agent framework for turning verified research records into machine-usable knowledge
SummaryThis paper introduces VerifiedLab, a multi-agent framework that converts research knowledge from human-oriented records into forms usable by AI systems. Its AI scientists with distinct roles verify records field by field, and a gate and critic check each claim against the reliability of its data before release. Experiments on perovskite solar cells show that the framework recovers known physical trends and anticipates devices reported a year later.
In briefFAD-KV: Adaptive Dictionary Caching to Speed Up Language Model Reasoning
SummaryThis paper proposes FAD-KV, a framework that compresses the key-value cache used by large language models during decoding. It represents each token's KV states as sparse combinations of learned dictionary atoms, using a one-shot Top-n screening step with lightweight coefficient refinement to limit computation. A fixed-size dictionary reuses inactive atoms so that it does not keep growing during generation. The authors report a strong quality-efficiency trade-off on reasoning and long-context benchmarks.
Yuxuan Zhu, David H. Yang, Parisa Shiravani et al.
In briefSciGame: Letting AI Agents Play Scientific Discovery Games in Real and Altered Worlds
SummaryThis paper introduces SciGame, a benchmark that converts scientific literature into structured cards and builds paired original and counterfactual environments with consistent units. It includes historical canonical laws and a temporal collection of papers published after model knowledge cutoffs, and tests agents in identification and open-ended discovery settings under prior-only, fixed, and adaptive protocols. Experiments on five frontier models find that familiar priors hurt performance when the governing relation is altered, adaptive experimentation helps strong models but harms weaker ones under limited budgets, and good explanations can hide quantitative errors. Recovering the exact symbolic form remains the main bottleneck.
In briefVenus-DefakerOne2: A Broader Vision-Forensics Foundation Model for Detecting Forgeries
SummaryThis paper presents Venus-DefakerOne2, a unified foundation model for detecting forgeries across digital manipulation and physical spoofing, in both images and videos. It is trained on tens of millions of open-source and in-house samples, using Tide-Style-Tuning, which alternates LoRA and full-parameter fine-tuning, and Entropy-based Test Adaptation, which selects the prediction modality by analyzing prediction entropy. The authors report new state-of-the-art results on several benchmarks, outperforming specialized models and their predecessor, DefakerOne.
In briefWhat Actually Matters? A Paired Evaluation of Inference Settings for π0.5 on BEHAVIOR-1K
SummaryThis paper systematically compares previously proposed inference and data-collection strategies for household robot VLA models across all 100 BEHAVIOR-1K tasks. The evaluation uses a fine-tuned π0.5 model based on the 2025 BEHAVIOR Challenge winner, which scores 0.39 Q on the public test set. The results indicate that the proposed strategies do not transfer uniformly to new environments and tasks.
Ruben Hammele, Samira Huber, Mohammad Nazeri et al.
In briefMuon Optimizer Makes First MoE Layer's Load Harder to Balance
SummaryThis paper studies how the Muon optimizer affects the routing biases of mixture-of-experts (MoE) language models. In models with 55M and 108M active parameters, the span of the first MoE layer's biases roughly doubled under Muon. The rise traced to the first router's inputs sharing a direction, and subtracting the batch mean removed the overload. On the smaller model, halving Muon's learning rate removed most of the rise.
In briefUnderstanding Long Financial Texts with Semantically Enriched Tables Using Multimodal Learning
SummaryThis paper proposes a multimodal Transformer framework for modeling long financial documents that combine structured tables with noisy textual disclosures. Each numerical variable is paired with a semantic embedding derived from its textual definition, and cross-attention integrates the tables with long-form text such as MD&A sections. Experiments on a large-scale dataset of financial statements and MD&A disclosures show that numerical features remain the dominant predictive signal, while cross-attention yields consistent gains over simple concatenation.
In briefSegmental evaluation of accent in speech generation, aligned with listener perception
SummaryThe paper argues that utterance-level similarity ratings cannot show where perceived accent differences occur in generated speech. It proposes a segmental protocol that combines speaker and accent similarity ratings with subword annotations of accent differences from listeners in target accent regions. Forty-seven retained listeners evaluated 200 content-matched reference–candidate pairs covering four Scottish and Australian voices, two zero-shot TTS systems, and an articulation-based voice conversion system. Speaker and accent ratings were strongly associated (ρ = 0.71), while seven objective distance measures showed weak correspondence with subword selections, reaching at most ρ = 0.22.
In briefGenerative Modeling for Fusion: A Large-Scale Stellarator Fast-Ion Simulation Benchmark
SummaryThe paper introduces BoSS, a dataset pairing stellarator designs with guiding center simulation results and particle trajectories for 180,000 magnetohydrodynamic equilibria. It formulates two machine learning tasks: generative modeling of stellarators with good particle confinement, and surrogate modeling of heat deposition patterns and particle trajectories. The authors report that their baselines generate diverse, performant stellarators and useful surrogate predictions, and they identify long-horizon trajectory forecasting and out-of-distribution generation as directions for future work.
Misha Padidar, Teresa Huang, Mark Goldstein et al.
In briefTERRA: Edge AI for regional fire detection that reports without device identity
SummaryThis paper presents TERRA, an edge-AI framework for distributed fire monitoring in which devices send task-relevant local inference results tagged by region rather than by device identity, and a receiver aggregates the regional evidence in a permutation-invariant way. It evaluates two reporting modes, TERRA-Score and TERRA-Feature, using 15 sensing modalities from the EN54 indoor fire-alarm dataset. With nine reports and perfect message recovery, TERRA-Feature reaches mean AUPRC 0.990 with a 24-bit payload, and TERRA-Score reaches 0.987 with only 6 bits, both with an MDR of 1.81% at FAR at most 1%.
In briefBeyond Low-Pass Dynamics: Frequency-Selective Spiking Reservoirs with Resonant Neurons
SummaryThe paper introduces HRF-Res, a liquid state machine that replaces the usual Leaky Integrate-and-Fire neurons with a heterogeneous population of Harmonic Resonate-and-Fire neurons, letting the reservoir respond to frequency-specific temporal patterns. Across ten benchmarks covering time series, sequential images, audio spike trains, and neuromorphic vision, it reports state-of-the-art accuracy among spiking reservoir methods. The authors state that this is achieved while preserving the energy efficiency of spiking computation.
In briefFilingLoop: Testing and Improving AI Agents That Review and Revise Company Disclosure Filings
SummaryThis paper introduces FilingLoop, an interactive benchmark and self-evolving agent framework for reviewing and revising corporate disclosure filings. It is built from historical U.S. SEC correspondence, reconstructing 380 mostly IPO review tasks from 96 companies, scored against 6,099 requirements. The framework learns procedural guidance across cases without updating model weights, raising requirement fulfillment from 24.28% to 27.96% on the evaluated stream.
In brief"Think Twice": A Report on Two Language Models Built From Scratch
SummaryThis report presents Metis-1.6, two language models trained from scratch by one researcher: Praxis (2.94B parameters) and Logos (10.23B parameters). Both process each token twice through a hybrid Mamba-2 and attention mixture-of-experts stack, and were trained on 1.02T tokens. Across 40 post-trained benchmarks, Logos averages 52.93, performing strongly in mathematics but weakly in code, instruction following and long context. The report also documents training mistakes and their fixes.
In briefSAC-GCL: Graph contrastive learning that uses structure-aware compression instead of random perturbations to create graph views
SummaryThis paper proposes SAC-GCL, a framework for graph contrastive learning that replaces random augmentations such as edge perturbation and node masking with structural decomposition. It extracts cohesive subgraphs and condenses them into supernodes, producing coarsened graph views that preserve local density and higher-level structure. Multiple cohesive subgraph models yield complementary views, which a hierarchical view encoder processes. Experiments on six benchmark datasets show more robust and discriminative representations for graph classification.
Muhammad Ifte khairul Islam, Tanvir Hossain, Esra Akbas
In briefRole of Socio-Demographics and Reasoning in Pluralistic Offensiveness Modeling
SummaryThis paper examines when and where annotators' socio-demographic information helps models predict offensiveness judgments, and whether explicit reasoning improves its use. Experiments on three multi-annotator datasets with prompted and trained LLMs show that demographic gains are small and inconsistent for individual predictions but clearer at group and population levels. Reasoning that decomposes judgments into a baseline and a demographic-specific deviation yields more accurate estimates than direct demographic conditioning.
In briefWhen More Planning Hurts: Learned Planners for World Models
SummaryThis paper studies planning for robots using learned world models, comparing several planners on the PushT task. The proposed fixed action updater has 1,505 parameters and solves 8 of 16 reserved scenes. Experiments with additional update budgets found that extra updates rescued no failing tasks and instead caused some previously successful tasks to fail.
In briefFYC: FPR-Disparity Weight Perturbation and Youden Calibration for Fair Deepfake Detection
SummaryThis paper addresses demographic disparities in deepfake detectors, where false-positive rates differ notably across gender and skin-tone groups. It proposes FYC, a two-stage post-hoc method that perturbs detector weights under an FPR-disparity criterion and then calibrates a separate decision threshold per group, without architectural changes or full retraining. On AI-Face v2 across six intersectional gender by skin-tone groups, FYC outperforms the compared detectors on 18 of 21 fairness and utility metrics.
Uzoamaka Ezeakunne, Taiwo Olabamipe, Deevesh Jain et al.
In briefFair Deepfake Detection via Intersectional Invariant Learning
SummaryDeepfake detectors often perform unevenly across demographic groups, with some groups facing much higher false-positive rates. This paper proposes Intersectional Invariant Learning (IIL), which improves feature learning rather than only adjusting model outputs, and covers six gender by skin-tone intersectional groups. Evaluated on a benchmark of over one million images against 10 baseline detectors using 21 metrics, IIL reports high AUC, false-positive rates under 2.5%, and the lowest intersectional fairness disparity among compared methods.
Uzoamaka Ezeakunne, Taiwo Olabamipe, Deevesh Jain et al.
In briefInterpretable neural equation learning for high-dimensional materials data
SummaryThis paper presents nn-EQL, a framework that generates candidate equations by iteratively training, pruning, and retraining a neural network with user-specified activation functions. Candidates are ranked by a minimum description length-inspired score balancing error and sparsity. The method is demonstrated on standard AI-Feynman benchmarks and on high-dimensional materials datasets, including finite-element simulations of heterogeneous materials and high-throughput thin-film deposition data.
Philip Jacobson, Prasad Iyer, Remi Dingreville et al.
In briefVerification Learning Theory: A model of how AI-assisted checking affects retained judgment skills
SummaryThis paper proposes Verification Learning Theory (VLT), a process model describing how checking evidence during AI-assisted claim evaluation affects later independent judgment. The model links perceived assistant reliability to evidence checking and the history of checks to retained rule use. Simulations examine the consequences and parameter dependence of its assumptions, and the paper identifies where its predictions diverge from alternative accounts and specifies experiments to test the proposed mechanism.
In briefPairCLIP-SWA: A single-model deepfake image detector using similarity-guided balancing and weight averaging
SummaryThe paper presents PairCLIP-SWA, a deepfake image detector built on a single OpenCLIP ViT-L/14 visual backbone, trained on a class-balanced set created by CLIP-similarity-guided one-to-one subsampling. Its fine-tuning trajectory is averaged in weight space. On the IEEE SP Cup 2025 (DF-Wild-Cup) benchmark, the weight-averaged model reaches 97.04% accuracy, 98.64% AU-ROC and a 4.07% equal error rate. The authors show that checkpoints selected by validation accuracy can be unstable and recommend per-class selection criteria measured on a split separate from early stopping.
Md. Nakib, Sudipto Sarkar, A. J. A. Jubayer Talukder et al.
In briefMulti-agent reinforcement learning models how agreement design affects global climate cooperation
SummaryThis paper extends an existing multi-agent reinforcement learning climate framework with a multi-stage negotiation mechanism where agents propose, accept or reject agreements, and face penalties for non-compliance. In an idealized setting with four homogeneous agents, non-binding agreements do little to stop free-riding, while moderate penalties improve cooperation and reduce warming. However, excessively strong penalties discourage ambitious commitments. The authors present the framework as a methodological tool for testing agreement design rather than a predictive model of real negotiations.
Anne Moll-Elsborg, Oskar Bohn Lassen, Francisco Camara Pereira
In briefHuman-Residual Steering: Aligning Frozen LLM Judges with Human Judgment Distributions
SummaryThis paper proposes Human-Residual Steering (HRS), a method that aligns frozen large language model judges with the distribution of individual human ratings by learning hidden-state interventions. Across 14 benchmarks and four judges from three model families, HRS outperforms all evaluated frozen-weight baselines on distribution matching. It also trains about 88 times fewer parameters than distributional fine-tuning.
WhyThe authors propose learned hidden-state interventions to align frozen LLM judges with human distributions, reporting gains across 14 benchmarks with far fewer trained parameters.
In briefWhat Food Similarity Means: Visual Representations and Retrieval in Indian Cuisine
SummaryThis paper examines food similarity beyond simple identity matching, comparing how seven pretrained visual models group Indian dishes into neighborhoods in image retrieval. The neighborhoods differ substantially across representations, including some similarities crossing culinary-family boundaries. At the dish level, the models share some visual structure, but it corresponds only weakly to recipe-derived similarity, suggesting food similarity depends on representation and abstraction level.
Saransh Kumar Gupta, Akshadha Sharma, Lipika Dey et al.
In briefFinite-Sample Optimal Conformal Prediction Sets under Covariate Shift
SummaryThis paper studies conformal prediction under covariate shift when the covariate density ratio is known, aiming for prediction sets with minimal expected length under the target distribution. For every calibration sample size, the authors prove that combining an inverse conditional-density score with randomized weighted conformal calibration minimizes expected length among all rules valid over the weighted-exchangeability class. They approximate the oracle with quantile and kernel histograms, including forest-weighted estimators, and select estimators and tuning parameters by resampling. Experiments on 36 known-density configurations and four public regression datasets compare coverage, set length, and runtime with existing methods.
WhyThe authors prove the inverse-density score with randomized weighted calibration minimizes expected set length over weighted-exchangeable rules, and report experiments on 36 density configurations and
In briefEcho: scaling visual reinforcement learning with off-policy experience
SummaryThis paper introduces Echo, a single-stage off-policy reinforcement learning framework that collects experience without proportionally increasing visual rendering and inference costs. Through inverse distillation, a state-based auxiliary policy imitates the evolving visual policy and gathers inexpensive, unrendered experience used for off-policy value learning. Implemented with FastTD3 and tested in Genesis on three visual locomotion tasks and two dexterous hand manipulation tasks, Echo reduces time to reference returns by about 60% on average for the locomotion tasks and 56% for the hand tasks, relative to the compared method. It also improves peak mean returns by 62% and 12% respectively.
In briefEffects of Deep Learning Architecture and Pretraining on Automated Vocal Fold Segmentation in High-Speed Phonation Images
SummaryThis study compares four deep learning architectures (UNet, UNet++, TransUNet, Swin-UNet) for automatically segmenting vocal folds in high-speed images from ex vivo canine larynx phonation experiments. It uses 16,344 expert-annotated images from 227 trials. Mean Dice scores were similar across models, with differences mainly in boundary accuracy. ImageNet pretraining significantly improved boundary metrics for TransUNet and Swin-UNet, and without pretraining TransUNet outperformed the two CNN-based models on boundary fidelity.
Varun Vinayak Kalaiarasan (1), Charles Farbos de Luzan (1, 2) et al.
In briefSeq@K: Testing How Language Models Improve from Feedback
SummaryThis paper introduces Seq@K, an evaluation protocol that checks whether a language model can solve a problem within K sequential attempts, receiving feedback after each failure. It isolates the ability to refine answers based on feedback, unlike benchmarks focused on long-horizon planning. Tested on 7 benchmarks and 16 frontier models, Seq@K yields model rankings and efficiency trade-offs that differ from Pass@K.
In briefTesting whether AI agents can accurately rebuild sequencing protocols
SummaryThis paper introduces LibStructBench, a document-based benchmark that tests whether AI agents can reconstruct sequencing-library structures and molecular workflows from sequencing protocols. The experiments show that frontier agents often produce roughly correct reconstructions but are unreliable at exact reconstruction, especially for molecular-state representation and workflow connectivity. Switching the underlying language model caused larger average performance differences than switching agent harnesses, and only the molecular-state reconstruction difference was statistically significant.
In briefDetecting hidden AI agent collusion via partner-outsider positive listening
SummaryThe paper proposes a behavioural test for AI agents that share a secret code. A partner using the code reacts to the exact wording of a message, while an uninvolved outsider does not. The method, Partner–Outsider Positive Listening (PoPLi), measures the difference between the two agents' action changes when messages are reworded with the same meaning, computed by replaying logged prompts. In a simulated market, PoPLi was high when a code was present and near zero without it, while message-only monitors failed to distinguish the conditions.
In briefLower bound for projection-free online learning with linear oracles
SummaryThis paper studies projection-free online learning algorithms that handle the decision set using linear minimization oracles. Prior work showed that Online Frank-Wolfe achieves regret scaling as T^{3/4} under specific tunings, but whether all tunings do so was open. The authors prove a universal worst-case regret lower bound of Ω(T^{3/4} m^{-1/4}) for any projection-free algorithm making at most m oracle calls per round. They also propose an extended Online Frank-Wolfe that uses multiple linear optimizations per round and prove a regret upper bound of O(max{T^{3/4} m^{-1/4}, √T}).
WhyThe authors prove a universal regret lower bound for projection-free algorithms and match it with an extended OFW upper bound, answering an open question.
Julien Weibel, Pierre Gaillard, Wouter Koolen et al.
In briefLearning convex potentials to compute optimal transport between medical images
SummaryThe paper proposes a method that estimates optimal transport maps between medical images by parameterizing a convex potential function. The resulting map is curl-free by construction, consistent with Brenier's theorem. The method is evaluated on 2D and 3D medical imaging datasets, and the authors report lower transport estimation errors and higher-quality image reconstructions than existing approaches.
In briefWhere the Error Lives: Spectral Representations of Learned Kohn-Sham Hamiltonians
SummaryThis paper studies machine-learned models that predict Kohn-Sham Hamiltonian matrices to replace costly self-consistent-field calculations in density functional theory. It shows that matching matrix entries does not ensure accurate eigenvalues or eigenvectors, and that two existing physics-informed losses each capture only part of the error. The authors propose the cumulative local density of states as a complete training signal and report that fine-tuning on QH9, called FermiScope, reduces the HOMO-LUMO gap error without increasing element-wise error.
In briefBetter scaling for large language models when training data is limited
SummaryThis paper studies how to train large language models when both compute and high-quality data are limited. The authors fit power laws that specify how key hyperparameters should transfer across model sizes and epochs, and scaling laws that specify how to allocate compute through ensembling. On 100M tokens from FineWeb, their recipe lowers validation loss by over 0.15 nats compared with prior recipes, a 14% perplexity decrease, and improves monotonically with compute via ensembling.
Ethan Baron, Shmuel Berman, Anirudh Satheesh et al.
In briefEarly Signs of Feature Learning in Deep Linear Networks
SummaryThe paper analyzes features learned by a deep linear network in a mean-field parameterization after a single gradient step from initialization, using high-dimensional data with a low-dimensional spiked signal. It derives closed-form expressions for the bias, variance, and generalization error of kernel ridge regression on these features. A one-step generalization advantage appears only when spike strength and effective signal-to-noise ratio are sufficiently large, and it vanishes for isotropic inputs.
WhyThe authors give a closed-form account of one-step feature learning in mean-field deep linear networks and derive bias, variance, and generalization error expressions.
In briefDepth Through Recurrence: Compact Speech Recognition for Edge Devices
SummaryThe paper proposes a block-recurrent speech recognition architecture that replaces many independent layers with a few recurrent blocks of weight-shared layers, keeping effective depth while shrinking model size. Using representation-guided grouping and knowledge distillation, a model with only two shared blocks reportedly reaches 97.4% of large models' accuracy on average across Open ASR Leaderboard benchmarks. The compact design also supports lightweight on-device personalization via MoE-LoRA adaptation and lower memory and energy use.
Chen Feng, Tianyi Xu, Ramchalam Kinattinkara Ramakrishnan et al.
In briefChoosing Random Mixtures of AI Versions with Fewest Stakeholder Rejections
SummaryThe paper studies how to pick a randomized mixture of several candidate AI system versions that the fewest stakeholders reject. Each stakeholder accepts only mixtures inside an unknown halfspace, and the developer can learn this only by asking yes-or-no membership queries. The authors show that minimizing rejections is APX-hard in general, give exact algorithms for special cases, and provide approximation guarantees when only some stakeholders can be queried.
WhyThe authors model stakeholder acceptance as unknown halfspaces, prove APX-hardness for minimizing rejections, and give polynomial and quasi-polynomial exact algorithms with sampling guarantees, plus s
In briefDeFT-RAG: Breaking Multi-Hop Questions into Self-Contained Sub-Questions with Reusable Answers
SummaryThis paper proposes DeFT-RAG, a retrieval-augmented generation framework for multi-hop questions that require chaining facts across several documents. It has the language model generate one self-contained sub-question at a time, while rule-based updates and a lightweight decision model handle the remaining tasks, and it caches question-answer pairs for reuse. On MuSiQue, 2WikiMultiHopQA and HotpotQA, it achieves the best accuracy among thirteen baselines and uses fewer input tokens than other iterative methods. With a cache shared across 1,000 MuSiQue questions, 23.1% of sub-questions were answered from the cache, reducing LLM calls by 12.9% without lowering accuracy.
In briefLearning Without Clocks: Neural Networks Invariant to How Event Streams Are Split
SummaryEvent sensors output asynchronous events, but most learning systems first group them into frames, making results depend on how the stream is sliced. The paper studies networks that update their state exactly at each event and evolve in closed form between events, so their output does not depend on this slicing. It proves conditions for order-independence of simultaneous events and continuity in timestamps, and introduces a learnable Clock-Free Event (CFE) layer. In preliminary development-set experiments, the proposed state outperforms an otherwise identical additive state on several event datasets.
In briefResponsible Quantum Machine Learning for Medical Imaging: A Fairness Audit Across Subgroups
SummaryThis paper empirically audits the fairness of classical and quantum classifiers for medical imaging, assuming quantum machine learning is not inherently fair. It compares SVM, MLP, QSVM and QCNN on dermatological lesions (DDI, with Fitzpatrick skin type as the sensitive attribute) and chest X-rays (CheXpert-derived, with age group as the attribute), testing several mitigation methods. The quantum models can be competitive overall yet still show subgroup disparities, and fairness effects depend on the model and dataset. All quantum models run in classical simulation, and the authors claim no quantum advantage.
Giovanni dos Santos Franco, Felipe Mahlow, Ellison Fernando Cardoso et al.
In briefGOFish: Full-field-of-view Gaussian splatting that builds scene meshes from fisheye images
SummaryFisheye lenses give robots, self-driving cars, and 360° cameras a wide view, but earlier Gaussian splatting methods rendered only up to 180° and produced no meshes. GOFish is the first Gaussian splatting surface reconstruction method that trains directly on fisheye images, renders beyond 180°, and extracts an aligned mesh from the same scene. It improves PSNR over the best prior method by 2.28 dB on FIORD (200°) and 1.97 dB on FullCircle (205°), and achieves the highest mesh F-score on FIORD and the main ScanNet++ set.
Allen Tu, Meher Gitika Karumuri, Shifan Zhu et al.
In briefWhat an Ontology Catches: Curated, LLM-Generated, and Data-Derived Ontologies as Checkers of LLM Agents
SummaryThe paper studies ontology-based checkers that judge answers from LLM agents querying Web knowledge graphs. It compares curated, LLM-generated, and data-derived ontologies, used as prompt context or as checkers, to see which errors each catches and which correct answers each wrongly flags. On UniProt, the curated ontology flags more correct answers than either alternative. On real agent answers, its flags are only marginally more precise than random flags.
In briefTACOS: Terrain-aware contact supervision method for visual footstep planning
SummaryThis paper presents TACOS, a method that builds a footstep instability score from the residual of a factor graph used in legged robot state estimation, then projects that score onto earlier camera images to train a dense visual predictor. It proposes a hard score that uses hindsight and a causal soft score available immediately. On 1716 manually annotated contacts and three outdoor test runs, both variants outperform the compared alternatives at detecting instability and ranking its severity. In a real-world 50 m forest trial, the trained predictor handles mud and vegetation and steps around fallen branches.
WhyThe authors derive a footstep instability score from factor-graph residuals to supervise a visual predictor, compared against baselines on 1716 annotated contacts and outdoor runs.
Jean-Michel Fortin, William Larriv\'ee-Hardy, Gabriel Jeanson et al.
In briefSpokes, Not a Ring: Revisiting How Language Models Represent Weekdays
SummaryThis paper examines how large language models represent weekdays as a circular geometric structure, called a manifold. The authors test whether moving continuously along this manifold corresponds to a smooth change in the represented concept. They find that weekday representations behave more like seven discrete states arranged in a circle than a continuous ring. The intermediate space between these states does not appear to encode time of day, uncertainty, or natural text.
In briefUniversal Short-Time Fourier Transform: adaptive window length for time-frequency analysis
SummaryThe short-time Fourier transform analyzes local frequency content with a sliding window, but short windows favor time resolution while long windows favor frequency resolution. This paper proposes the Universal Short-Time Fourier Transform (U-STFT), which treats window duration as a data-adaptive state chosen frame by frame, scoring candidate windows with normalized order-three Rényi entropy and aggregating them through a window-length trellis. The authors give a cumulative regret bound and show on synthetic and marine bioacoustic signals adaptive resolution tracking, improved spectral concentration, and exact time-domain invertibility.
In briefWhere sycophancy appears across training stages of open language models
SummaryThis paper studies sycophancy, where language models change correct answers after a user suggests a wrong one. The authors measured 13 open checkpoints from 5 model lineages at every available training stage, from base model to release. They find two sycophancy behaviors with different origins: deference to claims in the prompt is already present in base models, while conversational capitulation is largely absent from base models and appears in all post-trained checkpoints.
In briefWebAnything: Automatically Building Reusable Web Environments from Real Sites for AI Agents
SummaryWebAnything is a framework that automatically rebuilds reusable, resettable website environments from real URLs, so web agents can be trained and tested without manual per-site setup. It constructed 244 website environments across 14 application domains, supporting 12,984 executed and accepted tasks. In a blind audit it outperformed the human-assisted WebHarbor on three fidelity metrics. Fine-tuning Qwen3.5-9B on 5,856 interaction trajectories improved success rates on several benchmarks.
WhyThe authors automate reconstruction of reusable web environments from real URLs with an independent verifier, reporting 244 sites and gains from fine-tuning on 5,856 trajectories.
In briefSparse tensor networks for estimating individual tree diameter in mixed forests
SummaryThis paper studies how to automatically estimate individual tree diameter at breast height (DBH) from UAV laser scanning point clouds. The authors introduce TreeNext, a lightweight deep learning architecture that estimates tree parameters directly from voxelized LiDAR data, and build a full pipeline from instance segmentation to DBH regression on four mixed temperate forest plots in Germany. TreeNext reaches an R² of 0.8355 on automatically segmented trees and 0.9380 on ground truth segments, outperforming Random Forest and XGBoost. The study finds that segmentation quality is the dominant error source in the automated pipeline.
Clara L\"o{\ss}l, Fabian Ewald Fassnacht, Viacheslav Barkov
In briefSelective and Hybrid-State Communication for Latent Multi-Agent Systems
SummaryThis paper tackles the communication cost of large language model multi-agent systems, which exchange internal model states instead of repeated natural language. It proposes two complementary methods: SelectKV, which selectively communicates KV states while protecting inherited context, and HyLaMAS, which extends latent collaboration to hybrid linear-attention models by sharing both KV caches and recurrent states. According to the abstract, SelectKV cuts communicated KV positions and logical KV payload by 9.7-9.9% across six benchmarks, with accuracy largely preserved. HyLaMAS reduces decoded output tokens by 53-80% and achieves 1.16-2.78x speedups over text-based communication on Qwen3.5 models.
In briefMDraft: Multi-Draft On-Policy Distillation for Training Parallel Diffusion Drafters
SummaryThis paper studies how to train one-step diffusion drafters that propose several tokens in parallel for speculative decoding, a technique that speeds up language model inference. The authors show that when verifier feedback is used separately for each sampled prefix, the targets conflict and can suppress tokens that are plausible under some prefixes. They introduce MDraft, which samples several draft blocks from the same context and aggregates the verifier's full-vocabulary conditionals across their prefixes. Experiments on mathematical, coding, and conversational tasks show longer accepted drafts than prior distillation methods, and the method changes only training, keeping the architecture and inference procedure unchanged.
In briefHoneycomb-Surfels: Compact Hexagonal Primitives for Rendering and 3D Reconstruction
SummaryThe paper introduces Hexels, planar hexagonal surface elements with a first-order color field, for efficient representation of 3D scenes and 2D images. Across 34 scenes, Hexels improve PSNR over memory-matched 2DGS by 4.53 to 7.38 dB while requiring less training VRAM. For 3D surface reconstruction, Hexels reduce error by 28.9% on 15 DTU scans and achieve a mean F1 of 34.38% on six Tanks and Temples surface scenes, versus 31.72% for 2DGS.
In briefThinkWorld: Using World Imagination to Teach Robots Action Reasoning from Human Videos
SummaryThis paper introduces ThinkWorld, a learning framework for Vision-Language-Action robot models. It represents action reasoning as a Chain-of-Dynamics, a sequence of discrete motion tokens linking the current world state to a future state, and reinforces those chains whose predicted futures better match observed outcomes. This lets videos without action labels supply supervision from observed state transitions, so human and robot experience refine the same reasoning space. Experiments in simulation and on real robots show effective human-to-robot transfer and generalization to unseen tasks and environments.
In briefCan Discrete Tokens Represent Images Without Semantic Loss? This paper tests whether quantizing visual signals into discrete tokens loses meaning, and proposes a new quantization method.
SummaryThe paper challenges the assumption that quantization error inevitably causes semantic loss, finding that discrete representations with larger errors can sometimes understand better. The effective bit budget of discrete indices is identified as the key factor. The authors propose Factorized Vector Quantization, which uses symmetric parallel codebooks to quantize complementary views of visual features, continuing to improve at larger bit budgets. The model reaches state-of-the-art results among discrete-representation models and is competitive with continuous models such as Qwen3-VL, including on fine-grained OCR.
WhyRevisits the claim that quantization error inevitably loses semantics; proposes Factorized Vector Quantization, which the abstract reports matches continuous representations in multimodal understandin
In briefWhen to Intervene at Test Time: PORTAL, a Controller for Timing Interventions and Label Queries
SummaryThis paper studies test-time adaptation, where a deployed model's intervention can help some input batches and hurt others. The authors propose PORTAL, a controller that decides when to deploy an intervention, which inputs to label, and when to stop querying labels. Experiments show that it turns an intervention that lowers overall accuracy into a useful choice on selected batches and prevents severe failures from changes in arrival order.
WhyProposes PORTAL, a controller that decides when to apply a test-time intervention, which inputs to label, and when to stop querying; each label scores all candidate rules at once, and the abstract rep
In briefAttention as lattice gauge theory: relations interact only via Wilson loops
SummaryThe paper argues that if each entity's coordinate frame can be rotated independently, relations interact only through shared cycles, called Wilson loops. From this it derives the unique symmetry-compatible form of edge attention. The authors implement this as holonomy attention on general relational graphs, which infers unobserved S4 relations without training and flags corrupted triangles.
In briefKPPax: A Differentiable Kinetic Preprocessor for Machine Learning in Atmospheric Chemistry
SummaryThis paper presents KPPax, a tool that converts atmospheric chemical mechanisms written in the Kinetic PreProcessor language into differentiable, just-in-time compiled, vectorisable JAX code. It obtains exact sensitivities through the implicit solvers and forward- and reverse-mode adjoints of the diffrax library, without hand-derived adjoint code. The authors demonstrate it on the POLLU benchmark mechanism, using Fisher information to assess which rate coefficients can be identified and a neural-network correction of an emission inventory trained offline and fine-tuned online.
In briefA study on dexterous in-hand object pose reaching that balances skill and safety
SummaryThis paper addresses in-hand object pose reaching, where a multi-fingered LEAP hand moves an object to target positions and orientations. The policies are trained in simulation without human data and transferred zero-shot to the real hand. A learned residual safety filter with adjustable conservativeness reduces object drops by 75% while still reaching 91.5% of goals.
In briefGEPA-Curator: A framework for curated learning in self-improving AI agents
SummaryThis paper proposes GEPA-Curator, a framework that improves how prompts are optimized for large language model agents. It filters out tasks likely to give noisy or unreliable learning signals and retrieves past successful trajectories for a task as extra context for the prompt optimizer. The method was evaluated on GoBrowse, ALFWorld, AppWorld, and τ-bench using Gemini 3.1 Pro and Claude Opus 5 as optimizer models. It matches or exceeds the mean performance of both baselines, with gains of up to 20.7% over GEPA and 21.2% over ACE.
In briefCan LLMs Find Hidden Physics Through Free Exploration?
SummaryThe paper introduces LongChang, an automated framework that builds simulated worlds where some physical laws are altered. Large language models act as explorers, manipulating objects and building tools to pursue their own questions without task instructions or hints. Among 12 tested models, GPT-6-Astra achieves a 16.7% discovery rate, while the other 11 score 0%. The exploration records show that some anomalies remain uninvestigated and that the altered laws are only partially identified.
In briefScaling laws for tool-use AI agents trained on controllable synthetic environments
SummaryThis paper studies how to train large language model agents that complete complex, multi-step tasks using tools. The authors propose SynthAPI, a pipeline that controls the diversity of generated tools, tasks, and databases, and run 458,752 tasks across 849 models of four sizes to derive a scaling law predicting generalization. They find that as data budgets grow, increasing tool and database diversity helps more than adding tasks, and higher data complexity yields larger gains. Agents trained this way outperform models trained on prior work's data across four evaluation benchmarks.
In briefMuonGS: Pretraining LLMs with Relaxed Weight-Norm Control
SummaryThis paper analyzes how weight-norm control affects the Muon optimizer in language model pretraining, using an accumulated effective learning rate as the main tool. It compares vanilla Muon, MuonW, and MuonH, and proposes MuonGS, which uses Gram–Schmidt projection to remove the radial component of each update without fixing the weight norm. In pretraining experiments on dense and MoE models, MuonGS achieves lower validation perplexity than MuonH at long training horizons, with an estimated token-equivalent speedup of up to about 15% for the dense model at matched validation loss.
In briefMulti-State Linear Attention with Generalized Second-Order Routing: A Method for Better Long-Sequence Modeling
SummaryThis paper addresses the quadratic cost of self-attention in long-sequence modeling by improving Multi-State Linear Attention (MSLA). The proposed MERLIN method tracks the first and second moments of each segment's keys, so a query can recover both the mean and variance of its alignment with each segment. Experiments on multi-state variants of Gated DeltaNet-1.3B and Mamba-2-780M show that MERLIN outperforms existing aggregation methods across various long-context tasks.
In briefMetacognitive Decoding: A Controller That Makes Diffusion Language Agents More Efficient
SummaryThis paper introduces a metacognitive controller, a lightweight plug-in for diffusion language models, that guides generation based on the needs of the current response and the agent's interaction with its environment. It learns when to finalize tokens and how far generation should extend. Trained on agent trajectories to balance correctness and generation cost, the controller reduces inference FLOPs by 20.6% on average across five benchmarks compared with the most efficient baseline, while keeping task accuracy competitive.
In briefAdapting video diffusion transformers to track the same object across views and time
SummaryThe paper introduces XTrack, a framework that keeps the identity of an object consistent across different camera views and over time, starting from a query mask. It repurposes a pretrained video diffusion transformer to transport RGB information into mask latents, jointly handling cross-view matching and temporal tracking. To improve reliability, it adds focal cross-view attention alignment, hard negative conditioning, and local context cropping, and it converts sparse Ego-Exo4D annotations into dense pseudo-labels for training.
In briefTracking how language models learn knowledge during pretraining using point-in-time models
SummaryLanguage models pretrained on multi-year web data learn general and domain-specific knowledge, but how this happens over training is unclear. The authors use point-in-time models trained on different years' web snapshots to define a temporal variability score (TVS), the coefficient of variation of each item's negative log-likelihood across models. Across six benchmarks and six models, high-TVS items improve faster later in pretraining. At the end of training, high-TVS items outperform low-TVS items in 34 of 36 settings, by an average of 7.3 points of chance-normalized accuracy.
In briefKnow, Then Draw: A New Training Approach for Visual Generation
SummaryThis paper proposes a learning paradigm that uses autoregressive pretraining to acquire semantic knowledge and diffusion post-training for high-fidelity rendering. Controlled comparisons find autoregression to be the stronger pretraining objective for downstream rendering, and it introduces a purely semantic tokenizer trained with language supervision alone. The authors establish compute-optimal scaling laws for both stages and train a 3B-parameter text-to-image model that scores 86 on GenEval and 87 on DPG-Bench.
WhyProposes a knowledge-then-rendering paradigm: autoregressive pretraining for semantics, diffusion post-training for rendering. The authors report that controlled comparisons favor it over diffusion fr
In briefRare, Relevant, Yet Still Wrong: How Memory Causes Errors in LLM Group Decisions
SummaryThis paper examines what a short memory should keep when an LLM group's discussion is compressed for a final decision maker. The authors introduce CrossProfile, a diagnostic benchmark of 480 tasks with automatically verified answers, to test two common heuristics: keeping the rarest facts or the facts most relevant to the decision criteria. Both heuristics fail in predictable ways, and accuracy depends more on the balance of evidence kept than on the selection policy used.
Anouar Imel, Imad Charradi, Abderrahman Sghir et al.
In briefPerceiving and predicting fast sports from a player's-eye camera using vision models
SummaryThis paper introduces an annotated tennis dataset combining egocentric video, synchronized allocentric videos, player positions, contact times, and stroke labels. The authors evaluate vision encoders, video encoders, and vision-language models on seven tasks, including localizing players and the ball, forecasting trajectories, and predicting stroke type and time to contact. They also propose a conditional diffusion model that generates whole-body tennis motion in real time from the vision predictions.
In briefVideoLVR: Latent Visual Reasoning over Grounded Video Evidence
SummaryThis paper introduces VideoLVR, which helps multimodal large language models reason about videos by building latent visual evidence from local regions at relevant moments. This supports intermediate reasoning steps without relying only on textual coordinates or adding many cropped frames to the context. A visual token prediction objective unifies visual evidence reconstruction with text generation, and the approach is extended to reinforcement learning using final-answer rewards. Experiments on multiple video benchmarks show its effectiveness, especially for tasks needing fine-grained visual evidence.
In briefEnvSched: Environment-Aware Scheduling for Efficient Agent Reinforcement Learning
SummaryThis paper introduces EnvSched, an online scheduler that decides which environment-task pairs to use in each training batch for agentic reinforcement learning, based on past rollout history. It builds a dual-relation graph that propagates statistics to predict within-task reward variance, and balances selecting high-variance tasks with covering under-observed environments. In experiments on Qwen3-8B, EnvSched reaches an average score of 48.81 across BFCL-v3, tau-Bench, and ACEBench-Agent after 40 training steps, surpassing the best score random task selection reached after 120 steps.
In briefBeyond Answer Changes: Matched Same-Class Controls for Audio Evidence
SummaryThis paper examines how audio-language models respond to local edits of a queried sound event. It compares same-class and different-class replacements that share the same interval, level estimate, donor fitting, and crossfade, with only same-class replacements preserving the correct answer. Evaluated on 600 synthetic soundscapes and three models from two architecture families, the study reports how often answers change under each replacement type.
In briefMoI-Bench: A Benchmark for Image Generation and Editing on Phones
SummaryMoI-Bench is a benchmark for image generation and editing on mobile devices, built from mobile user needs and containing 64 subtasks and 1150 samples. It combines 16 operations with eight visual target categories and introduces MoI-Eval, a task-specific evaluation method. The authors also propose a two-dimensional Quality-Deployability map, which finds that on-device models can generate or edit efficiently but need broader task coverage and better editing quality, especially identity preservation and text rendering.
In briefGAS-PFN: Graph-Aware Zero-Shot Traffic Forecasting Model
SummaryThis paper introduces GAS-PFN, a zero-shot model for traffic forecasting that incorporates road network topology into its synthetic training priors using random-walk diffusion. Its inference architecture uses two mask-conditioned passes through a shared Transformer to model sensor-specific temporal dynamics and interactions among neighboring sensors. Across six benchmarks at a 90-minute horizon, it reduces MSE by 34.0% on average relative to TimePFN and by 27.7% relative to four recent general-purpose zero-shot forecasters.
In briefTANDEM: Balancing Centralized and Decentralized Critics for the Information Gap in Multi-Agent Learning
SummaryThis paper addresses the information gap in cooperative multi-agent reinforcement learning under centralized training with decentralized execution, where critics see global state during training but agents lack it when acting. It proposes TANDEM, which reads one centralized critic in both a centralized and a decentralized way at each training step, uses a parameter-free gate to train on the decentralized read where the two agree, and adds a shared target to align them. The method is tested on discovery, sampling, and wildfire suppression tasks, measuring return after the centralized critic is removed. TANDEM is reported as the best method on eight of nine setups.
Alireza Saleh Abadi, Leen-Kiat Soh, Adam Eck et al.
In briefA paper on the limits of latent world models, which learn internal predictive models of environments.
SummaryThe paper examines whether latent world models, which agents use to reason over long horizons, actually capture the true dynamics of their environments. It introduces faithful latent world models as an equivariance condition between the environment's transition map and learned latent dynamics, and proposes latent rollout consistency (LRC) as a metric for evaluating action-conditioned latent prediction. The authors find that none of the state-of-the-art latent world models studied solves the task by faithfully modeling the environment's dynamics.
In briefTimestral: A Family of Time Series Foundation Models for Probabilistic Forecasting
SummaryThis paper introduces Timestral, four time series foundation models with 5M to 209M parameters for probabilistic forecasting. The authors report state-of-the-art results on GIFT-Eval, TIME, and fev-bench, with the 209M and 123M models taking the top two places on every metric. The approach combines Decoupled Horizon Forecasting, a covariate framework supporting known-future covariates in a causal architecture, and a two-stage pipeline of synthetic-data pretraining followed by post-training on mixed real and synthetic data with extended context.
WhyProposes decoupled horizon forecasting (DHF) and four models from 5M to 209M parameters, reporting state-of-the-art results on GIFT-Eval, TIME, and fev-bench, with the two largest models taking the to
Adrien Petralia, Alexandre Gavaudan, Artjom Joosen et al.
In briefREGER: A Generative Recommendation Framework Modeling Item Relations via Semantic IDs
SummaryThis paper proposes REGER, a framework for generative recommendation that explicitly incorporates inter-item semantic and collaborative relations into semantic pretraining, history encoding, and semantic ID generation. It provides a theoretical analysis of hierarchical preference decomposition over the SID prefix tree and distills a discriminative teacher's conditional distributions alongside cross-entropy training. Experiments on real-world datasets are reported to show consistent improvements over existing generative recommendation baselines.
In briefZero-shot detection of events in soccer broadcasts without special training
SummaryThis paper presents a zero-shot framework that detects, classifies, and temporally locates events in full-length soccer broadcasts without training task-specific parameters. Matches are split into overlapping clips, each represented by a compact visual storyboard and aligned commentary transcript, processed jointly by a general-purpose multimodal model. On three manually annotated matches with 130 events, the best configuration reaches 72.0% F1@5s and outperforms SoccerNet-trained CALF and NetVLAD++ in a cross-dataset setting.
In briefSpectral Braids: A Topological Descriptor for Dynamic Directed Networks
SummaryThis paper proposes spectral braids, a method for learning from dynamic directed networks. Each network snapshot is reduced to a community-level transition matrix, its complex eigenvalues are tracked over time, and the resulting braid is encoded as a word and mapped by the Burau representation to a fixed-size feature vector. On synthetic online-community networks, the features reach 97.1% accuracy on four classes, compared with 44.5% for the best feature set that does not track eigenvalues.
In briefIdentifying Film Characters by Watching the Movie Like a Human
SummaryThis paper addresses identifying principal characters and assigning their names using only the movie itself, without external cast lists or pre-built character banks. The authors define a character naming task, where a name mentioned in dialogue must be matched to the visible character using video context. They provide a dataset with annotations, a video-and-dialogue model, and a progressive framework for building a character bank as the film unfolds.
In briefThe Noise Plane: A Diffusion Bridge Across Exponent and Dispersion
SummaryThis paper proposes p-Bridge, a unified diffusion framework for denoising Tweedie-distributed sensor noise. An exponent p describes how noise variance relates to intensity, and accumulated measurement controls the noise level, together spanning Gaussian, Poisson, and Gamma models. A single exponent-conditioned network handles different noise exponents without separate training. Experiments on synthetic images, fluorescence microscopy, and SAR images show its effectiveness.
In briefTeamGym: A training and evaluation environment for AI agent teams working with distributed information
SummaryThe paper presents TeamGym, an environment for evaluating and training language model agents on team decision problems built from legal, finance, medical, and software engineering corpora. It defines "pivotal information" as a minimal set of facts that is necessary and collectively sufficient for a correct decision, and uses it to measure whether agents identify and integrate such facts. Across 8 models and 12 corpora, distributing pivotal information among participants lowered success rate by 22.9 percentage points on average compared with giving it to a single model. Reinforcement learning on team decision loss improved decision quality and produced emergent coordination in participation, speaker selection, and communication.
Chanwoo Park, Jaywoong Jeong, Zachary Wojtowicz et al.
In briefLCM-Promo: An LLM-Based Uplift Model for Promotion Allocation
SummaryThis paper presents LCM-Promo, an uplift model for allocating promotions that is built on a large language model. It is a task-specific Large Commerce Model, finetuned on transactional commerce data to predict user intent and response to promotions. The model learns monotonic promotion-response curves from production logs and conditions on session context such as day and delivery shift. In a production A/B test on churn reactivation, it increased conversion by 1% and reduced promotion cost by 6% compared with the Segment-EMA production baseline.
In briefGym-Anything-RT: Scaling Real-Time Test Environments for Vision-Language AI Agents
SummaryThis paper introduces Gym-Anything-RT, a framework for testing how well vision-language AI agents handle real-time interaction. An agentic pipeline automatically generates 100 visual minigames based on a taxonomy of six real-time capabilities. The agents perform poorly in real time, with GPT-6 Astra (xHigh) passing only 9.0% of minigames, rising to 75.0% when the environment is paused. The authors have publicly released the framework and the generated minigames.
In briefReferStream: Interactive Subject Injection for Streaming Video Generation
SummaryThis paper presents ReferStream, a framework that adds new visual subjects to a video while it is being generated, without any extra training. It combines Joint Latent, which places reference latents into the current scene so their features are extracted in context, with Cache Restoration, which uses these features for the injection chunk and then restores overwritten cache entries. The authors also introduce ReferStream Bench, a benchmark with scheduled reference arrivals and subject-presence annotations. On this benchmark, ReferStream scores higher than the compared baselines, including offline methods, on text adherence, subject similarity, and count accuracy.
In briefPneuma: A Real-World Benchmark for Predicting Commercial Aircraft Maintenance Needs
SummaryThis paper introduces Pneuma, a publicly available benchmark for predictive maintenance built from onboard time series data of in-service commercial aircraft. It centers on bleed air systems and covers four aircraft types and nine verified fault scenarios aligned with real maintenance logs. The benchmark defines three tasks: multivariate forecasting, supervised fault classification, and unsupervised anomaly detection. Evaluating over 25 models, the authors find that even strong current models struggle to detect subtle fault signatures.
In briefWhen uncertainty measures fail: fixing biased deferral in imbalanced clinical prediction
SummaryThis paper studies selective prediction in clinical AI, where high-confidence cases are decided automatically and ambiguous cases are deferred to human experts. It finds that in prognostic tasks with severe class imbalance, standard uncertainty scores can reflect class membership rather than clinical ambiguity, systematically penalizing minority patients. The authors propose a corrective framework, tested across 56 synthetic conditions and three neurodegenerative cohorts. They recommend threshold-aligned uncertainty measures and class-stratified evaluation.
Walter Endrizzi, Flavio Ragni, Raffaele Marchesi et al.
In briefHow Models Collapse When Self-Rewarding Across Many Problems in Label-Free RL
SummaryThe paper studies label-free reinforcement learning with verifiable rewards (RLVR), where a language model is rewarded for agreeing with its own majority vote. Reasoning first improves but eventually collapses, sometimes to the same answer for every problem. The authors explain the dependence on dataset size as a collective instability in a model where problems share parameters, deriving a critical number of problems above which collapse occurs. Simulations and experiments with a small Transformer support these predictions.
In briefJudging Articulated Object Manipulation by Joint Constraints
SummaryThis paper argues that manipulation policies are usually judged only by binary task success, which cannot show whether a commanded motion matches the object's kinematic constraints. The authors propose Constraint-Referenced Strategy (CRS), which projects the commanded end-effector twist onto the object's joint axis to separate progress along the joint from off-constraint motion. CRS requires a joint model and a joint-following scripted controller derived from it, but no recorded demonstrations or learned judge.
In briefWhen, Where, and What to Correct: Human-Guided Robot Policy Refinement
SummaryThis paper introduces Adaptive Residual Correction (ARC), a lightweight post-training module that learns a residual correction on top of a frozen robot policy using reinforcement learning. Sparse human corrections guide autonomous exploration by determining when residual control activates and where in the residual action space refinement should explore, while task reward determines the final residual executed. On three precision-manipulation tasks on a real robot, ARC raised average task success from 33.3% to 75.0%.
Anukriti Singh, Maxime Caniot, Pratap Tokekar et al.
In briefGroundControl: Generating new views of outdoor scenes from street and satellite images without 3D reconstruction
SummaryThis paper proposes a multi-view diffusion model that synthesizes novel views of outdoor scenes by conditioning on both satellite images and street-level input images. It uses cross-attention conditioning to keep pretrained multi-view diffusion priors while following the scene layout shown in the satellite views, and it does not rely on 3D scene reconstruction. Experiments on outdoor datasets report better novel-view synthesis than satellite-only or street-view-only baselines, especially when target and source views overlap very little.
Akshay Krishnan, Matias Turkulainen, Jamie Wynn et al.
In briefWhich Hand Was Late? Diagnosing Robot Failures in Two-Handed Tool Use
SummaryThis paper studies how a robot can diagnose failures in bimanual tool-use tasks, where a tool may slip or one hand may act late. Using a simulated task whose failure causes are read from recorded traces, the authors compare what different sensing channels reveal about these causes. They find that a repair policy that escalates when the cause cannot be identified repairs about as many failures within two attempts as one given the true cause.
