In briefFixing a Model That Learned Reversed Cancer Risk: Monotonic Constraints for Bladder Cancer Recurrence
SummaryUsing data from 435 patients in a UK multicentre bladder cancer trial, the authors found that an unconstrained XGBoost model learned that higher tumour stage and carcinoma in situ predicted lower recurrence, contrary to clinical expectation. Standard checks such as discrimination, calibration, and SHAP did not detect this. The authors introduce a counterfactual direction test and monotonic constraints, which removed the inversion without reducing discrimination, and they recommend such testing before clinical deployment.
In briefLang3DSeg: Annotation-Free Open-Vocabulary 3D Segmentation with Point Transformers
SummaryThis paper presents Lang3DSeg, which uses a point transformer, trained from scratch without geometric pre-training, for annotation-free open-vocabulary semantic segmentation of outdoor LiDAR point clouds. To handle depth ambiguity in 2D-to-3D label projection, it composites masks with an explicit class-priority rule and truncates each projected instance at the first gap in its depth distribution. The method reaches 52.8% mIoU on nuScenes validation and 41.4% on SemanticKITTI, and runs in real time on a single LiDAR sweep without running vision-language models at inference.
Cigdem Kokenoz, Amir Salarpour, Alkim Domeke et al.
In briefPhysics-based evaluation of AI models for Earth science, beyond forecast accuracy
SummaryAI foundation models pretrained on weather and climate data are increasingly fine-tuned for Earth science tasks beyond weather forecasting, faster than the community can evaluate them. Current assessments rely almost entirely on benchmark skill against reference products, which does not show whether a model represents the physical processes governing the system. The paper identifies five priorities for physical evaluation, spanning training data, fine-tuning, behavioral testing, mechanistic interpretability, and output validation. It recommends open AI-ready evaluation datasets, a shared reporting standard for physics-based evaluation, and a dedicated research program on model safety over the coming decade.
Makoto Kelp, Amirhossein Arzani, Patricia Castellanos et al.
In briefChild-adapted structured phonological representations for interpretable speech sound analysis
SummaryThis paper adapts the PhonoQ-2.0 model to child speech using CHILDES-aligned data, comparing three alignment-supervision conditions and two initialization strategies. Evaluated against manual annotations on 1,352 consonant targets from 58 typically developing children, child-speech adaptation raises voicing recognition macro-F1 from 0.922 for Adult PhonoQ to 0.972–0.987. Manner is more sensitive to alignment supervision, while place remains comparatively strong across systems.
Abner Hernandez, Tom\'as Arias Vergara, Andreas Maier et al.
In briefSmoothOperator: Improving Open-Set Recognition by Adjusting Label Smoothing per Sample
SummaryThis paper studies spherical representation learning objectives used in open set recognition, which aim to classify known classes while rejecting unseen ones. It shows that label smoothing acts as an alignment control but applies the same coefficient to every sample. The authors propose SmoothOperator, a plug-in method that sets each sample's smoothing coefficient based on its prominence, meaning how clearly its class stands out from the strongest competing class. Integrated into four existing methods with minimal training overhead, it generally improves results on the Semantic Shift Benchmark, with gains of up to 4.7% in AUROC, OSCR, and closed-set accuracy.}
Thiru Thillai Nadarasar Bahavan, Yu Xia, Sachith Seneviratne et al.
In briefLearning Multiple Timescales for Goal-Conditioned Reinforcement Learning
SummaryThis paper addresses the difficulty offline goal-conditioned reinforcement learning has with long-horizon tasks, where discounting erases value differences between distant states. It introduces Generalized Implicit Temporal Abstraction (GITA), which conditions one value function on the abstraction step k and trains a single policy by aggregating advantage-weighted supervision across multiple k values. On OGBench, GITA raises average success rate by 25 points over HIQL and by 7 points over the fixed-k method OTA.
Pedro Robles Dutenhefner, Dikshant Shehmar, Wagner Meira Jr. et al.
In briefGeometric Similarity in VLM Low-Level Vision Representations: Checking Whether Tasks and Architectures Share Representation Structure
SummaryThis paper studies how vision-language models (VLMs) organize hidden-layer representations for low-level vision tasks, comparing autoregressive models and diffusion transformers. The authors propose GeoSim, a four-level framework that analyzes representations across 24 tasks using global similarity, local geometry, sparse feature decomposition, and topological verification. The results reveal organizing principles of low-level visual representations and show their limits in agreement across tasks and models.
In briefMintFlow: Minimal Intervention for Constrained Flow Matching Samples
SummaryThis paper introduces MintFlow, a training-free sampling framework for flow matching models that enforces constraints by minimally perturbing an intermediate state of the pretrained flow trajectory. The pretrained flow field is kept unchanged, and an adjoint formulation yields a closed-form perturbation without iterative optimization. The method also adaptively chooses the intervention time to balance perturbation size against its amplification. Across generative vision and physical system modeling tasks, the authors report competitive constraint satisfaction with better preservation of the pretrained distribution than prior constrained methods.
In briefTracking how word meaning shifts as sentences unfold in language models
SummaryThe paper introduces token-wise incremental trajectories, recomputing each token's contextual embedding as words are added to a sentence to show how representations evolve. Using garden-path sentences as a test case, the trajectories reproduce known processing disruptions and reliably distinguish them from disambiguated controls. Ambiguity-related information was found in both the sentence-level CLS representation and ordinary vocabulary tokens.
In briefSHARPO: Segment-level credit assignment for training AI agents with reinforcement learning
SummaryThis paper introduces SHARPO, a method for training language model agents that carry out multi-step tasks. It refines GRPO by computing teacher-student log-probability gaps within each environment-facing segment and using them to set a bounded multiplier on the advantage, so credit can differ across segments. Using Qwen2.5-7B-Instruct, the authors report that SHARPO outperforms baselines including GRPO, SDAR, RLSD, and StepOPSD on the ALFWorld and WebShop benchmarks.
In briefPLDiv: Measuring dataset diversity and semantic collapse using persistence landscapes
SummaryThis paper proposes PLDiv, a diversity metric based on persistence landscapes that describes the structural organization of a dataset from its geometric representation. It summarizes how data neighborhoods change as nearby samples gradually connect, yielding an interpretable scalar measure of structural diversity without task-specific supervision. The authors report that PLDiv performs competitively against baselines on text generations, image embeddings, instruction datasets, and reasoning trajectories, and aligns strongly with diversity labels during semantic collapse.
In briefTrueMuse: A Benchmark for Testing Which Training Data Influenced Text-to-Music Models
SummaryTrueMuse is a controlled dataset and benchmark for data attribution in text-to-music models. Three diffusion-based models were fine-tuned on curated samples whose inclusion is known, providing reliable ground truth for evaluation. Existing black-box attribution methods were tested across four settings, and the results show that attribution remains challenging, with performance varying substantially across settings.
In briefKepler: Auditable World Models for ARC-AGI-3
SummaryThis paper presents Kepler, an open-source harness that represents hypotheses about environment rules as executable world models and checks them with retrospective and conditional prediction tests. Under one frozen Claude Opus 5 configuration, it reports a server-verified 100.00 RHAE across all 25 public games. The authors also describe three evaluation failures and argue that public-set scores alone have limited discriminative value, recommending first-attempt, cost, and verification-aware reporting.
In briefVERITYGATE: A Four-Step Checker for Whether AI Explanations Stick to Structured Evidence
SummaryThe paper introduces VERITYGATE, a four-gate checker that tests whether LLM explanations match declared evidence IDs, entities, numbers and claim types in a structured system. It checks a fixed schema rather than every fact in the prose. Tests on 900 instances with three models found many claims failing before repair, and one repair pass raised claim survival for some models with mixed effects on output volume. The code and data are released.
In briefAnyJev Technical Report: A Training-Free Method to Read Probabilistic Choices from One Language Model Pass
SummaryThis report presents AnyJev, which reads a probabilistic choice among fixed options from a single prefill of a pretrained instruction-tuned language model, without gradient steps or parameter changes. It corrects label bias by dividing out a prior estimated from unlabelled inputs, and position bias by averaging log-probabilities over cyclic rotations of the option list. A stopping rule reduces the number of prefills needed. On two 20-option tasks, the authors report lower order-flip rates and higher accuracy on most models, and the code is open source.
In briefWhether a language model's stated confidence matches its internal probabilities
SummaryThis paper examines whether a large language model's internal sampling probabilities align with the confidence it verbalizes in words or numbers. The authors intervene on uncertainty sources in training and in-context data to see how both readouts change. They find that both internal and verbalized probabilities are affected by distributional and asserted uncertainty in the training data. The two readouts are aligned more than independent tracking of the same sources would predict, suggesting verbalized probabilities can probe a model's internal distribution.
In briefAlign Then Reason: A Lip-Sync Judge for Checking Dubbed Lines
SummaryThis paper proposes Align Then Reason (ATR), a method for judging whether a candidate dubbing line matches a speaker's lip movements in content and timing, using only silent video and text. ATR first aligns frame-level lip features with the candidate line's phonetic units, then has a language model reason over that alignment. On a seven-language benchmark, it improves mean AUC over Qwen3.5 SFT baselines by 59.4%, 50.2%, and 50.8% for 2B, 4B, and 9B reasoners.
In briefSparse Components and Dimensions in Transformer Refusal Mechanisms
SummaryThis paper examines how activation steering changes refusal behavior in large language models by intervening on internal activations. Across four open-weight models, the authors break refusal steering into attention and MLP components and find that a sparse subset reproduces the full effect. These mechanisms comprise 28–48% of upstream components and retain 88–101% of steering effectiveness. Within them, about 50% of residual stream dimensions carry most of the signal.
Vincent Siu, Glenn Grant-Richards, Vlad Pavlovich et al.
In briefRotate the Loop: Blockwise Conformal Injections for Looped Transformers
SummaryLooped transformers reuse one shared block across many recurrent steps, allowing depth to be allocated at runtime. The paper argues that whether the loop learns to track state depends on the geometry of the linear map applied to the residual between steps. It proposes blockwise conformal injections, which split the residual into coordinate-aligned groups and rotate within each group so that each has its own timescale. Experiments report gains on state tracking, measured by last-token accuracy when the rest of the prefix is already correct, and on language modelling that requires retrieving an earlier-written signal at the end of a sequence.
Alexey Zemtsov, Albina Klepach, Vadim Lomshakov et al.
In briefOn-the-fly Weight Generation: A Proof of Concept for Generating Model Parameters from Examples
SummaryThe paper asks whether specialist models can be generated directly from a few demonstrations rather than trained for each task. Using ARC-1D as a controlled testbed, it shows that individual transformations can be represented by tiny specialist models and that a hypernetwork can generate their parameters from context. The generated specialists show partial compositional generalisation and generalisation to transformations not seen during training.
In briefPI-AMFM: Learning to Split Biomedical Signals into Variable Numbers of Oscillating Parts
SummaryThis paper proposes PI-AMFM, a permutation-invariant neural network for decomposing nonstationary physiological signals into AM-FM modes whose number can vary. The method combines a multiscale temporal encoder, a Mamba backbone, and component-presence estimation, trained with Hungarian matching. The authors report lower decomposition, frequency, reconstruction, and mode-count errors than compared methods on synthetic signals. Trained only on synthetic data, it also recovered cardiac and respiratory dynamics from photoplethysmographic recordings.
In briefMeasuring How Ambulance Handover Delays Affect Victoria's Emergency Services
SummaryThis paper analyses about 2.85 million ambulance attendances in Victoria, Australia, from January 2020 to March 2024, focusing on ramping, the delay before patients are handed over at hospitals. Cumulative lost ambulance hours reached about 1.49 million, with losses peaking in 2022 while transported demand fell. Handover time varied little with patient acuity but increased with the number of ambulances arriving at the same hospital in the preceding hour. The authors argue these findings support hospital-state aware ambulance routing.
Ayesha Tanveer, Khandakar Ahmed, Assefa Teshome et al.
In briefTabJoinBench: A Benchmark for Finding Joinable Tables
SummaryThis paper introduces TabJoinBench, a benchmark for evaluating methods that find tables in data lakes which can be joined to a query table. It covers semantic, relational, and hybrid scenarios, building query-candidate pairs and applying composable perturbations that change structure, representation, and semantics while keeping reliable ground truth. The authors evaluate set-based, feature-based, and learned methods along with language-model embedding baselines, and release the datasets, annotations, and generation pipeline publicly.
In briefSelecting Synthetic Training Data by Target Coverage and Training Effects (TATC)
SummaryThe paper studies when synthetic data actually help LLM training, since more synthetic data does not always yield better models and errors can offset gains. It develops a linear theory describing this tradeoff and introduces TATC, a selection method that picks candidates beneficial to the target task and expands coverage of uncovered target-relevant directions. On mathematical reasoning, TATC outperforms other synthetic-data selection methods on GSM8K across selection budgets.
In briefSurvey of how pretrained video generation models are post-trained and aligned to human intent
SummaryThis paper is the first comprehensive survey of post-training and alignment for video generation models. It notes that pretrained video models often fail to reliably follow human intent, keep motion temporally coherent, or meet physical and safety constraints. The authors organize existing methods into four categories: supervised fine-tuning, self-training and distillation, preference- and reward-based methods, and inference-time methods. It also reviews datasets, benchmarks, and open challenges such as reward design and long-horizon consistency.
In briefCan We Trust This Evaluation? Instance-Level Reliability of LLM-as-a-Judge
SummaryThe paper notes that LLM-as-a-judge evaluations usually aggregate several judges' decisions by majority vote, which does not ensure each individual item's evaluation is reliable. The authors propose SCUR, a label-free method that estimates instance-level reliability from the judge panel's behavior, combining panel uncertainty, measurement cross-entropy, and perturbation sensitivity into a risk score. The approach is tested on five real-world benchmarks. It is reported to better detect high-consensus risk items that other methods often miss.
In briefCutting Token Costs: Simple Heuristics for Valid, Cheaper Multimodal LLM Video Annotation
SummaryThis paper evaluates cost-saving heuristics for annotating short-form videos with vision-language models, such as sampling frames, using image grids, or relying on a single modality. The authors test these methods on sentiment and topic classification tasks, measuring accuracy, validity of downstream conclusions, and token cost. They find that the most accurate configuration can still yield wrong conclusions, and text alone can perform strongly. A single 2×8 image grid achieves agreement close to full-video understanding at about 15% of the token cost, and the authors derive guidelines for cost-aware annotation.
In briefSpatially gated diffusion for editing chest X-rays while keeping unrelated areas intact
SummaryThe paper studies a latent diffusion model for editing chest radiographs, mixing two generation trajectories with a learned gate and compositing the result with the source image through an image-space mask to preserve unrelated content. On 2,400 MIMIC-derived requests, 93.5% of outputs met the joint target, preservation, quality, and coverage rubric, with 97.4% target completion. Joint validity exceeded a composition-only control by 1.7 percentage points.
In briefCan large language models extract data from eye-exam PDF reports?
SummaryThis study tested two GPT-4o-assisted methods for pulling structured data from 20 deidentified ophthalmic diagnostic PDF reports across four report types. Both methods reached high value accuracy when compared with a manually reconciled ground truth. Extraction time fell from about 56.5 seconds per report for manual review to about 5 seconds, a roughly 92% reduction. The authors suggest that the complementary strengths of the two approaches justify further evaluation of hybrid, validation-aware workflows.
In briefNeural network and Euler-based inference for SDEs driven by weighted sub-fractional Brownian motion
SummaryThis paper studies estimating the drift, diffusion, and noise covariance of stochastic differential equations driven by a Gaussian process from discrete observations. The method reconstructs driving increments via the Euler approximation and builds a trajectory likelihood from their joint density, with neural and radial-basis models for the coefficients and temporal weight. The covariance exponent and diffusion scale are estimated through a likelihood profile. The approach is compared with two neural alternatives on the same simulated trajectories across twenty coefficient settings.
In briefEnterprise Representation Simplification: Reducing Complexity for Enterprise AI
SummaryThis paper proposes Enterprise Representation Simplification (ERS), which removes unnecessary representational complexity while preserving required information in a defined scope. It introduces Enterprise Representation Complexity (ERC), a representation-neutral model measured across four dimensions: Objects, Interactions, Behaviors, and Supporting Sources. The authors discuss lifecycle costs of representational structures and how reducing task-level ERC may ease AI reasoning, citing Text-to-SQL research as evidence. The paper states that ERC is not a universal complexity, performance, or cost metric.
In briefVTV-FM: A flow matching method that learns curved generative paths using minimum-acceleration velocity estimates
SummaryThe paper proposes VTV-FM, a second-order flow matching framework that learns generative transport from a simple source distribution to data. It derives the terminal velocity missing from static datasets by minimizing acceleration energy, yielding a closed-form closure that also defines the optimal transport pairing cost and training targets. The authors report that experiments on low-dimensional data, PDE-governed physical fields, and CIFAR-10 show improved transport geometry and generation quality over first-order and higher-order flow matching baselines.
In briefPaper on why selecting synthetic training data needs group-level signals, not just individual ones.
SummaryThe paper argues that curating synthetic data for LLM training based only on each sample's individual utility is insufficient, because samples interact when used together. Across pre-training and post-training settings, it finds that group-level signals distinguish datasets that individual-level signals cannot, and that only curation methods using them improve over baseline as pipelines become more synthetic. The authors also propose a cheap diagnostic that identifies which synthetic data groups most need group-level estimation.
In briefA method for learning goal-reaching quasimetric geometry using finite-horizon reachability
SummaryThis paper proposes ReQRL, a goal-conditioned reinforcement learning method that constrains the critic's value gradients using finite-horizon reachability. Building on state-constrained optimal control, it separates dynamical reachability from boundary geometry and estimates both from data. The authors report that on the OGBench benchmark, the method outperforms or matches existing quasimetric approaches and other offline GCRL methods.
In briefLocalizing Transfer Between Memorization Tasks: How Pre-Training on One Random Mapping Helps Learning Another
SummaryThis paper studies transfer learning between memorization tasks that map random inputs to random outputs. The authors report two transfer patterns: equivalent transfer, where each extra pre-training epoch saves roughly one fine-tuning epoch, and non-equivalent transfer, where a mismatched pre-training task can be more efficient than training directly on the downstream task. Ablation experiments split the transfer into a trivial magnitude-driven effect in the last layer and a non-trivial structure-driven effect partly tied to covariance in other layers.
In briefGrafting beliefs onto post-trained models: a faster way to test pre-training interventions
SummaryThe paper proposes grafting, which trains a synthetic document fine-tuning (SDF) adapter on a pre-trained checkpoint and then adds the learned weight update to a post-trained model. The authors test it by installing false facts, training misaligned model organisms, and applying a constitutional mid-training intervention across model families up to 284B parameters. They report that grafting installs the target belief about as strongly as SDF applied directly to the post-trained model, while reducing reality drift and loss of preference coherence by more than half on average.
Peter Nutter, Dani Roytburg, Cl\'ement Dumas et al.
In briefScalable Multi-Task Inverse Reinforcement Learning: Learning Transferable Rewards From Pooled Agent Data
SummaryThis paper proposes a multi-task inverse reinforcement learning method that pools data from multiple agents with different rewards in the same environment, under a low-rank assumption. The approach relaxes state coverage requirements and scales planning with rank rather than the number of tasks. The authors provide finite sample guarantees on reward recovery and policy learning in new environments. Experiments report robustness to limited coverage and lower transfer regret than baselines.
In briefVideo Evidence Indexing: Learning Where to Look in Long Videos Under a Visual Token Budget
SummaryLong-video question answering is constrained by the high cost of visual tokens and fixed VLM context widths. The paper proposes Video Evidence Indexing (VEI), which uses a dense low-resolution Video Preview for global coverage and builds a compact high-resolution Evidence Set for final reasoning. It applies privileged self-distillation, where an answer-aware teacher guides the student's indexing traces. Experiments report improved accuracy under limited visual budgets, and self-distillation further improves QA accuracy and temporal evidence localization.
In briefLearning to set electricity prices that shape demand using contextual signals
SummaryThis paper proposes a neural-network algorithm for contextual energy pricing that models pricing as a Stackelberg game with a mean-field solution representation. It learns constrained mappings from contextual features, such as weather forecasts, sunrise and sunset times, and day-of-week patterns, to feasible time-varying price signals. The approach is validated by simulating energy grids in several US cities. The authors report that incorporating contextual information can considerably increase the value of demand response programs.
In briefWider networks let greedy layer-wise training match end-to-end backpropagation in self-supervised learning.
SummaryThis paper studies how network width and depth affect convolutional networks trained with greedy layer-wise versus end-to-end self-supervised learning. The authors find that as networks get wider, the advantage of end-to-end backpropagation shrinks, and in shallow, very wide networks greedy layer-wise training can even perform better. Analysis indicates that these greedy-trained networks form more favorable representational geometry, a possible explanation for their improved performance.
In briefSignal-noise factorization: separating nuisance variation into removable subspaces
SummaryThis paper proposes regularizers that reinforce signal-noise factorization (SNF) and signal-signal factorization (SSF) in neural network representations during training. On CIFAR-100, enhancing SNF improved performance over an L2-regularized baseline, while enhancing SSF did not. Gains were larger on BloodMNIST with MedMNIST-C corruptions at five severity levels. Projecting out estimated noise directions further raised accuracy, suggesting that isolating nuisance variables matters more than preserving factorized categorical representations.
In briefWhat verifiable rewards teach video-language models about time
SummaryThe paper fine-tunes four open video-language models with reinforcement learning from verifiable rewards, comparing synthetic data, self-supervised real videos, and a mixture. Training only on synthetic questions markedly degrades accuracy on real video, and adding real verifiable questions removes that loss. Much of the gain is non-visual, and no training recipe produces grounding in temporal event order.
In briefWhat Builds the Scene? Luminance Dominates Geometry Formation in 3D Gaussian Splatting
SummaryThe paper studies how luminance and chroma each contribute to geometry formation in standard 3D Gaussian Splatting. The authors train under different channel supervision, freeze non-appearance parameters, and compare held-out reconstruction across eleven scenes. Geometry learned from luminance alone scores 0.085 dB below RGB-trained geometry on average, while chroma-only supervision yields geometry 3.9-5.5 dB worse than luminance-only supervision.
In briefStabilityArc: Turning Protein Sequence Embeddings into Transferable Stability Maps
SummaryThis paper presents StabilityArc, a shared decoder that maps frozen ESMC-600M residue representations to a matrix of mutation effects on protein stability, with a contact-aware term for simultaneous substitutions. In 66 strict leave-one-protein-out evaluations covering 134,794 ProteinGym variants, it reaches a Spearman correlation of 0.7134, above the strongest zero-shot baseline, ProSST-2048 (0.6526). Used as a prior for Kermut, it achieves 0.8280 across three supervised split schemes, improving on Kermut's reported 0.8167.
Aaron L. Feller, Andrew D. Ellington, Claus O. Wilke
In briefQ-MINO: a minimal-norm optimizer for training with quantization
SummaryQuantization-aware training often uses the straight-through estimator, whose surrogate gradients can mismatch the quantized objective and cause noisy updates and oscillations at very low bit-widths. This paper proposes Q-MINO, which combines gradient consensus, state-drift regularization and an alignment constraint to build stabilized minimum-norm update directions from recent optimization states. The constrained subproblem is solved with a warm-started Frank-Wolfe method with a feasible fallback initialization. The authors show asymptotic neighborhood convergence using a stochastic Lyapunov Kurdyka-Łojasiewicz framework and report numerical experiments at various quantization levels.
In briefPersonalized image generation using reasoning and reflection on user history
SummaryExisting personalized image generation mostly relies on curated visual examples rather than a user's broader history. The paper introduces the first unified benchmark for generating images from user histories, with two tasks, Personalized Scene Generation and Personalized Creative Generation, and a multi-axis evaluation protocol. It also proposes PEARL, which pairs a multimodal reasoner with a frozen image generator in an interleaved reason-reflect loop. The authors report that PEARL outperforms strong baselines, with an average improvement of 15% across personalization metrics.
In briefReMILP: Self-supervised learning of MILP representations via equivalent reformulations
SummaryThis paper studies mixed-integer linear programs (MILPs) and uses equivalent reformulations of the same problem as a self-supervised learning signal. The authors separate reformulations that leave variables unchanged from those that transform variables predictably, and propose ReMILP, which jointly trains a graph neural network and a hypernetwork to predict how variable embeddings change. The authors report that the learned representations show the intended invariance and equivariance on unseen problem classes, without solver-derived labels. They also report that the frozen representations are useful for predicting binary solutions, constraint activity and integrality gap, and provide a good initialization for fine-tuning.
Ousema Bouaneni, Mathis Le Bail, Cl\'ement Elliker et al.
In briefReward as Observation: Learning reward-based policies that adapt quickly to new environments.
SummaryThis paper proposes a policy conditioned only on rewards and actions, allowing zero-shot transfer to environments with completely different observations. The method is trained in Pointmass, Cartpole, and 2D Car Racing, and transferred without further training to different color palettes, 3D rendering, and Stretch robot navigation in Habitat-Sim. The authors also show that the reward-based policy can guide training of an observation-based policy in the target environment.
In briefBenchmarking generative models for near-surface weather analysis on real station data
SummaryThis paper presents the first controlled benchmark of generative data assimilation for near-surface weather analysis, using real observations from 11,849 NOAA MADIS stations across the contiguous United States. The dataset, observation operator, and architecture are held fixed while comparing diffusion versus flow matching, pixel versus latent-space designs, and several conditioning strategies against a classical 3D-Var baseline. The best generative methods outperform 3D-Var in RMSE improvement over ERA5 (35.7% vs. 33.3%), and full-gradient guidance works best among the conditioning strategies. The advantages widen when stations are sparse.
Ruizhe Huang, Qidong Yang, Jonathan Giezendanner et al.
In briefCF-JEPA: Making Visual World Models Robust to Distractions by Splitting Controllable and Uncontrollable Latents
SummaryThe paper introduces CF-JEPA, a JEPA-style world model that splits its latent space into controllable and uncontrollable subspaces. Distractor information is placed in the uncontrollable part, while control-relevant information is used for the task. The model performs comparably to baselines on 2D and 3D control tasks under nominal conditions and performs better under distraction, where it is the only model that avoids latent collapse. The authors also validate it on a simulated robot task with distractions.
In briefReason in Style: Finding and Steering Language Model Writing Styles
SummaryThe paper presents an unsupervised algorithm that separates content and style representations in language model outputs, validated on math questions. Applied to over 100K verified traces from nine teacher models, it finds six recurring but imbalanced styles. Student models fine-tuned with importance weighting to follow these styles improve Pass@k over standard fine-tuning across six math reasoning benchmarks, and style choice affects correctness.
WhyUnsupervised style discovery and control improves Pass@k across six math benchmarks over standard fine-tuning, with fairly solid evidence and clear novelty.
In briefJEPA-TTT: Persistent Test-Time Training of Latent World Models for Planning Under Shifting Dynamics
SummaryThis paper proposes JEPA-TTT, which keeps updating the latent dynamics predictor of a pretrained action-conditioned world model during test time, while the visual encoder and reward head stay fixed. It uses dense replay with a growing buffer and is evaluated on eight dynamics shifts across four continuous-control environments. After 500 test-time episodes, the authors report an average 83% reduction in autoregressive latent prediction error and a 153% improvement in planning performance over the frozen JEPA world model.
SummaryA Condorcet domain is a set of rankings where every election using only votes from the domain yields a transitive majority relation. This paper studies outer diversity, a measure of the expected swap distance from a random vote to the closest ranking in the domain. The authors numerically analyze maximal Condorcet domains with few candidates and establish asymptotic behavior for several special domains, mostly through theoretical results.
Piotr Faliszewski, Jan Jabrocki, Mateusz S{\l}uszniak et al.
In briefSequential structured Tucker compression of LLM attention that needs no fine-tuning
SummaryThe paper proposes FTC, a post-training method that compresses LLM attention layers by jointly exploiting the Q/K/V head structure under a fixed storage budget. The output projection is handled separately to account for representation changes caused by earlier compression. FTC needs no fine-tuning or gradient-based recovery. Across seven decoder-only LLMs from 6B to 32B parameters, it achieves the lowest WikiText-2 perplexity at every tested keep ratio on five GQA models, with the largest gains under aggressive compression.
In briefRobust Nash Alignment for Uncertain Preferences
SummaryThis paper addresses the fragility of preference-based alignment methods when pairwise preferences are noisy, heterogeneous, or shift after deployment. It proposes Robust Nash Alignment, a game-theoretic framework that seeks policies with high worst-case win rates against an adversarial competitor and preference kernels in an ambiguity set. The authors develop a four-player primal-dual proxy game solved by a single-loop optimistic mirror descent-ascent algorithm, prove an O(1/√T) average-iteration convergence rate, and report experiments on tabular games and LLM alignment showing gains over nominal baselines.
In briefParameter-free regret bounds over intervals for online optimization with heavy-tailed noise
SummaryThis paper studies online convex optimization where each round yields one unbiased stochastic subgradient, and the conditional p-th noise moment (1<p≤2) is unknown. The authors present a learner that uses none of the problem parameters and proves an expected interval regret bound against comparator paths. The interval adaptation adds a logarithmic cost, and a change-of-measure lower bound is also provided under explicit conditions.
WhyGives parameter-free interval dynamic regret bounds under unknown heavy-tailed noise, plus a matching change-of-measure lower bound.
In briefWOMBAT: A whitebox benchmark for testing molecular GNN explainers
SummaryThis paper presents WOMBAT, a benchmark of 14 whitebox graph neural networks whose message-passing weights are set by hand to detect specific SMARTS motifs, so each model's decision rule is known. The authors validate these models on millions of PubChem molecules and use them to evaluate post-hoc explainers such as GNNExplainer, PGExplainer, and Integrated Gradients. They also construct a model that causes Integrated Gradients to spread attribution across the graph even though the model reliably detects the intended motif. The dataset, models, and evaluation code are released publicly.]
Dominik Matuszek, Bartosz Zieli\'nski, Tomasz Danel et al.
In briefReLiveGym: testing long-running AI agents on replayed weeks of real-world news and markets
SummaryThis paper introduces ReLiveGym, a diagnostic environment where LLM agents act sparsely over simulated weeks of chronologically replayed real-world news, market, and social-media streams. The authors test eight base language models and find that how agents decide when to act is an important design factor, with the best design varying by task and sometimes by model. They also evaluate whether continuous learning from hindsight feedback improves performance and addresses observed failure modes.
In briefBeyond Single-Shape Bases: A Pullback Geometry for Multimodal Data
SummaryThis paper proposes a Riemannian pullback geometry for data lying on mixtures of manifolds, defined using a latent Gaussian mixture with a metric built from responsibility-weighted component precisions. The method reduces to the existing single-Gaussian construction when only one component is present. It is implemented in a normalizing flow with adaptive mixture learning, and experiments on synthetic data, a multi-view image setting, and MNIST report reduced transport distortion and improved interpolation realism.
Honglei Brinkmann, Lucas Ng, Georgios Batzolis et al.
In briefStarting with parallel exploration boosts LLM-driven discovery
SummaryThis paper studies how large language models iteratively optimize objectives to discover algorithms, theorems, and other results. The authors build 12 'Modular' harnesses and test them on 5 tasks, finding that success is sensitive to harness design and often suffers from mode collapse. They propose an initialization step of parallel exploration before iterative optimization, which gives consistent gains across many harnesses and tasks.
In briefAnchorPrompt: Self-distilled soft prompts that make audio-language models more robust to distorted sound
SummaryLarge audio-language models are sensitive to noise, waveform corruption, and adversarial injections. The paper proposes AnchorPrompt, which keeps the model frozen and learns a single block of prompt vectors inserted at the decoder input, trained by self-distillation over diverse audio and text perturbations. Evaluated on three models and three benchmarks, it improves answer consistency in most conditions and reduces hallucinations under severe audio corruption.
In briefMeta-learning for quickly adapting multi-agent driving policies to new tasks
SummaryThis paper proposes a meta-multi-agent reinforcement learning (meta-MARL) framework that lets interactive policies in multi-agent systems adapt quickly to new tasks. The authors model these problems as Markov games and define a new solution concept, meta-NE, with sufficient conditions linking it to stationary points of a gradient-play algorithm. Experiments on autonomous-driving tasks show faster adaptation than pretrained MARL baselines.
Huiwen Yan, Kyriakos G. Vamvoudakis, and Mushuang Liu
In briefSkillSpec: Consensus-Gated Evolution and Structuring of LLM Agent Skills
SummaryThe paper proposes SkillSpec, a two-phase framework for evolving natural-language skills that help LLM agents reuse task knowledge without changing model weights. Updates are committed only when paired evaluations reach consensus on improvement, and the retained knowledge is then structured as a flat, graph, or hybrid representation based on signals from the optimization trajectory. Across six benchmarks and three target models, SkillSpec improves average success rate over SkillOpt by 6.89%.
Huancheng Chen, Xiaodi Sun, Zhaoqiong Huang et al.
In briefR-GroundBench: A Diagnostic Benchmark for R-Group Grounding in Markush Molecular Editing
SummaryThis paper introduces R-GroundBench, a diagnostic benchmark built from real patent Markush structures, which encode molecular families with variable R-group placeholders. It includes a multiple-choice VQA track with controlled difficulty and modality splits, plus an open-ended generation track. Models score above 90% on easy VQA but only 56–66% on hard VQA, and generation exact match is below 20% for most models. The authors conclude that current AI systems lack reliable grounding and execution for Markush editing.
In briefTRACE: A Reproducible Benchmark for Electricity Price Forecasting Using Official Operational Text
SummaryThe paper introduces TRACE, a reproducible benchmark of 7,300 zone-day instances from five zones in a major U.S. market, pairing prices with official operational text available at the forecast cutoff. The text is reconstructed at each cutoff to prevent post-cutoff information leakage. Semantic checks show the text aligns with central price movements and both tail risks, most strongly for upper-tail risk. Across time-series foundation models, it yields a median 7.4% reduction in upper-tail pinball loss, but a cross-day text-mismatch ablation reverses these gains, falling below the no-text baseline.
In briefHow Divergence Becomes Decision Flips: Total Variation vs. KL in Compressed Language Models
SummaryThis paper examines how the output distribution gap between a compressed language model and its dense original turns into changed decisions. Across 802 compressed or perturbed copies of 19 open models, the rate at which the top predicted token changes tracks total variation at a stable ratio. KL divergence maps to these flips less reliably, so the authors argue total variation better indicates how many decisions change.
In briefFedMAD: Federated learning that weights client updates by modulation alignment for remote sensing image classification
SummaryThis paper proposes FedMAD, a personalized federated learning framework for classifying remote sensing images without sharing local data. It separates globally shared representation parameters from client-specific adaptation parameters, and uses a modulation-aware directional aggregation strategy that adjusts each client's contribution based on the alignment of its local modulation updates. Experiments on BigEarthNet-S2 and EuroSAT show that FedMAD outperforms state-of-the-art federated learning methods under heterogeneous data distributions.
In briefSoundwich: Layered, Controllable Audio for Video Generation
SummarySoundwich is a training-free framework that adapts a frozen joint audio-video generation model to produce multiple synchronized audio stems for one video. Each stem has explicit control over its timing, and a shared scene representation keeps the stems coherent while preserving their separation. Experiments and human evaluations report better temporal control, source separation, and naturalness, and stems can be retimed, muted, replaced, or remixed independently.
Zhuo Ning, AmirHossein Naghi Razlighi, Sagi Polaczek et al.
In briefMaking Numerical Claim Verification More Robust
SummaryLarge language models are widely used to verify claims but struggle with numerical reasoning, where small changes in values can sharply reduce accuracy. The paper shows that frontier models have this weakness too, and that adversarial fine-tuning on numerically perturbed examples can mitigate it. Small Qwen3 models fine-tuned with parameter-efficient methods reach 98.7% accuracy on label-flipping perturbations, outperforming larger zero-shot and frontier systems, and the gains extend to unseen perturbation types and Spanish.
In briefLeto: fast in-place recovery of LLM training on hardware that still works
SummaryThis paper presents Leto, a fault-tolerant training system for large language models that recovers from failures on the same hardware without reset, repair, or replacement. It keeps the working model and reusable process state, and prepares the remaining state in a shadow trainer, using two-tier erasure protection and chunk-level transactional updates. On 6- and 72-GPU NVIDIA A100 clusters, Leto recovers 3.6 to 6.5 times faster than the best checkpointing baselines and raises productive training time by up to 13.7 percentage points. Simulation on a 131,072-GPU cluster shows over 95% productive training time.
WhyProposes in-place recovery on surviving GPUs with two-tier erasure coding, achieving 3.6–6.5× faster recovery on real clusters and simulated large-scale validation.
In briefSemanTok: Predictable Semantic Tokens for Efficient Autoregressive Video Generation
SummaryThis paper introduces SemanTok, a flexible-length video tokenizer that feeds frozen DINO features into its encoder and uses lightweight heads to reconstruct them from each token prefix alone. The authors report that a 201M SemanTok autoregressive model matches or beats a VideoFlexTok model 3.4 times its size. Short token prefixes are cheaper to predict and yield better generation fidelity, with pixel detail deferred to later tokens.
Mikhail Dereviannykh, Vikram Voleti, Simon Donne et al.
In briefGrand Canonical Generators: a generative sampler for systems with variable particle number
SummaryThis paper introduces Grand Canonical Generators (GCG), extending Boltzmann generators to the grand canonical ensemble where particle number can vary. It presents two designs: a variable-size model conditioned on chemical potential, and a factorization into a particle-number distribution and a canonical Boltzmann density. Tests on a Lennard-Jones fluid and methane adsorption in a zeolite show GCG reproduces grand canonical observables and can be corrected via self-normalized importance sampling and grand canonical Monte Carlo.
In briefOntology-Based, Reasoner-Verified Benchmark for Testing LLM Reasoning in Scientific AI
SummaryThis paper proposes a pipeline that automatically generates multiple-choice question benchmarks from OWL 2 ontologies, with correct answers grounded in the ontology. Wrong options are created by perturbing class definition axioms, and an OWL reasoner formally verifies that they are incorrect. The authors generated 112, 2,491, and 15,216 questions from three ontologies, and six LLMs scored 41.1% to 76.8% zero-shot, well above the 25% random baseline.
Nishtha N. Vaidya, Stephan Grimm, Thomas Hubauer et al.
In briefCurvature under attack in a compact Vision Transformer: how curvature and scale affect robustness
SummaryThis paper studies hZACH-ViT, a compact Vision Transformer with Euclidean, Poincare, and spherical prototype heads, tested on three MedMNIST datasets across five seeds. At curvature c=1, the Poincare head shows the lowest attack success rate but also the lowest clean MacroF1. Lowering Poincare curvature to c=0.1 improves clean MacroF1 yet raises strong attack success on OrganAMNIST from 89.7% to 99.3%. The authors conclude these results do not establish intrinsic hyperbolic robustness, but show that curvature, scale, and proximity to the Poincare boundary jointly shape recognition and adversarial behavior.
In briefBayesian fine-tuning makes language models act Bayesian, within the limits of their beliefs
SummaryThis paper examines whether language models can act as Bayesian decision makers on tasks requiring reasoning about hidden variables. The authors test models on a flight recommendation task across behavior, representations, and computations, comparing fine-tuning on outputs of a Bayesian model with standard fine-tuning on true answers. The Bayes-trained model behaves more Bayesian, encodes Bayes' rule quantities in its middle layers, and partly uses these beliefs for recommendations. The oracle-trained model differs in both the beliefs it holds and whether it turns them into recommendations.
Polina Tsvilodub, Andreas Waldis, Linlu Qiu et al.
In briefA method that condenses large pathology slide images into compact data while preserving the features needed for learning.
SummaryWhole-slide histology images are very large, often several gigabytes each, which makes learning expensive. The paper reformulates slide condensation as a distribution-matching problem and proposes NICER, an approximation framework using a nonparametric prior with slide-adaptive capacity. Across five histopathology datasets, NICER reports an average accuracy gain of 7.44% and better efficiency-accuracy trade-offs.
WhyReformulates WSI condensation as distribution matching with a nonparametric prior; validated on five datasets with a 7.44% average accuracy gain.
Duong M. Nguyen, Trong Nghia Hoang, Hang Thi Nguyen et al.
In briefUsing vision-language models to judge AR visual quality and automatically adjust virtual content
SummaryThis paper proposes an automated framework that uses vision-language models (VLMs) to evaluate and refine the visual quality of augmented reality (AR) scenes. The authors introduce RateAR, a benchmark of AR images and videos, and evaluate eleven commercial VLMs, whose predictions reach Spearman correlations of up to 0.8695 with human ratings. A 21-participant user study found that more than 90% of participants felt the automatic adjustment system improved the placement and size coherence of virtual content.
Elias Rotondo (1), Lin Duan (1), Yanming Xiu (1) et al.
In briefLearning reusable robot skills that transfer to new environment layouts
SummaryThe paper studies unsupervised skill discovery from reward-free datasets to pretrain general-purpose policies. The authors propose an objective for learning action-aware temporal representations that satisfy functional equivariance while preserving local temporal structure. They then use bisimulation so that skill behavior depends only on the state features that directly affect its execution. Empirical results indicate that skills learned in one environment can solve downstream tasks in different layouts, showing strong out-of-distribution generalization.
Mohammad Amin Abbasfar, Farbod Azimmohseni, Mohammad Hossein Rohban
In briefLabBook: Using Experiment History to Make LLM-Driven Discovery More Efficient
SummaryThe paper presents a single-agent discovery framework built around LabBook, a memory the agent maintains to retrieve relevant evidence from a full experiment log and to guide generating new programs. On 49 Frontier-CS problems, it improved the observed quality-cost trade-off over the evaluated evolutionary baselines with two backbone models. It also stayed competitive on nine additional mathematical, systems, and heuristic-design tasks.
In briefClosing the Loop: Practical Training Recipes for Looped Language Models
SummaryThis paper presents practical training methods for looped language models, which repeatedly apply a shared block of layers to increase effective depth. The authors reduce the from-scratch training budget from 7.7T tokens (Ouro) to 310B tokens while retaining strong reasoning performance. In controlled comparisons, a 1.4B looped model outperforms a parameter-matched dense model on all 12 evaluated benchmarks, and the authors also propose a minimal method for converting pretrained dense models into looped ones.
Andrei Marchenko, Viacheslav Bezrukov, Oleg Kashurin et al.
In briefORBIT-FMIB: Tracking How ESM-2 Retains Higher-Order Gene Interaction Information
SummaryThis paper introduces ORBIT-FMIB, a diagnostic tool for checking whether the protein model ESM-2 keeps higher-order epistatic information (effects from interactions among several mutations), tested on the GB1 fitness landscape. An initial run suggested weaker retention of higher-order information, but an independent replication shrank the difference and its sign was unstable. The authors therefore conclude that the directional question remains unresolved.
Maryam Rahimimovassagh, Ivan Garibay, Niloofar Yousefi
In briefA Simple Logic for Norm-Guided Decisions Under Uncertain Beliefs
SummaryThis paper introduces a simple doxastic deontic logic for autonomous systems that must decide when norms conflict and information is incomplete or unreliable. It combines a classically reducible fragment of Chellas' Minimal Deontic Logic, extended with explicit conditional norms, with multi-agent KD45, so norms can depend on agents' beliefs. The authors define the Doxastic Norm Compliance Optimization Problem and distinguish subjective optimization from objective optimization. They give conditions under which the two coincide and show that optimal decisions can then be reduced to weighted partial MaxSAT in polynomial time.
In briefVisionQ: A Benchmark for Vision-Language Models Judging Qualitative Computer Vision Results
SummaryThis paper introduces VisionQ, a benchmark built from peer-reviewed CVPR and ICCV comparison figures, testing whether vision-language models can pick the better output according to a named visual criterion. It includes 1,409 papers, 3,911 hand-annotated data points, a 51-leaf taxonomy of visual criteria, and a fine-tuned judge model called VisionQ-Judge. Evaluating 20 models, the strongest reaches only 63.1% accuracy against a 32.2% chance baseline, and reliability varies sharply across criteria.</summary>
WhyIntroduces criterion-conditioned VLM judging benchmark, grounded in 1,409 CVPR/ICCV papers and 3,911 annotations, evaluating 20 models with per-criterion accuracy.
Vu Dinh Xuan, Duc-Hai Nguyen, Minh-Dung Dao et al.
In briefAnalysis of Quantized and Efficiently Adapted Protein Language Models
SummaryThis paper tests 4-bit quantization and QLoRA fine-tuning on several protein language models for prediction and sequence generation. Many model-task pairs kept more than 90% of full fine-tuning performance, and GPU memory savings approached 90% for the largest models. The authors suggest QLoRA as a first-pass option under memory limits and recommend distributional checks for generative models.
Ilan Yaniv Zeisler, Sebastian Clancy, Pouriya Bayat et al.
In briefPhysicsMate: A Bengali Secondary Physics Question-Answering Benchmark Aligned to the Curriculum, with Small-Model Adaptation
SummaryThe paper addresses the lack of curriculum-aligned benchmarks for Bengali secondary physics question-solving. It introduces PhysicsMate, with 1,834 question-answer pairs built from the NCTB Grade 9-10 physics syllabus and grounded in a knowledge graph of 1,760 nodes and 2,600 edges. Low-rank adaptation of 0.6B, 1.7B, and 4B models raised closed-book accuracy by 5.5, 15.0, and 23.3 percentage points respectively. Structured curricular knowledge such as physical quantities and named laws benefited most, while loosely specified entity-level knowledge benefited least. The adapted 4B model was quantized into a small offline binary for local use.
In briefExploring More, Reasoning Better: A stepwise risk-sensitive training method for diffusion language models
SummaryThis paper studies reinforcement learning for diffusion large language models, which generate text by denoising sequences. The proposed StepRS-GRPO varies the risk coefficient of the group-advantage transformation across denoising states while keeping the underlying trainer. Across several dLLM backbones and math reasoning benchmarks, it improves pass@1 accuracy and pass@k coverage over centered GRPO while increasing answer diversity.
In briefLingtai: What Concept Geometry Reveals and Doesn't About LLM Inference
SummaryThe paper introduces Lingtai, a training-free method that projects a language model's hidden states onto named concept anchors at each generation step, yielding a per-step concept signal. On code generation and grade-school math, this signal is associated with predictive uncertainty, but the authors find no evidence it provides a stable correctness coordinate, so correctness is treated as externally supplied. They also report execution-specific trajectory identity and a 0.7-1.6% per-token decoding overhead.
In briefWhen More Data Is Not Enough: Context Sufficiency in Generative AI Personalization
SummaryThe paper proposes a theory of context sufficiency, arguing that for generative AI personalization, how relevant situational context is to a customer's current intent matters more than its volume. It identifies four states (insufficiency, sufficiency, saturation, interference) and introduces the Context-Sufficiency Frontier to locate the minimal relevant set. In a full-factorial experiment with a generative recommender at a large home-furnishing retailer, relevant context improved appropriateness, while irrelevant context reduced it and destabilized retrieval.
In briefAgent Evaluation Reliability: Why More Tasks Won't Always Fix Agent Leaderboards
SummaryThe paper proposes a Bayesian variance-decomposition framework for sparse, imbalanced agent leaderboards and applies it to 22 benchmarks from the Holistic Agent Leaderboard and Harbor Index. It finds that reliability depends on the measurement goal: fixed model-scaffold rankings are reliable, while rankings of underlying models are much less so. It also reports that adding similar tasks yields limited gains, whereas pooling diverse benchmarks raises projected cross-task reliability from 0.44 to 0.75 at the same task budget.
In briefRuleEvolve: Self-evolving coding rules that improve AI coding agents automatically
SummaryThis paper observes that AI coding agents depend heavily on their coding rules, which are usually hand-written and fixed. The authors propose RuleEvolve, a framework that keeps a pool of candidate rules and iteratively improves them using an LLM-based mutator to create variants and a judge to keep the best ones. Tested across two coding-agent frameworks, four backbone LLMs, and three benchmarks, it reportedly outperforms manual rule design and existing prompt optimization baselines in functional correctness, code length, and/or generation cost.
In briefTesting whether quantum kernel methods stay reliable for detecting fake speech with little training data across different datasets
SummaryThe study compares a quantum support vector machine, a classical SVM, and a small neural network for detecting synthetic speech, using only 200 training samples and four reduced features. Under a severe shift from ASVspoof 2019 to ADD 2023, the neural network falls to near-random performance while the quantum kernel method keeps meaningful discrimination. However, this advantage does not hold consistently across other transfer directions. The authors frame the results as a characterization of inductive bias under distribution shift, not evidence of quantum advantage.
In briefIncident-Arena: a benchmark testing whether AI agents can handle production outages reliably
SummaryThis paper introduces Incident-Arena, a benchmark of 20 human-built tasks for evaluating AI agents on production incident response, grounded in real open source applications. Each task deploys an app to an ephemeral Kubernetes cluster, injects a fault, and applies a sustained load. The authors also propose functional verifiers that go beyond static checks, and report that frontier models score below 64.3%, with failures in diagnosis, incomplete repairs, and unsafe regressions.
In briefGroup Symmetry in Word Statistics Shapes Learned Embedding Geometry
SummaryLanguage models often learn circular and saddle-shaped geometries for concepts like months and weekdays. This paper proves that when a word family's co-occurrence statistics are invariant under a group G, the learned embeddings consist of matrix elements of G's irreducible representations, with cyclic groups yielding circular Fourier structure. The authors verify this in three settings: the cyclic group Z12 for months, a dihedral group acting on major and minor triads, and a spherical representation of celestial objects in LLMs.
In briefLearning Linear Systems with Heavy-Tailed Noise from One Trajectory
SummaryThis paper analyzes how accurately least-squares can estimate vector autoregressive models of exponentially stable systems with heavy-tailed noise, using a single observed trajectory. Under i.i.d. noise, bounded noise covariance, and persistent excitation, it derives non-asymptotic error bounds for bounded p-th moments with p greater than 2. It also presents a unified analysis approach applied to sub-exponential and sub-Gaussian noise, and shows the dimension factor is independent of model order for autoregressive models with exogenous inputs.
In briefCompMat-Bench: Testing AI Agents on Computational Materials Science Tasks
SummaryThis paper introduces CompMat-Bench, a benchmark of 94 tasks taken from published computational materials studies, where AI agents complete one step toward each study's goal. The research steps were reproduced in advance so expensive simulations are not repeated during evaluation, and answers are graded with fixed rules rather than an LLM judge. With full guidance on single tasks, agents based on three LLMs achieved pass rates of 66.0% to 90.4%. Most failures were attributed to scientific errors rather than software usage errors.</summary>
In briefMACTS-EM: Multi-agent collaborative forecasting with emergent memory
SummaryThis paper proposes MACTS-EM, a framework in which specialised forecasting agents collaborate on time series prediction, coordinated by a meta-cognitive layer and supported by an emergent memory for cross-domain pattern transfer. The authors evaluate it on financial markets, climate, energy consumption and pandemic propagation. They report that it outperforms existing approaches in most scenarios, with 8-12% higher forecasting accuracy and 22-27% better zero-shot transfer.
In briefHAWK: Rethinking Multimodal Drafting for Faster Vision-Language Model Decoding
SummaryThis paper proposes HAWK, a method to speed up large vision-language models using speculative decoding, where a lightweight drafter proposes tokens that the target model verifies. HAWK selects informative target layers via representation similarity, gives the drafter compressed visual hidden states instead of raw visual tokens, and trains it to model how target predictions shift after its own proposals. On SmolVLM-256M across ten multimodal benchmarks, HAWK improves average acceptance length and speedup over EAGLE-3 under both greedy decoding and sampling.
In briefUsing Large Language Models to Address Challenges in Interpretive Structural Modeling
SummaryInterpretive Structural Modeling (ISM) is a multi-criteria decision method that models causal relationships and produces hierarchical structures, but its traditional process relies on repeated expert interaction and does not scale to studies with hundreds of variables. This work explores an integrated LLM-ISM approach that treats large language models as imperfect experts. It compares pairwise, k-wise, rowwise, and full-graph causal discovery methods. Rowwise (SHD=160, F1-score=0.77) and full-graph (SHD=135, F1-score=0.73) approaches performed best.
In briefA mixture of decoders that makes chatbot replies more varied
SummaryThe paper argues that sequence-to-sequence dialog models give low-diversity responses, possibly because they learn a single-mode response distribution. It adds a mixture of decoders to such models within a conditional variational autoencoder framework, so each decoder specializes in different topics. On an open-domain chat corpus, the approach improves over strong baselines on quantitative measures and human evaluation.
In briefMisaligned Low-Loss Regions Drive Grokking
SummaryThis paper studies grokking, the delayed emergence of validation generalization after a model has already overfit the training set. Using mode connectivity and the geometry of low-loss regions, the authors build an analysis framework. Under a symmetry-preserving train/validation split, they observe a stable anti-grokking case where validation performance does not recover. They conclude that grokking arises when low-loss regions induced by the training and validation partitions are misaligned.
