In briefCausalWAM: Learning What Causes Actions in World Action Models
SummaryThis paper studies world action models (WAMs), which learn latent variables from observations to both predict scene changes and generate actions. The authors note that only part of the scene causes actions, so a model may mix action-relevant variables with irrelevant ones. They prove that under a sparsity constraint, the variables used for actions are an invertible function of the action-relevant latent variables alone. They propose CWAM, which adds a sparse routing path to a pretrained WAM's action head, and experiments on synthetic data and real robots show improved closed-loop success.
In briefRethinking asymmetry in self-supervised learning of 3D geometry from unposed images
SummaryThis paper studies self-supervised 3D geometry learning from unposed images, where models predict cameras and scene structure and render held-out views. It proposes imposing the asymmetry in the 3D representation space rather than the input space, excluding held-out views only from the reconstructed scene while a shared backbone uses all views. The authors add training signals by aligning sparse-view and dense-view features and features from a frozen vision foundation model, and report improved depth and pose estimation.
In briefCost-aware stopping rule for freeze-thaw Bayesian optimization: when one-step lookahead is enough
SummaryThis paper studies freeze-thaw Bayesian optimization, where candidates are trained in resumable increments and the stopping decision uses a budget-free utility of best score minus cost per increment. It analyzes a one-step expected-improvement (EI) stopping rule and constructs a certificate showing when stopping is premature, with correcting it never lowering expected utility under the surrogate. On three hyperparameter optimization benchmarks, EI stopping is noninferior to the correction in most high-cost settings, while the correction helps at low costs on two of the three.
Long Vu, Jongseong Brad Choi, Binoy Ravindran et al.
In briefRangeSWA: Training-free length extrapolation for continuous diffusion language models
SummaryContinuous diffusion language models can suffer from repetition, semantic drift, and quality loss when generating text longer than their training length. This paper proposes RangeSWA, a training-free method that combines local sliding-window attention, remote block summaries, and dynamic content retrieval to keep local detail and access long-range context. It further reduces degradation through positional remapping, conditioning attention correction, and denoising-stage-adaptive window scheduling. Experiments on multiple continuous diffusion language models show it lowers generative perplexity and repetition rates in long-text generation.
In briefThink in 360°: Benchmarking and training AI to reason over panoramic video
SummaryThe paper introduces PanoVideoBench, a manually annotated benchmark of 1,000 questions for understanding panoramic video across all viewing directions and time. It also builds PanoVideoCoT, a reasoning dataset of 13,465 QA pairs grounded in real panoramic video. The resulting model, PanoVideoReasoner, reaches 55.7% on the benchmark, improving its backbone and the previous state of the art.
In briefPrivacy-preserving synthetic data via noisy sliced Chebyshev moments
SummaryThis paper introduces SPLICE, a differentially private synthetic data algorithm that measures noisy sliced Chebyshev polynomial moments of a target distribution and fits a distribution to match them. For a dataset of n examples and constant privacy parameters, the authors prove the output's Sliced Wasserstein-1 distance to the target is Õ(1/√n), a bound independent of data dimension. The method reportedly achieves state-of-the-art performance against prior differentially private synthetic data methods on an extensive benchmark.
WhyThe authors present SPLICE, a dimension-independent Sliced Wasserstein bound for DP synthetic data from noisy Chebyshev moments, compared against prior methods on a benchmark.
Lucas Rosenblatt, R. Teal Witter, Christopher Musco
In briefRAMA: Memory-Augmented Multi-Agent Routing for LEO Satellite Networks
SummaryThis paper proposes RAMA, a multi-agent deep reinforcement learning routing framework that augments each satellite's decisions with experiences retrieved from a per-satellite episodic memory. Each satellite combines local observations, recent routing history, neighbor state, and retrieved memory entries through a Transformer-based decision network. In a 140-satellite Kepler constellation, RAMA reduces mean latency by 13.4% and 99th-percentile latency by 66.7% compared with the DDQN baseline. On the 1,584-satellite Starlink shell 1, the mean latency reduction reaches 93.2%, and online updating lowers it by a further 25.5%.
In briefKnowledge Layer Model (KLM): Visible Evidence for AI Answers, Routing Tests, and Failed Results
SummaryThe paper proposes the Knowledge Layer Model (KLM), an open specification for recording AI inferences with versioned assessments, evidence links, and explicit unknowns. It uses the SEDIM architecture from Nage AI OS as a reference implementation and reports internal experiments, including a platform-domain routing benchmark with 137 of 150 classifications correct (about 91.3%) alongside several negative or inconclusive results. The author stresses that these findings are internal and that evidence visibility does not establish correctness, causal reliance, security, or regulatory compliance.
In briefPersonaChatBench: Testing Chat Assistants With Varied Simulated Users
SummaryThis paper introduces PersonaChatBench, a framework that evaluates chat assistants by pairing the same task with multiple simulated users, each guided by a persona card describing distinct needs, preferences, and interaction behaviors. It tests 20 language models across six domains, measuring task outcomes, fulfillment of persona needs, and simulator fidelity separately. The results show that task performance and need fulfillment vary across personas and domains, and that strong task performance does not necessarily satisfy individual needs.
In briefCRoS: Rewarding Causal Reasoning on Scrambled Chains of Thought
SummaryThis paper proposes CRoS, a metric for measuring whether a model's chain of thought causally affects its final answer. It scrambles the order of reasoning steps without changing their content, avoiding the confounding effects of earlier perturbation methods, and uses the metric as a training reward. On Reasoning Gym with Gemma-3-4B, CRoS reaches 46.8% accuracy on five unseen logic and game datasets versus 43.8% for the strongest baseline, and averages 43.4% CoT-Pass@K versus 41.4%. Its hint acknowledgment score is 34.0% versus 20.2% for baselines.
In briefREA: A Routed Embodied Agent Runtime for Asynchronous Brain-Cerebellum Control
SummaryThis paper presents REA, a unified runtime in which a deliberative high-level reasoning model, an embodied router, and fast reactive robot policies cooperate without any running policy waiting on the reasoning model. It uses a route memory of pre-measured policy success rates, a judge gate triggered at policy execution boundaries, and a shared-budget escalation ladder for recovery. On 3,000 episodes across 60 LIBERO-Pro tasks, REA reaches 95.9% success, above the best single-agent baseline (94.7%), while cutting control steps by 25% and control time by 17%.
Pengfei Zhou, Yixing Ma, Zhengfeiyang Zhang et al.
In briefExact Theory and Optimal Stepsizes for Stochastic Gradient Descent in Smooth Convex Minimization
SummaryThis paper studies how to choose stepsize schedules for stochastic gradient descent on smooth convex problems, assuming zero-mean gradient noise with bounded variance. It gives an exact worst-case analysis of the expected suboptimality of the final iterate for any N stepsizes in [0,1.5]. When the variance bound is at least order 1/√N, it identifies the minimax optimal schedule, proves its uniqueness, and shows the optimal stepsizes decrease monotonically.
WhyThe authors give an exact worst-case analysis of SGD's final-iterate suboptimality, prove uniqueness of the minimax optimal stepsize schedule, and rule out fractal accelerations.
In briefA Species-Level Acoustic Biodiversity Index: Classifier Benchmarks, False-Positive Control and Comparison with Standard Soundscape Indices
SummaryThis paper revises an Acoustic Biodiversity Index (ABI) built from species-level sound detections. Its main form is the Shannon diversity of detections divided by the regional species pool expected for the site and season, with temporal stability reported separately. The index was tested at an African savanna and at zoo habitats in Abu Dhabi. It correlated strongly with species richness (Spearman rho = 0.93), while six standard acoustic indices showed weaker correlations.
Muhammad Saeed, Dmitry Mikhaylov, Mohamed H. Alhosani et al.
In briefNear-optimal second-order convex finite-sum optimization under individual smoothness
SummaryThis paper studies the stochastic second-order oracle complexity of smooth convex finite-sum optimization with n components. Under individual smoothness and convexity, it proves a lower bound of Ω(n+n^{5/7}ε^{-2/7}). The authors propose a Taylor-corrected lazy Newton method combined with Monteiro-Svaiter acceleration, achieving Õ(n+n^{5/7}ε^{-2/7}) complexity, which matches the lower bound up to logarithmic factors.
In briefSame Question, Different Answers: Auditing How Six LLMs' Smoking-Cessation Advice Varies by User Persona
SummaryThe authors asked six widely used large language models from four providers the same four smoking-cessation questions, posed by twelve user personas differing in life stage, self-described expertise, and smoking history, yielding 2,880 responses. Models mentioned FDA-approved prescription cessation drugs 32 to 67 percentage points less often for teen personas than adult personas, and five of six models gave fewer facts to personas claiming healthcare expertise. The study also tested a system prompt specifying a factual floor, which narrowed persona differences but produced effects that varied across models.
WhyThe authors audit six LLMs with persona variation on smoking-cessation advice, test a factual-floor system prompt across 2,880 responses, and release all responses and analysis code.
Artin Yousefi, Aizierjiang Aiersilan, Lorien C Abroms et al.
In briefLearning clean image generation from only noisy measurements
SummaryThis paper studies how to train a generative model of clean images using only noisy observations. The authors show that Noise2Noise regression can learn posterior predictions from independent noisy views, but only at or above the training noise level. To cover lower noise levels, they impose a consistency condition and derive when it uniquely determines the path to the clean distribution. Experiments on CIFAR-10, photon-counting microscopy, and synthetic aperture radar show that the consistency relation repairs the sampling process beyond the data's noise level.
WhyThe authors use a consistency constraint to learn posterior predictions below the training noise level and derive conditions for uniquely recovering clean image distributions, reporting tests on CIFAR
In briefAny Forcing: Flexible-step, stable one-step autoregressive video generation
SummaryThis paper proposes Any Forcing, a framework for autoregressive video generation that aims to support flexible sampling step counts while keeping one-step generation stable. A unified adversarial discriminator performs real-distribution alignment, inter-step consistency alignment, and first-frame quality enhancement. A training-free adaptive long-term memory mechanism selects historical states using global semantic relevance and local value contribution, and remaps their temporal positions. The authors report 20.9 FPS for short videos and 24.7 FPS for long videos on a single H100 GPU.
In briefNo free loop: whether looping a trained model is worth its extra inference cost
SummaryThis paper examines whether re-applying a block of layers in an already trained model, which adds loops without adding parameters, justifies its extra cost per generated token. The authors compare looped models against an unlooped model with equal parameters and an unlooped model with equal inference FLOPs. Across Qwen3 and Qwen3.5 models from 0.8B to 9B, only the looped 0.8B model beats both baselines. At 4B and 9B, the looped models do not beat the equal-parameter baseline, so adding loops after pretraining pays off only for small models with a fixed parameter budget.
In briefCausalXRL: Explainable reinforcement learning via causal graph reasoning
SummaryThis paper introduces CausalXRL, a framework that uses causal graph reasoning to explain reinforcement learning agents' decisions through transparent, structured, multi-level explanations. The authors validate it with a quantitative analysis of explanation fidelity and causal-structure learning efficiency on benchmark RL environments, and a qualitative expert study on the MicroRTS real-time strategy benchmark. The results indicate faithful explanations with efficient causal-structure learning, and experts found it useful for inspecting high-level RTS strategies.
In briefA neural network method for estimating knot point density in real time for a nonlinear quadrotor
SummaryThe paper proposes a learning-based method for adaptively placing knots in B-spline approximations of trajectories generated by ordinary differential equations. A neural network maps the system's initial condition to quantities describing the local variation of the trajectory. These estimates are used to build nonuniform knot distributions that roughly equalize a local error indicator, reducing the number of control points needed for a prescribed accuracy.
Jasmine Ratchford, Raffaele Romagnoli, Mark H. Klein
In briefGhostPrefix: Universal Audio Prefix Attacks That Manipulate Speech Recognition Without Syncing
SummaryThis paper introduces GhostPrefix, a universal adversarial attack against automatic speech recognition (ASR) systems. It learns a short, reusable audio segment that is prepended to live speech, so it can influence ASR outputs without knowing the upcoming utterance or aligning with it in time. The authors test three attack modes across 13 open-source ASR models and report high success rates in both digital and over-the-air settings. They also show that the resulting errors affect downstream tools such as voice assistants and AI note-taking applications.
WhyThe authors present the first synchronization-free universal audio prefix attack, evaluated on 13 ASR models in digital and over-the-air settings across three attack modes.
In briefDetecting When an Equation Is Wrong: Structural Uncertainty in LLM-Guided Symbolic Regression for Extrapolation
SummaryThe paper addresses uncertainty when large language models are used to discover equations from data, since many methods return only one best equation. The authors propose LLM-SR-UQ, which keeps competing equation structures and uses their disagreement to measure uncertainty about the functional form. Experiments on LLM-SRBench show the approach improves coverage under extrapolation while preserving in-distribution coverage, and it also generalizes to other symbolic-regression methods.
In briefAsynchronous flow map language models that build token order into one-step generation
SummaryThis paper introduces asynchronous flow map language models (aFMLMs), which learn flow maps of asynchronous flows to incorporate prediction order into one-step parallel language modeling. The flows specify an ordering through a tokenwise time schedule that controls each token's relative denoising progress, and can use either prescribed or learned, state-adaptive schedules. The models generate all tokens in a single forward pass, and they set new state-of-the-art one-step results on Sudoku solving and code generation for GSM8K math word problems.
Chanhyuk Lee, Wonjung Kim, Patrick Pynadath et al.
In briefWhen Does Jacobian Fine-Tuning Stabilize Neural PDE Emulators?
SummaryThis paper examines why neural emulators of chaotic PDEs, trained on one-step prediction, accumulate errors when run autoregressively. A first-order decomposition separates error growth that the model's own dynamics amplify from growth that is physical and cannot be trained away, summarized as an amplification ratio. Across 21 configurations of four PDEs, five architectures and two model sizes, Jacobian fine-tuning markedly reduced error when the amplification ratio was high.
In briefTurnTurk-GPT: Text-based turn-taking prediction for Turkish
SummaryThis paper introduces TurnTurk-GPT, a text-based model that predicts turn-taking in Turkish conversation, built by fine-tuning a Turkish GPT-2 model on synthetic and real dialogue. It is compared with rule-based baselines, a zero-shot Turkish language model, and a multilingual end-of-turn detector on real Turkish conversation. TurnTurk-GPT outperforms the multilingual detector in both evaluation settings, but beats the zero-shot model only when test speakers appear in training. The model and evaluation protocol are publicly released.
Ahmet Tu\u{g}rul Bayrak, Fatma Nur Korkmaz, Nur Ya\u{g}mur Tuncer et al.
In briefTradeWar: A Benchmark for Testing AI Negotiators on Real Trade Outcomes
SummaryThis paper introduces TradeWar, a multi-country simulation benchmark that turns negotiation dialogue into concrete tariff and trade-share policies, applies capacity and network constraints during settlement, and scores results with an Economic Health Index. It also proposes TACTIC, an LLM agent that refines its negotiation strategy across episodes using past experience, opponent behavior, and settlement outcomes, without updating model weights. The authors report that TACTIC achieves a higher final EHI than baselines and degrades less under technology-control and energy-crisis shocks.
In briefGeoFlip: A Benchmark for Geometry-Sensitive Physics in Video Generation
SummaryGeoFlip is a benchmark testing whether video generation models produce correct physical outcomes when the initial scene geometry changes. It contains 482 scene pairs and 964 reference videos across six physical categories, with reference outcomes verified through physics simulation. Evaluating 16 models, the best average event correctness reached only 27.98%.
In briefAutoAnalysis: A Self-Updating Meta-Analysis of Human-AI Collaboration
SummaryThis paper introduces AutoAnalysis, an automated, self-updating meta-analysis that examines when humans and AI perform better together and when they perform better alone. Its most recent run synthesizes 1,955 effect sizes from 420 papers. On average, human-AI combinations outperform humans alone but perform significantly worse than the better of humans or AI working alone. As new studies appear, the system will incorporate their findings.
Michelle Vaccaro, Abdullah Almaatouq, and Thomas W. Malone
In briefLess Loss, Fixed Weights: Fixed Logit Corrections for Language Model Forgetting
SummaryThis paper examines whether the loss increase that language models suffer after instruction tuning or further pretraining on a new domain, known as forgetting, can be undone with simple fixed adjustments to final token predictions. The authors fit per-token vocabulary biases, alone or with a temperature, across 14 base and derived model pairs ranging from 1B to 7B parameters. Bias alone recovers under 10% of the loss, while bias combined with temperature recovers up to 36%, and math-pretrained models recover almost nothing.
In briefCAPER: Circuit-aware fine-tuning for LLM-based analog circuit verification
SummaryThis paper proposes using large language models as assistants to verify analog circuits, flagging unmet specifications and reliability issues under process, voltage, and temperature (PVT) variation. It introduces CAPER, a parameter-efficient fine-tuning method that adds a branch beside LoRA to read each circuit stage's properties and combine them across relevant stages. The authors report that CAPER correctly judges whether a feedback loop is negative and learns to flag PVT-induced failures in comparator latches, where generic adapters perform near chance.
In briefDAMI: Dual-Axis Memory Initialization for Better Long-Context Retrieval in Gated Linear Attention
SummaryThis paper proposes Dual-Axis Memory Initialization (DAMI), which jointly configures gate parameters of gated linear recurrences before training to improve long-context retrieval. It assigns geometric timescales across heads and a low-high-low input-sensitivity profile across layers, calibrated with a small activation sample without adding trainable parameters. Across GDN, KDA, GDN2, Mamba-2, and RWKV-7, DAMI raises average RULER-style retrieval accuracy by 5.29 to 16.71 percentage points over random-initialization baselines, while language modeling, commonsense reasoning, and 2K real-world retrieval remain broadly comparable.
In briefMoETok: Sparse Expert Routing for 1D Visual Tokenizers
SummaryThis paper introduces MoETok, a 1D visual tokenizer that routes image tokens through sparsely activated experts. The authors find that learnable queries spontaneously specialize under semantic supervision, with a small set capturing high-level content and a larger set preserving fine visual details for reconstruction. Jointly optimized with reconstruction and semantic objectives, MoETok improves reconstruction, semantic representation, and autoregressive generation over dense tokenizers under matched conditions.
In brief"Neurons Are All You Need": a new neural layer that stores persistent memory within units so learned abilities can transfer between model instances
SummaryThe paper proposes NURESA (Neural Unit with Retentive Self-Adaptive Architecture), a neural layer that adds persistent memory inside each unit, separate from computation, which accumulates and consolidates into transferable memory registers. Experiments on MLP, GPT-2 and Qwen2.5 report zero-shot transfer of learned capability into a naive model instance without retraining or the original training data. The authors also report that the method protects against domain loss during fine-tuning, and that transfer takes about one minute with files far smaller than the original model.
Joseph B. Alzagatiti, Alain C. Glanzman, Perry A. Demsko et al.
In briefID2Pattern: Pretraining Reusable Identity Behavior Patterns for Transaction Risk Models
SummaryThis paper proposes ID2Pattern, a pretraining framework for transaction risk foundation models in online payments. It encodes card, device, and account IDs with a Local Anonymous ID tokenizer that captures how each identity appears and recurs within a transaction sequence, not its specific value. A multi-scale pattern codebook is then pretrained on unlabeled sequences to turn these identity sequences into discrete behavior patterns. Experiments on an industrial dataset and six public datasets show consistent gains over strong baselines, and the method remains effective when all identity values are unseen during training.
In briefFoAViS: Studying learned gaze strategies in a human-like active vision system
SummaryThis paper presents FoAViS, a machine vision system that samples input in a foveated, log-polar manner and accumulates information in a 2D spatial memory usable for both classification and segmentation. A fixation policy is trained jointly with the task to choose where to look on a 2D map. The learned policies select more informative locations than non-learned heuristics, and models trained with them outperform those trained on random locations. The resulting scanpaths partly resemble characteristics of human gaze behavior.
Valentin Hassler, Jonathan Henrich, Yoha\"i-Eliel Berreby et al.
In briefTwinPast: Testing Whether Streaming Video Models Use Historical Evidence
SummaryThis paper argues that high scores on streaming video benchmarks do not show whether models actually use visual history. It introduces TwinPast, a benchmark of 350 human-reviewed video pairs that keep recent context, question, and options fixed while changing key historical evidence and the correct answer. A model passes only if it answers correctly for both histories. Across 34 model and input configurations, the strongest streaming system reaches 49.14% PairAcc, versus 66.57-91.43% for general-purpose models.
In briefNew lower bounds on sample needs for private PAC learning
SummaryThis paper studies realizable PAC learning under approximate differential privacy and presents a framework for proving lower bounds on sample complexity. It constructs classes with identical VC and Littlestone dimensions but different private sample complexities, showing these dimensions alone do not determine the sample complexity. The authors introduce the Independent Class Packing parameter and prove that any private PAC learner, even an improper one, needs Ω(ICP/(αε)) examples. Using this framework, they obtain tight bounds for axis-aligned rectangles, unions of intervals and linear subspaces, and improved bounds for halfspaces and convex polygons.
WhyThe authors introduce Independent Class Packing and prove lower bounds for private PAC learning, reporting tight bounds for rectangles, unions of intervals, and subspaces.
In briefDoes vendor image processing matter? Learning how echocardiography displays look from raw beamspace data.
SummaryMany echocardiography datasets contain images processed by vendors, but that processing is often undocumented and may affect model performance and robustness. The researchers trained a small convolutional neural network to reproduce General Electric display appearance from native beamspace brightness-mode ultrasound, using 8,406 recordings from 605 patients. The beamspace CNN cut test mean absolute error by 43% relative to the optimal pointwise brightness map, outperformed an equivalent Cartesian CNN, and runs in real time on a CPU.
In briefGradual pruning with sparse adaptation: an efficient way to shrink large language models to extreme sparsity
SummaryThis paper proposes ETAPS, a method for pruning weights from large language models. It tightens a global sparsity constraint in progressive stages while training a sparse delta under a small fixed budget, aiming to keep quality at high sparsity. Across six models from 125M to 13B parameters, ETAPS achieves lower perplexity than ELSA at every sparsity from 50% to 90%, with reductions of up to 42%. At 90% sparsity it also needs fewer GPUs and less persistent state and GPU-hours than ELSA.
Jean Lucien Randrianantenaina, Hung Nguyen, Nasib Ullah et al.
In briefWhen is a game a Markov decision process? Links between game theory and reinforcement learning
SummaryThis monograph gives theorems showing that game theory and reinforcement learning describe the same objects, treating a game as a Markov decision process in three distinct senses. It covers standard game classes and solution concepts such as Nash equilibrium, the soft Bellman operator and the Shapley value, and maps each to suitable training procedures and solution complexity. It also identifies correspondences that fail, including non-stationarity for independent learners, impossibility results for uncoupled convergence to Nash equilibrium, and the gap between coarse correlated and correlated equilibria.
In briefRobust learning of multi-index models under logconcave distributions using weak partial learners
SummaryThis paper presents a new algorithmic framework for learning Boolean concept classes under logconcave distributions, applicable to broad families of multi-index models and extendable to partial classification and testable learning. Its core is a boosting approach built on weak partial learners, which uses a recursive moment-based partitioning procedure and a method for detecting moment discrepancies to construct the required weak learner. The authors obtain faster robust learners for concept classes such as functions of halfspaces and multiclass linear classifiers under logconcave marginals.
WhyThe authors propose a boosting framework using weak partial learners plus a moment-based partitioning and discrepancy-finding method, reporting faster robust learners for multi-index models under logc
In briefWASABI: Wasserstein Sampling via Bridge Interpolants for Boltzmann Distribution Sampling
SummaryThe paper introduces WASABI, a method that trains a diffusion model to generate samples from a Boltzmann distribution using only the energy function. It minimizes the KL divergence between the model's distribution and the target using Wasserstein gradient flows, avoiding computation of the model's own log-density gradient. Once trained, new samples require no further energy or gradient evaluations. On alanine dipeptide, the authors report outperforming the best baselines with an estimated 3.3 times fewer energy evaluations, and competitive results on interacting particle system benchmarks.
Sebastian Salazar, Dilum Fernando, Michal Kucer et al.
In briefChronos-2: A foundation model for forecasting many related time series at once
SummaryThis paper introduces Chronos-2, a pretrained foundation model for zero-shot forecasting that models interactions among multivariate time series and external covariates. It infers relationships among series from historical data and produces probabilistic forecasts without task-specific training. The authors report state-of-the-art average accuracy on various benchmarks and much faster forecast generation than specialized methods. They also show it matches or surpasses domain-specific methods in infectious-disease, macroeconomic, carbon-flux, and runoff forecasting.
Abdul Fatir Ansari, Oleksandr Shchur, Jaris K\"uken et al.
In briefShock loading reaches reaction pathways beyond accelerated thermolysis.
SummaryThis paper develops EMFF-MACE, a reactive machine-learning potential for C/H/N/O energetic materials, and uses it to simulate thermolysis of 20 materials and shock loading of 10 on a common potential energy surface. At matched normalized reaction progress, the two loading types show divergent chemical states and local structural environments. The authors conclude that loading history affects the structural states and subsequent reaction pathways accessed during decomposition, not just the rate of thermolysis.
In briefAligning multimodal AI models with human cognition using behavior-derived semantic dimensions
SummaryThis paper proposes a framework that aligns multimodal large language models (MLLMs) with the multidimensional mental representations behind human similarity judgements. It builds on a behavior-derived semantic embedding space with 66 interpretable dimensions, derived from 4.70 million human Odd-One-Out judgements. The approach improves consistency with human similarity judgements while staying competitive on standard multimodal benchmarks. Searchlight Representational Similarity Analysis on an independent fMRI dataset also shows increased representational alignment between the model and humans.
In briefOptimizing stochastic policy geometry for few-step flow reinforcement learning
SummaryFew-step flow models generate samples through a small number of large denoising steps, and applying GRPO to them requires adding randomness to those steps. This paper proposes Geo-FSRL, which selects, from a family of coefficient-preserving isotropic Gaussian transitions, the policy that minimizes conditional KL divergence, and derives a unique closed-form optimizer. The authors report higher final reward scores than Flow-GRPO and Flow-CPS on image and video generation models in the 4-step setting.
In briefSensorium: An open-source library for collecting and organizing multimodal surgical data
SummaryThis paper introduces Sensorium, an open-source Python library built on ROS and ROS2 for recording, processing, annotating, and organizing multimodal data from configurable robot and sensor streams in surgery. It supports operator-facing data acquisition, configurable processing, semantic annotation, and generation of episodic datasets. The authors evaluated the library on 1516 recordings from phantom, ex vivo, and in vivo porcine experiments and report preliminary usability feedback.
In briefObservable-specific validity tests for scientific AI: controlled approximations and intervention response
SummaryThe paper proposes observable-specific validity records for scientific AI that link a simplified model to its reference model, error measure, tolerance, parameter domain, and transformations or interventions. It gives explicit displacement and velocity certificates for a singularly perturbed oscillator and uses exact Hubbard dimer relations to show how energy and response criteria accept different coupling domains. In 16 repeated neural fits, imposing dimensional homogeneity reduced energy scale transfer error from 52.3% to 0.0768% and raised the median fraction satisfying both gap and response tolerances from zero to 100%.
In briefAdaptive coverage algorithms for mobile networks in dynamic environments
SummaryThis paper extends an existing framework for coordinating mobile agents in unknown spaces to handle dynamic instability. It adds two settings, Agent Loss and World Change, to the original problem. The authors improve the centralized CADENCE algorithm and its decentralized version DADENCE. On a benchmark of over 4500 test cases, the adapted algorithms outperform another state-of-the-art benchmark algorithm.
In briefBinary Sparse Codes: How Far Have We Come in Energy-Efficient Representations?
SummaryThis paper studies binary sparse codes, which represent information in artificial systems using only a few active units, inspired by sparse activity in biological brains. The authors propose metrics that formalise information preservation as similarity preservation, measuring how faithfully a code reproduces the relational structure of its input. Comparing four representative methods under matched conditions, they find a trade-off between high information preservation and high sparsity. Competitive Sparse Codes is the only method reported to keep near-maximal sparsity while retaining meaningful information preservation.
In briefA family of prior forces for implicit denoiser priors and how they guide decisions in linear inverse problems
SummaryThis paper studies implicit priors built from Gaussian denoisers, used in methods such as plug-and-play (PnP), regularization by denoising (RED), and score-based approaches. It defines a prior-force family indexed by noise variance so these implicit priors can be compared with explicit ones like MAP priors. The authors compare the resulting energies for MMSE denoisers and conclude that the noise level acts as a prior parameter rather than a step size. Experiments with exact priors support these predictions.
WhyThe authors define a prior-force family for Gaussian denoisers and compare PnP, RED and MAP energies, reporting that noise level acts as a prior parameter, with experiments on exact priors.
In briefVariance reduction for decentralized optimization with multiple local updates
SummaryThis paper proposes a primal-dual decentralized optimization algorithm that combines variance reduction with multiple local updates under stochastic gradients. Its stepsize is independent of the network size and inversely proportional to the number of local updates, allowing larger stepsizes and faster convergence. In the strongly convex setting, the authors establish the first linear convergence guarantee in expectation for this setting. The method exchanges only one variable per round, halving communication cost compared with gradient-tracking methods, and numerical experiments show gains in speed and accuracy.
In briefIGDA-ICMH: A Framework That Adapts Features and Networks by Importance for Image Compression Serving Both Machines and Humans
SummaryThis paper addresses image compression intended for both machine vision and human vision. The authors first quantitatively examine how different features and network layers affect machine vision task performance. Building on this, they propose the IGDA framework, which uses implicit neural representation and adds low-rank adaptation modules that adjust key network layers and exploit non-local feature properties. Experiments on image classification and object detection show better performance than existing implicit neural representation-based methods.
In briefStructure-Aware Skill Retrieval and Arbitration for Optimization Modeling
SummaryThe paper presents ArchOR, an archetype-centric skill agent that extends OptSkills with two retrieval modules: Centroid-Aware Representation Learning (CARL) and Signature-Guided Retrieval (S³R). Candidates are ranked by weighted Jaccard overlap of structural signatures, and an arbitration rule decides when the structural route replaces the original selector. Using the same DeepSeek-V4.1-Flash backbone, ArchOR reaches 65.00% Micro-Avg. Pass@1 across five benchmarks, improving over OptSkills.
In briefRate-distortion optimized skip mode coding for AVS neural video compression
SummaryThis paper studies skip mode coding in neural video compression under the AVS standard, examining how rate and distortion are distributed across spatial positions and temporal reference qualities. It proposes RDOSkip, a framework that models skip-mode rate and distortion using statistical distributions predicted by a hyperprior and Bayes' theorem, with an adaptive decision method based on joint spatial-temporal rate-distortion optimization. The method requires no training, adds negligible complexity, and yields BD-rate savings of 0.90% on AVS and 1.29% on HEVC datasets.
In briefHow completely and cheaply retrieval methods find all the tools an LLM agent needs for multi-step requests
SummaryThe paper compares six retrieval structures for selecting tools when an LLM agent must call several of them for one request, across two multi-tool testbeds. It measures whether the top-k results contain every required tool, along with the LLM calls, tokens, and latency each method incurs. Methods that are untrained on the testbed and make at most one LLM call per query complete at most 10% of AppWorld queries at k = 16. A set-level retriever trained on in-domain queries does better with no LLM calls, while an LLM-navigated hierarchy also does better but needs many calls and about five minutes per query.
In briefStatistical foundations of reinforcement learning: a finite-sample analysis (monograph)
SummaryThis monograph systematically analyzes reinforcement learning sample complexity from a non-asymptotic perspective, based on tabular Markov decision processes. It covers three sampling settings, namely RL with a generative model, online RL, and offline RL, along with model-based and model-free algorithms. The emphasis is on minimax-optimal algorithm design and sharp statistical characterizations of widely used RL methods. The work also briefly discusses extensions such as robust RL, multi-agent and federated RL, uncertainty quantification, and RL from human feedback.
In briefReX-GG: LLM ensemble pipeline for relation extraction and graph building
SummaryThis paper presents a lightweight workflow in which several large language models work together, using JSON-structured outputs and anonymous peer-review ranking to reduce hallucinated outputs and single-model failures. The pipeline is applied to extracting four types of relations (strong causation, weak causation, contrastive, and compositional) from English popular science articles, producing interactive, colour-coded HTML causal graphs. An anonymous user study reports an average perceived accuracy of 0.778 against a human-annotated gold standard.
In briefUsing machine learning on learners' interaction traces to measure self-efficacy in technology-enhanced learning
SummaryThis paper explores whether low self-efficacy can be detected from learners' interaction traces while they solve exercises in a technology-enhanced learning system. The authors ran an experiment at two universities, collecting interaction data alongside self-reported self-efficacy measures. They built a machine learning pipeline with multiple classification models and argue that closely monitoring self-efficacy, which changes dynamically during learning, improves prediction performance.
In briefGeneral AI models outperform chemistry-specific models at predicting molecular properties
SummaryThis paper compares general-purpose foundation models with chemistry-specific ones on prediction tasks across chemistry. With only 60 experimental data points, the general model GPT-6 Astra outperforms existing chemical models on five physicochemical and quantum-chemical regression tasks. Chemical models only matched its performance with about ten times more data. The authors conclude that learning basic chemical structures does not replace training on relevant human knowledge.
Stepan N. Ostarkov, Gregorii A. Semakin, Ivan A. Bespalov et al.
In briefMean squared consistency of k-nearest-neighbor regression in metric spaces
SummaryThis paper studies mean squared consistency of k-nearest-neighbor regression when features take values in a metric space. It analyzes how neighbor selection redistributes probability and gives separate conditions for the approximation error and the conditional response variance. Under a finite fourth response moment with k_n tending to infinity and k_n/n tending to zero, a uniform L2 bound on the selection densities suffices. The authors also construct a space satisfying sigma-finite Nagata dimensionality on which the mean squared error of a fixed noiseless response with all finite moments diverges, showing that this dimensional condition does not guarantee mean squared consistency for unbounded responses.
WhyThe authors generalize k-NN regression mean squared consistency to metric spaces, give selection-density bounds, and construct a counterexample with divergent error.
L\'aszl\'o Gy\"orfi, Vladimir G. Pestov, Harro Walk et al.
In briefUsing LLMs to automatically generate search keywords for virtual datasets
SummaryThe paper addresses the difficulty of finding virtual datasets in a data virtualization service (DVS), where many versions of datasets are created for different tasks. The authors propose using large language models to generate keywords for each virtual dataset, and design five strategies for selecting text data to feed the models. They tested these strategies on virtual datasets from five case studies, using three LLM generators and two embedding techniques. The results show that strategy performance depends on the quality of the text descriptions, supporting a multi-strategy keyword generation service.
Saiful Khan, Yiwen Xing, Joyraj Chakraborty et al.
In briefLearning long-range interactions in large heterogeneous metamaterials with scientific machine learning
SummaryThis paper introduces domain-decomposition attention neural operator (DANO), which learns configuration-dependent interactions in heterogeneous systems governed by linear field equations using field data alone. Using large-scale optical metasurfaces as a case study, the authors show full-field prediction under geometric heterogeneity and varying drive conditions, transfer from small to large domains, and multifunctional inverse design. The learned interactions are validated through physics-based interrogation.
In briefUsing Question Text as an Index for Zero-Shot Cross-Domain Knowledge Tracing
SummaryThis paper proposes TIZSKT, a method for knowledge tracing that predicts a student's next response on new platforms that have question texts but no response logs for training. It uses text as an index for building question-specific historical correctness summaries, rather than as a directly transferable representation. In leave-one-domain-out experiments on three public datasets, TIZSKT achieved the highest mean AUC and accuracy on all target domains among the evaluated baselines.
In briefA fuzzy model for forecasting how customer preferences change over time
SummaryThe paper forecasts how retail chain customers' preferences evolve, representing each preference as a triangular fuzzy number. The evolution is modeled by a Mamba-type selective state space model computed in a two-component fuzzy number algebra. On a synthetic example with competitor promotions and customer switching, a 107-parameter model achieves a lower forecast-centre error than two baselines and gives an interpretable estimate of forecast vagueness.
In briefMovens-Zero: scaling humanoid robot motion models using large-scale human videos
SummaryThis paper addresses the shortage of diverse, robot-executable data for training humanoid motion foundation models by building Movens-5M from Internet videos, containing 15 million human-motion sequences and 5 million physically executable humanoid sequences. It introduces Movens-Zero, an architecture that jointly learns from text, speech, music, and video and supports real-time closed-loop execution. The authors also release Movens-Intent, a four-modality benchmark of 8,192 expert-verified condition-motion pairs, and report consistent gains as data and model size increase.
In brief"Learning to Live": A Step Toward Lifelong Embodied Intelligence
SummaryThis paper introduces LIFE, a framework that lets robots observe continuously, remember selectively, and turn experience into new capabilities. It combines RoboBeats, lightweight streaming gates that track conditions and discover events; Plastic Memory, a flat store of images and structured records; and Explore-Distill-Mount (EDM), which turns validated problem-solving solutions into reusable tools and skills. Under matched protocols, RoboBeats improves condition tracking over its base models and retains more useful evidence than uniform sampling at the same writing budget. A limited case study replaying 24 hours of real-robot video shows the system building memory from a day of experience and answering everyday questions.
In briefAI screeners favor cover letters written by certain AI models, regardless of candidate quality
SummaryThis paper examines whether the choice of large language model used to write job applications and to screen them affects hiring outcomes beyond candidate qualifications. The authors built a multi-model audit framework pairing eleven writer models with nine evaluator models across ten job descriptions and fifty candidates per job. They find a consistent global ranking of writer models, with locally hosted models disadvantaged, and evaluators showing self-preference for their own outputs. In the most adverse pairing, about 1.8 of 25 qualified candidates were filtered out solely due to model mismatch.
Nicol\`o Pagan, Michael Andreas Ebling, Stefania Ionescu et al.
In briefVideo director AI agents that learn to interact and direct together
SummaryThe paper introduces LIT, a framework with a director agent, a simulated user, and a meta engineer that improves multistage video production and user interaction. The meta agent proactively learns external knowledge to better discover problems and propose solutions, forming a co-evolutionary closed loop with the other two components. On the Test and OOD Test splits, LIT improves production and interaction scores by 19.1% and 32.0% respectively over the initial pipeline.
In briefSimXAI: A Drop-in Conversational Layer Improving Dialogue Quality in Explainable AI
SummaryThis paper presents SimXAI, a lightweight conversational layer centered on a large language model that replaces the separate intent recognition, dialogue management, and template-based response modules of GUS-style explainable AI agents with a single tool-calling LLM. It also introduces an evaluation suite covering single- and multi-turn parsing, faithfulness, and contextualization. Across existing and new datasets, SimXAI achieves strong function-selection accuracy, high faithfulness, and better handling of context-dependent requests. In a user study replacing only the conversational layer of TalkToModel, users completed tasks more accurately and in fewer turns.
In briefWord-level and character-level similarity via path signature embeddings
SummaryThis paper proposes building sentence embeddings from the path signatures of Word2Vec trajectories, using iterated integrals from rough path theory to capture word order and lexical structure. Ablations varying Word2Vec dimensions, signature depth, and PCA compression correlate embedding distances with five lexical string metrics. Average pooling almost entirely fails to capture character-level edit distances, while path signatures align strongly with them, with depth-2 signatures best balancing character and word-level sensitivity.
In briefHow to Simplify a Line Chart: A Taxonomy and Feature-Based Evaluation
SummaryThis paper proposes a taxonomy for simplifying line charts based on how methods preserve visual features such as trends and extrema. It evaluates 19 algorithms across three families (Transformers, Reducers, and Aggregators) using 12 defined features and 21 preservation measurements. A Feature–Complexity score balances fidelity and simplicity, and results across 80 datasets show clear differences between algorithm families.
In briefDynamical Memory Transformer: a hybrid model with fixed-cost memory for long sequences
SummaryThis paper proposes the Dynamical Memory Transformer (DMT), a hybrid architecture that separates sequence encoding from information reading. Its encoder uses several memory units built from dynamical systems with complex diagonal recurrent matrices, updated by an input-dependent leak rate, while its decoder uses attention to read their outputs instead of the token history. Compared with Mamba and Transformer on cognitive and language modeling tasks, DMT performs best on several cognitive tasks and comes close to Transformer on language modeling but trails Mamba. It also shows stronger length extrapolation than both on selective copy.
