In briefOptimizer-dependent training dynamics still reach the same optimal one-third data scaling
SummaryThe paper studies how neural network loss falls as a power law with dataset size, separating the per-step dynamic exponent of a training run from the optimal data exponent after tuning. In an online teacher-student model, SGD and Adam have different dynamic exponents, yet with optimally tuned learning rates both reach a loss scaling of D^(-1/3). A stochastic-dynamics relation, 2α_r + α_t = 1, explains this, and measurements across seven optimizers, including Muon, are reported to agree.
In briefContextRender: Managing Agent Context Using Execution Dependencies
SummaryThis paper presents ContextRender, a method for managing the context of LLM agents that perform long-horizon tasks. It uses a persistent graph of execution dependencies and Tool-Flow Analysis to detect how later steps reuse earlier tool results, combined with recency and semantic relevance to select results within a fixed history budget. In experiments on AppWorld and 8-objective QA with a 6K history budget, it outperforms evaluated baselines and achieves task performance close to or above full-history passing, while reducing mean inference cost by 10.2%–32.2%.
Savini Kashmira, Jayanaka L. Dantanarayana, Lingjia Tang et al.
In briefHyDI: a hybrid deep learning and inductive logic programming ensemble for multi-label classification
SummaryDeep learning models achieve strong results but are hard to explain. This paper introduces HyDI, a hybrid ensemble architecture for hierarchical multi-label classification that combines a deep learning model with rule-based classifiers generated by inductive logic programming. For leaf classes of the label hierarchy, the rule-based classifiers replace the deep learning model, making results more transparent. HyDI is applied to the ChEBI ontology, providing rules for 314 classes, and can produce global explanations as well as local explanations combining visual and text descriptions.
In briefThe Camera Inside the Editor: Reading Image Editors' Implicit Camera from Painted Patterns
SummaryThe paper studies which camera an instruction-based image editor implicitly assumes when it paints into a photograph. The authors ask the editor to paint a checkerboard and read pitch, roll, focal length and other parameters from the projective structure using classical vanishing-point geometry. On 120 rendered cameras with exact ground truth, Qwen-Image-Edit-2511's implicit camera beats GeoCalib except in roll.
WhyNovel method reading implicit camera from painted checkerboards, compared against calibrators on 120 renders with ground truth; solid evidence and moderate impact on editor analysis.
In briefWeights That Read and Write Features: A Scalable Way to Decompose Model Parameters
SummaryThe paper introduces Activation-Supported Parameter Decomposition (ASPD), which decomposes both activation and parameter spaces together and links each learned weight component to the activation features it reads or writes. The method is demonstrated on the pretrained large language model Qwen-3-8B. The learned read-write components can be composed into parameter-level mechanism circuits, and the authors use ASPD to recover the mechanisms of the classic IOI circuit.
In briefSpatiotemporal Hyperedges for EEG Seizure Detection and Prediction
SummaryThis paper addresses detecting and predicting seizures from EEG, which are rare and spread across multiple channels. The proposed HyBrain model summarizes spatiotemporal EEG evidence with a small set of soft hyperedges instead of pairwise channel edges. A per-channel Mamba backbone produces tokens that are pooled into shared group embeddings, and one encoder serves window-based detection, point-wise detection, and preictal prediction.
In briefContext Language Models: Models That Manage Their Own Context
SummaryThe paper introduces Context Language Models (CLMs), which treat the context as a file that the model can freely update, so it learns what to keep. Across tasks including BrowseComp-Plus, a 12-hour EdgeBench, and a 24-hour multi-repository agent-swarm task, the authors report higher accuracy and lower FLOPs than state-of-the-art context management strategies. The paper also describes natural-language steering, an online reinforcement learning method, and Suffix Cache Reuse for serving.
Rulin Shao, Shannon Zejiang Shen, Junjie Oscar Yin et al.
In briefCogWAM: Aligning semantic reasoning with world-action modeling for robot control via event-driven interfaces
SummaryThe paper introduces CogWAM, a robot policy model that links task reasoning with world-action learning through a persistent Semantic State recording completed task events and the active subtask. This state is updated only when observations indicate a semantic transition. Reported results include performance on RoboDojo and BiCoord benchmarks and closed-loop real-world dual-arm manipulation experiments.
In briefPredictive geometry of hidden states in transformers: output-relevant directions induced by terminal loss
SummaryThe paper studies layerwise loss-to-go functions in decoder-only transformers, the terminal loss obtained by continuing a hidden state through the remaining blocks. Near successful validation trajectories, it shows that their local second-order geometry is governed by a pullback Fisher operator on hidden-state space, which separates output-sensitive directions from approximately prediction-null ones. Evaluated on WikiText, OpenWebText, and FineWeb, this geometry predicts perturbation sensitivity, guides layerwise rank allocation, supports token pruning, and improves low-rank student recovery in distillation.
Timur Mudarisov, Mikhail Burtsev, Tatiana Petrova et al.
In briefVolatility-Clustering Adaptation for Financial Time Series Forecasting
SummaryThe paper argues that fine-tuning time-series foundation models on more financial data does not always improve forecasts, since individual price changes are hard to predict while large moves cluster together. It proposes Volatility-Clustering Adaptation (VCA), which adds a differentiable penalty on the autocorrelation of squared returns to the standard next-token loss. Tested on three asset sets and two evaluation conventions, VCA improves over the pre-trained model, mainly by reducing variance error.
Manh Nguyen, Minh Hoang Nguyen, Huu Hiep Nguyen et al.
In briefBilliger.de Products: A Bilingual Entity Matching Benchmark
SummaryThis paper introduces Billiger.de Products, a bilingual German and English benchmark for matching product entities, built from the price comparison site billiger.de. It covers thirteen consumer product categories, including clothing and furniture, and provides aligned English translations plus cross-language test sets. Validation with six supervised matchers and zero-shot GPT-5.2 shows the benchmark is harder than existing English benchmarks such as WDC Products and Abt-Buy.
Aaron Steiner, Ksenia Elagin, Ralph Peeters et al.
In briefPolyOCR-Venus: Unified OCR Foundation Models for Visual Text Understanding
SummaryThis paper presents PolyOCR, a family of unified OCR foundation models of different sizes that aim to handle recognition, parsing, and reasoning over text in images. It combines a large-scale data engine that converts diverse visual resources into quality-verified OCR training data with a training method, Competence-Guided Policy Optimization, mixing verifier-based reinforcement learning and on-policy distillation. The authors also release OCRBench v2.1, a revised benchmark with corrected annotations, and report state-of-the-art or highly competitive results across several OCR benchmarks.
GuangJian Team: Kaili Huang, Yongshuo Zhang, Bingtao Fu et al.
In briefVIF-Bench: A Benchmark for Visual Instruction Following in Multi-Reference Image Generation
SummaryThe paper introduces VIF-Bench, a benchmark of 1,241 tasks that tests image generation models using multiple reference images together with several heterogeneous visual instructions such as layouts, arrows, and pose cues. It also covers cases where references conflict with instructions and compares visual instructions against text descriptions of varying detail. The authors report an adherence-artifact trade-off, lower adherence when references show the controlled attribute, and that direct visual constraints outperform text, with moderate text detail working best. The benchmark is released openly.
Yuta Oshima, Masakazu Yoshimura, Masahiro Suzuki et al.
In briefGenerative Interactions: Multiparty Human Motion via Bilevel Latent Dynamics
SummaryThis paper introduces BRAID, a hierarchical latent-variable model for generating motion of multiple interacting people. It separates a group-level latent state capturing shared interaction dynamics from person-level states capturing individual behaviour, and frames social motion generation as meta-transfer learning across datasets and partial observations. The model is evaluated on social forecasting, tracking and in-filling, and response generation using an SMPL-based representation. The authors also show that the latent space separates group-level and individual-level structure.
Ojas Shirekar, Yash Surange, Agustinas Ju\v{c}as et al.
In briefDynamic neuron expansion for learning new classes without forgetting old ones
SummaryThis paper addresses class incremental learning, where a model must learn new classes over time without access to past data or task identity. The authors propose a dynamic width expansion method that adds neurons to existing layers based on a normalized loss criterion, combined with an attention mechanism using persistent key-value memory. Experiments on Split MNIST and Split CIFAR-100 show that progressive width expansion outperforms fixed architectures, with the best results from combining expansion and attention. The authors note that uncontrolled growth may increase overfitting and computational cost.
In briefFinding Where Frozen LLMs Signal Answer Correctness
SummaryThe paper examines whether a frozen language model's internal signals, read in one forward pass, can predict if its answer is correct. It compares hidden states, token probabilities, residual-stream features, attention, and their fusion, in both closed-book and with-context settings. Correctness concentrates in the answer span, and the signal families are complementary, so fusion helps most out of distribution. The approach works across two backbones and is applied to gate a retrieval controller.
In briefTesting whether shape features help AI judge lung nodule cancer risk on 3D CT scans
SummaryThis study uses 3D CT scans from the LIDC-IDRI dataset to compare a single-task neural network that predicts malignancy with a multi-task model that also predicts spiculation and lobulation. Adding these morphological features did not clearly improve malignancy classification, with the two models performing similarly. The authors note that class balance, label formulation, and reader-level annotation structure matter in multi-task pulmonary CT analysis.
In briefGARDiff: Graph-aligned diffusion model for probabilistic multivariate time-series forecasting
SummaryThe paper finds that dependency graphs derived from deterministic forecasts often misalign at the edge level with residual dependency structures, making direct use as diffusion conditions unreliable. It proposes GARDiff, which estimates residual uncertainty to refine these graphs and sparsifies edges across diffusion timesteps, shifting from broad dependency aggregation to localized refinement. According to the authors, experiments on six real-world benchmarks show improved probabilistic forecasting and uncertainty calibration over strong baselines.
In briefHoneycomb: A Fixed-Size Scene Memory for Video World Models
SummaryThis paper presents Honeycomb, a video world model that stores scene information in HexMemory, a fixed-size representation made of six spatial and spatiotemporal planes. New video chunks are written into these planes and fused with existing features, and a reader retrieves latents to condition later video generation. Experiments on WorldScore and RealEstate10K report strong generation quality and consistent revisits while memory storage stays constant during generation.
In briefHow Reader Proficiency Shapes Where LLM Surprisal Predicts Reading
SummaryThis study uses MECO L2 eye-tracking data to compare how high- and low-vocabulary readers' first-pass and total gaze durations relate to surprisal from large language models across layers. The authors measure where predictive power concentrates using Predictive Depth, across 12 LLMs. Readers with lower vocabulary proficiency show deeper Predictive Depth for first-pass gaze duration, with a smaller difference for total gaze duration.
In briefWISE-ATTA: Deciding When to Request Labels in Budget-Limited Test-Time Adaptation
SummaryThe paper introduces budgeted active test-time adaptation, where labels are available for only some test batches. It proposes WISE-ATTA, which uses lightweight online signals to decide when to apply supervision over the test stream, and a drift-based criterion to select samples so a single labeled example can drive effective updates. Evaluated on ImageNet-C and ImageNet-R/K/A, it achieves competitive or better performance than recent methods while using substantially fewer labels.
Muhammad Huzaifa, Lea Sch\"onherr, Thorsten Eisenhofer
In briefEngiWorld: What Can Frontier Agents Deliver in Professional Engineering Environments?
SummaryThis paper introduces EngiWorld, a benchmark of 1,301 expert-curated tasks covering six engineering domains (CAD, CAE, CAM, BIM, EDA, and DCC) and 26 professional software applications. It uses an artifact-centric evaluation built on domain verifiers that check geometric validity, physical feasibility, and design-rule compliance, and scores how well optimized designs meet specifications. Testing seven frontier models, the strongest scored only 44.7/100, and just 3.6% of runs on cross-software tasks succeeded.
In briefPAIQ: Rotation-Based Patch-Level Fusion of Two Vision Encoders
SummaryThis paper introduces PAIQ, a framework that fuses features from a language-aligned encoder (SigLIP) and a self-supervised encoder (DINOv3). DINOv3 patch features serve as the spatial base, and SigLIP differences are injected through a shared orthogonal transformation that preserves residual norms and pairwise angles. Only the projection and fusion parameters are trained, and fusion keeps 196 visual tokens. The authors report gains in judge-assessed correctness and reduced hallucination severity over single-encoder interfaces on image description and visual question answering.
In briefLearning from Driver Overrides: Context-Driven Acceleration Profiles for Personalized Overtaking
SummaryAdaptive Cruise Control is usually calibrated for an average driver, so its behavior during highway overtaking can differ from what individual drivers expect. This paper treats drivers' throttle overrides as human-in-the-loop feedback and proposes CoP-ACC, a data-driven framework that combines hierarchical clustering, a context classifier, and a residual regressor to produce personalized acceleration profiles. Evaluated on real-world public-road data against a forced-ACC baseline, the approach shows high reconstruction fidelity and profiles that tend toward expected driver behavior in override contexts.
Ruizheng Xu (Heudiasyc), Lounis Adouane (Heudiasyc), Javier Iba\~nez-Guzm\'an et al.
In briefMed-RADIO: Distilling Multiple Medical Expert Models into One Vision Foundation Model
SummaryThe paper proposes Med-RADIO, a multi-teacher distillation framework that compresses knowledge from several domain-specific teacher models into a single medical vision foundation model. It selects generalist and specialist teachers, aligns distillation streams with imaging modalities to reorganize pretraining data, and uses a balanced loss so no single teacher dominates. On classification benchmarks covering five modalities, it outperforms strong medical generalist models under linear probing and is competitive with representative specialists on most modalities.
In briefWhen Models Don't Manipulate Manifolds: How an LLM Compares Numbers
SummaryThis paper studies how the Qwen2.5-7B-Instruct language model internally represents and compares numbers. Although numbers form curved geometric structures inside the model, the model largely relies on linear representations to perform the comparison. It combines the two numbers' vectors in a shared space via attention and residual connections, then uses MLP neurons to compare them locally. The same linear reliance holds when comparing three numbers.
Sai Sumedh R. Hindupur, Hadas Orgad, Thomas Fel et al.
In briefSpectral skills: a compositional motion interface for humanoid control
SummaryThis paper introduces spectral skills, a latent representation that links a planner to a humanoid controller, learned by predicting subsequent motion rather than reconstructing the input. On a 29-DoF humanoid, a controller conditioned on these skills reduces global tracking error by 62% relative to the state of the art. A frozen controller can chain independently encoded skills and compose new behaviors by adding orthogonal directions to compatible base skills, demonstrated on Unitree G1 hardware.
In briefLEMON-ZEST: Evolution-Informed Tokenization for Efficient Protein Language Modeling
SummaryThe paper introduces ZEST, a protein vocabulary built from conserved regions of multiple sequence alignments, so biological priors are embedded at the tokenization stage. It then presents LEMON, a 200M-parameter model for detecting remote homology between protein sequences, trained on a single H100 GPU for one week. The authors report that LEMON outperforms models from 600M to 3B parameters, suggesting evolution-informed tokenization can substitute for large-scale parameter scaling.
Biswajit Banerjee, Claudia Alvarez Carreno, Anton S. Petrov
In briefKUPAS MASTER: Turning Expert Tacit Knowledge into Corpora AI Agents Can Use
SummaryThis paper presents KUPAS MASTER, a platform that converts work records and practitioner interviews into traceable, reusable experience corpora for LLM agents, and packages them as callable skills. Using authorized samples from 20 practitioners, it processed 1,576 source files into 23,024 individual experience records and 13,113 organizational assets. In the evaluation, the base model, raw-corpus RAG, and the KUPAS MASTER agent scored 70.63, 79.75, and 89.58, with the agent outperforming raw-corpus RAG on all seven scoring dimensions.
In briefMeanFlowAdvantage: Stable Reward Fine-Tuning for Fast Average-Velocity Generators
SummaryThe paper introduces MeanFlowAdvantage, a signed advantage-weighted least-squares objective for reward fine-tuning of average-velocity generators. It uses a shared, detached MeanFlow derivative correction so that reward training matches the average-velocity network used at inference while keeping the few-step sampler. On SD3.5-Medium, it improves all eight reported metrics over a matched four-step baseline and, with four NFEs, matches or exceeds a 40-step baseline on six of eight metrics. The objective is also applied to DNA promoter design.
In briefRetrieve, Reproduce, Reveal: Analyzing Retrieval-Augmented Software Vulnerability Detection
SummaryThe paper studies six open-source retrieval-augmented systems for detecting software vulnerabilities, reproducing their experiments with open-weight models and comparing them on a unified dataset, metrics, and model pool. It also breaks each pipeline into input abstraction, knowledge retrieval, and detection to analyze components separately. The results show that reproducibility varies widely and published performance does not fully transfer under controlled evaluation, depending strongly on the model used. Even with oracle knowledge retrieval, pairwise accuracy stays at 0.51, indicating that good retrieval alone is insufficient for reliable detection.</summary>
Sabrina Kaniewski, Tim Kr\"amer, Julius B\"achle et al.
In briefWhere Privacy Belongs: Placement Diagnosis and Certified Selection for Private Graph Counterfactuals
SummaryThis paper studies counterfactual explanations for graph neural networks, arguing that both privatizing the graph before explaining and explaining then perturbing leak sensitive structure or corrupt results. It proposes PrivCFS, which selects counterfactuals from a fixed, data-independent candidate set and releases them via the exponential mechanism with pure ε-differential privacy. The authors report retaining most of the non-private utility at ε=8 and discuss the monotone cost of public disclosure.
In briefSemantic map sharing and capability-aware coverage planning for AI-native 6G robot teams
SummaryThis paper proposes an edge-centric, semantic-aware coverage planning framework for heterogeneous aerial and ground robot teams in search and rescue. Aerial observations are converted into compact semantic grid maps, and areas are assigned to each robot only where its capability profile allows traversal. Sharing semantic corrections instead of periodic full maps reduces the application payload by about 82 times in the high-update case. In matched benchmarks, the method reached 91.5% coverage with no capability-infeasible allocations, versus 78.8% coverage and a 21.5% infeasible-allocation rate for LS-MCPP.
In briefLearning causal normalizing flows from incomplete data using observed-data likelihood
SummaryThe paper introduces MissCNF, a method that trains causal normalizing flows directly on incomplete data by maximizing the marginal likelihood of each partially observed sample, avoiding row deletion or dataset completion. Experiments on eight synthetic causal benchmarks with three missingness mechanisms and missing rates up to 90% report lower KL divergence and counterfactual error than listwise deletion and impute-then-fit baselines in most nonlinear settings. On linear SCMs, MissCNF ranks in the top two in most settings, though linear imputation performs best there.
Trung-Dung Hoang, Alceu Bissoto, Tim Fl\"uhmann et al.
In briefGraph-based retrieval method that spreads query relevance through entity neighborhoods
SummaryThe paper proposes NexusRAG, a retrieval-augmented generation method that adds a corpus-level entity neighborhood structure, built from entity co-occurrence and semantic similarity, to a relation-free Tri-Graph. It uses two propagation paths, neighborhood-constrained semantic propagation and direct structural propagation between neighboring entities, and also uses the propagated weights to initialize Personalized PageRank. On three multi-hop QA benchmarks and a domain-specific GraphRAG-Bench subset, it outperforms existing approaches, achieving the highest evidence recall in all question categories on the subset, exceeding baselines by 4.2-8.1 points.
In briefLearning Service Rates While Scheduling Jobs Without Preemption
SummaryThis paper studies a single-server queue where each job has a context vector and its departure probability follows an unknown logistic model, so the server must learn while deciding which job to serve or whether to idle. Jobs cannot be interrupted once started. The authors propose Learn–Clear–Plan (LCP), which uses a finite-horizon Bellman recursion on estimated parameters, achieving queue-length regret of Õ(√(d/T)) and a matching lower bound up to polylogarithmic factors when T≥d². For unknown horizons, they propose an estimated-SEPT algorithm with tracking error Õ(√(d/t)).
In briefAre In-Context Images Worth 10 Dimensions? A study of shared low-dimensional geometry in vision-language models
SummaryThis paper studies how Large Vision Language Models represent labeled images provided in context for few-shot classification. The authors find that early layers reduce vision representations to a low-dimensional space shared across image classification tasks, where in-context images become linearly separable. They show analytically that linear self-attention can perform this reduction by projecting onto principal components, and provide evidence that trained models use a similar mechanism.
Adhemar de Senneville, Xavier Bou, J\'er\'emy Anger et al.
In briefEnterpriseBench: a benchmark testing LLM agents on enterprise strategy and decisions
SummaryThis paper introduces EnterpriseBench, a benchmark for evaluating LLM agents from static question answering to dynamic decision-making in enterprise settings. It combines reorganized enterprise and financial QA datasets with three interactive settings: management-consulting cases, a supply-chain inventory simulation (Beer Game), and an enterprise digital twin for workforce, risk, and project planning. Experiments with nine agent methods and four backbone models show that current agents are not yet stable or reliable across tasks in enterprise scenarios.
In briefTracing Evidence: VTrace, a Faithful Token-Attribution Method for Vision-Language Reasoning
SummaryThis paper addresses the difficulty of identifying which image and text tokens support a vision-language model's answers. It proposes VTrace, which traces input contributions through intermediate reasoning paths and calibrates attribution scores across modalities to rank tokens together. The authors report superior attribution faithfulness over seven baselines across six visual reasoning benchmarks.
In briefExemplar2VQA: A scalable framework that generates visual question-answer data for AI models by using multiple coding agents guided by example questions.
SummaryThe paper addresses the shortage of complex 3D question-answer data needed to improve spatial understanding in multimodal large language models. Exemplar2VQA uses collaborating agents that run deterministic code with a library of geometric tools in simulated environments, starting from static spatial question templates as exemplars. Fine-tuning Qwen2.5-VL (3B/7B) on the synthetic indoor data improves performance on various benchmarks, and the gains also extend to outdoor and mixed-scene benchmarks.
In briefTexture Space Material Diffusion: generating high-quality 3D object materials directly in texture space
SummaryThis paper presents a method that generates high-quality materials for 3D objects entirely in texture space. The authors finetune a video diffusion transformer for text-guided material generation, multi-view material generation, and material upscaling. By using the known projection from image space to texture space, the approach generalizes across arbitrary geometries and texture parameterizations and avoids view consistency issues common in video and multi-view diffusion. The authors report scaling to 8K resolution, 100+ input views, and neural material representations, with state-of-the-art results in quantitative and qualitative evaluations.
In briefRACE: Relation-Level Counterfactual Explanations for Heterogeneous Graph Neural Networks
SummaryThis paper presents RACE, a method that explains predictions of heterogeneous graph neural networks by identifying which relation types drive them. For each instance, it exhaustively searches relation subsets for the minimum deletion set that flips the prediction, then refines it into typed edges verified on the model. The paper reports that RACE improves counterfactual success rate over the strongest baseline on several datasets while deleting fewer edges.
In briefA method for embedding directed data using Finsler geometry
SummaryMany datasets have intrinsic directionality, but standard spectral embeddings symmetrize them and lose this information. The authors model directed data as sampled from a Finsler manifold, whose distance depends on travel direction, and study a kernel operator built from this asymmetric distance. They show its symmetric part captures geometry and its antisymmetric part captures directionality, and prove almost sure uniform convergence of the graph operators to these limits. For Randers metrics the vector field is explicit, enabling an algorithm that recovers both manifold structure and drift.
Gwendal Debaussart-Joniec (CB, ENS Paris Saclay), Th\'eau Blanchard (HeKA | U1346 et al.
In briefVoxelSage: A Tool-Augmented System for 3D CT Liver-Tumor Analysis and Simulator-Checked Resection Planning
SummaryThe paper presents VoxelSage, a multimodal system that supports liver-tumor analysis from 3D CT scans, including visualization, measurement, and sequential resection planning. Its dual-port design keeps physical measurements in an image-processing component, separate from the language model that selects tasks. In 256 unseen simulator scenes, it reduced mean simulated time and simulated blood loss compared with a deterministic baseline. These results concern system integration and simulator performance only, not clinical efficacy or safety.
In briefBenchmarking Jev, a Commercial Non-Generative Model That Returns Choices and Probabilities
SummaryJev is a commercial model from TypeSafe AI that does not generate text. It returns a fixed-option choice, a rubric position, or a probability that a statement is true. The authors evaluated jev-1.13.0 zero-shot on 37 datasets with 346,009 requests and compared it with Qwen3.8-27B and Gemma-4-E4B. Jev beat Qwen on 27 of 37 datasets and Gemma on all 37, while all three models degraded on low-resource languages, noisy labels, and rubric-based judgments.
In briefA dual-latent world model that plans over short and long time scales for robot-like tasks
SummaryThis paper introduces Dual-WM, a world model that uses separate latent representations and dynamics models for local action execution and long-range planning. The high-level model plans with learned macro-actions and generates latent subgoals for the low-level model to carry out. The authors also propose LoRe, a training method that weights self-generated predictions across horizons. On five goal-conditioned visual control tasks, mean success rose from 75.9% to 84.4% at a 50-step offset and from 61.4% to 69.5% at a 100-step offset.
In briefFlattening the connectome spectrum: a spectral filter that yields a pretraining target for fMRI encoders
SummaryThis paper examines why kernel ridge regression on functional connectivity matrices predicts individual traits well, and argues that the matrix eigenvalues are miscalibrated for this task. The authors apply a spectral filter to recalibrate each subject's eigenvalues, matching or exceeding the baseline across 5 datasets, 11 parcellations and 6 targets. They then pretrain a small fMRI encoder on about 4,000 hours from 162 open datasets, aligning embedding similarities with the recalibrated connectomes, and report performance on par with the best of six published brain foundation models using far fewer parameters.
Giovanni Marraffini, Victoria Shevchenko, Carlo Alberto Barbano (UNITO) et al.
In briefZero2Repo: Can Coding Agents Build Repositories from Scratch?
SummaryThis paper introduces Zero2Repo, a benchmark where coding agents receive a requirements document, an interface contract, and an empty workspace, and must build a complete repository. Tasks are generated from real open-source projects and judged by hidden acceptance tests with no LLM judge. Even on 11 tasks likely seen in training, the strongest agent solved only 10, and failing submissions passed 90–99% of the hidden tests.
In briefTargeted visual counterfactual explanations for contrastive vision-language models (MACE method)
SummaryExisting explanation methods for contrastive vision-language models such as CLIP mainly identify important regions but do not show how to change an input to reach a target prediction. This paper introduces MACE, which builds an adaptive, mask-guided editable region and uses a latent diffusion inpainting model with a frozen CLIP to make targeted edits. Across ImageNet, Food-101, Oxford Pets, and CUB-200, the source-mask variant achieves the highest target top-1 success, while the difference-mask variant yields the smallest pixel and perceptual changes and the best realism. Both variants improve proximity and realism over a Stable Diffusion-only baseline, and the study reveals a tradeoff between counterfactual validity and source-image preservation.
Van Bach Nguyen, J\"org Schl\"otterer, Christin Seifer
In briefOPMD: On-Policy Mechanism Distillation Beyond Output Supervision
SummaryThis paper proposes On-Policy Mechanism Distillation (OPMD), a method for improving distillation of language models. Standard on-policy distillation supervises only the student's outputs, whereas OPMD transfers the teacher's intermediate computation to the student. It first builds a student-compatible teacher that converts teacher representations into intermediate states usable by the student, then has the student internalize these states from its own trajectories. The authors report the highest average accuracy among evaluated baselines in three teacher-student settings, and generalization to multi-teacher distillation and cross-family supervised fine-tuning.
In briefCo-Linguistics: AI as a Partner in Building Linguistic Theories
SummaryThis paper proposes using AI as a co-scientist to help build and evaluate linguistic theories, a approach it calls Co-Linguistics. AI could make existing formal theories fully explicit, compare competing theories, propose new ones, and speed up the testing of key predictions. The authors stress that linguists remain central, providing scientific direction and conceptual evaluation, while human participants are still needed to test empirical predictions beyond LLMs' reach.
Emmanuel Chemla, Benjamin Spector, Alexandros Kalomoiros et al.
In briefRLTLDR: Self-Improvement by Internalizing Self-Generated Feedback
SummaryThe paper proposes RLTLDR, in which an agent receives the verifier's output after each failed attempt and writes its own one-line insight, which conditions later attempts until a solution is found. The method also backpropagates through these in-context insights so the model internalizes a task-to-insight mapping. On hard tool-calling and coding datasets, standard GRPO stays at 0% to 1% Pass@1, while RLTLDR reaches 14% to 31%, and 12% to 13% without insights at evaluation time.
Michael Kirchhof, Eleonora Gualdoni, Andrew Szot et al.
In briefProCTI: Using learned prototypes to add global context to diffusion-based time series imputation
SummaryThis paper introduces ProCTI, a framework for filling in missing values in time series using diffusion models. Unlike existing methods that rely mainly on nearby local windows, it retrieves global dataset-level priors through learned prototypes and combines them with local signals during reverse diffusion. Experiments on several benchmarks report that it outperforms strong baselines under random missingness and stays competitive under attribute-wise missingness. The authors also provide a latent-regime model and theoretical analysis of when prototype-derived global conditioning improves imputation.
In briefProcedural Core: a compact recurrent starting point for Vision Transformers
SummaryThis paper proposes Procedural Core, which trains a small recurrent transformer on procedurally generated abstract data and then expands its weights to initialize transformers of arbitrary width and depth. The authors report gains on image classification, DINO self-supervised learning, and language and code modeling. They also report improvements in zero-shot segmentation, object localization, and depth estimation, and argue that recurrence is key to learning transferable compact weights.
Zachary Shinnick, Christian Intern\`o, Hemanth Saratchandran et al.
In briefSPLASH: Switching attention parallel layouts during LLM serving without interrupting requests
SummaryThis paper presents SPLASH, a system that switches the parallel layout of attention in large language models while requests are still running, rather than fixing one layout at launch. It exploits the fact that in modern attention, KV cache placement is decoupled from attention weight sharding, which makes switching nearly free. The authors also introduce a new layout, Decoupled Ownership Parallelism (DOP). On B200 GPUs serving GLM-5.3, SPLASH reports 1.3-1.73x higher end-to-end throughput than fixed-layout deployments.
In briefCorrect, Don't Delete: Fixing Harmful Training Data to Reduce Emergent Misalignment
SummaryThe paper studies emergent misalignment, where fine-tuning on harmful medical advice makes a language model misaligned on unrelated questions. Using Qwen2.5-14B-Instruct, it finds that replacing poisoned training rows with corrected answers cut the misalignment rate by about a third and improved held-out medical answers, while deleting the same rows had little measurable effect. The authors report that the advantage held across a second base model and a second misaligned model organism.
In briefAS²D: Faster On-Device Audio Understanding on Phones
SummaryThe paper proposes AS²D (Audio Speculative Speculative Decoding), a method that speeds up audio language models on mobile devices. A small audio-conditioned drafter generates candidate tokens independently of the large target model's evolving text, while the target verifies and corrects them, so drafting and verification run concurrently. Tested on four phones across seven datasets and three tasks totaling 12.2 hours of audio, it improved pooled ASR throughput by 42–76% over target-only decoding.
In briefAuthority Bias in Language Models: Source Deference vs. User Agreement
SummaryThis paper examines whether language models accept wrong answers more readily when they are attributed to a verified source. Across five open-weight model families and three closed APIs, a single verified-source note endorsing a wrong answer flipped 45-88% of baseline-correct responses in seven of eight models, with compliance rising as the source sounded more authoritative. The authors find that source deference and user agreement can be selectively affected by internal interventions, so they require separate evaluation.
WhyShows source deference and user agreement separate internally, validated by causal interventions across many model families and APIs.
Abhinav Rajeev Kumar (Lossfunk), Paras Chopra (Lossfunk)
In briefLEGO-OPD: Splitting Language and Vision Teacher Signals for Multimodal Distillation
SummaryThis paper presents LEGO-OPD, a method for multimodal on-policy distillation that combines a language expert's prior with a grounding expert's visual likelihood into a single teacher distribution. This lets language reasoning and visual grounding be controlled separately. An adaptive calibration step sets how strongly visual evidence updates the language prior at each decoding step. According to the authors, experiments with Qwen3 models show gains over the evaluated single- and multi-teacher baselines on multimodal and text-only reasoning tasks.
In briefPHASE: A single model for simulating incompressible plasma (MHD) flows across different physical settings
SummaryThis paper presents PHASE, a machine-learning surrogate that models incompressible magnetohydrodynamics across varying physical parameters using a single model. It combines transfer learning, regime-aware adaptation, physics-centered learning, and residual refinement. On two-dimensional MHD turbulence, it reduces relative L2 errors on physical fields by more than an order of magnitude versus prior neural-operator baselines, and generalizes to unseen parameter values without retraining. The authors also show it can simulate the Kelvin-Helmholtz instability.
Radhika Achikanath Chirakkara, Rajdeep Haldar, Zezheng Song et al.
In briefCoEvoke: Eliciting Multi-Agent Collaboration to Collect Better Training Trajectories
SummaryThis paper introduces CoEvoke, a framework that induces checking, feedback, and repair during the collection of multi-agent task trajectories by using controlled perturbations and restricted reference guidance. The resulting dataset contains 4,567 reconstructable trajectories over 4,130 software-engineering and reasoning tasks. Experiments report that perturbations increase feedback and that perturbation-enriched judge training raises ACPBench-Hard success from 35.38% to 42.12%.
In briefTomoTransformer: Toward a Foundation Model for CT Reconstruction
SummarySparse-view CT reconstruction models trained with supervision are often brittle and need retraining when projection counts, angles, detector sizes, or data distributions change. This paper proposes TomoTransformer, a transformer that treats each local filtered projection as a token, predicts missing views through self-attention, and operates in a back-projection space. The authors report that it is a single model handling any number and angles of input and target projections, outperforms ViewTrans on benchmarks, and generalizes zero-shot to real synchrotron X-ray nanoscale brain data.
In briefGraphVQ: Structure-Aware Autoregressive Graph Generation with Quantized Graph Tokens
SummaryGraphVQ is a method for generating graphs as token sequences, addressing the difficulty autoregressive models have with edges that span beyond the serialization window. It quantizes node contexts into a shared codebook with a VQ-VAE and uses a second-stage structure-aware decoder conditioned on token-derived pair features to emit the global adjacency. Across four datasets, pair conditioning improves orbit MMD, and GraphVQ ranks first among learned generators on PROTEINS.
In briefIndependent Verification Paths Are Not Independent: Common-Mode Failure in a Satellite Catalogue Pipeline
SummaryThe paper reports a failed verification gate in a data pipeline that cross-checks two open registers of Earth-orbiting objects. Two computation paths built on different technologies shared the same constants, which misread the source's status vocabulary, so the gate reported agreement despite a wrong count. The authors also found that their correction still contained 42 artefacts among 261 disagreements, which their three checks did not flag. In a controlled replication with three models, most independently generated check paths reproduced the defective count.
In briefWhen to Adapt: Multi-Signal Shift Detection for Efficient Training-Free Segmentation Adaptation
SummaryThis paper addresses how robots performing long-term missions can lose segmentation accuracy when environmental conditions change and vision foundation models fail to adapt. It proposes a method that monitors visual change, adapter mismatch, and semantic drift across consecutive frames to trigger training-free adaptation only when needed. The approach is validated on indoor and outdoor benchmarks and real robotic data, where it maintains accuracy while substantially reducing the number of adaptations.
Michele Antonazzi, Alejandra C. Hernandez, Jos\'e Araujo et al.
In briefThe Weakest Link: Worst-case constrained RL for distilling LLM reasoning
SummaryThe paper argues that distilling reasoning from large language models into smaller students is hampered by reward hacking, where correct answers come from flawed logic, and by soft teacher-divergence penalties that let good steps offset bad ones. The authors formulate distillation as constrained reinforcement learning, requiring the teacher log-likelihood to satisfy a worst-case constraint along every prefix of a reasoning trajectory. Experiments on mathematical reasoning and code generation show the method keeps answer accuracy of pure RL while sharply reducing teacher constraint violations, achieving the highest Reasoning Success Rate.
In briefBlenDAgger: Blended Shared Control for Interactive Imitation Learning
SummaryThis paper proposes BlenDAgger, which collects training data for imitation learning by blending the policy's and demonstrator's actions through shared control during interventions. On two real-world manipulation tasks, it reports higher autonomous performance than the typical human-gated correction method HG-DAgger. It also reports smoother transitions between policy control and human intervention, and higher trajectory similarity to the training data.
In briefLLM math transfer follows reasoning method rather than topic
SummaryThis paper examines whether, when choosing math training data for LLMs, matching by topic or by reasoning method yields better transfer after fine-tuning. The authors use two counterbalanced 2×2 designs comparing sources that share the target's solution method but differ in topic against sources sharing the topic but differing in method. Across all 40 seed-pooled model-target comparisons, same-approach sources outperformed same-topic sources. The authors conclude that reasoning approach is a more effective matching criterion than topic for mathematical transfer.
Sajad Goudarzi, Samaneh Zamanifard, Seyed Amin Seyed Haeri et al.
In briefXU-RS: Explaining Uncertainty Width in Random-Set Language Models
SummaryThis paper proposes XU-RS, a framework that attributes a language model's uncertainty about an answer, measured as credal width (the gap between lower and upper probabilities), to the input tokens. It uses Expected Gradients to estimate each token's contribution. On the MedQA dataset with SmolLM3-3B and Llama-2-7B, zero-masking tokens ranked highly by XU-RS changes credal width more than masking random tokens. The authors also examine normalisation effects, how token attribution can mask numerical errors, and diagnostic checks for attributions.</summary>
David Achara, Maryam Sultana, Alexander D. Rast et al.
In briefCredit-Guided Policy Improvement for Test-time Adaptive Vision-Language Navigation: a method that infers decision credit from observation changes caused by actions and keeps only supported policy updates
SummaryThe paper addresses off-course decisions caused by distribution shifts when vision-language navigation policies run in unseen environments. It proposes Credit-Guided Policy Improvement (CGPI), which, with the pretrained policy frozen, derives signed, reference-relative decision credit from the observation transitions induced by each executed action, without external outcome feedback. CGPI uses this credit to propose lightweight adaptation updates, retaining them only when verified against prior credit-supported interactions and rolling them back otherwise. The authors report consistent gains across evaluated benchmarks and navigation backbones, and qualitative robot trials illustrating feasibility of zero-shot sim-to-real transfer.
In briefFOCUS: Training-free context compression that preserves agent decisions
SummaryThe paper addresses how LLM agent interaction histories grow with task length, raising inference cost and degrading performance through attention dilution. It proposes FOCUS, a training-free framework that runs entirely at test time and compresses context by identifying which past interactions causally shape future decisions. It attaches as a modular layer to any closed-API frontier model. Across API/tool-calling, QA, web, and multi-turn dialogue benchmarks, the authors report up to 48% lower peak context, 73% lower dependency, and up to 8.9 points higher task success.
In briefAssistive agents deciding when to ask clarifying questions via value-of-information reasoning
SummaryThis paper examines whether a language-based assistant should act on an ambiguous request or ask a clarifying question. It proposes REVOIR, which decides on clarification at inference time by estimating the expected improvement in task reward from a question's answer. On ambiguous question answering and preference-aligned household planning tasks, REVOIR reportedly succeeds with fewer questions than several baselines and requires no training. It also adapts to cases where cheap user corrections after acting are available.
T. Duy Nguyen-Hien, Yee Whye Teh, Wee Sun Lee et al.
In briefReLMem: Learning Recurrent Memory for Longitudinal EHR Modeling
SummaryReLMem is a framework that keeps a fixed-size patient memory for a frozen language model, updating it recurrently with each new visit instead of rereading full records. Lightweight compression adapters and multi-granularity training align compressed memory with full-history attention and predict answers from the final memory. On EHR-based medication prediction, it approaches full-history F1 while cutting retained historical storage by 97.1%. Under the same memory budget, it outperforms the strongest compressed-memory baseline by 4.66 and 4.75 points in macro- and micro-F1.
In briefLearning as Deepfakes Evolve: RF-Prompt for Continual Audio Deepfake Detection
SummaryThis paper addresses continual audio deepfake detection, where models must learn newly emerging deepfake methods while still recognizing previously seen speech. It proposes the RAMI evaluation protocol, which uses real speech as a recurring reference while new deepfake mechanisms arrive incrementally. It also introduces RF-Prompt, a method that keeps a shared real-speech prompt and expands mechanism-specific fake experts. On RAMI, RF-Prompt reports 10.110% average EER and 10.370% pooled EER, outperforming all evaluated continual-learning baselines.
In briefFedSocket: A method for exchanging knowledge that recipients can directly run in mixed-type federated learning.
SummaryThe paper presents FedSocket, a method that makes knowledge exchanged in federated learning directly usable by recipients who have different modalities, private architectures, and tasks. A shared module Q connects heterogeneous private models through a common prediction interface, and only Q's parameters and counts are exchanged. Across six datasets, the authors report accuracy gains of 14.44 and 15.51 percentage points over local training for missing-modality recipients on MELD and UCF-51.
In briefTReVS: Pruning Vision-Language Model Visual Tokens Using Text Relevance and Visual Saliency
SummaryThis paper addresses the high inference cost of Vision-Language Models caused by large numbers of visual tokens. It proposes TReVS, a training-free framework that prunes tokens in two stages: before the language model, it combines textual relevance with vision-encoder saliency, and inside the language model, it uses high-variance attention heads to remove task-irrelevant tokens. On LLaVA-1.5-7B, TReVS prunes 94.4% of visual tokens while retaining 92.8% of unpruned baseline performance, outperforming prior state-of-the-art methods.
In briefOrnstein-Uhlenbeck diffusion is hard to beat, but superlinear drift lowers transport costs in experiments
SummaryThis paper tests a claim by Brešar and Mijatović, who showed that Ornstein-Uhlenbeck diffusion is hard to beat in forward convergence when superlinear drift is excluded. The authors instead apply superlinear Langevin diffusions to score-based image generation, computing conditional scores numerically from a Fokker-Planck equation. Their experiments show the superlinear models beat the Ornstein-Uhlenbeck baseline on empirical Wasserstein distance across nearly the whole tested grid, with less variation across diffusion horizons. The authors conclude the earlier "hard to beat" verdict does not hold universally.
In briefPair Difficulty Matters: Rethinking Pairwise LLM-as-a-Judge Evaluation
SummaryThis paper argues that three common proxies for judging LLM reliability, namely position bias, transitivity, and pairwise agreement, can be misleading. Under the Bradley–Terry model, these proxies are dominated by pairs with close ranks, while the ranking signal mainly comes from pairs with large rank gaps. The authors validate this in simulations and on two human-rated corpora, and recommend rank-gap-conditional evaluation, ideally against human rankings.
In briefEvaluating the Evaluators: Testing AI Models That Judge AI-Generated Images
SummaryThis paper introduces SQUARE-Bench, a benchmark that tests how well large multimodal models judge AI-generated images across semantics, quality, authenticity, and responsibility. It uses 38 sub-dimensions, nearly 10K AI-generated images from 22 models, and over 3K real images. Experiments on 23 models show that the top proprietary model, Gemini-3-Pro, outperforms an individual human expert baseline, though large gaps remain between models. The authors also show that model feedback can guide image editing, yielding selective gains in some dimensions with a trade-off in visual quality.
In briefA paper on how task scope, not receiver type, determines what information a source codec can discard
SummaryThe paper argues that splitting codec design into Coding for Machines and Coding for Humans is misleading for deciding what information can be dropped. The required rate depends on task scope, including the predictions supported, their losses, the encoder observation, and allowed decoding. It extends source and feature coding to finite task families and derives when restricting the encoder observation keeps the minimum rate unchanged.
In briefMERGE: A Multi-LLM Ensemble Method for Retrieval Through Generative Query Enrichment
SummaryThis paper presents MERGE, a framework that rewrites user queries to improve BM25 retrieval. Three open-source 7-8B language models each generate candidate expansions, and a larger model synthesizes them into one query. The framework also adds an automatic prompt optimization loop that scores candidates by their downstream retrieval performance. On five BEIR benchmarks, BM25 nDCG@10 improved by 2.1 to 14.9 points over the original queries.
In briefBreaking Chinese characters into components for more accurate text-image rendering reward
SummaryThis paper addresses the difficulty text-to-image models have in rendering accurate Chinese characters. Existing OCR-based reinforcement-learning rewards compare whole characters and overlook their component structure. The authors propose IDSpect, which decomposes target text into Ideographic Description Sequence tokens, uses a trained recognizer to match components and spatial relations in rendered images, and combines this with a whole-character semantic reward for GRPO post-training of Qwen-Image. The paper reports leading structural quality and semantic alignment on LongText and GenTextEval.
In briefDevils in Question Relay: a training-free method to reduce cross-modal hallucinations in audio-visual AI models
SummaryThis paper examines source-confused grounding hallucination in audio-visual large language models, where cues from an unused modality lead to answers the required modality does not support. Through path-intervention and representation analyses, the authors identify a question-relay mechanism in which question states carry both interfering cues and required-source evidence. They propose SECRET, a training-free method that steers question states toward required-source evidence, and report consistent gains over prior training-free methods on the CMM and AVHBench benchmarks across three models.
In briefRAVEN: Receiver-conditioned action-value encoding for compact multi-agent messages
SummaryThis paper proposes RAVEN, a method for multi-agent communication using a four-symbol, one-step-delayed channel that preserves the action-value distinctions a receiver needs within its own context. It offers an offline version with a teacher and an online version built into a QMIX learner. Across eight navigation settings and other benchmarks, RAVEN reports higher returns and normalized scores while each message costs only 2 bits.
WhyProposes receiver-conditioned action-value encoding for finite-alphabet communication, with bounded codebook error and experiments across navigation and SMAC/MPE benchmarks.
In briefA model-agnostic physics-guided adapter for few-shot transfer of coastal flood models
SummaryDeep learning surrogates can generate high-resolution coastal flood maps much faster than physics-based simulators, but transferring them to new regions requires many costly simulations for fine-tuning. The authors propose the Physics Adapter (PA), a compact, architecture-agnostic module that predicts peak water level via a differentiable wet/dry response using terrain elevation and a learned water level, blended with a data-driven branch through a learned gate. PA adds very few trainable parameters. In the low-shot setting (K=3), averaged across backbones and transfer settings, PA reduces RMSE by 11.5%, 15.4%, and 22.9% for output-head-only adaptation, PEFT, and full fine-tuning, respectively.
In briefOrthogonal yet coupled: separating geometric parts for model merging
SummaryThis paper proposes DiGA, a disentangled geometry-aware framework for merging pretrained models. Using the pretrained weights as a shared reference, it orthogonally decomposes each task vector into components tied to distinct geometric attributes, aggregates them independently, and recombines them into one update. The authors report that DiGA can be added to many existing merging methods and improves merged-model performance while reducing capability degradation across diverse models, tasks, and methods.
In briefAPM-Bench: A Benchmark for Cross-Session Persistent Memory in Egocentric Streaming Video Assistants
SummaryAPM-Bench reformulates real-world streaming video interaction as multi-session life trajectories, with 549 sessions, 104 trajectories, and 2,719 candidates. It evaluates whether models can use persistent memory to answer questions about past sessions, respond proactively, and acknowledge when evidence is missing. The results show a utility–latency–storage trade-off, with existing methods struggling to achieve reliable long-term recall, low overhead, and effective cross-session assistance at once.
In briefBenchmarking graph-based models for computer-simulated toxicity prediction in drug discovery
SummaryThis paper proposes a unified, standardized benchmark for graph deep learning models that predict chemical toxicity. It systematically evaluates more than 20 representative methods under consistent conditions across multiple datasets and partitioning strategies, and complements this with a literature analysis. The benchmarking framework is released as open-source software.
In briefRisk-aware semantic grounding to make LLM-based robot planning more trustworthy
SummaryThis paper addresses the unreliability of large language models used as high-level planners for robot navigation when instructions are ambiguous, unsupported by the environment, or semantically inconsistent. The authors propose a framework that estimates ambiguity, hallucination, and semantic-conflict risks before planning, so the system can execute, request clarification, or reject an instruction. They also introduce TRUST-NAV, a benchmark combining standard navigation tasks with risk-inducing instruction scenarios. Experiments show the framework substantially improves ambiguity detection and semantic-conflict rejection compared with conventional LLM planners, which perform well on valid tasks.
{\L}ukasz Sobczak, Nur Kele\c{s}o\u{g}lu, S{\l}awomir Piotr Nowak
In briefEnd-to-end optical semantic communication over nonlinear WDM fiber, sending only task-relevant image information
SummaryThis paper proposes an end-to-end optical semantic communication system for joint image classification and reconstruction over a nonlinear wavelength-division multiplexed fiber channel. Images are mapped directly to fixed-length channel symbol sequences without explicit source compression or channel coding. On MNIST at 0 dBm, classification accuracy stays between 98.92% and 99.31% across all tested link lengths and modulation orders, using fewer symbols than an LDPC-coded JPEG baseline.
In briefHow recommendation feedback can evolve agent memory, and the TIDE framework for updating memory from delayed signals
SummaryThis paper studies how content-generation agents can evolve their external memory using delayed and noisy outcome feedback from recommendation systems. The authors propose TIDE, which estimates contextual memory fitness through temporal and semantic credit assignment, distributes responsibility credit to memories referenced during generation, and then reinforces, crosses over, mutates, or evicts memories. They also introduce Memory Evolution Gain (MEG), which measures the utility improvement of evolved memory over a no-memory baseline on strictly future tasks.
In briefRunyaNER: A Named Entity Recognition Benchmark for Runyankore and Auxiliary Language Selection
SummaryThe paper introduces RunyaNER, the first publicly available named entity recognition benchmark for Runyankore, an East African language, containing over 237k annotated words across 30k sentences. The dataset was built with a semi-automated pipeline and fully manually verified, and pretrained models evaluated on it show it is of sufficient quality and size for effective NER. The authors then study which auxiliary languages to use in cross-lingual zero-shot and multilingual fine-tuning. They find that embedding-based measures from labelled training spans correlate more strongly with transfer performance than metadata or typological features.
Prosper Arineitwe Asiimwe, Francois Meyer, Jan Buys
In briefSkillGym: Training LLM agents to use skills via auto-generated verifiable environments
SummarySkillGym is an automatic pipeline that crawls skills from the internet, keeps those whose workflows run reproducibly offline, and builds difficulty-controlled tasks with reference solutions and executable verifiers. The authors built 6.8k environments and collected 19k verified successful trajectories for supervised finetuning. Finetuning improved models from 2B to 122B parameters across four skill-use benchmarks, and a finetuned Qwen3.5-9B outperformed an untrained 397B model on two of them. The training also raised how often agents read the relevant skill from 28% to 96%.
In briefRemoving Bad Seeds: A Noise-Robust Unlearning Method for Text-to-Image Diffusion Models
SummaryThe paper finds that leading unlearning methods for text-to-image diffusion models are fragile: suppressed concepts can reappear under certain random initial noise, which the authors call "probabilistic forgetting." They attribute this to standard Gaussian sampling during unlearning being misaligned with the unlearning objective, and propose an adaptive, concept-conditioned sampling strategy that focuses gradient updates on noise regions where the target concept appears. The method is integrated with six existing unlearning approaches across four diffusion backbones and is reported to improve robustness against adversarial attacks.
Arian Komaei Koma, Seyed Amir Kasaei, Aida Aryafar et al.
In briefWhy Adaptive Optimizers Underestimate Rare Tokens
SummaryThis paper analyzes why coordinate-wise adaptive optimizers such as Adam, RMSProp, and sign descent underestimate rare tokens in the softmax output layer. It identifies which optimizers preserve the mean output embedding and derives a closed-form fixed point for RMSProp in a unigram model, showing that rare tokens' equilibrium probabilities fall below their data frequencies. In a small language model trained from a known distribution, optimizers with the biased fixed point also fit the generating distribution less well.
WhyShows adaptive optimizers systematically underestimate rare-token outputs, with closed-form fixed points, validated in unigram and small language model experiments.
In briefE-MoE: Enhanced Mixture-of-Experts for Diffusion Language Models
SummaryMasked diffusion models generate text by unmasking several tokens per step, but their reverse process is usually factorized across positions, which hurts quality when only a few steps are used. The paper proposes E-MoE, which builds the reverse process as a mixture of factorized distributions over a discrete shared latent given by a Mixture-of-Experts backbone's routing decisions, without increasing active parameters. The authors report that E-MoE improves few-step generation over factorized baselines on synthetic multi-modal benchmarks, binarized MNIST, and LM1B.
Arseny Ivanov, Alexander Kolesov, Alexander Korotin et al.
In briefDScale: Faster LLM Decoding via Adaptive Verification for Block-Diffusion Speculative Decoding
SummaryThis paper addresses efficiency problems in block-diffusion speculative decoding for large language models at high concurrency, such as verification padding, rejected candidates, and incompatibility between variable prefixes and fixed-shape graphs. The proposed DScale keeps the drafter's architecture and weights, adds a separate 112K-parameter predictor, and uses path-aware tiles, dynamic verify-length allocation, and fixed-address workspaces. On A100-40GB with Qwen3-8B and Qwen3-4B, it reports higher geometric-mean throughput and lower request latency than DFlash, DSpark, and Domino.
In briefRawVLA: A learned camera-processing module that helps robot manipulation policies cope with poor imaging
SummaryVision-language-action models usually take RGB images produced by a fixed camera image signal processor, which sits outside training and evaluation. The paper studies five ISP factors (gain, sensor noise, chromatic response, tonal response, bit depth) and finds they materially affect action prediction and manipulation success. It introduces RawVLA, a streaming neural ISP that adaptively renders RAW observations for frozen VLA policies, along with the RawVLA-Bench benchmark. Results show performance is maintained under standard conditions and robustness improves substantially under degraded imaging.
In briefPrincipled MAP estimation for inverse problems that links convergence guarantees with strong performance
SummaryThis paper proposes an algorithm for severely ill-posed image inverse problems that combines denoisers at decreasing noise levels under a schedule designed to ensure convergence. From a Bayesian perspective, the authors prove that the method converges to a Maximum a Posteriori (MAP) estimate under suitable assumptions. Applied to various inverse problems, it outperforms existing convergent methods while remaining competitive with state-of-the-art empirical approaches.
Alexandre Lagier, Valentine Tosel, Anne Gagneux et al.
In briefRaw satellite imagery and onboard AI: how image degradation affects detection.
SummaryThis paper studies how image degradation affects onboard AI object detection when models process raw or minimally processed satellite imagery. Signal-to-noise ratio, modulation transfer function, and ground sampling distance are varied on Maxar imagery, and three lightweight detectors are tested. Impact depends on the degradation mechanism, and severe combined blur and noise cause the largest losses.
