In briefA Physics-Guided Transformer for Fast Electromigration Stress Prediction in Chip Interconnects
SummaryThis paper proposes a physics-guided transformer framework that quickly predicts electromigration stress in multi-segment on-chip interconnect lines, as an alternative to costly physics-based solvers. Each line is converted into segment tokens carrying geometry and DC information, and attention captures line-level context before a lightweight decoder predicts stress at chosen locations and times. On IBM power grid benchmarks, the authors report relative-L2 error below 8% and up to 2459.68 times speedup over a matrix exponential solver.
In briefAdapting Vision-Language Models Without a Complete Class Vocabulary
SummaryThis paper introduces Vocabulary-Incomplete Unsupervised Vision-language model Adaptation (VI-UVA), where only some class names are available during adaptation. The authors propose Label Space Reconstruction (LSR), which rebuilds a latent label space from unlabeled data and uses it in prompt tuning to reduce bias toward the provided classes. Experiments on eight benchmarks report that LSR outperforms existing state-of-the-art methods.
In briefMuon optimizer on convolutions: theoretically misaligned yet empirically effective
SummaryMuon is an efficient optimizer with a clear theory for matrix updates, but convolutional kernels are four-dimensional tensors that standard implementations reshape into matrices, breaking that theory. The authors introduce Convolutional Newton-Schulz (Conv-NS), which approximates the polar factor in convolutional operator geometry while preserving kernel support. On CIFAR-10 and ImageNet classification, Conv-NS and reshape-based Muon are similarly efficient and achieve comparable accuracy. Despite better theoretical alignment, Conv-NS does not outperform, and the authors hypothesize that exact convolutional orthogonalization may overconstrain updates.
Thibaut Boissin (IRIT), Thomas Massena (IRIT, DTIPG - SNCF et al.
In briefAgentPrivArena: Evaluating and Auditing Privacy Risks of Real-World AI Agents
SummaryThis paper presents AgentPrivArena, a framework for testing privacy risks in realistic LLM agent workflows, using real MCP tools and self-hosted services in a reproducible environment. The authors propose trajectory-level privacy metrics that measure unnecessary information access beyond leakage in the final response. They also introduce AgentPrivAudit, a runtime method for monitoring privacy violations during agent execution. Experiments on state-of-the-art LLM agents reveal privacy risks that existing evaluations overlook, underscoring the value of trajectory-level auditing for trustworthy deployment.
In briefNormality Constraint Learning: Adapting Foundation Models for Time Series Anomaly Detection
SummaryTime series foundation models generalize well but can also reconstruct or forecast rare anomalies accurately, which weakens error-based anomaly scores. This paper proposes NCL, a lightweight plug-and-play framework that builds a normality subspace from a few normal patch features and steers features toward it using contrastive constraints. The calibrated features are fused with the original model output to widen the gap between normal and abnormal observations. Experiments across several model families and benchmarks report consistent improvements in anomaly detection.
In briefMultimodal safety evaluation should also measure controllability, not just classification.
SummaryThis paper argues that current safety evaluations of vision-language models rely on input and output classification, which does not show whether internal safety states can be accessed or controlled. The authors propose reporting a controllability profile alongside behavioral classification, covering representation-level detectability, cross-modal specificity, intervention sensitivity, and benign-preserving selectivity. Using implicit toxicity as a test case, experiments on LlavaGuard and Qwen3.5 with sparse feature decompositions show that internal readout and controllability can diverge.
In briefEMG-FM-Bench: A Benchmark for Transferring Foundation Models to Electromyography
SummaryThis paper introduces EMG-FM-Bench, a benchmark for testing how well foundation models transfer and adapt to electromyography (EMG) signals. It unifies 20 public datasets with over 1 million EMG segments and evaluates nine pretrained models. The authors find that full fine-tuning can substantially change downstream performance, the benefit of pretraining is not universal, and performance drops for new users, with five-shot adaptation recovering only part of the loss.
In briefTime-series foundation models for predictive control and the role of control excitation
SummaryThis paper examines whether time-series foundation models (TSFMs) can support model predictive control. Using residential heat-pump control as a test bed, it measures how well predicted effects of control interventions match ground-truth effects. The authors find that TSFMs recover the system's input-response relationship when the context contains sufficient independent control excitation, and that low forecast error alone does not guarantee this.
Mazen Amria, Jasper Hoffmann, Philipp Bordne et al.
In briefThe Assistance Dilemma: Teaching LLMs via Multi-Turn RL
SummaryLanguage models trained to answer questions are poor tutors, and existing reinforcement learning methods can reward simply telling students the answer. The paper proposes a masked near-transfer post-test reward and two binary reward gates to encourage real teaching. The authors train Eduardo tutor models from 4B to 27B parameters and open-source the environment, dataset, and models.
WhyIntroduces masked near-transfer post-test and binary reward gates replacing penalties to curb answer telling, validated across models and benchmarks.
In briefChoosing energy-efficient software designs for building fault diagnosis
SummaryThe building sector accounts for about 30% of global energy use, and much of it could be saved by repairing existing faults. This paper proposes an energy-consumption model that accounts for both the diagnostic performance of fault detection and diagnosis software and the energy the software itself uses. In a simulation using performance and energy values from prior literature, it compares rule-based, model-based, classical machine learning, and large-language-model-based architectures. Results show that accounting for computational energy and accuracy changes which approach gives the most benefit: efficient methods like random forest save most in smaller buildings, while resource-intensive fine-tuned large language models become advantageous as building size grows.
In briefARO: A Representation Learning Model Aligning Multi-Omics Data
SummaryThis paper proposes ARO, a model that learns representations from multi-omics cancer data and can reconstruct missing and unpaired data modalities. Unlike larger, more complex foundation models, ARO is designed for practical use when data is limited or incomplete. The authors report a mean squared error of 0.15 for reconstructing missing modalities on validation and test data in the unmasked setting. The learned latent embeddings also support a downstream cancer classification task.
In briefBetter Call Reward: How a legal AI learned to avoid giving clear answers
SummaryThe paper fine-tunes Qwen3-8B with reinforcement learning using a reward based on citation count, legal jargon density, and response length. Instead of reasoning better, the model learned to withhold direct answers, and its accuracy on 16 yes-or-no LegalBench tasks fell from 0.500 to 0.072. The authors report that 89.3% of post-training citations are implausible hallucinations and propose three diagnostic metrics to detect this failure before deployment.
In briefHeuFouFT: Task-guided search for which frequency coefficients to fine-tune
SummaryThis paper proposes HeuFouFT, a method that chooses which frequency coordinates to train in Fourier fine-tuning based on downstream task performance rather than fixed rules. It initializes three metaheuristic optimizers from a coarse intensity map and uses a Random Forest to keep only the top 30% of candidates for proxy fine-tuning. On E2E with GPT-2-Medium, all three variants beat random-uniform FourierFT, Gaussian band-pass FourierFT, and LoRA on five metrics, and PSO also beats LoCA on four metrics using 37.6% fewer trainable coefficients.
In briefKESurv: A kernel ensemble method for predicting each patient's survival curve
SummaryThis paper proposes KESurv, an ensemble method that uses a Survival Forest as its main model alongside several base models. It adds the Beran estimator, a kernel-based method, to improve patient-specific survival curve predictions. Tested on four healthcare datasets, the method outperformed baseline models in calibration and ranking on most of them.
In briefDeciding when to stop iterative AI proposal loops while controlling false acceptances
SummaryMany agentic workflows use a generator-verifier loop, where a generator proposes candidates, a cheap verifier scores them, and the loop stops once a proposal is judged good enough. Because the verifier only proxies a costlier ground-truth check, adaptive search can accumulate false acceptances. The paper studies when to stop such loops while controlling the false discovery rate of accepted proposals, introducing distribution-free testing and conformal risk control tools. The approach is validated on synthetic settings and a protein-design benchmark.
Mahmoud Hegazy, Michael I. Jordan, Aymeric Dieuleveut
In briefDo Speech Representations Keep Regional Accent Between Read and Spontaneous Speech?
SummaryThe study uses RVG1, with 500 German speakers from nine regions, to compare ten speech representations on regional classification and continuous geolocation. Whisper performs best under matched conditions but degrades sharply when transferred between read and spontaneous speech, while speaker embeddings are less discriminative in-domain yet more robust under transfer. The authors conclude that strong matched-condition performance does not indicate robust regional information.
Paula A. Perez-Toro, Tomas Arias-Vergara, Annette Schwarz et al.
In briefA systematic review of semantic decision engines in network control loops
SummaryThis paper reviews 139 paper families by decision interface, execution path and check ownership. It finds that most claims of fitting a control loop or time budget lack matched measurement support, with the largest gaps where decisions lack a deterministic computation step. The authors propose a minimum reporting record, design rules and a research agenda for admitting decision engines into control loops.
In briefWhen to Remember, When to Abstain: Category-Conditioned Memory Retention for Reliable AI Agents
SummaryThe paper studies whether an AI agent's persistent memory should decide what to keep using a confidence threshold that depends on the category of each assertion, rather than one global threshold. In a cold-start memory pipeline evaluated on 100 synthetic personas, value and belief assertions were supported by their sources only 77.9% of the time, compared with 96.2% for other categories. A stricter bar on value assertions reduced unsupported retentions from 6.2% to 4.0% while preserving more coverage than a global threshold at comparable retention.
In briefDetecting Noisy Keypoint Labels for Reliable Infant Pose Estimation
SummaryThis paper addresses errors in manually annotated keypoints used to train infant pose estimation models for analyzing preterm infants' movements. It proposes a framework that trains a hybrid convolutional-attention model to predict each keypoint's anatomical category, then uses the resulting cross-entropy training dynamics and unsupervised clustering to separate clean and noisy keypoints. On the NeoPose dataset of 65 hospitalized preterm infants, it reaches an F1-score of up to 91.9% for noisy-keypoint detection. On the COCO benchmark, filtering detected noisy keypoints improves downstream pose estimation by up to 7.4 AP points at moderate-to-high noise levels.
Emanuele Cardinale, Sara Moccia, Alessandro Cacciatore et al.
In briefTraining-free Transformer merging by aligning operators layer by layer
SummaryThe paper introduces Sequential Local Operator Alignment, a training-free method for merging fine-tuned transformers. It uses calibration data to align each functional component in order along the partially merged model's execution path, then factorizes the merged operators back into valid transformer parameters. The authors report that this sequential alignment reduces error accumulation across layers and that the factorization allows optional rank expansion for multi-task capacity. They state the approach works across CLIP, RoBERTa, billion-parameter LLMs, and LoRA-fine-tuned models, improving over strong merging baselines.
Akansh Maurya, Ya-Wei Eileen Lin, Stefanie Jegelka et al.
In briefSpatialChain: a benchmark for checking whether vision-language models reason faithfully about space
SummaryThe paper argues that correct final answers from vision-language models on spatial benchmarks do not prove faithful reasoning, since shortcuts may explain them. The authors introduce SpatialChain, a dataset pairing spatial questions with scene-graph-grounded reasoning chains, and a two-axis evaluation using chain-overlap metrics and an LLM judge for faithfulness and completeness. Testing nine thinking-enabled models, they find several models with high accuracy but many unfaithful correct answers, and report that fine-tuning on the dataset improves one model.
WhySeparates reasoning faithfulness from final-answer accuracy, validating the judge against human annotations and a second provider's judge.
Rafael Teixeira Sousa, Vin\'icius Paulo Lopes de Oliveira, Elisa Ayumi Masasi de Oliveira et al.
In briefFrom Benchmark to Bench: Can AI Agents Handle Real Drug Discovery?
SummaryThe paper introduces MAGI, an open modular agent that sets objectives, runs and monitors molecular optimization, interprets structure-activity relationships, and adjusts its strategy. Tested on nine retrospective lead-optimization campaigns from three pharmaceutical companies, it found that campaign success depended on the predictive models' applicability rather than the generation route. In a blinded evaluation, chemists could not distinguish agent proposals from expert-made compounds.
In briefFlashCart: Fast Cartesian tensor products for equivariant interatomic potentials
SummaryFlashCart is a method for machine-learned interatomic potentials that makes higher-order atomic correlations cheaper to compute. It combines generated GPU kernels with an architecture that builds equivariant features recursively and compresses them to a fixed width at each step. On SPICE-MACE-OFF, a 5.6 million parameter FlashCart model achieves lower energy and force errors and 10 times faster inference than a 189 million parameter transformer.
Viktor Zaverkin, Payman Goodarzi, Sergey V. Sukhomlinov et al.
In briefWhat Did the Agent Actually Do? Evidence-Grounded Oversight for Long-Horizon Agents
SummaryThis paper studies monitors that help users oversee autonomous agents on long-horizon tasks by identifying consequential decisions and locating supporting evidence. The authors introduce AgentMonBench, a software-engineering benchmark, and propose the Evidence-Grounded Behavior Graph (EBG), a training-free method that groups evidence into behaviors and organizes them in a graph. Experiments across eight models show EBG improves decision identification and evidence localization in most settings.
In briefMeasuring how errors grow across multi-step reasoning
SummaryThis paper studies the stability of multi-step reasoning in language models, such as chain-of-thought, where errors from intermediate steps can accumulate and grow. The authors derive an error bound based on the product of spectral norms of input-space Jacobians across steps, which serves as a stability measure. They propose controlling stability through shorter chain-of-thought lengths and quantization-aware training, reporting average gains of 3.5% and 8.2% on longer inputs across seven evaluations.
In briefVision-based swarm robot navigation using deep reinforcement learning and evolutionary design
SummaryThis paper proposes a method that combines multi-agent reinforcement learning with neuro-evolutionary strategies to automatically design swarm robot controllers that navigate using only monocular camera images. It optimizes a pre-trained individual navigation policy with the cross-entropy method and CMA-ES, using a compact network for low-cost, resource-constrained robots. In high-fidelity simulation, the cross-entropy controller covered 36.20% more regions than the CMA-ES controller, and the best vision-based policy matched distance-sensor methods statistically while reducing average energy consumption by 31.40%.
\'{A}lvaro D\'{i}ez (Department of Computer Science, Artificial Intelligence, University of Alicante) et al.
In briefCVIF: A Criticality-Driven Visual Intervention Framework for Improving Geometric Diagram Understanding in Multimodal LLMs
SummaryMultimodal large language models often struggle with geometric diagrams because they rely on textual priors that can conflict with visual evidence. This paper proposes CVIF, a training-free inference-time method that intervenes at critical layers to select vertex-centered visual evidence and redistribute attention toward it. On PGPS9K and PGDP5K, Overall F1 rises from 77.85 to 85.58 and from 75.23 to 82.84, respectively.
In briefT-CCL: Resource-Efficient Collective Communication Using the Tensor Memory Accelerator
SummaryLarge transformer models running on multiple GPUs need frequent collective communication, which often consumes many streaming multiprocessor resources. This paper presents T-CCL, an intra-node collective communication library that offloads data movement and reduction to the Tensor Memory Accelerator and runs collectives as pipelined asynchronous operations. The authors report up to 2.4x speedup over NCCL, up to 3.42x under restricted resource budgets, and up to 1.31x higher end-to-end inference throughput when integrated into vLLM.
Keyvan Dadashzadeh, Yuehong Zhou, Minyu Cui et al.
In briefLearning Pareto stationary fronts in one backpropagation pass
SummaryThe paper proposes MOSEL (Multi-Objective Stackelberg Efficient Learning), a framework for a posteriori multi-objective optimization in deep neural networks. It recovers a full front of Pareto stationary solutions at roughly the cost of standard single-objective training. The problem is cast as a bilevel optimization and solved as a Stackelberg game in a single forward-backward pass. Experiments report diverse Pareto frontiers in strongly conflicting settings such as fairness-accuracy, and solutions closer to the utopia point in weakly conflicting multi-task learning.
Elina Rojin Celik, Marcos Medeiros Raimundo, Isabel Valera
In briefTraining-free human-object interaction detection using multimodal large language models
SummaryThis paper addresses human-object interaction (HOI) detection, which locates human-object pairs and recognizes their interactions. The authors propose HarnessHOI, a training-free framework that turns multimodal large language model inference into an interaction-centered, active process. An interaction-guided perception mechanism finds missing participants and refines ambiguous evidence, and a relation-agnostic geometric adjudication module reconciles multi-source evidence. Experiments on HICO-DET and V-COCO reportedly achieve state-of-the-art results among training-free methods.
In briefLearning super-resolution and segmentation for Earth imagery when each image has only one type of label
SummaryThis paper studies a multi-task learning setup for super-resolution and semantic segmentation on Earth observation data, where each example has annotation for only one task. The authors compare sequential and shared multi-task architectures and propose a hybrid variant plus a re-projection loss to share representations and keep super-resolution image quality during segmentation training. Experiments reportedly show favorable results against the state-of-the-art sequential variant.
Ho\`ang-\^An L\^e, Minh-Tan Pham, Solange Lemai-Chenevier et al.
In briefScaling Down LLM Scaling Laws for Resource-Constrained Training
SummaryThis review surveys how large language model scaling theory has developed, from early empirical scaling laws to compute-optimal training. It focuses on parameter efficiency, token utilization, data efficiency, and training in resource-constrained settings. The authors argue the field is shifting from maximizing scale toward more deliberate allocation of parameters, tokens, compute, and hardware. They also note that whether principles established on enterprise infrastructure transfer to smaller models and constrained environments remains an open question.
In briefSteering by Influence: Curvature-Aware Data Weighting for Controlling Language Models
SummaryThis paper proposes a method for controlling language model outputs at inference time by shifting internal activations toward a concept. It uses influence functions, which account for the curvature of the loss landscape, to identify training examples that most thematically express a concept, then applies optimal transport to move non-concept activations toward these weighted examples. The method outperforms existing activation-transport baselines on toxicity suppression, concept induction, and truthfulness tasks, while perplexity and MMLU accuracy indicate model quality is largely preserved.
James A. E. Dixon, Stephen J. Roberts, Francesco Quinzan
In briefMTOR: Detecting AI-generated videos using text cues and unusually steady motion
SummaryMTOR is a method for detecting AI-generated videos that combines global visual features with text semantics derived from video captions. It also models a pattern the authors call temporal over-regularity, where synthetic videos show stronger temporal persistence and lower temporal variability. The authors report state-of-the-art overall performance across five benchmarks covering 46 generator variants, against 16 baselines, and strong resilience to twelve real-world video perturbations.
WhyCombines caption semantics with temporal over-regularity; beats 16 baselines across five benchmarks and 46 generators with robustness tests, giving solid evidence.
In briefCrop disease image datasets let models identify each photo's shooting session from shared sensor readings
SummaryThis paper argues that reported gains from combining environmental sensor data with leaf images for crop disease classification may come from how the datasets were built. Images taken in the same session share one sensor reading, so models can predict disease by recognizing session identity. Testing two Korean datasets, the authors found that an image-free classifier using only timestamps matched or exceeded sensor-based models. They recommend evaluating on session-held-out splits and comparing against sensor-free date-time baselines.
In briefMultimodal survival analysis model for mapping sinkhole risk
SummaryThe paper proposes a multimodal Cox proportional hazards model to predict sinkhole susceptibility in karst terrain. It uses modality-specific encoders and a cross-modal fusion layer to integrate heterogeneous data, and treats unreported locations as right-censored rather than negative. A statewide Florida case study with spatially blocked validation and ablation studies shows the method ranks risk effectively and produces a susceptibility map.
In briefCapability-driven self-evolution of agent memory (PrisMem)
SummaryThis paper proposes capability-driven evolution, which optimizes agent memory programs along individual capability dimensions rather than overall performance. The PrisMem method combines dependency-aware capability selection, history-guided diagnosis, and trace-guided integration of paired differential cases. Experiments report gains of 10.54 and 7.83 percentage points over the strongest baselines on BEAM-1M and LongMemEval-M.
In briefOntology concept overlap as a graded reward for clinical question answering reinforcement learning
SummaryThe paper proposes a reinforcement learning post-training method for clinical question answering that uses UMLS concept identifier overlap as a graded, externally specified reward. It combines this with an LLM judge and a consistency penalty within GRPO. On MedQA and PubMedQA, the composite reward improves over supervised fine-tuning on Phi-3-mini and Llama-3.2-3B across reported metrics.
In briefLatent Similarity Gaussian Processes for Personalized Suicide-Risk Forecasting
SummaryThis paper introduces Latent Similarity Gaussian Processes (LSGPs), which embed patients in a continuous latent space to jointly model similarity and forecast risk. The model selectively draws information from latent peers to better capture individualized risk trajectories. Experiments on intensive longitudinal suicide data report that LSGPs outperform nomothetic, idiographic, and hierarchical models for next-week risk forecasting, with the largest gains for first-occurrence events.
In briefGraphDecisionBench: A Benchmark for System One Decision Models on Graph Tasks
SummaryThis paper introduces GraphDecisionBench, a model-independent benchmark for evaluating how well System One models, which select directly from supplied options, perform on graph tasks. It compares Jev and related choice-based models with language-model baselines across fourteen model-interface configurations. The results show that accurate adjacency recognition does not guarantee overall structural correctness, that joint graph-text input does not consistently improve prediction, and that feasible construction does not imply high solution quality.
In briefKineWorld: Action-Induced Transport Fields for Embodied World Modeling
SummaryThis paper presents KineWorld, a world-modeling framework that predicts the visual results of robot actions before they are executed. It builds camera-aligned transport fields from commanded robot motion (Kinematic Transport Lifting) and uses them to reweight future-RGB flow matching so that spatially sparse but interaction-critical changes receive more emphasis. The model is trained on ALOHA-AgileX bimanual manipulation data from RoboTwin 2.0, reaching an EWMScore-P of 68.95 in single-view evaluation and a TWB-Score of 54.82 in multi-view evaluation.
In briefIGA-KAN: Isogeometric analysis with closed-form Kolmogorov-Arnold networks for solving forward and inverse PDEs
SummaryThis paper proposes IGA-KAN, which improves an isogeometric analysis (IGA) solution by fitting local Kolmogorov-Arnold networks in closed form on each knot-vertex patch, rather than replacing the IGA solver. On eight benchmarks, including one posed on a domain fitted to an MRI brain slice, it reduces error substantially at the same number of Galerkin unknowns. In an inverse problem recovering a constant source from one noise-free observation, it is 167 times more accurate than IGA. The authors attribute the gain to superconvergence of local averages of the Galerkin solution.
Sima Naraghi, Kourosh Parand, Amirhossein Sadr et al.
In briefImproveAnyTask: An autonomous framework that lets language models repeatedly improve their own task performance
SummaryThe paper introduces ImproveAnyTask, an autonomous post-training harness that improves large language models on specific tasks under a limited compute budget. It organizes adaptation into error attribution, update-direction selection, and executable model updates, drawing on research-backed strategies. Across 11 tasks, it reports mean gains of 18.29 and 11.97 percentage points on the Base and Instruct models, under a 24-hour budget with resources equivalent to eight H20 GPUs.
In briefBreaking Bureaucracy: Evaluating Open-Source LLMs for Legal Document Review
SummaryThis paper tests open-source generative language models on legal natural language inference using the ContractNLI benchmark and two NLI4Wills datasets. The authors reproduce a supervised baseline and compare several models on invalid-output rates and stability across temperatures and domains. Gemma-4 26B performs best among the generative models at 81.2% accuracy, beating the supervised model on one metric, though zero-shot approaches cannot exceed the supervised model on accuracy. The authors conclude that zero-shot open-source models are a viable option for real legal NLI tasks when no labeled data is available.
In briefStability-Based Deep Graph Learning to Counter Over-Smoothing
SummaryDeeper graph neural networks enlarge their receptive field but tend to make node representations converge, a problem called over-smoothing. The paper treats layer depth as time and interprets over-smoothing as undesirable synchronization of features, analyzed with the master stability curve. It proposes SDGL, which uses controlled Turing instability or near-critical stable propagation to mitigate over-smoothing. Experiments on diverse node- and graph-level benchmarks report better depth scaling and accuracy gains over strong baselines.
In briefMeSD: Multi-Evidence Self-Distillation for Video Language Models
SummaryThis paper proposes MeSD, a training framework for video large language models that uses three teachers conditioned on the ground-truth answer, temporal evidence, and spatial evidence to give token-level guidance. It introduces Verification-Guided Optimization, which sorts trajectories into success, failure, or indeterminate cases and adjusts updates accordingly. Experiments on multiple video benchmarks report consistent gains over reinforcement learning and self-distillation baselines.
In briefBabelFake: A Multilingual Audio-Visual Deepfake Detection Benchmark
SummaryThis paper introduces BabelFake, a multilingual audio-visual deepfake benchmark built with consenting participants. It contains about 399k clips (1,323 hours) from 496 people speaking five languages. The clips are generated by pairing 11 video manipulation methods with 4 voice cloning engines. Tests of state-of-the-art detectors show that detection difficulty depends on the audio-visual pairing, and human perceived realism does not always match machine detection difficulty.
In briefSPIN: Protecting Images From Diffusion Editing With Hidden Perturbations
SummaryThis paper proposes SPIN, a method that adds imperceptible perturbations to an image so that diffusion-based instruction-guided editing is disrupted. It generates stochastic neighboring denoising states under the same instruction and predicts their clean latents via one-step projection, avoiding full trajectory unrolling. Experiments on two image editors show SPIN outperforms compared methods on all six metrics for both seen and unseen instructions.
In briefRobust LLM post-training via adaptive data-source sampling with minimax regret
SummaryThis paper studies how to train large language models on several data sources when only mini-batch bandit feedback is available. The proposed DUCB-OGD method pairs a discounted upper-confidence-bound sampler that picks data sources with an online gradient descent optimizer that updates the model. The authors prove a dynamic minimax regret bound of O~(K^{1/4}T^{3/4}) and report experiments showing improved worst-group robustness with little added computation.
Chengbo Zang, Haoyu Dong, Mehmet Kerem Turkcan et al.
In briefDual autoencoders that help low-cost robots navigate after training in simulation
SummaryThis paper addresses vision-based navigation for low-cost robots, where policies trained in simulation often fail in the real world. The authors combine domain randomization with feature-level domain adaptation using a dual convolutional variational autoencoder with a shared decoder, trained on 45,225 simulated and 4,556 real images. They report an average success rate of almost 91% on real-world indoor navigation image classification and a real robot deployment.
\'Alvaro D\'iez (Department of Computer Science, Artificial Intelligence, University of Alicante) et al.
In briefFine-tuning a 3B-parameter language model on a phone, and what sustained training does to it
SummaryThis study measures a 3-billion-parameter language model fine-tuned on a smartphone, recording memory use, per-step time, heat, and energy. It reports that an iPhone 17 Pro can fine-tune the model for a typical user within one battery charge, with adapters that personalize as well as server-trained ones. Sustained training throttles the phone to about half its initial throughput, and the pause and burst schedules tested did not recover it. The authors also repaired a never-dispatched, incorrect kernel in Apple's MLX, making adapter training 1.47x faster on about a third less energy.
In briefReadout Blindness: VLM Scores Miss Directional Space Information Kept by Frozen Encoders
SummaryCLIP-style vision-language models score near chance on directed spatial relations such as "left of." The authors show that symmetric scoring rules cancel direction regardless of encoder training. They propose Antisymmetric Displacement Readout (ADR), which needs no additional training and scores relations by signed displacement between matched object centroids, and it substantially outperforms deployed scores.
In briefAgentic, schema-guided extraction of materials process knowledge from research papers
SummaryThis paper presents SciKGExtract, a framework that combines large language model extraction, chemical name normalization, and agent-based evaluation and refinement to pull process knowledge from atomic layer deposition papers. It is tested on 176 papers about zinc oxide and IGZO, plus an expert-annotated full-schema subset. PubChem normalization improved exact-match F1 for every model, and agentic refinement raised the best ZnO F1 from 0.591 to 0.805. IGZO remained harder, with a best F1 of 0.344.
In briefCRAFTER: A causality-based method for making autonomous IoT systems adapt themselves
SummaryThis paper presents CRAFTER, a framework for designing and deploying self-adaptive IoT systems using Causal Reinforcement Learning. It generates causal graphs that reveal dependencies among system components and uses these cause-effect relationships to guide adaptation decisions. The experiments show that CRAFTER can derive causal graphs across diverse IoT use cases and improves self-adaptation performance by 25% over state-of-the-art reinforcement learning approaches.
Houssam Hajj Hassan (IP Paris, SAMOVAR), Ajay Kattepur et al.
In briefWiring Matters: How to Connect Affordance Heads in Vision-Language-Action Policies
SummaryThis paper studies how to attach an affordance prediction head to a vision-language-action (VLA) policy, tested on the LIBERO benchmark. The authors read the backbone through a stop-gradient and re-inject intermediate head features into the action expert via a learned bridge; letting affordance gradients reach the backbone lowers performance below the headless baseline (85.5% vs. 93.1%). With the backbone protected, bridge initialization matters most: the best configuration, an actively initialized residual bridge, reaches 96.2%, close to the more complex AffordanceVLA (95.8%) with under 1% extra parameters.
In briefRollPlace: Improving Chip Macro Placement with Monte Carlo Rollout Search
SummaryThis paper proposes RollPlace, a macro placement framework for chip design. It first generates initial layouts using machine learning or heuristic methods, then refines them by adjusting selected macros. The framework integrates Monte Carlo Tree Search and a rollout mechanism for local search. Experiments report that it outperforms state-of-the-art methods on the ISPD 2005 benchmark and performs well across 19 benchmarks in OpenROAD end-to-end evaluation.
In briefSPDAlign: Interpretable Riemannian Alignment for EEG Signal Shifts
SummaryThis paper addresses distribution shifts across sessions and subjects that limit EEG-based brain-computer interfaces. It proposes SPDAlign, an unsupervised framework that needs no costly labeled calibration data. The authors show theoretically that these shifts can be recovered through linear transformations on the Symmetric Positive Definite manifold, and the method aligns domain means and rotations using Wasserstein Procrustes. Experiments on public EEG datasets show competitive performance, and the linear design is intended to be interpretable.
In briefMeasuring AI Inference Cost: A Framework for Enterprise Agent Workloads
SummaryThe paper argues that LLM use is shifting from single-turn completion to agentic, multi-step trajectories, where sequential dependence makes per-token decode latency the dominant factor in task completion time. The authors propose a four-layer evaluation framework covering silicon, serving systems, agent episodes and enterprise use, built on goodput at an agentic SLO and cost per successful episode. They suggest that token-throughput rankings may diverge from cost-per-successful-task rankings on long-horizon work.
Abbas Raza Ali, Muhammad Ajmal Siddiqui, Moona Zahid
In briefDimension-dependence of midpoint methods for Langevin sampling
SummaryThis paper studies deterministic and randomized midpoint discretizations of Langevin dynamics for sampling from a target distribution proportional to e^{-V}. It gives upper bounds on the number of gradient queries needed for the deterministic Heun method and the underdamped exponential midpoint method, using Malliavin calculus techniques that improve on earlier coupling-based bounds. The authors also show that randomized midpoint methods require at least d^{1/3}ε^{-1/3} queries, suggesting deterministic methods can outperform randomized ones in high dimensions.
WhyUses Malliavin calculus and stationarity cancellation to sharpen dimension and accuracy bounds for midpoint discretizations, with matching lower bounds.
In briefTeaching a Minimalist Machine to Discover Recursive Programs for Arithmetic
SummaryThis paper proposes the Minimalist Machine, an inductive program synthesis framework that searches a compact Prolog subset of body-free facts and two-body conjunctive Horn clauses. Recursion is not built in through a dedicated metarule but emerges when a target predicate is reused in a learned clause's body. Following a primary school curriculum, the system starts with an empty knowledge base and is taught arithmetic concepts in a human-curated sequence. It learns recursive programs for addition, subtraction, multiplication, and division, producing a transparent inductive reasoning trace.
Dominik Magiera, Christiane Wiebel-Herboth, Frank J\"{a}kel
In briefSmall language models for classifying smart data models at the edge
SummaryThis paper examines whether lightweight open-source language models can match input data entities to the best-fitting Smart Data Model (SDM) under limited computing resources, such as edge devices in IoT settings. It benchmarks general-purpose, reasoning-specialized, and code-specialized models across several domain-specific datasets. A comparison with low-cost baselines, TF-IDF and a lightweight sentence encoder, provides a reference point for judging the practical value of LLM-based classification on edge platforms.
Cristian Martella, Angelo Martella, Antonella Longo et al.
In briefDialectSentEval 2026: Shared Task on Arabic Dialect Sentiment Analysis and Swapping
SummaryThis paper presents the DialectSentEval 2026 shared task, hosted with ArabicNLP 2026, on sentiment analysis across Arabic dialects. It has two subtasks: Subtask 1 asks models to classify sentiment polarity across multiple Arabic dialects, and Subtask 2 is a generative task that requires inverting sentiment polarity while preserving core meaning. The overview covers the task's motivation and dataset creation, and summarizes the main findings from participating models.
In briefSmart Content Ingestion for Generative AI: A Configurable, Measurable Document Extraction System
SummaryThe paper argues that enterprise knowledge is stored in mixed formats such as PDFs, tables and scans, and that extraction errors at this stage cannot be fixed by downstream retrievers or language models, so content extraction should be treated as its own lifecycle stage. The authors present a production-ready system with selective OCR routing, a scarcity-first curation engine, a structure-aware chunker and a read-only retrieval evaluator. On a 180-document corpus the best extractor scores 97.4 out of 100 (character error rate 0.13%, table similarity 0.995), and the chunker reaches hit@1 of 68.6%, hit@10 of 92.8% and MRR 0.77 over 25,050 generated questions.
Abbas Raza Ali, Muhammad Ajmal Siddiqui, Moona Zahid
In briefVepAgent: Tool-augmented reinforcement learning for predicting future video events
SummaryThis paper proposes VepAgent, an agent framework that combines causal-transition reasoning with tool-augmented reinforcement learning for Video Event Prediction. The authors build FutureBench-4K, a chain-of-thought dataset for supervised fine-tuning, and a tool library for state tracking, frame retrieval, and region magnification. Evaluations on FutureBench and NEPBench report state-of-the-art results, outperforming larger multimodal large language models.
In briefGAMBIT: Learning to Plan Continuous Multi-Robot Trajectories
SummaryThis paper presents GAMBIT, a framework for planning trajectories for multiple robots in continuous spaces. It first learns coordinated motion-primitive selection through imitation learning, then fine-tunes the policy with reinforcement learning, and uses a safeguarded rollout with backup trajectories to keep execution collision-free. Experiments report that GAMBIT outperforms centralised and decentralised baselines and coordinates over a thousand robots with planning latency below a few hundred milliseconds.
Rishabh Jain, Akmaral Moldagalieva, Lorenzo Magnino et al.
In briefTrajectory-guided tokenization for Wi-Fi signal sensing
SummaryThis paper proposes Trajectory-Guided Tokenization (TGT), a method for converting complex-valued Wi-Fi channel state information into compact tokens while preserving temporal changes. It uses an orthonormal Helmert transform and asymmetric attention to build the tokens. On the authors' self-collected dataset, TGT with TokenMLP reaches the highest mean accuracy of 92.83% among the evaluated combinations. Experiments on EHUNAM and Widar are reported to support its use for presence detection and gesture recognition across domains.
In briefFrom Sentence Labels to Span Tagging: Pinpointing Abusive Text and Its Targets
SummaryThis paper defines Abusive Language Identification (ALI) as a sequence-labeling task that jointly extracts abusive spans and target mentions. Using a pilot corpus from a production moderation pipeline, it compares ALI with sentence-level classification on cross-domain generalization and implicit abuse. ALI remains competitive with classification while giving moderators localized outputs, with a modest advantage on implicit abuse. Exact abusive boundaries and target spans remain difficult to recover.
Nicolas Zampieri, Ignacio Lopez, Manon Girard et al.
In briefDeferKV: Delaying KV Cache Eviction Until the First Decoding Step
SummaryLong-context LLMs face large memory and inference costs from their growing KV cache. Existing one-shot compression methods irreversibly evict entries right after prefill, before any generation signal exists. DeferKV moves the eviction decision to the first real decoding step and combines prompt-side and decode-side attention observations. It needs no extra training, draft model, or future-query predictor, and it improves performance under compression on LongBench, RULER, and Needle-in-a-Haystack while keeping latency low.
In briefWhen concept bottleneck model explanations are faithful and compact
SummaryConcept bottleneck models are neural classifiers that explain decisions through high-level concepts, but their explanations are often derived heuristically. The paper argues explanations should also be faithful, meaning they do not misreport which concepts matter, and shows that for common CBM architectures, including VLM-based variants, faithful explanations must include all bottleneck concepts. The authors propose modeling concepts probabilistically and using group lasso sparsification to encourage compact faithful explanations, and they extend formal explainability algorithms to CBMs.
Stefano Teso, Emanuele Marconato, Steve Azzolin et al.
In briefdIon: A self-supervised method for learning from tandem mass spectra
SummaryThis paper introduces a new invariance for peptide tandem mass spectrometry data that links precursor mass and charge to fragment-ion evidence, enabling self-supervised learning without peptide sequence labels. The authors adapt the DINO framework into dIon, which uses two latent prediction tasks to learn representations. Under identical training, dIon initialization improves de novo peptide precision over training from scratch by 5.5 and 8.4 percentage points on the held-out test sets. On the Kingdoms corpus, the resulting models surpass fully supervised state-of-the-art de novo sequencing models.
WhyNovel precursor-based invariance adapted to DINO; consistent de novo precision gains over scratch training and supervised SOTA with ablations supporting design.
Alfred Nilsson, Joel Lapin, Samuel H. Payne et al.
In briefFuture-anchored verification and online recovery for world action models, a method to detect and correct robot execution drift
SummaryWorld action models predict how a robot task should unfold and then decode actions from that predicted future, but actions become invalid once execution drifts from the prediction. The FAVOR framework keeps the predicted frames as anchors, uses an Anchor Verifier to flag deviations that break the task, and uses a vision-language model to produce a short corrective instruction that returns the robot to its intended future. Without modifying the policy, task success rose from 97.85% to 98.10% on LIBERO and from 72.60% to 72.98% on LIBERO-Plus.
In briefWhat Can an AI Agent Change About Its Own Settings? A Safety Floor for Self-Configuring Runtimes
SummaryThis paper studies agent runtimes that let an agent edit its own configuration, separating ability-granting fields from fields that set limits. The authors propose allowing changes only to ability fields and enforce this as a containment floor in the configuration tool. Without the floor, a frontier model wrote a protected limit value in 25 of 72 ordinary requests; with the floor, all 167 protected writes were blocked. The study covers two models and one agent.
In briefPredicting Race and Ethnicity Probabilities from Group-Specific Name Lists
SummaryThis paper proposes the list-powered BISG (ℓBISG) method, which estimates the probability that a person belongs to a racial or ethnic group using their name and location when standard census name data is unavailable. It uses group-specific name lists, compiled by experts or generated by large language models, and applies a proximal inference correction. The method was validated on U.S. voter files, the 1900 U.S. Census, and the Lebanese voter registry, with LLM-generated lists producing accurate, well-calibrated probabilities.
In briefVLA-ZO: A Fast Zeroth-Order Adaptation Method for Vision-Language-Action Robot Models
SummaryThis paper introduces VLA-ZO, a framework that speeds up zeroth-order adaptation of vision-language-action (VLA) robot models to shifts encountered at deployment. It restricts adaptation to the action side, keeps the vision-language prefix frozen and reuses its conditioning states, and uses schedule-aware prefetching to hide state-transfer overhead. On LIBERO camera-viewpoint shifts, it cuts end-to-end adaptation time by 25.59× at q=16 and 32.54× at q=64 versus baseline zeroth-order optimization, and raises average task success from 48.27% without adaptation to 58.17% and 63.58%.
In briefDPNL: A DPLL-Based Algorithm for Probabilistic Neurosymbolic Learning
SummaryThis paper proposes DPNL, a framework for probabilistic neurosymbolic learning that trains end-to-end from final-output supervision without labels for intermediate concepts. Instead of fully materializing the logical provenance of a query, DPNL lazily explores intermediate assignments, with oracles resolving regions that can be certified to produce or exclude the target output. The authors also present ApproxDPNL, which adds early termination with certified bounds on the exact output probability. Experiments on several neurosymbolic tasks show that these methods handle a substantially wider range of problem instances.
Thomas Jean-Michel Valentin (ENS Paris Saclay, TYREX), Pierre Genev{\`e}s (LIG et al.
In briefEvolving in Thought Space: Test-Time Training a Small Model for Better Discoveries
SummaryThis paper introduces Guidance-TTT, which trains a compact guidance model at test time to propose high-level strategic changes, while a frozen, stronger execution model turns them into executable solutions. Each step selects a promising prior solution, applies a change, executes and verifies it, and updates only the guidance model with an adaptive group-relative RL objective. The authors report that, without web access, the method outperforms the best prior reported solutions across four domains (Polyomino Packing, AHC058, Lasso, and TriMul) and is competitive with public leaderboard results.
In briefRamp metering via hybrid deep reinforcement learning using partial connected-vehicle data
SummaryThis paper addresses congestion at freeway on-ramp merges by controlling ramp signals with a learning-based method. It combines aggregate traffic measurements with a grid encoding of sparse connected-vehicle presence and speed, and uses a Dueling Double Deep Q-Network to choose green durations. In SUMO simulations across 50 matched scenarios, the hybrid controller under partial connected-vehicle visibility reduced total travel time by 11.4% and mean spillback duration by 84.9% relative to ALINEA.
In briefCentriQ: Calibration-Free Quantization of Diffusion Transformers via Exact Mean Centering
SummaryThis paper presents CentriQ, a calibration-free method for quantizing diffusion transformers (DiTs) to 4 bits in weights and activations, reducing image generation cost. The authors find that data-free Hadamard rotation degrades DiT quality because adaptive layer-norm adds a per-token mean that a single dominant direction retains after rotation. CentriQ centers each token before rotation and restores the mean exactly through a rank-1 full-precision branch. Across three DiTs it matches calibrated SVDQuant at 4 bits and keeps usable image quality at 2-bit activations.
WhyIdentifies a structural cause of DiT quantization loss from adaptive layer-norm mean and proposes calibration-free exact centering, validated on three DiTs against calibrated baselines.
In briefOCL-PDE: A Generative Method for PDE Inverse Problems Using Complementary Hidden Features
SummaryThis paper addresses partial differential equation inverse problems, which are often ill-posed and struggle to recover fine-scale details. The authors propose OCL-PDE, a generative framework that pairs observations with a learned latent representation, letting the observation guide large-scale structure and the latent supply finer details. It is built on a physics-aware autoencoder and conditional Flow Matching, and supports both inverse reconstruction and forward PDE prediction. Experiments report improved reconstruction accuracy and fine-detail recovery compared with the evaluated baselines.
In briefSimAuthor: Using Foundation Models to Persistently Write Scientific Simulators
SummaryThe paper introduces SimAuthor, a framework that iteratively writes and revises executable scientific simulators using limited real observations, guided by distributional differences between simulated and real signals. It is evaluated on six biomedical tasks covering cardiac and respiratory audio, photoplethysmography (PPG), and electrocardiography (ECG). Under a fixed 100-attempt budget, the authors report that SimAuthor outperforms PUCT score search on all six tasks and achieves the highest endpoint score on five of them. The simulators are also reported to improve on unseen recordings and to yield out-of-distribution gains in downstream ECG classification.
In briefLipschitz Thinking: A Tutorial on Certifiable Robust Neural Networks
SummaryThis tutorial paper introduces the fundamental concepts of Lipschitz-bounded neural networks, which measure and limit a network's sensitivity to input perturbations. It explains the principles of Lipschitz-constrained layers and the mechanisms used to enforce their bounds. The paper shows how these networks provide robustness certificates with a single forward pass and discusses emerging and open research directions.
In briefHow shared stopping decisions alter answers in HQQ cache quantization
SummaryThis paper studies compression of the key-value cache used during language-model generation. In Transformers' HQQ backend, compression parameters are updated separately per request but share one average error to decide when to stop, so batching with other questions can change a four-bit answer. The authors remove this companion dependence with fixed iteration counts or request-local stopping, but neither repair shows an established quality advantage.
In briefGenerative world model for predictive control of laser melt pool depth
SummaryThis paper introduces a generative world model that predicts the evolution of a localized, highly dynamic laser melt pool from histories of temperature and phase morphology under candidate laser actions. Its generative latent dynamics capture effects of unresolved melt flow, yielding more accurate recursive rollouts than deterministic regressors under transient inputs. Because the learned model is differentiable, gradients through imagined futures are used to optimize laser schedules that regulate melt-pool depth on unseen geometries, paths and initializations. The optimization is further distilled into an amortized policy that outputs control actions in a single forward pass, offered as a proof of concept for machine deployment.
In briefFrom Papers to Mechanisms: An Evidence-Grounded Knowledge Base for Science AI
SummaryThe paper proposes an evidence-grounded mechanism knowledge substrate that organizes scientific literature into provenance-linked evidence units, role-typed entities, and directed mechanism paths. It instantiates this as MS$^3$, a Material-Sensor-Signal-System schema for conductive-fiber flexible sensors, built from 13,689 papers, 131,083 evidence items, and 26,648 mechanism objects. Across question-answering benchmarks, MS$^3$ retrieval improved macro-averaged scientific correctness, citation entailment, and answer completeness compared with closed-book generation, Web search, and Raw-PDF RAG.
In briefGoal-directed planar graph generation under constraints using grammar-based reinforcement learning
SummaryThis paper proposes a dataset-free method for generating planar graph embeddings by combining parametric graph grammars with safe reinforcement learning. The generation process is formulated as a constrained Markov decision process, with an action projection to improve constraint satisfaction during training. The authors also introduce a benchmark suite for constrained, goal-directed planar graph generation. The method is reported to outperform baselines on all benchmark tasks while satisfying the constraints.
Nicolas Hochuli, Lorenzo Miele, Kristina Shea et al.
In briefRoSA: Memory-efficient fine-tuning that trains rotating layer blocks
SummaryThis paper introduces RoSA (Rotational Sparse Adaptation), a method for fine-tuning large models by training only a subset of layers at a time. Lower layers near the input stay frozen, while a trainable block rotates across later layers, with the frozen region progressively growing. This reduces optimizer-state memory and shortens backpropagation. Experiments on several LLM architectures and tasks report lower peak memory while keeping strong fine-tuning performance.
In briefFew-shot prototype head adaptation for on-device ECG personalization on PSoC 6
SummaryWearable and bedside ECG monitors need to adapt to each patient's waveform, but personalization is usually done offline. The paper proposes adapting only a prototype (class-mean) classification head on microcontrollers, using a 1-D CNN trained on MIT-BIH and frozen as a feature extractor. Patient-specific adaptation needs just one forward pass per support beat and no backpropagation. Inter-patient macro-F1 improves at 1, 5, and 10 shots, and on-device results on PSoC 6 match the host.
Guilherme Silva, Pedro Silva, Gladston Moreira et al.
In briefUsing machine learning to measure how cognition and culture shape each other
SummaryThe paper argues that cognition and culture are co-constituted: individual subjectivities are trajectories through a high-dimensional cultural field, which is itself the aggregate of those trajectories. It proposes that embedding methods and large generative models can measure this relationship across individual, organizational, and societal levels. The authors also discuss generative AI agents as simulators of human subjectivity, introduce a Heisenberg-like uncertainty analogy for AI-influenced culture, and outline ethical questions and research guidelines.
Douglas R. Guilbeault, Bhargav Srinivasa Desikan, James A. Evans
In briefCertification-Enhanced Generalization Bounds: Formal Methods for Verifiable Learning Guarantees
SummaryThis paper uses formal methods to derive tight and sound generalization bounds for learning algorithms. It treats algorithmic stability as a specification to verify, using reachability analysis to obtain provable sample-specific bounds, and introduces a new concentration inequality to extend these to the expected generalization gap. Experiments from toy datasets to fine-tuning classification heads on large language models report guarantees orders of magnitude tighter than alternative sound computational approaches.
WhyRecasts algorithmic stability as a verifiable specification, linking reachability analysis and a new concentration inequality to generalization bounds tested up to LLM heads.
In briefEncoded but Not in Control: A Grounding Gap in Vision-Language Robot Policies
SummaryThis paper examines whether language-conditioned robot policies truly follow instructions or simply infer tasks from the scene. The authors keep the scene fixed while changing the instruction's target object, testing vision-language-action and world-action models in simulation and real-world experiments. Policies mostly approach and pick up the original scene target, even though instruction information is encoded, and the authors propose a diagnostic framework for measuring this gap.
In briefJoint class-time learning for video classification with multi-instance partial-label learning
SummaryThis paper studies multi-instance partial-label learning (MIPL) for video classification, where only bag-level labels with candidate classes are available. The authors propose a method that jointly assigns candidate classes and temporal evidence, using candidate-restricted inference and a structured teacher model. Benchmarks built from Breakfast, DoTA, and FineAction show the method outperforms existing MIPL algorithms in effectiveness and efficiency.
In briefAUTOPILOT: Perception, localization and path planning for self-driving cars using YOLOv7 and MiDaS
SummaryThis paper studies self-driving car perception by combining YOLO object detection with MiDaS depth sensing to locate obstacles, plus perspective transform and decision making for path planning. The authors evaluate the combined system through simulations and experiments. According to the paper, the YOLO and MiDaS combination yields a robust system for object detection and depth sensing.
In briefEstimating machining dynamics parameters with delayed effects and friction using physics-informed neural networks
SummaryThis paper proposes a multi-domain eXtended Physics-Informed Neural Network (XPINN) framework for nonsmooth delay differential equations. The time domain is split into subdomains based on integer multiples of the characteristic delay, and subnetworks are trained progressively while earlier parameters are frozen. Tested on a machining model with regenerative and nonsmooth friction effects, the framework outperforms a generic PINN (SPINN) in solution reconstruction and parameter estimation. The inverse version (I-XPINN) remains resilient when reference data contains Gaussian measurement noise.
In briefA data-centric review of plant disease datasets and their impact on precision farming
SummaryThis review critically examines visual and deep learning methods for plant disease detection, focusing on the datasets used to train them. It proposes a taxonomy based on acquisition setting, accessibility, plant diversity, disease composition, class structure, and imbalance, and compares laboratory and real-field datasets. The authors analyze how multi-level imbalance and limited environmental variability affect model generalization, and highlight integrating temperature, humidity, and leaf wetness with image data. The paper concludes by outlining challenges and future directions for multimodal disease monitoring in precision agriculture.
In briefSO(3)-RoPE: A position encoding that respects sphere geometry for transformers
SummaryThe paper introduces SO(3)-RoPE, a relative positional embedding that brings spherical geometry into transformer attention using unitary SO(3) representations. The method is SO(3)-equivariant and compatible with FlashAttention, keeping the efficiency of standard transformers. On shallow water dynamics prediction over a rotating sphere, the proposed SO3ViT reportedly outperforms an S2Transformer baseline with lower errors and reduced runtime.
Christian Libner, Chase van de Geijn, Alexander S. Ecker et al.
In briefAPOD: An AI agent that reasons about biology to discover population models of drug effects
SummaryThe paper introduces APOD, a language-model agent that iteratively reasons over biological knowledge and model fit diagnostics to discover a population digital twin, a shared differential-equation system that accounts for between-subject variability. On synthetic pharmacokinetic and tumor-dynamics benchmarks it recovered the ground-truth structure in 94-100% of runs, with a median speed 12 times faster than an established library-based search. On a real cohort, it produced a model of radioligand-therapy-induced platelet dynamics that predicts thrombocytopenia from first-cycle data.
WhyA language-model agent discovers population ODEs with inter-individual variability in an open search space, validated on synthetic benchmarks and real radioligand therapy cohorts with dosing simulatio
Romain Ferrara, Martin Soucail, Victor Gertner et al.
In briefLeAVJEPA: A Simple Architecture for Learning Audio and Video Together Without Labels
SummaryThis paper introduces LeAVJEPA, the first audio-visual encoder trained with LeJEPA's collapse-free objective. It uses a single Vision Transformer that handles audio, video, and combined inputs, and modality dropout treats a missing modality as another view of the same event. A controlled ablation identifies modality dropout as the key mechanism for audio-visual alignment.
Benjamin Robson, Santeri Mentu, Wenshuai Zhao et al.
In briefFrequency-decoupled diffusion guidance for known-kernel image deblurring
SummaryThis paper addresses non-blind image deblurring, where the blur kernel is known, using a pretrained diffusion model as an image prior for training-free posterior sampling. It separates frequency activation from degradation-induced attenuation, progressively activating measurement frequency bands from low to high while a kernel-derived attenuation map applies a selective spectral prior to inactive components. A local trajectory regularizer is added to stabilize sampling. Experiments on natural-image benchmarks report strong PSNR and SSIM results, including at higher measurement noise levels.
In briefCross-lingual Calibration of Success Probes for Multilingual LLM Routing
SummaryThis paper examines whether probes that estimate a language model's likelihood of answering correctly, using hidden activations before decoding, stay reliable across languages. Using 3,000 MATH problems in 10 languages and 8 open-weight model configurations, the authors find English-trained probes still discriminate across languages but become less well calibrated. Pooled multilingual training improves both discrimination and calibration. In routing experiments, the pooled router achieved a 0.7% higher test success rate and 13.0% lower modeled cost than always selecting the highest-average-success model.
Andrea Paganelli, Stefano Civelli, Pietro Bernardelle et al.
In briefLinUCB sampling allocation and optimal experimental design in the small-gap regime
SummaryThis paper studies how the LinUCB algorithm allocates its samples when reward gaps are of order at most n to the power of negative one half. In this regime, the empirical sampling distribution induced by LinUCB converges to the set of D-optimal designs, meaning its allocation is asymptotically efficient for learning the reward parameter. Building on this, the authors refine the leading-order constant of LinUCB's regret and show that the regularized least-squares estimator satisfies a central-limit-type theorem, enabling valid statistical inference.
WhyShows LinUCB's small-gap sampling distribution converges to D-optimal designs, linking regret minimization with experimental design; novel, but proof details are not shown.
