In briefHow Much Fresh Data Should CTR Recalibration Use? A Temporal Evaluation of Post-hoc Calibration
SummaryThis paper examines offline evaluation and calibration of click-through rate (CTR) models, arguing that random data splits understate calibration error caused by temporal drift. The authors test logistic regression, DeepFM, and DCN-v2 on the Avazu dataset, applying Platt scaling and isotonic regression recalibration over rolling time origins. They find that the time span of the fitting data strongly affects results, with wider windows outperforming the freshest few hours. The authors note that other logs would be needed to confirm this finding.
In briefGaussian Episodic Memory for Long-Term Spatiotemporal Reasoning: a robot memory framework for semi-static scenes
SummaryThe paper presents GEMS, which pairs a persistent spatio-temporal 3D Gaussian Splatting map with an agentic reasoning pipeline, so robots can render novel views on demand. The map is updated only where the scene changes, storing just those modifications. The authors also introduce GEMS-Bench, a benchmark of 473 question-answer pairs across 3 real and 12 synthetic semi-static environments. GEMS outperforms recent baselines on 3D object grounding and spatiotemporal question answering, with rendering quality comparable to dedicated scene-update methods.
Ling-Chia Chen, Abdelrhman Werby, Kaan Aytekin et al.
In briefThe Language of Instruction: Genre and Generative AI — argues instruction-tuned language models implement a distinct “instructional-conversational” genre
SummaryThis paper argues that instruction-fine-tuned language models should be understood as implementing a genre called the instructional-conversational genre. The genre is described as an amalgamation of discursive language from prior conversations between generative models and humans, recursively processed through large neural networks. The author says it is realized in applications through instruction fine-tuning, a pretext called a system prompt, and formalized inputs in markup languages using special tokens and role assignment in turn-taking structures.
In briefSceneBloom: Generating realistic, varied indoor 3D scenes for training robot agents
SummarySceneBloom is an agentic framework that generates indoor environments for training embodied agents by combining diffusion-based layout sampling with contract-guided 3D compilation. A hierarchical sampler proposes furniture arrangements within functional zones, and a compiler uses extracted relations to fit geometry and correct poses. Across five room types, it achieves the highest visual quality and physical stability, with no detected mesh collisions or out-of-bounds objects, and its layouts are closer to real room distributions than the baselines. Object navigation policies trained in these scenes reach a higher success rate on held-out real scans than policies trained on an equal number of real scans.
In briefUniAction: A unified action-conditioning method for world models
SummaryThe paper introduces UniAction, a framework that represents any action as a sparse set of action elements with semantic identity and spatiotemporal coordinates, mapped into a pretrained video model's positional space so diverse actions can condition one shared world model. It also proposes cardinality-calibrated attention to address bias from differing action representation granularities. Evaluated on force, camera control, hand pose, and robot pose, UniAction improves action following over alternative conditioning methods, and a single jointly trained model outperforms action-specific experts.
In briefFlexGlue: A neural network that matches image points densely yet quickly, and rejects unrelated image pairs.
SummaryThe paper introduces FlexGlue, a neural network for matching points between two images. Given points in the first image, it estimates their corresponding locations in the second image and rejects the image pair if the images do not match. According to the authors, experiments on multiple datasets show improved relative pose estimation and significant gains over existing methods for Structure-from-Motion and visual localization.
WhyCombines sparse and dense matching with disambiguation, a substantive improvement; multi-dataset experiments against SfM and localization baselines give solid evidence.
Philipp Lindenberger, Paul-Edouard Sarlin, Ondrej Miksik et al.
In briefWebUniverse: Scaling Verified Training Data for Agentic Web Development
SummaryThe paper presents WebUniverse, a pipeline that builds verifiable front-end development tasks and verified trajectories from open-source repositories and real web pages. It covers both text-grounded and vision-grounded tasks. The authors release WebUniverse-Env with 15K executable task environments and WebUniverse-Mix with 15K corresponding development trajectories and verification records. Fine-tuning on WebUniverse-Mix improves several benchmark scores and outperforms a public-data baseline trained under the same configuration.
WhyA verifiable data pipeline for web development agents yields substantial benchmark gains, supported by matched-budget baselines and ablations, useful for a subfield.
In briefRobustRLlib: A Unified Library and Benchmark for Robust Reinforcement Learning
SummaryThis paper introduces RobustRLlib, an open-source, algorithm-centric library and benchmark that brings 17 representative robust online, offline, and safe reinforcement learning algorithms under one interface. It is evaluated across 273 perturbation configurations, producing a capability map of current methods. The results show where these algorithms remain effective and where notable robustness gaps persist.
In briefDexPort: A Handheld Skill Interface for Collecting Dexterous Hand Demonstrations
SummaryDexPort is a portable data-collection interface for high-DoF robot hands that uses the same hand hardware for collection and deployment and records both commanded joint targets and measured joint states. The operator moves the handheld device to set wrist motion and selects or adjusts skills that coordinate finger motion. The authors collected over 2,500 demonstrations in lab and in-the-wild settings and trained visuomotor policies with behavior cloning across four tasks.
In briefMiD-VLM: Detecting and Locating Global Semantic Misinformation in Image-Caption Pairs
SummaryThe paper notes that an authentic image paired with a caption that distorts its narrative can be misleading, and pixel-level forgery detectors may miss such cases because the image shows no visible manipulation. Using the DGM4+ dataset, the authors propose a multimodal model that combines SigLIP and DINOv2 features with a Qwen2-VL-7B model adapted via LoRA. The model jointly predicts authenticity, identifies the manipulation type, and grounds the misleading caption tokens. On the DGM4+ validation split, it reaches an F1 score of 0.9679 for binary authenticity classification.
In briefViT-H: Generating Humanoid Robot Motions From Videos, Refined by Motion-Tracking Feedback
SummaryThe paper presents ViT-H, a framework that uses a diffusion model to generate humanoid robot reference motions directly from videos, skipping separate human motion reconstruction and retargeting steps. The generator is post-trained with direct preference optimization using motion-tracking outcomes to improve executability. Experiments on the Unitree G1 robot report better motion quality and executability than existing reconstruction-retargeting-tracking pipelines.
In briefFlownomaly: Single-Step Normal-Feature Recovery for Few-Shot Anomaly Detection
SummaryThis paper addresses few-shot industrial anomaly detection, learning what normal parts look like from only a few defect-free images. It uses conditional flow matching in frozen DINOv2 feature space to learn a velocity field, then recovers each query patch's normal counterpart in a single step. Residuals between queries and their recovered features serve as anomaly scores for localization and image-level detection. On MVTec-AD and VisA in the 1-shot setting, reported image- and pixel-level AUROC reach 97.8% and 98.0% on MVTec-AD and 95.3% and 98.7% on VisA.
WhyApplies single-step flow-matching posterior mean recovery to few-shot anomaly detection, with multi-benchmark comparisons on MVTec-AD and VisA showing solid gains.
In briefMeasure before training: comparing typed decision models, a classical baseline, and a generative LLM for routing personal multilingual email
SummaryThe paper presents Foyer, a self-hosted personal email agent, and a protocol for obtaining ground-truth labels before training. Using 739 real Italian and English emails labelled by the owner, it compares nine systems under thread-grouped cross-validation, including production keyword rules, a TF-IDF logistic-regression baseline, local and hosted decision models, and generative LLMs. The production rules perform far below the majority-class baseline, and logistic regression is most accurate on each field. However, the most accurate system often ignores emails needing action, and a cost-chosen noise threshold halves that rate.
In briefAligning where rounding occurs to stabilize LLM reinforcement learning
SummaryThis paper studies why reinforcement learning for large language models can become unstable when inference and training use separate systems in BF16 precision. Small implementation differences can turn an intended on-policy update into an off-policy one. The authors align training-side rounding locations with inference in three regions (RoPE, SwiGLU, and residual RMSNorm) by keeping selected intermediates in FP32 until their BF16 outputs. On four Qwen-family models, aligning all three regions prevented the collapse seen with the default BF16 baseline, and aligning only the RoPE path also sufficed in the studied runs.
In briefBeyond Discrete Tokens: Keeping Numbers Continuous in Medical Vision-Language Models
SummaryThe paper introduces HAD-VLM, a hybrid framework combining autoregressive text generation with diffusion-based continuous prediction for prompt-conditioned quantitative ultrasound measurement. On four echocardiographic measurement tasks in EchoNet-LVH, it achieves mean absolute errors of 1.36, 2.77, 3.05, and 1.16 mm for IVSd, LVIDd, LVIDs, and LVPWd, lower than a Vicuna-based autoregressive baseline. Medical question answering performance is preserved on the evaluated dataset, and on X-ray vertebral measurement the mean absolute error falls from 11.4 to 6.7 pixels.
In briefMonetizing Intelligence: Advertising and Subscription Design on LLM Platforms
SummaryThis paper studies how a monopolistic large language model platform jointly sets response quality, advertising intensity, and subscription prices. It finds that when partly ad-supported and ad-free plans coexist, advertising can fund useful service for users who would otherwise receive none, though utility can become polarized across user groups. The authors also show that an intermediate advertising cap may maximize consumer surplus, and that cheaper compute can raise total advertising revenue.
In briefKnowledge distillation across item and code spaces for parallel generative recommendation
SummaryThis paper studies generative recommendation with semantic IDs, where each item is a sequence of discrete codes predicted in parallel. The authors propose two complementary distillation objectives, Item Distribution Matching (IDM) and Code Marginal Matching (CMM), to transfer a sequential teacher's soft item-level knowledge to the student. Experiments on benchmark datasets report that both methods consistently outperform existing state-of-the-art approaches, with IDM improving Recall@10 by 15.39% and CMM improving NDCG@10 by 8.28% on average.
In briefRetry from earlier: a lightweight recovery method for frozen robot VLA models
SummaryThe paper presents Backtrack, a lightweight failure detection and recovery mechanism for frozen Vision-Language-Action (VLA) robot models that requires no retraining. A small prediction head reads the model's internal representations to estimate whether the current episode will fail. When the predicted failure probability exceeds a threshold, the controller pauses the policy, retraces recent arm motion, and returns control to the unchanged policy to retry. On LIBERO-Long, the long-horizon suite, the method reduces failure rates by roughly 19% to 48% across the evaluated policies.
In briefLearning positive ReLU combinations in fully polynomial time
SummaryThis paper studies PAC learning of positive linear combinations of k arbitrary ReLU activations on R^d under the Gaussian distribution, using L2 loss. The authors present the first fully polynomial-time algorithm for this task, with sample and running time poly(d, k, 1/ε). The approach uses a boosting procedure built on a list-decodable weak learner, implicit tensor computations based on Hermite tensors, and a pruning step.
WhyGives the first fully polynomial-time algorithm for learning ReLU positive combinations via a novel list-decodable boosting method; strong theory with impact mainly within learning theory.
Ilias Diakonikolas, Jelena Diakonikolas, Puqian Wang et al.
In briefYalda-Embedding: An open, competitive Persian text retrieval model
SummaryThis paper presents Yalda Embedding, a 595.8-million-parameter Persian retrieval model adapted from Qwen3-Embedding-0.6B using over one million query-document pairs. On eight Persian retrieval tasks from FaMTEB, it achieves the highest macro-average among evaluated 0.6B models (0.5488) and approaches Qwen3-Embedding-4B. An ablation study examines contrastive batch size, optimization stabilization, and relational distillation from F2LLM-v2-8B, and the model weights, code, and datasets are released.
Yalda Taheri, Mohammad Hassan Heydari, Armon Rasooli
In briefModeling Liquid-Cooled Data Centers in EnergyPlus
SummaryAs AI workloads push data centers toward dense racks cooled directly at the chip, existing tools struggle to assess whole-building, multi-year efficiency. The paper adds a native EnergyPlus object that models chilled-water and warm-water direct-to-chip liquid cooling as plant-loop components. Validation against an independent Modelica implementation across 21 operating conditions showed loop-temperature RMSE at or below 0.021 K. A case study reports that warm-water cooling lowers annual PUE from 1.43 to 1.25 and cuts total facility energy by 37.9% versus an air-cooled baseline.
In briefZENO: A neural operator method for nonlinear filtering via the Zakai equation
SummaryThis paper introduces ZENO, a neural operator trained with a physics-informed loss based on the Zakai equation to approximate the unnormalized filtering distribution of latent stochastic dynamical systems. Within an expectation-maximization framework, it performs both state inference and parameter learning. The authors test it on double-well, Van der Pol, and Lorenz systems, and show it recovers limit cycles for forward and backward locomotion in whole-brain imaging data from C. elegans.
Aaditya Prasad, Steven W. Flavell, Brian DePasquale
In briefAdaBPE: adaptive byte pair encoding for mixed text and LaTeX corpora
SummaryThis paper introduces AdaBPE, a tokenization method that improves on standard Byte Pair Encoding for language models. Standard BPE treats whitespace as a fixed boundary, which misses multiword expressions and formal notation such as LaTeX. AdaBPE adapts its minimum frequency thresholds based on token length and type, distinguishing general language from LaTeX patterns. Across five datasets covering LaTeX and natural language, it reports better encoding efficiency and downstream performance on domain-specific content without hurting general-language performance.
In briefSemantEq: A Stricter Test for Whether Formal Math Statements Match
SummaryThis paper addresses checking whether a formal Lean 4 statement faithfully matches its natural-language source. It argues that plain logical equivalence is too weak, since a strong prover can prove unrelated statements and pass them as equivalent. The proposed SemantEq relation requires each statement to transform into the other only through allowed steps such as logical rearrangement and library rewriting, judged by Lean's proof term. On the ProofNetVerif test split it reports 100% precision and 65.6% recall, compared with 95.8% and 61.8% for BEq+.
WhyProposes a chain-of-allowed-steps equivalence replacing trivial logical equivalence, judged on proof terms; benchmarked against BEq+ with model-audit experiments, giving solid evidence and practical v
Xinrui Li, Pawan Sasanka Ammanamanchi, Siddharth Bhat et al.
In briefA tutorial on physics-constrained machine learning for hybrid modeling, optimization, and control
SummaryThis tutorial presents KKT-HardNet, a modular physics-constrained machine learning framework, and shows how it can be used for hybrid modeling, process simulation, optimization, and model predictive control. It guarantees satisfaction of algebraic constraints through differential projection and uses the implicit function theorem to backpropagate through the projection layer, avoiding costly automatic differentiation. The authors also provide engineering benchmark problems and code, reporting large gains in constraint satisfaction and inference time.
In briefRecommender regret for better regularised RL estimates with bandits and MCTS
SummaryThis paper studies how estimates of policy, value, or Q-values in regularised reinforcement learning, such as those from Monte Carlo tree search, can be improved. The authors frame estimate generation as a recommendation process and define regulariser-aware recommender regrets, analysing their lower bounds. They show that simple methods like uniform sampling can reach these bounds. Experiments on synthetic trees, arcade games, and scientific discovery tasks report faster target convergence and better downstream performance.
Danilo Vucetic, Alexandre Larouche, Mohammad Pedramfar et al.
In briefNeural-DISCO: Editing Neural Activity to Test How Labels Shape Brain Signals
SummaryThis paper introduces Neural-DISCO, a disentangled representation learning framework for counterfactual modeling of neural population activity. It splits the latent space into components tied to observed behavioral or stimulus labels plus a residual capturing unlabeled variability, so one label can be edited while other content is preserved. The authors test it on simulated and real neural recordings and pair it with feature attribution to identify neurons important for each label.
In briefLoCA: Making Geospatial Location Encoders Work Across Multiple Image Types
SummaryThe paper shows that location encoders trained contrastively on satellite imagery are limited to an intrinsic dimension of at most 2, and their embeddings tend to collapse to low rank. The authors propose LoCA, a two-stage method that first jointly trains Sentinel-1 and Sentinel-2 image encoders into a co-aligned, high-rank target, then freezes them and trains location and tabular metadata projection adapters. On 17 tasks against five state-of-the-art encoders, the authors report that LoCA achieves the most task wins and wins 68.8% of per-task comparisons.
WhyProves a dimension bound for location encoders and proposes a two-stage multimodal method, validated across 17 tasks against five baselines.
Panagiotis Koromilas, Eleanna Vali, Diogenis Antonopoulos et al.
In briefGreen Code or Costly Code? Measuring the Energy Cost of AI-Refined Code Optimization
SummaryThis paper examines whether code optimized by large language models through iterative self-refinement is also energy-efficient. Testing 11 open-source models on all 1,000 EffiBench problems, the authors found that 42 of 44 configurations produced code using more energy than canonical human reference solutions, with overheads from +4% to +886%. Mixture-of-Experts models used substantially less inference energy than similarly sized dense models, suggesting architecture matters more than decoding settings.
WhyMeasures energy of LLM self-refined code against human references, showing performance gains often increase energy use. Large-scale experiments on 1,000 problems and 11 models support it, with impact
Sadra Emami, Benjamin Weigell, Mohammad Reza Mousavi et al.
In briefNarrowing the modality gap with simplex-constrained embeddings
SummaryMultimodal models such as CLIP and CLAP learn shared representations but keep a persistent gap between text and image embeddings. The paper proposes changing the embedding space from the unit sphere to the probability simplex, using lightweight softmax adapters and Total Variation similarity. Across five multimodal retrieval benchmarks, the centroid ℓ2 gap is greatly reduced, with retrieval performance matching or exceeding standard unconstrained models.
Shubham Gupta, Siva Reddy, Perouz Taslakian et al.
In briefHow human corrections travel: safety in federated VLA adaptation for robots
SummaryThis paper examines combining vision-language-action (VLA) robot policies, which adapt to local human intervention, with federated learning, which shares learning without pooling raw data. The authors argue that a shared update improving average task success may suppress rare safety interventions or cause unsafe behavior on another robot. They propose three requirements: preserving the context and uncertainty of multimodal corrections, testing candidate updates in each receiving robot's closed loop before deployment, and evaluating leakage alongside retention of safety-critical evidence. The work offers counterexamples and a falsifiable evaluation agenda rather than experimental results.
In briefReinforcement learning with multi-step transition predictions, a Bellman theory study
SummaryThis paper studies finite Markov decision processes where, at the start of each K-step block, the decision maker receives multi-step probabilistic transition predictions for some actions. The authors show the optimal value satisfies a Bayesian Bellman equation on the original state space. They propose BF-BOLA, a model-based algorithm, with sample-complexity bounds that match up to logarithmic factors.
In briefVTRL-Med: Topology-Guided Progressive Visual Refinement for Medical Image Clustering
SummaryThis paper presents VTRL-Med, a framework that clusters unlabeled medical images by progressively refining visual-structural features. It adapts the Segment Anything Model to extract features with structural cues, then generates pseudo-labels through prototype selection and k-nearest neighbor consistency checks. An auxiliary branch built from persistent homology descriptors adds geometric regularization during training. The authors report state-of-the-art clustering results across 13 medical image benchmarks, with a 4.48% average accuracy gain over existing methods.
In briefLearning pathology-sensitive features for label-free retinal OCT anomaly detection
SummaryThis paper presents an unsupervised framework for detecting and localizing anomalies in retinal OCT scans without disease-specific labels. It adapts pretrained visual representations to normal retinal anatomy using anatomy-aware reconstruction and contrastive triplet learning with synthetic structural perturbations. The authors report that adapting to normal structure improves sensitivity to abnormalities, and that pixel-space reconstruction errors and feature-space deviations provide complementary signals.
Tania Haghighi, Sina Gholami, Behafarin Emam et al.
In briefNgramRL: Speeding up RL post-training for structured generation with n-gram speculative decoding
SummaryThe paper addresses the cost of reinforcement learning post-training for language models, where autoregressive rollout generation dominates training time. The authors propose NgramRL, a per-prompt n-gram drafter built from recent rollout history, which needs no auxiliary model and keeps the rollout distribution through target-model verification. Across two model families and three workloads, it commits more tokens per target forward than prior methods, and on Spider a prefetcher improves measured rollout throughput by 1.38x.
Nilesh Prasad Pandey, Jangseon Park, Jason Kong et al.
In briefMultimodal AI Assists Full Echocardiogram Interpretation
SummaryThis paper presents EchoChat, a generative vision-language model that analyzes echocardiograms to support clinician interpretation, using chain-of-thought reasoning grounded in visual evidence. Trained on instruction pairs from 46,358 examinations, it covers 33 cardiac conditions and findings across six international centers. In a source-blinded reader study, reports co-generated with EchoChat improved diagnostic F1 score from 72.78% to 82.75%.
WhyIntegrates clinical chain-of-thought reasoning into a vision-language model, validated across six centers and a reader study showing clear diagnostic gains.
In briefImproving SVD-based LLM compression with layer-wise error model search
SummaryThis paper proposes two methods for compressing large language models using low-rank SVD: LEMS for global rank allocation and KFAC-SVD for better decomposition quality. The authors test them on Mistral, Qwen3, and Llama 3 model families. LEMS improves zero-shot accuracy, and KFAC-SVD achieves an average perplexity improvement of 19%.
WhyProposes layer-wise error surrogate with ILP rank allocation and token-wise KFAC-SVD, validated across several model families up to 70B.
Moritz Thoma, Maximilian Groezinger, Maximilian Forstenh\"ausler et al.
In briefDo "True" and "False" Add Up to One? Checking Consistency of LLM-Agent Opinion Simulations
SummaryThis paper studies opinion simulations in which LLM agents report the probability that a claim is true, given their own and neighbors' previous probabilities. The authors test whether rewording a claim as its complement and replacing each probability p with 1−p yields reports that sum to one. Experiments with three small open models show paired reports summing to about 1.18 on average, and the same failure appears in a controlled benchmark with two possible outcomes. The authors conclude that consistency under rewording must be tested before simulated probabilities are read as beliefs.
In briefJ-Comm: Jacobian-Guided Latent Communication Between LLM Agents
SummaryThis paper presents J-Comm, a framework that lets LLM agents communicate through latent representations instead of text. It combines a training-free reasoning method (GEAR) with an H+J interface that passes hidden-state and Jacobian-based cues between models without sharing key-value caches. In Qwen3-8B to Qwen3-4B experiments, J-Comm reaches the highest mean score across seven benchmarks and improves over TextMAS by 1.29 to 8.59 points.
Xiangwei Wang, Wei Wang, Sachith Seneviratne et al.
In briefA five-level architecture for communication with strong language-based superintelligence, built on deliberately constructed representations of activity
SummaryThis paper proposes a five-level structure describing how messages are formed and understood in thinking and activity involving artificial superintelligence. The author argues that a language model's context alone cannot show or verify the foundations of utterance formation, which are established in advance by acquired schemas and connections accepted as true. The five levels distinguish systemic representations of the world of activity, a situation drawn from them, its narrative description, the full sense of a message, and text. The theory draws on Shchedrovitskiy's systems methodology, Korzybski's structural differential, and Mel'chuk's Meaning-Text theory.
In briefTopKLoRA: interpretable sparse fine-tuning for finding backdoor circuits in models
SummaryThis paper presents TopKLoRA, a LoRA variant designed to be interpretable, where only the k largest adapter latents are active at each token. It is trained purely on downstream loss, aiming to learn functionally causal latents. In a sleeper-agent backdoor study on gemma-2-2b, Qwen2.5-1.5B, and Qwen2.5-7B, the authors identify and ablate circuits that are necessary and sufficient for the backdoor. The sparse circuits are up to an order of magnitude smaller than those from a dense LoRA baseline, while preserving instruction-following performance.
Marek Masiak, Andrzej Szablewski, Nicola Cancedda et al.
In briefPoKER: Machine-Learning Model Estimating Air-to-Air Missile Kill Probability
SummaryThis paper presents PoKER, a probabilistic model that estimates the kill probability of air-to-air missiles in Beyond Visual Range (BVR) combat. Unlike conventional Weapon Engagement Zone models that use static distances such as maximum, minimum, and no-escape ranges, PoKER uses machine learning to incorporate target stochastic behavior and missile miss distances. It also accounts for warhead lethality, target and shooter orientation, and engagement conditions to quantify success probabilities.
Joao P. A. Dantas, Andre N. Costa, Diego Geraldo et al.
In briefA foundation model (PICLE-PFN) that estimates longitudinal causal effects of dynamic treatment policies without per-dataset training
SummaryThis paper introduces PICLE-PFN, described as the first foundation model for longitudinal causal effect estimation that supports dynamic treatment policies. It represents target policies through in-context history-decision demonstrations and is pretrained on synthetic data from longitudinal structural causal models. Given a new observational dataset and policy, it predicts conditional average potential outcomes in one forward pass without task-specific training or tuning. The authors report competitive RMSE and better uncertainty calibration on synthetic and semi-synthetic benchmarks, with much lower deployment runtime than separately tuned conventional estimators.
In briefCarbon Cost of Time Series Foundation Models and Pruning to Cut It
SummaryThis paper examines the carbon cost of time series foundation models (TSFMs) compared with supervised standalone models that must be retrained periodically. It finds that TSFM pretraining costs are repaid within months for large or frequently queried deployments, while classical models stay cheaper indefinitely for small or rarely queried ones. The authors also show that removing one encoder block (7.9% of parameters) with a gradient-free pruning criterion reduces FLOPs, CO2 emissions, and latency, with essentially no loss in zero-shot forecasting performance on a 25-task benchmark.
In briefRegularised reinforcement learning for better sim-to-real robot transfer
SummaryThis paper studies how regularised reinforcement learning affects transferring robot policies from simulation to reality. The authors introduce SMM-AC, a new algorithm based on a soft mellowmax value function. On MuJoCo benchmarks it is competitive with Soft Actor-Critic, and on real robots it achieves higher return and greater robustness to perturbations.
Danilo Vucetic, Mohamad H. Danesh, Florence Clerc et al.
In briefThe singular geometry of low-rank adaptation (LoRA) for fine-tuning neural networks
SummaryThis paper studies how hard it is to learn low-rank adaptation (LoRA), a common method for efficiently fine-tuning large neural networks, using the real log canonical threshold from singular learning theory. It models LoRA as a low-rank regression problem and uses a forward-backward covariance to identify which low-rank changes are statistically observable. The authors also propose a curvature-corrected multi-temperature estimator and run experiments on multilayer perceptrons and transformers.
In briefNear-optimal list replicability bound via Littlestone dimension
SummaryThis paper studies the list replicability number in the realizable PAC learning setting. For concept classes over a countable domain with Littlestone dimension d and dual VC dimension v*, the authors prove an upper bound of 2^{O(d·min{log d, v*})}. This nearly matches the 2^{Ω(d)} lower bound and greatly improves the previous best bound. The new bound does not depend on the accuracy parameter ε.
WhyExponentially improves the list replicability upper bound to single exponential, nearly matching the lower bound, with rigorous proofs significant for learning theory.
In briefReimagining Enterprise RAG: A Knowledge Layer That Organizes Before It Retrieves
SummaryThis paper proposes a knowledge layer for enterprise multi-agent systems that organizes artifacts by provenance metadata such as producer, service, and domain before retrieval. The authors applied a seven-agent reverse-engineering pipeline to 241 repositories across 18 services, producing thousands of artifacts describing scope, journeys, rules, and acceptance criteria. Queries are resolved to artifact identifiers that constrain vector search, so identical rules from other services are excluded rather than merely ranked lower. The authors report it supported root-cause analysis of over 100 production defects and attributed 35 release regression failures.
In briefNeural LoFi with backward coupling: a spectral theory of how layers learn features together
SummaryThe paper extends Neural LoFi, a layer-wise spectral method for feature learning, to model the feedback between layers that backpropagation creates during joint training. It derives a local coupling between neuron input weights and the next layer's corresponding column, and proposes a backward-coupling algorithm. Experiments on binary CIFAR-10 with fully connected and convolutional networks report improved accuracy over forward Neural LoFi at matched widths.
Yatin Dandi, Matteo Vilucchio, Hugo Tabanelli et al.
In briefOpenStripe: open-set identification and stripe profiling of wild Amur tigers from camera traps
SummaryThe paper presents OpenStripe, a multi-agent framework that reconstructs canonical 2D tiger torso patterns and encodes four interpretable stripe structures. Evaluated on 100 camera-trap records from 28 wild Amur tigers, it achieved 0.90 matched accuracy and 0.85 adjusted Rand index for automatic identity grouping, outperforming 49 baseline configurations. The results also indicate that stripe similarity declines with geographic distance and that certain traits are associated with geographic proximity.
In briefReaching Everyone: Machine Translation for Emergency Alerts and Warnings
SummaryThe authors surveyed emergency managers nationwide to learn their current workflows, views, and operational challenges. They then evaluated commercial translation tools against these needs. The paper shows where these tools are useful and gives practical guidance on when and how they can be safely used in emergency response.
In briefSPOOK: Sparse pixel edits probe how vision encoders behave off the natural-image manifold.
SummaryThe paper studies how modern vision encoders behave far from natural images. In a black-box setting, it applies sparse pixel edits to blank canvases using a simple pixel-level hill-climbing method. On ImageNet and Flickr30k, the edited blank images could retrieve chosen targets in OpenCLIP and DINOv3 encoders. The authors conclude that visually non-semantic inputs can reach target semantic regions in both language-aligned and self-supervised representation spaces, suggesting a shared off-manifold failure mode.
Ankit Gupta, Akshat Pandya, Christoph Adami et al.
In briefOptimal correction sets for provably redressable contestable causal discovery
SummaryThe paper addresses conflicting conditional-independence tests in constraint-based causal discovery by formulating repair of these tests as a minimum-cost correction problem. It implements this in Answer Set Programming, building on Causal ABA with a compact encoding that has soundness and completeness guarantees and a two-way correspondence between optimal stable models and minimum-cost correction sets. Empirically, the approach retains more correct facts, reduces ambiguity in outputs, and improves graph reconstruction over Causal ABA on common benchmarks.
Fabrizio Russo, Giuseppe Mazzotta, Carmine Dodaro et al.
In briefTheoretical limits of sparse linear regression under shuffled differential privacy
SummaryThis paper studies high-dimensional sparse linear regression under shuffled differential privacy (SDP), which sits between central and local privacy models. For non-interactive protocols, it proves an error lower bound that reveals a dimensionality barrier and gives a matching-style upper bound algorithm. In the interactive user-level setting, it proposes a two-round algorithm that first recovers the support via heavy hitters and then estimates parameters, with error free of polynomial dependence on dimension, and shows a near-optimal lower bound up to a log(d/k) factor.
WhyProves a dimensionality barrier for non-interactive shuffled DP and near-optimal user-level bounds via a matching lower bound, a substantive theoretical advance for the subfield.
In briefKrishokTech: A Bengali Farm Advice System With Clear Authorization Controls
SummaryThis paper presents KRISHOKTECH, a Bengali agricultural advisory system that places an explicit authorization boundary around answer generation. It blocks high-risk requests, asks users to specify a crop, and removes dosage claims not supported by BARI/BRRI evidence. On 200 farmer queries and 400 stress queries, the controls cut cross-crop retrieval violations from 36.5% to 0% under the tested alias filter, and detected 453 of 454 text-image conflicts.
Khan Raiyan Ibne Reza, Sanjana Aktar Maria, Shakil Ahmed et al.
In briefLocating Surgical Tools by Matching Their Sound to Their Appearance
SummaryThis paper investigates whether a surgical instrument that is making sound can be localized using only one microphone and one camera, by learning the link between how tools look and sound without position labels. The model was trained on recordings of simulated sawing, drilling and chiseling, with added terms to suppress silence and to pick the sounding tool among two. On an independent corpus recorded with different equipment and optically tracked positions, it separated tool activity from silence and background noise and correctly attributed sound when two tools were visible.
In briefHarnessMM: Recursive Self-Improvement for Lower Token Use in Agent Harnesses
SummaryThis paper introduces HarnessMM, a method that reduces repeated reasoning in agent harnesses by reusing already-resolved decisions under guarded conditions and refining harness configurations through recursive self-improvement. A controller at the model-tool boundary replaces later reasoning callbacks with stored decisions while actor weights stay fixed. On HumanEval, Codex success is preserved at 90.82% while tokens per task drop from 66.72k to 47.02k; on DeepSeek Harness, success rises from 71.43% to 85.71% while tokens per task fall from 96.28k to 11.62k.
In briefU-NeXt: A General Event Detection Method for Neonatal EEG
SummaryThis paper presents U-NeXt, a semantic segmentation framework for detecting multiple event types in neonatal electroencephalography (EEG), including low-voltage intervals, artefacts, and seizures. The model was trained on large neonatal EEG datasets and evaluated on two independent seizure datasets annotated by experts. It outperformed the ConvNeXt baseline and ran substantially faster during inference. The authors suggest it could support continuous monitoring and clinical decision support.
Robert Hogan, Sean R. Mathieson, Aurel Luca et al.
In briefFrontier-OPD: Finding Where Agents Struggle and Teaching Them to Go Further
SummaryThis paper introduces Frontier-OPD, an on-policy distillation method for multi-turn agents. It uses a shared-state trajectory graph to locate the student's learning frontier, where brief teacher guidance lets the student continue autonomously, and adds accepted teacher responses to distillation data. Experiments on ALFWorld, WebShop, and ScienceWorld report higher mean autonomous performance than the compared distillation baselines.
In briefComparing multiplicative and additive solver-state coupling in learned optimizer schedules
SummaryThe authors learn schedules for a continuous penalty relaxation of MAX-CUT by backpropagating through an unrolled classical solver. A state-blind schedule generalizes to much larger instances than those used in training. A bilinear, state-adaptive schedule also generalizes out of distribution, while an additive ablation with identical inputs and outputs performs worse than its baseline there. The authors conclude that how solver state enters the schedule, rather than state adaptivity alone, distinguishes out-of-distribution transfer.
M. Harper Langston, Pierre-David Letourneau, Dalton Jones et al.
In briefCompMVR: Diffusion-based multi-view restoration for compressed images, keeping 3D reconstruction geometrically consistent
SummaryReal-world 3D reconstruction often uses multi-view images degraded by lossy compression, and the resulting artifacts can break geometric consistency across views. This paper introduces CompMVR, a compression-aware multi-view diffusion framework that jointly restores compressed views while preserving cross-view consistency, and builds compressed datasets for evaluation. Experiments across datasets, codecs, and compression levels report gains in restoration quality and in downstream tasks such as novel-view synthesis, camera pose estimation, and view matching.
In briefCARES: Coverage-Aware Replay Exemplar Selection for Continual Learning
SummaryThe paper proposes CARES, a method for building the replay buffer in continual learning that does not use training signals from the learner. Instead, it uses the geometry of a frozen pretrained feature space with class-aware, density-weighted coverage to select samples. On Tiny-ImageNet, CARES is 2.3 to 8.8 times faster than existing coreset-based selectors, and it outperforms the strongest baseline by up to 7.58 percentage points under class imbalance.
Arundhati S. Shanbhag, Brian B. Moser, Federico Raue et al.
In briefParaTaxe: Open-source tools for processing and publishing multi-parallel text corpora
SummaryThis paper presents an open-source toolkit developed within the ACR Grimm Tradalign and ParaTaxe projects for working with parallel text corpora. It includes AIlign, an aligner based on multilingual embeddings, a Python pipeline for converting and aligning corpora, a web interface for managing parallel corpus collections, and an exploration interface built with TEI-Publisher. The authors note that many academic parallel corpora become inaccessible online due to lack of maintenance or funding, which the toolkit aims to address.
In briefConsommé: Conditioned Token Merging for Efficient Split Vision Transformer Inference
SummaryThis paper studies reducing the number of tokens sent between an edge device and a server when a Vision Transformer is split across them, using a frozen backbone with only a small adapter shipped. The proposed method, Consommé, compresses the CLS token into a compact latent via an information bottleneck and uses it to condition a patch gate that merges patches into exactly k tokens in one step. On DeiT-S/T, it outperforms other methods at aggressive token budgets, works at shallow splits where bipartite matching fails, and matches fine-tuned methods at gentle budgets at a fraction of the deployment cost.
In briefSREdit: scene-response-guided appearance editing for projection mapping
SummaryProjection mapping can change the visible appearance of a physical object by projecting light onto it. Pretrained image editors can generate a desired look, but they lack accurate scene geometry and response, so their outputs may misalign or fall outside the projectable range. SREdit estimates a scene response map from the reconstructed scene and uses it, together with the source appearance, to control how much editing each surface region receives. Experiments on two pretrained editors show better projected appearance and reproduction fidelity, with under one second of added time per image, and participants preferred its results.
In briefToken perceptron: how embeddings and vocabulary shape generalization
SummaryThe paper studies how the geometry of token embeddings affects learning in large language models, using a minimal solvable linear model called the token perceptron. The authors compute its generalization error with the replica method, covering both continuous spherical embeddings and vocabulary embeddings. For vocabulary embeddings, they find that the Bayes-optimal learner first resolves the directions of largest vocabulary variance, even when the teacher is isotropic.
Dario Bocchi, Matteo Negri, Flavio Nicoletti et al.
In briefCounting the Bandages: a paper on measuring the time AI saves
SummaryThis paper examines what it means for AI assistance to save time, arguing that time saved should be read alongside the errors the work lets through rather than taken from a stopwatch alone. It defines model-free measures of time saved and excess escaped defects, and re-analyzes public data from METR's developer study and published GDPval results. A small pre-specified experiment then tests how AI reviewers change their checking when told the drafter's error rate is higher or lower than it actually is.
WhyProposes a ledger view of time saved with a break-even defect cost, novel framing; supported by reanalysis of METR data and a small pre-registered experiment, though limited in scale.
In briefVideo-MemComp: Sensitivity-aware memory management for streaming video understanding
SummaryThis paper proposes Video-MemComp, a training-free framework that manages memory for streaming video understanding under tight memory limits. It uses a first-order Taylor expansion to split compression error into local approximation error and layer-wise structural sensitivity, guiding token aggregation and KV cache budget allocation. The authors report performance matching or exceeding full-cache and other compression methods on MLVU, VideoMME, and EgoSchema. They also state that Qwen2.5-VL-3B's accuracy rose from 62.3% to 64.1% while saving over 90% of KV memory.
In briefEvery Edge Needs a Quote: A Verified Graph of Healthcare Software Customers and Integrations
SummaryThe paper presents a pipeline that builds a knowledge graph of healthcare software customer and integration relations from the open web. Every relation must be supported by a verbatim quote from its source page, checked through deterministic rules and a voting panel of model calls. On 61 vendors, it reports 0.94 precision and 0.71 recall against a page-level key.
Juan Felipe Duran, Arthur Tobler, and Martin Valli\`eres
In briefAI Alignment for Fairer Voter Representation in Digital Democracy
SummaryThis paper studies demographic bias when large language models emulate voters, finding that some voter groups are represented more accurately than others. Using three U.S. presidential elections, the authors compare a closed-source and an open-source LLM and test alignment strategies. Removing gender or education from voter profiles reduces the largest group error, while removing self-reported race increases mean group error by over 40%, without lowering accuracy. Common strategies such as repeated queries or fairness instructions did not reduce these errors and could double them.
In briefComputational de novo design of miniprotein candidates to neutralize IL-23 p19
SummaryThis paper computationally designs de novo miniprotein binders targeting the IL-23R-binding face of the IL-23 cytokine subunit p19, which is implicated in inflammatory bowel disease and psoriasis. Designs are scored with ipSAE, a length-robust interface metric, and a conventional pipeline is compared with an interface-optimized method. Five all-alpha-helical leads are reported to engage six targeted hotspots, though all results require experimental validation.
In briefDeep Learning Prediction of HER2 Status from Whole-Brain MRI in Brain Metastases
SummaryThis study examines whether deep learning can predict HER2 status directly from whole-brain contrast-enhanced T1-weighted MRI in breast cancer brain metastases, and whether adding tumour segmentation improves results. Using the public BCBM-Radiogenomics dataset of 135 patients, a segmentation-free network outperformed a conventional radiomics model. Adding an expert tumour mask gave a further modest gain, but overall performance remained limited.
Sara Early, Chris Kang, Vibujithan Vigneshwaran et al.
In briefToken-level credit assignment using reward-model gradients for GRPO
SummaryThis paper addresses how GRPO shares one response-level advantage across all tokens, leaving each token's contribution to learning unclear. It proposes Direction-Aware Token Credit Assignment (DATC), which maps frozen reward-model gradients into policy logit space and weights tokens by how well the reward direction aligns with the policy update direction. The response-level advantage's sign and mean are preserved. The authors report that DATC outperforms GRPO in all four settings across two model families.
In briefPARM: Particle-arrival-rate matching for denoising very low-count images
SummaryThis paper addresses low-count imaging, where limited illumination or dose yields only a few detected particles and Poisson shot noise dominates. It proposes PARM, a self-supervised denoising framework that models particle accumulation as a Poisson counting process and uses Poisson thinning to create complementary lower-exposure views that serve as rate targets for each other. Experiments on synthetic data, Poisson-corrupted natural videos, and real low-dose electron microscopy show leading or competitive performance while preserving fine structures. No clean reference images are required.</summary>
In briefSelf-training language models without human-made data or answers
SummaryThis paper proposes a self-training framework in which a model generates its own question-answer pairs and privileged reasoning traces, guided only by coarse high-level topic descriptions. No human-curated prompts or ground-truth answers are used. The authors report that self-distillation on fully synthetic privileged information improves performance across multiple domains while preserving capabilities in domains not self-trained on, with consistent gains on text benchmarks and Visual Question Answering.
Leon Sick, Lukas Hoyer, Yannick Str\"umpler et al.
In briefSCDM-HTD: A spectral-conditioned diffusion model for finding targets in hyperspectral images
SummaryThis paper addresses hyperspectral target detection, which finds objects in fine-grained remote sensing images, by proposing a spectral-wise conditional diffusion model. The method treats the detection task as generating a detection map, using pixel spectra as conditions to learn the distribution difference between background and target. It uses a lightweight multilayer perceptron denoising network and a deterministic reverse non-Markov chain to speed up generation, and a composite loss combining constrained energy minimization and binary cross-entropy. Experiments on four benchmark datasets reportedly show improvements in quantitative metrics and visual results.
In briefGame-theoretic study of clean-label poisoning attacks and defenses in federated vehicular networks
SummaryThis paper studies clean-label poisoning attacks and defenses in federated learning for vehicular networks, where Basic Safety Message classification for anomaly detection is modeled over a dynamic topology. The authors introduce a feature-space attack that perturbs benign samples toward malicious feature extremes while keeping their labels, and propose a feature-consistency relabeling defense to detect and relabel suspicious samples. A game-theoretic framework characterizes equilibrium strategies based on attack effectiveness and cost. Results show the attack grows more effective with more compromised nodes, while the defense keeps attack success low without reducing clean accuracy.
Yalin E. Sagduyu, Tugba Erpek, Nathan Stephenson et al.
In briefARENA: Attention-based method for interpretable solar flare prediction
SummaryThis paper presents ARENA, a neural network that predicts solar flares from a variable number of active-region image cutouts and learns which region matters most. It is trained with weak supervision, needing only flare or no-flare labels rather than flare-origin annotations. Its attention mechanism localizes the active region driving each prediction. ARENA achieves a Heidke skill score of 0.45, comparable to other flare prediction approaches.
Luis Bauer, Annalena Kofler, Andr\'es Asensio Ramos et al.
In briefSemantic Infrastructure for AI Agents to Safely Self-Play in Automated Wet Labs
SummaryThe paper introduces VERID, a persistent semantic digital twin of a laboratory workcell designed to link AI agents' high-level reasoning with error-prone physical experiments. It offers queryable introspection, time-compressed simulation for protocol validation, trace persistence, and a framework for safe execution. Three case studies using aqueous dye and deionized water mixtures show pre-emptive error detection, mid-run recovery, and post-hoc identification of physical failures.
In briefSparse Bayesian neural networks sampled with sticky piecewise deterministic Markov processes
SummaryThis paper studies how to obtain sparse neural networks within a Bayesian framework, using spike-and-slab priors to incorporate pruning. It uses Monte Carlo sampling based on piecewise deterministic Markov processes (PDMPs), where sticky versions set some weights exactly to zero during sampling. The authors examine the predictive performance and convergence of these samplers. They report that sparse Bayesian networks sampled this way can be more robust to input noise than dense ones.
August Arnstad, Geir Olve Storvik, Filippo Pagani et al.
