In briefFOAM: a lighter, more scalable variant of the SOAP second-order optimizer
SummaryThis paper introduces FOAM, a lightweight variant of the SOAP second-order optimizer for training transformers. It applies preconditioning only on the input side and replaces SOAP's dense rotated second moment with a factored estimator based on row and column statistics. According to the authors, FOAM matches SOAP and Muon validation loss in large language model pre-training from 124M to 760M parameters, uses half of SOAP's optimizer state, and runs at Muon-class wall-clock time.
In briefTokenStream: Multi-subject personalized video generation using token-wise LoRA and guided attention
SummaryThis paper presents TokenStream, an optimization-based framework that combines independently learned subject identities in text-to-video generation without needing reference-video training data. It uses Token-wise LoRA, which applies low-rank updates to subject-associated text tokens, and a T2V-guided stream that reuses the frozen base model's layout and motion priors. The authors also introduce Multi-DB, a benchmark of 30 subjects and 122 prompts. At 1.3B scale on Multi-DB, identity preservation rises from 0.422 (DreamBooth) to 0.467 and textual alignment from 0.652 to 0.688.
In briefHuman or Synthetic Preference Data? Cost-Matched Comparison for LLM Post-Training
SummaryThis paper compares human-annotated and synthetic preference data for post-training large language models, under a fixed, dollar-matched annotation budget. The authors train and evaluate several base model families on helpfulness and summarization tasks. They report that synthetic data is far more cost-efficient at low-to-moderate budgets, reaching high win rates for as little as $200 to $400. Human data closes the gap on helpfulness only at budgets roughly two orders of magnitude larger, and does not close it on summarization at any tested budget. The paper also finds that one widely used human preference source underperforms another.
Shourik Banerjee, Ethan Mendes, Mohit Raghavendra et al.
In briefUsing LLMs to Assist Formal Verification of Algorithms in Dafny
SummaryThis paper presents a ten-step workflow for using large language models to help verify scientific algorithms with the Dafny verification system. The steps cover formalization, contract writing, invariant validation, proof construction, profiling, and proof optimization. The authors used Claude Opus 5 and Opus 5.5 and evaluated the approach on 37 graph theory and labelled transition system algorithms, including the Hopcroft-Tarjan triconnectivity test. They suggest such workflows let human verifiers focus on supervision, specification review, and strategic guidance rather than detailed proof construction.
Wieger Wesselink, Kees Huizing, Huub van de Wetering
In briefHow training data statistics shape how language models learn to compare entities
SummaryThis paper studies how language models learn to compare entities, such as deciding which of two countries is larger, across nine comparison tasks covering five entity types. Using controlled experiments with small 163M-parameter models, the authors identify three learning phases: selecting entities by frequency, then by position in the prompt, and finally choosing the correct answer. They find these phases are causally driven by statistical properties of the training data, and that the same three phases appear in the open-source OLMo-2 32B model.
In briefProtein language models lose contact prediction accuracy when amino-acid letters are relabeled
SummaryThis paper tests whether protein language models learn the same invariant coevolutionary statistics as classical alignment-based methods. The authors apply a bijective relabeling of the amino-acid alphabet, which leaves alignment information intact and leaves Potts and inverse-covariance models unchanged. Across ESM-2, MSA Transformer, and AMPLIFY, contact maps degrade under relabeling, and the authors argue these models rely on canonical residue identities rather than pure coevolutionary statistics.
In briefContext Graph: A Traceable Record of Which AI Agent Decisions Caused Network Problems
SummaryThis paper addresses the difficulty operators face in tracing which decision by one of several autonomous agents (xApps and rApps) caused an alert or degradation in AI-RAN networks that share radio resources. It proposes the Context Graph, a heterogeneous directed graph that links policies, measurements, decisions, and alerts into a queryable explanation layer, with Decision-Event nodes acting as causal junctions. Outgoing effects are progressively confirmed through counterfactual verification. A proof-of-concept over existing E2, A1, and O1 data flows shows compliance and diagnosis queries resolved at sub-millisecond latency on graphs of tens of thousands of nodes, distinguishing true causal impacts from spurious correlations.
In briefStochastic Kernel Mode Decomposition: a method for inferring hidden regimes from signal modes under noise
SummaryThe paper introduces Stochastic Kernel Mode Decomposition (sKMD), which fits joint regime distributions on signed KMD mode coordinates and converts the reference mode posterior into regime probabilities and conditional reconstructions. The authors characterize when fixed mode energies keep the full regime posterior as noise variances vary, and extend the analysis to nonzero means and correlated noise. They also derive bounds linking feature compression, independently learned mode spans and sensor-noise error to log loss. Two controlled experiments with 64 independent panels each compare signed information against energy summaries and a strong signed control.
In briefDPA: A decentralized learning method for highly varied local data
SummaryThis paper proposes Decentralized Primal Averaging (DPA), an algorithm for training deep learning models across networked agents while keeping data local. DPA combines primal averaging with a local quasi-global momentum buffer to address performance loss caused by heterogeneous local data. The authors give a convergence bound for smooth nonconvex objectives and introduce DPA-1G, which uses one communication round per iteration. In experiments on the GUT benchmark, DPA outperforms several baselines, with the largest gains under high heterogeneity on sparse rings.
WhyExtends centralized primal averaging to decentralized training with quasi-global momentum, avoiding tracking variables; the authors report gains over several baselines on GUT benchmarks, largest on sp
In briefRaysemble: Recovering True-Scale 3D Bones from One X-ray
SummaryRaysemble is a method that reconstructs eight lumbar and pelvic bones in 3D from a single X-ray image. It separates bone shape from placement, using predicted depth and camera geometry to scale and position the bones on a common millimeter grid. Tested on simulated radiographs from 43 CT scans of 23 held-out patients, it achieved an average Dice of 65.24% without rigid alignment to CT, leading 15 compared methods. It improved Dice over VolumeNeRF by 5.86 percentage points with identical inputs.
In briefVisual Cipher Bench: Testing AI Agents on Long-Chain Visual Puzzle Solving
SummaryThis paper introduces Visual Cipher Bench, a set of 40 original visual puzzles across four categories: spatial manipulation, logical planning, rule induction, and knowledge-based decoding. Each puzzle gives only one to three black-and-white line drawings, so agents must find the solving rules from the images alone. Among eight frontier models tested across 29 reasoning settings, the best reached 26.7% accuracy, while five models scored below 1%. Adding a single non-spoiler mechanism hint raised accuracy on a ten-task subset from 7.8% to 55.6%, suggesting that discovering rules from visuals is the main bottleneck.
In briefEnergy and carbon footprint of federated LoRA fine-tuning on several GPUs
SummaryThis paper measures the energy use and estimated CO2-equivalent emissions of federated learning with LoRA fine-tuning of language models. The authors tested 135M and 360M models on NVIDIA A40, RTX 6000 Ada, and A100 GPUs, varying local training steps between 10 and 100. Local steps accounted for 84.8 to 91.2 percent of the variation in energy and emissions. On the A40 and RTX 6000 Ada, runs with fewer steps reached comparable accuracy using roughly 6 to 8 times less energy.
Pedro Augusto Londe Ribeiro, Rodrigo Moreira, Larissa Ferreira Rodrigues Moreira et al.
In briefShortcuts or Memorisation? Re-examining robustness to data perturbation in LLMs on language understanding
SummaryThis paper re-runs perturbation experiments, such as removing nouns or shuffling words, on eight large language models and on encoders fine-tuned on GLUE and ANLI. Direct-answering models keep more of their performance without nouns on MNLI than on ANLI, while only a model that reasons before answering behaves differently. The authors also test 368 new human-validated items unseen by the models and find that robustness reflects dataset shortcuts rather than memorisation, telling us more about the benchmark than the model.
In briefKosmo-GFM: A production-ready geometry foundation model for scene reconstruction
SummaryThis paper presents Kosmo-GFM, a scene reconstruction system combining a geometry foundation model (GFM), a pose optimization engine (PoseEngine), and supervision from KosmoWorld. GFM predicts camera parameters, depth, correspondences, and dynamic 3D points from images and videos, while PoseEngine uses geometric optimization to recover scene-level camera poses and geometry. The authors report that GFM achieves the lowest depth error and highest pose AUC on the SpatialBench subsets they tested, and that PoseEngine's cumulative runtime is about 20% of GLUEMAP's across 34 scenes.
In briefA modular framework for predicting ESG scores by combining specialized language models with graph learning.
SummaryThis paper introduces ESGPred, a modular framework that predicts corporate ESG scores by combining domain-specific language modeling with graph-based learning over a network of the evaluated companies. It was tested on 5,266 sustainability reports from 2,134 publicly listed companies published between 2020 and 2023, using LSEG Data & Analytics ratings as ground truth. The authors report that it predicts overall ESG ratings as well as the individual environmental, social, and governance components, and that adding graph-based relational learning outperforms purely text-based baselines.
Sofia Baiocco, Gianluca Bonifazi, Christopher Buratti et al.
In briefCalibration and uncertainty checks for AI medical question answering in Arabic and English
SummaryThe paper evaluates two open-weight language models on English MedMCQA and Arabic MedAraBench multiple-choice medical questions, examining whether their confidence matches their accuracy. Raw option probabilities were overconfident in all settings, and a single temperature parameter removed most of this calibration error. Conformal prediction met its target coverage, but prediction sets were large when accuracy was low.
Mariya (Stanley College of Engineering, Technology for Women, Hyderabad et al.
In briefOutcome-Mass Alignment: fixing uneven trajectory weighting in multi-turn LLM agent reinforcement learning
SummaryThis paper examines step-wise group-relative training for multi-turn LLM agents, where each trajectory is expanded into multiple rows and tokens. It identifies two accounting issues: outcome statistics can count one episode multiple times, and token-level aggregation can give trajectories unequal objective weight. The authors propose Outcome-Mass Alignment (OMA), which computes outcome statistics over unique trajectories and controls aggregate objective mass. Experiments on ALFWorld and WebShop report improved success rates in several settings and compatibility with GiGPO and GSPO.
In briefDetection Is Not Early Warning: Token-Level Alarm Evaluation for LM Failures
SummaryThe paper introduces PREWARN, a protocol that scores whether a monitor can warn before a language model failure begins at its onset token. As a case study, it tests linear probes on activations and six untrained baselines on repetitive degeneration in three open language models. Response-level metrics fail to separate the methods, while under PREWARN the probes warn before the onset on most degenerate responses of two of the three models.
Marco De Negri (1), Luca Sartori (1), Lorenzo Baggi (1) et al.
In briefEG360: An efficient, generalizable model for 360° depth estimation
SummaryThis paper presents EG360, a model for estimating depth from 360° panoramic images that aims to be both efficient and generalizable. It combines an Adaptive Panorama-Perspective Distillation strategy with a Spherical Cross-Level Feature Fusion module. After reparameterization the model has 6.593M parameters, achieves strong zero-shot performance on real-world benchmarks, and runs at 35 FPS on a Jetson Orin NX 16GB in MAXN mode.
In briefA systematic comparison of how training data is selected for machine-learning predictions of molecular and material properties
SummaryThis paper systematically compares strategies for selecting training data for machine-learning models that predict molecular and material properties, from random sampling to various active learning methods. The strategies are evaluated across different model classes, representations, datasets, and target properties. The authors find that effectiveness depends strongly on the interplay between model, dataset, and prediction task, though uncertainty-based sampling generally performs best. They derive guidelines for choosing a data acquisition strategy under a given labelling budget.
In briefScrewSense: a study on detecting anomalies and classifying errors in screwdriving processes using deep learning
SummaryThis paper applies deep learning to detect anomalies and classify errors in screwdriving, using multivariate time series such as torque and rotation angle. Anomaly detection analyzes the embedding space of a vector quantized variational autoencoder, and four model architectures are evaluated for error classification. The models distinguish multiple error categories beyond binary classification, reaching accuracies of up to 100%, and show strong generalization on real-world data.
Tobias Lettner, Tatyana Ivanovska, Anwar Al Assadi et al.
In briefClass-Competence Aggregation: Output-Layer Routing for Federated DDoS Detection
SummaryThis paper addresses federated learning for multiclass DDoS detection when each attack class is held by only one client, a setting that causes standard federated averaging to fail on the concentrated class. The authors propose Class-Competence Aggregation, which weights each class's output-layer aggregation by clients' locally reported validation F1 while hidden layers use standard averaging. On CIC-DDoS2019, the method substantially recovers the concentrated LDAP class, and the results are validated across seeds and on CICIDS2017.
In briefTesting Video AI on Rare, Dangerous Driving Scenes
SummaryThis paper introduces DashBench, a benchmark that tests 25 Video LLMs on 1,709 carefully selected driving videos. The videos cover collision-related corner cases as well as rare non-collision events. Across prompts and open-ended questions, all evaluated models fell well short of the performance needed for safety-critical driving, often misjudging or missing dangerous events and hallucinating nonexistent agents or actions.
Gabriel Manalu, Minh Son Hoang, Hyeontaek Hwang et al.
In briefEfficient Periocular Recognition Through Model Scaling and Visible-Near-Infrared Cross-Spectral Verification
SummaryThis paper presents AdaEye, a systematic study of input and backbone scaling for periocular verification in visible-spectrum and cross-spectral (VIS/NIR) settings. The authors compare full-face, masked-face, and dedicated 80x40 periocular inputs across four protocols, measuring accuracy, computational cost, throughput, memory, template size, and training stability. Direct periocular processing cuts cost from 6.30 to 1.64 GMACs per image and roughly triples throughput, though its effectiveness depends on the evaluation protocol.
Fabian Stockhard, Juan E. Tapia, and Christoph Busch
In briefMulticlass Machine Learning for Diabetes and Prediabetes Risk Prediction Using NHANES
SummaryThis paper builds a supervised learning framework that classifies 15,356 adults from the National Health and Nutrition Examination Survey as normal, prediabetic, or diabetic. HbA1c is used only to define the labels and is excluded from the input features. Conventional machine learning and transformer models are compared, reaching accuracy up to 0.72 and macro F1 of 0.70 in the multiclass setting, with prediabetes the hardest class. Binary formulations, especially combining prediabetes and diabetes for screening, yield higher F1 scores.
In briefAgriSense: Vision-language decision support and robot intervention in a virtual greenhouse
SummaryThis paper presents AgriSense, a framework that combines a virtual greenhouse in Isaac Sim, a quadruped mobile manipulator, RGB and LiDAR sensing, environmental control, and supervised vision-language decision support. Frozen DINOv2 and BERT encoders with structured environmental inputs jointly predict twelve environmental references, a target zone, and an intervention flag. The authors report zone-decision accuracy, environmental prediction errors, and virtual-arm target evaluations. Navigation and integrated treatment rely on simulator-state inputs, and physical greenhouse performance and agronomic treatment efficacy remain unvalidated.
Lyes Saad Saoud, Oualid Doukhi, Reza Ghorbani et al.
In briefMachine learning tool that diagnoses why lead-free perovskite solar cells degrade, using temperature and light measurements
SummaryThe study examines lead-free RbSn0.5Ge0.5I3 solar cells using about 42,000 SCAPS-1D simulations, covering both performance prediction and the identification of limiting mechanisms. Replacing the Spiro-OMeTAD hole transport layer with CuI raises simulated efficiency from 19.6% to 21.6%, almost entirely through the fill factor. A multi-label classifier identifies which of six degradation mechanisms are present from temperature- and illumination-dependent open-circuit-voltage data, scoring a macro F1 of 0.938 versus 0.687 for a physics-rule baseline. It does not generalize to mechanism combinations absent from training.
In briefORCA: Orthogonal Residual Consensus Alignment for Multi-View Clustering
SummaryThis paper proposes ORCA, a deep multi-view clustering method that addresses redundant information accumulation during fusion and indiscriminate alignment across representation levels. Its semantic incremental recurrent fusion module extracts cross-view orthogonal residuals as non-redundant increments and adaptively weights recurrent states via divergence-guided gating. Its consensus-calibrated hierarchical contrastive alignment module weights contrastive alignment of view-level, pairwise, and recurrent representations using global consensus. Experiments on five benchmarks against fifteen competing methods show effectiveness, including a 5.25 percent accuracy gain over the strongest baseline on 100leaves.
In briefPhysLoop: Using LLMs' Physics Knowledge to Help Video Generation Models Improve Themselves
SummaryVideo generation models produce realistic-looking videos but often fail at physical motion and causal consistency. PhysLoop is a framework that guides and constrains large language models during video generation, selection, and correction, turning their physical knowledge into physically consistent videos. It supports controllable training-data construction and inference-time video repair. The authors release 100K videos and two evaluation datasets, PhysLoop-Gen and PhysLoop-Repair.
In briefScale-Induced Semantic Shift: How Image Resizing Changes Hidden Instructions for Vision-Language Models
SummaryThis paper studies a security problem in which resizing or compressing an image changes the text instructions that a vision-language model reads from it. The authors name this effect Scale-Induced Semantic Shift (SISS) and present a closed-loop black-box attack framework tested on four closed-source vision-language models. They also propose ScaleGuard, a consistency detector that reduces the success rate of such attacks.
In briefFog-FWL: A large-scale LiDAR dataset for recovering objects hidden by fog
SummaryThis paper introduces Fog-FWL, a dataset for recovering objects hidden by fog from full-waveform LiDAR signals, which retain weak returns that peak detection discards. It includes Fog-FWL-Real, with 7,527 real foggy frames, and Fog-FWL-Sim, with 12,000 frames generated by a physics-based simulator, both using the same format and annotations. Under the hardest condition, object returns fall within the three strongest peaks for only 34.9% of object rays, showing the need for waveform-level recovery.
In briefWhy Sequence-Level Supervised Learning Is Hard for Autoregressive Language Models
SummaryThis paper asks whether sequence-level losses that use inference-time decoding, such as greedy decoding or autoregressive sampling, can be designed for training autoregressive language models. It shows that two natural attempts yield loss functions that are unbounded below, and that the sampling-based loss equals the expected greedy loss under Gumbel perturbations. The authors analyze a moment matching update direction obtained with a stop gradient on sampling, prove it is a valid descent direction for the autoregressive negative log-likelihood, but note it has high variance. They conclude that teacher forcing may be sufficient for supervised learning of autoregressive models.
Mathieu Blondel, Micha\"el E. Sander, Germain Vivier-Ardisson et al.
In briefEgoX: A first-person framework for learning robot manipulation across different bodies using embodiment dreaming
SummaryThis paper presents EgoX, a framework that learns robot manipulation from mixed demonstrations collected from humans, real robots, and simulated agents by aligning their observation and action modalities. The authors introduce MXT+, a modular Transformer pretrained on this mixture with an embodiment-dreaming objective that predicts future latent visual and proprioceptive states, then fine-tuned on target-robot demonstrations. Across 13 task-embodiment pairs, cross-embodiment pretraining raised average normalized task progress from 0.63 to 0.80, and from 0.53 to 0.77 under out-of-distribution conditions.
In briefFrom General VLM to Driving VLM: A Triplet Incremental Training Method
SummaryThe paper studies how to adapt a general vision-language model to autonomous driving without heavy architectural changes. It proposes a Triplet Incremental Training Framework that trains on image, depth, and question-answer triplets through three parallel streams, with a depth stream adding 3D geometric awareness and a frozen twin stream preventing semantic drift. The depth branch is removed at deployment, so inference adds no overhead.
In briefFrom Data to Wisdom: A Framework for Rethinking AI Through Epistemology
SummaryThis paper proposes the Data-to-Wisdom Epistemic Architecture for AI (EAAI), which treats data, information, knowledge, understanding and wisdom as five recursively interacting epistemic functions. It notes that Bayesian reference books do not consistently distinguish prior information from prior knowledge, and maps the five functions onto Bayesian modeling, prediction and decision analysis. Three applications are discussed: quantitative financial risk, health insurance actuarial modeling, and agentic AI for medical reimbursement claims.
In briefLimits of Thermal Infrared Drone Detection: A Cross-Dataset Benchmark
SummaryThis paper examines how differences in cameras, image processing, and annotation across public thermal infrared drone datasets affect detection results, and classifies the resulting artifacts. Using long-wave infrared data, it tests how training data composition and model characteristics influence YOLO-based detection, generalization, and failure conditions. The study finds that training data composition has a stronger effect on performance than YOLO generation or model capacity, and that cross-dataset generalization is strongly asymmetric. Even the best configuration misses over 40% of drones, with especially high miss rates for very small targets, cluttered urban backgrounds, and thermal crossover.
In briefThe Accuracy-Approval Gap: How Users Judge Agreeable Versus Accurate AI
SummaryThis paper combines benchmark tests and two large-scale user studies to examine sycophancy in large language models, where AI tends to agree with and flatter users. Four personas were defined by system prompts: objective, soft-sycophantic, hard-sycophantic, and baseline. Objective AI was more accurate under conversational pressure but was rated less respectful, less satisfying, and less trustworthy than the other personas. Persona did not significantly affect how often participants changed their opinions.
In briefMusicRWKV: A Multimodal RWKV Framework for Music Representation Learning
SummaryThis paper proposes MusicRWKV, a music representation learning framework built on the RWKV architecture as an alternative to Transformers for long recordings. It jointly models waveform, spectrogram, and metadata within a unified self-supervised framework, using an invertible compression module to shorten sequences and FiLM-based conditioning for modality-aware learning. Experiments on diverse Music Information Retrieval benchmarks show the architecture adapts effectively to heterogeneous music understanding tasks.
In briefTowards Green Text Generation: Memory-Based Speculative Decoding
SummaryThis paper proposes Olifant, a draft model built on a memory-based classifier that speeds up large language model text generation through speculative decoding. It predicts the next token on CPU in under 0.3 ms, without requiring a GPU. On EU legal text, without a KV cache it delivers 2.43x throughput and 2.51x lower energy per token, and with a KV-cached verifier it still yields 1.25x throughput with the lowest energy in the study.
In briefPhone-based sound detection of breathing in and out phases
SummaryThis paper proposes a non-invasive method that uses sound recorded by ordinary smartphones to identify the inhalation and exhalation phases of breathing. It uses a hybrid CNN-Transformer model that combines local spectral features with long-range temporal modeling. The model was tested with leave-one-subject-out evaluation and on an independent unseen test set, showing it generalizes to new participants. An ONNX pipeline enables on-device inference with latency below 20 ms.
Julian Szymanski, Tomasz Sankowski, Piotr Sulewski et al.
In briefForecasting Sales Delay via Time-aware Analysis of Sales Events in Reports
SummaryThis paper addresses the problem that B2B clients' actual order dates often differ from scheduled order dates, raising development costs. It introduces sales delay forecasting, a task that predicts how many days an order will be delayed from its scheduled date. The method extracts future sales event dates anticipated by sales representatives from textual reports as confidence features, and adds status features based on current date and past events. Experiments on real sales data show a 16% lower mean absolute error than a feature-based baseline and better results than large language model baselines without these features.
In briefSprint3D: Speeding up end-to-end 3D asset generation by optimizing the pipeline
SummaryImage-to-3D generative models take minutes per asset, making them slow to iterate with and costly to deploy. The paper profiles the public Pixal3D pipeline and finds that most time (85%) goes to mesh post-processing, mainly CPU-based UV unwrapping, rather than the networks. The authors rebuild post-processing on the GPU, reduce flow overhead, and distill four rectified flows from 76 to 8 network evaluations. The resulting Sprint3D cuts median time per asset from 133 s to 4.05 s (33×) on one H200 GPU, with metrics on 2,653 Toys4K objects largely maintained.
In briefHow expert disagreement affects LLM decision revisions
SummaryThe paper introduces a controlled two-pass audit in which a language model's question, evidence, and initial answer stay fixed while only the revision context changes: neutral reflection, anonymous disagreement, or disagreement attributed to a domain expert. Across five models and 1,989 evidence-grounded decisions in medicine, science, and law, no model met the selectivity bar needed for revision to improve accuracy under expert disagreement, and accuracy fell by up to 16.1 points. The expert label increased reversals for four of the five models, but its accuracy effect ranged from a 5.2-point loss to a 2.4-point gain, so reversal counts are a poor proxy for harm.
WhyThe authors propose a paired two-pass audit with an exact selectivity threshold, reporting that expert labels raise reversals even with identical arguments across 1,989 decisions.
In briefPredicting future links in changing networks by reinforcement learning on edge changes
SummaryThis paper proposes EG-RL, a framework that explicitly encodes edge births and deaths in temporal graphs. A policy-driven reinforcement learning mechanism, guided by critic feedback, selects and augments the most informative edge snapshots. Experiments on multiple dynamic graph datasets show gains in AUROC and average precision, and ablations confirm the value of adaptive augmentation and the encoding modules.
In briefManifold Attracting Flow Matching: Continuous Geometric Sampling That Enforces Physical Constraints
SummaryThis paper proposes Manifold Attracting Flow Matching (MAFM), a method for generating scientific data that satisfies equality constraints not seen during training. It continuously evolves a persistent anchor through an ODE that drives it toward the constraint manifold while preserving model-guided tangent motion. Experiments on geometric and partial differential equation benchmarks show better generation quality than some constraint-guided baselines with low constraint residuals, without retraining the generative model.
In briefNatural gradient learning in predictive coding via lateral circuits
SummaryThis paper introduces Natural PC, which adds a lateral circuit to standard predictive coding so that natural gradient learning can be performed locally in space and time. The lateral circuit learns the activity correlations that define the Fisher information matrix and implicitly computes its inverse, avoiding costly matrix inversion. According to the paper, Natural PC achieves state-of-the-art sample and compute efficiency on image reconstruction benchmarks and scales to ImageNet, matching backpropagation trained with AdamW and various second-order optimizers at much lower computational cost.
In briefMuSeToM: A Benchmark Testing If AI Models Revise Beliefs Based on Sensory Clues
SummaryMuSeToM is a text-based benchmark, validated by humans, for testing whether large language models can revise beliefs in theory-of-mind reasoning. It contains 2,415 instances adapted from BigToM and OpenToM, with textual cues spanning six sensory categories. The benchmark checks whether models can decide when a protagonist should revise a belief and infer the updated belief state. Results show that models often notice a belief may change but struggle to identify the updated belief, and frequently over-update on perceptually salient cues that are epistemically insufficient.
Chunkit Chan, Kawai Chung, Sze Heng Douglas Kwok et al.
In briefShortening LLM agent batch completion time by coordinating inference and environments
SummaryThis paper addresses batch execution of LLM agent tasks, such as benchmark evaluation, synthetic-data generation, and RL rollouts, which alternate between model inference and environment interaction. It proposes coordinated optimizations, including GPU-memory allocation for hybrid LLMs, trajectory-aware routing and scheduling, and Chain-Advancing Action Scheduling (CAAS) for constrained environment capacity. Live evaluations with Ling-3.0-Flash and Ling-3.1-Flash on several benchmarks report makespan reductions of 10.4% to 65.0% while evaluation scores remain aligned.
In briefMuViRL: Learning view-invariant visual speech representations from audio and video using calibrated view aggregation
SummaryThis paper proposes MuViRL, a self-supervised framework for learning multimodal, view-invariant visual speech representations. It combines synchronized 3D Gaussian view synthesis with multimodal self-supervision, using calibrated scores from a decision model to aggregate information across views. Experiments on real and synthesized multiview data report improved representation consistency across views and better visual speech recognition performance.
In briefLearning Missing Derivatives for Inverse Volume Rendering with Neural Scattering Operators
SummaryThe paper predicts the multiple-scattering field of optically thick media with a neural operator and differentiates through it during inverse rendering, while computing transmittance and single scattering directly. To fix poor derivatives at density states visited during optimization, the operator is fine-tuned on states generated in its own inversions, using Monte Carlo scattering-field supervision. On held-out optically thick clouds, the method improves relighting over frozen-source tomography at equal time and outperforms differentiable path tracing given more time.
In briefGraph neural network-enhanced TCN algorithms for detecting anomalies in ECG signals
SummaryThis paper gives time series a directed graph structure, making graph neural networks usable in this setting. The authors develop a new pipeline with novel geometric deep learning models that combine temporal convolutional networks (TCNs) with graph neural networks. The pipeline is applied to anomaly detection on electrocardiogram signals.
In briefInertiaMPDD: Personalized Depression Detection from Speech, Video and Text Using Emotional Dynamics
SummaryThe paper presents InertiaMPDD, a method for depression screening that combines speech, visual behavior and language. Inspired by emotional inertia, it uses a world-model-style latent predictor to learn healthy next-state patterns and derives deviation, decay-rate and Jacobian spectral descriptors from a reference. These descriptors were evaluated across five cohorts from four datasets. Fusion achieved 98.41% F1 on D-Vlog and 84.98% on LMVD.
In briefPretraining on Calibrated Synthetic Data to Quickly Adapt UAV Channel Predictors
SummaryThis paper studies pretraining channel state information predictors for unmanned aerial vehicle links on synthetic data calibrated to flight measurements, so that a predictor can be adapted quickly from a short target-flight segment. The synthetic data is calibrated to power evolution, temporal shape, frequency selectivity, and phase disturbance, then combined with fine-tuning on measured source-flight data. On chronologically held-out flight splits, synthetic-to-real fine-tuning reduces the amount of measured data needed to reach useful prediction.
In briefRethinking Actions: Continuity-Aware Priors for Planning with World Models
SummaryThis paper examines whether actions in expert demonstrations have temporal structure, confirming the dependence across five environments through ordered-versus-shuffled comparisons. The authors propose LG-DKGP, a locality-gated deep-kernel Gaussian-process prior that combines recent and longer-range action history, predicts a neural continuation, and learns a temporal Gaussian-process residual. The resulting planner reaches 87.07% macro success across five simulated tasks, compared with 71.03% for INTACT, 61.66% for PRISM, and 58.66% for LeWM.
In briefDataset Distillation via Drifting: Learning a Few Synthetic Samples to Stand in for Large Datasets
SummaryThis paper proposes a new dataset distillation method that learns a small set of synthetic samples to replace a large real dataset in downstream tasks. Instead of optimizing for a classifier, the method moves the synthetic data in a feature extractor's latent space toward the real data and away from each other, so they approximate each class's data distribution. The authors report that the drifted samples cover the real distribution better than existing methods while remaining effective for classification training. They also show the method can serve as a model visualization tool and extend beyond images to other modalities.
George Cazenavette, Angelina Quan, Giannis Daras et al.
In briefAsymmetric Multimodal Learning with Decomposed JEPAs
SummaryThis paper argues that cross-modal alignment alone loses information when modalities are informationally asymmetric, as in visual question answering. It proposes Decomposed JEPA (D-JEPA), which splits latent representations into an aligned shared core and private modality-specific fibers that capture what alignment discards. On compositional reasoning benchmarks, D-JEPA substantially outperforms alignment-only and alignment-with-reconstruction baselines.
WhyThe authors prove an alignment redundancy bound and propose D-JEPA, splitting latents into shared and private fibers, evaluated on compositional reasoning benchmarks.
In briefMachine learning for coordinating many Wi-Fi access points under IEEE 802.11bn
SummaryAs Wi-Fi networks become denser, random channel access alone cannot guarantee quality of service. This paper studies how neighboring access points under the IEEE 802.11bn multi-AP coordination framework should be chosen and configured to transmit together. The authors develop machine learning methods, progressing from hierarchical multi-armed bandits to a generative pipeline using graph neural networks and flow matching. The methods are evaluated in simulation and on a hardware testbed, improving throughput over legacy access and scaling to dozens of access points.
In briefWhen generated images show what the prompt only mentions
SummaryThis paper examines a failure in text-to-image models where a target named in the prompt appears in the image in a way the prompt does not request, which the authors call over-realization. They build OverReal, with 698 prompts covering four situations and 2,841 human-labeled generated images used to measure the rate and verify an automatic judge. Evaluating eight systems, they find all of them over-realize, and conventional metrics such as CLIPScore reward rather than penalize the failure.
In briefTreating Safety as a Budget: Calibrating the Control Channel of Constrained GRPO for Vision-Language Models
SummaryThis paper studies how reinforcement learning can train vision-language models to refuse harmful image-text requests while keeping safety expressed as budgets on harmful assistance and over-refusal. Enforced with Lagrange multipliers inside GRPO, the approach fails because the scorer's mean bias decides whether the multiplier converges, saturates or switches off, and standard group-wise standardization cancels the multiplier. The authors propose two-channel constrained GRPO, which keeps raw scores for the base reward, calibrates penalties through a monotone map fitted on judge-labeled responses, and preserves the multiplier in the advantage. On Qwen3-VL-2B, held-out harm on 1,037 adversarial prompts falls from 8.9% under standard GRPO to 4.8%, with over-refusal at most 2%.
WhyThe authors identify two failure modes in constrained GRPO for VLM safety, scorer mean bias and within-group standardization cancelling the multiplier, and propose a two-channel fix. The abstract repo
In briefFolio: Recoverable Multi-Resolution Context Folding for Long-Horizon Agents
SummaryFolio is a context manager for long-running language agents whose history grows beyond the model's context window. Instead of having a model summarize old history, it records three linked levels while the agent runs: short turn notes, an event manifest of every call, and captured source records. It folds older turns into coarser views while keeping handles that let the agent recover the underlying evidence. On 75 LOCA-bench tasks, Folio solved 66 tasks, compared with 54 for the same runtime using model-written compaction and 49 to 59 for the other systems tested.
In briefDySR-GNN: A Dynamic Sensor-Relationship Graph Neural Network for Root-Cause Analysis in Automotive HiL Time Series
SummaryThis paper presents DySR-GNN, a graph neural network for identifying root causes of faults in automotive Hardware-in-the-Loop multivariate time series. The model combines temporal patterns within each sensor signal with a sparse sensor-relationship graph rebuilt for every input window. Under a recording-disjoint evaluation, it outperforms CNN-LSTM, static GCN, static GAT and context-based variants. The learned graphs reveal both persistent and state-dependent relationships among sensors.
Hamza Ouarrad, Mohammad Abboush, and Andreas Rausch
In briefSign Language Retrieval Research for Low-Resource Dictionaries
SummaryMost sign languages lack annotated corpora and have only small dictionaries with one or a few example videos per sign. This study applies a pose encoder trained on American Sign Language (ASL-Citizen) to six other sign language datasets without any training on the target languages. With one example per sign, it retrieves the correct sign more accurately than learning-free methods and an ASL-trained OpenHands model. Ablations indicate that the per-pair InfoNCE objective and blocking attention between body and hands matter most.
Ushnish Sarkar, Vineet Kumar Rakesh, Bhaswar Chattopadhyay et al.
In briefAn improved CNN-Transformer network that increases the resolution of neutron images
SummaryNeutron images often have speckle noise, blurred edges and missing details, which limits high-precision analysis. The paper proposes a super-resolution model combining CNNs and Transformers, using a fine-tuned restoration model to create reference images and an improved CRAFT-based network with several added modules. Simulation results show better PSNR, SSIM, AGM and SMD scores than bicubic interpolation, SRCNN and SRGAN-VGG.
In briefUnibuc - NLP team's approach to the PAN 2025 “Voight-Kampff” generative AI detection task
SummaryThis paper describes the Unibuc - NLP team's approach to the Voight-Kampff Generative AI Detection task at CLEF 2025 PAN. For Subtask 1, the team compared classical machine learning models and fine-tuned LLMs with a classification head, using embeddings from LLMs. A majority vote ensemble of classical models performed comparably to fine-tuning, reaching about 99.4% F1 score. For Subtask 2, on human-AI collaborative text classification, the team combined features from different transformer layers, with modest results.
In briefA Dual-Branch Graph and Convolutional Network for Interpretable EEG Movement Classification
SummaryThis paper proposes a two-branch deep learning model that distinguishes upper-limb movement from rest using scalp EEG. The graph branch groups 61 electrodes into anatomically informed regions and learns their spatial relationships, while the convolutional branch follows an EEGNet-style design. Evaluated with leave-one-subject-out cross-validation on the Ofner dataset, it reaches a mean F1-score of 0.74 ± 0.06. The strongest learned connectivity lies in the left parietal low region (electrodes P1 and P3), opposite the performed right-arm movements.
In briefMaking Low-Data Medical Image Classification Accessible: CPU-Trainable ZACH-ViT and Geometric Variants
SummaryThis paper studies medical image classification under limited data and computing resources. It compares four ZACH-ViT variants with different heads (linear, Euclidean prototype, Poincare, and spherical) against two compact multiple-instance-learning baselines, all trained from scratch with under 260,000 parameters. The models are evaluated on seven MedMNIST+ datasets and three external benchmarks. The results show that the best model depends on the task and training length, and that CPU-only training is feasible on the tested platform.
In briefSingularity-Bench: Can AI Agents Build Other AI Agents? A Benchmark for Autonomous AI Development
SummaryThis paper introduces Singularity-Bench, an open-ended benchmark testing whether AI agents can autonomously develop AI systems across six end-to-end tasks, including pre-training, post-training, inference, and cluster orchestration. Agents implement core systems from a minimal starting point and are compared against mature human-engineered baselines on identical hardware. The authors also found reward-hacking vulnerabilities in the Harbor framework and developed Breakwater, an extension that blocks the attacks they tested. Opus 5.5 with Claude Code matched or outperformed the human baseline on all six tasks with a Singularity-Index of 1.25, while the other agents averaged 0.58.
WhyThe authors evaluate agents on six end-to-end AI-development tasks against human baselines and report reward-hacking flaws in Harbor, plus a patched extension.
In briefAstroMemBench: A benchmark testing how spacecraft AI agents remember past operations
SummaryThis paper introduces AstroMemBench, a benchmark for evaluating memory systems in embodied spacecraft agents across planetary surface operations, on-orbit servicing, and deep-space exploration. It contains 16 memory subchallenges and 2240 instances across nine tasks. In six configuration comparisons, Mem0, Chain-of-Memory, LightMem, and RAG improved same-session scores over native Pi Agent, and after a session restart Mem0 outperformed RAG by 12.25 points. The benchmark and evaluation code are publicly available.
In briefUsing Large Language Models as Digital Twins to Answer Personality Questionnaires
SummaryThis paper tests whether large language models can answer a personality questionnaire the way a specific person would, using that person's social-media posts as input. Four models, including GPT-4, Gemma, Llama 3.1, and Mistral, are evaluated on the MyPersonality dataset by comparing their predicted Big Five scores with users' actual scores. The results show that adding user text does not consistently improve accuracy, and models perform better near the middle of the trait range than at the extremes. Overall, the evaluated models do not consistently reproduce individuals' questionnaire behavior.
Francesco Di Cursi, Chiara Boldrini, Andrea Passarella et al.
In briefReFineR: Pixel-level alignment of a single image to a dense point cloud
SummaryThis paper addresses fine registration of a single camera image to a LiDAR point cloud, starting from a roughly estimated camera pose. The authors propose ReFineR, an offline framework that refines the camera pose through analysis-by-synthesis optimization, using differentiable point splatting rasterization and a smooth Gradient Orientation Measure (GOM). The method is evaluated on a curated dataset of manually calibrated images from heterogeneous sources. The authors also show that a point cloud rescale factor can be estimated automatically before optimization.
Ugo Insalaco, Jean-Emmanuel Deschaud, Pierre Dellenbach
In briefDo Person Actions Help Group Intent? One-Way Transfer in Vision-Language Models on JRDB-Social
SummaryThis paper tests whether training vision-language models on both individual person actions and group intent in social scenes helps both tasks, using the JRDB-Social dataset. Qwen3.5 models were fine-tuned under several training arms. Writing actions before intent raised intent macro-F1 from 30.1 to 31.0, while action F1 barely changed. The intent gain requires correct, image-grounded actions, so information flows one way, from actions to intent.
In briefActWorld-JEPA: Action-conditioned latent world models for end-to-end driving planning
SummaryThis paper proposes ActWorld-JEPA, which predicts future scenes in latent space and uses the predictions to rerank candidate driving trajectories. It introduces PC-JEPA for plan-conditioned latent prediction and Relational Residual Reranking for bounded score corrections. The authors report results on NAVSIM v1, NAVSIM v2 and nuScenes.
In briefHALight: Multi-agent learning for traffic signals when lane data is missing
SummaryThis paper presents HALight, a reinforcement learning framework for controlling traffic signals when lane-level sensor observations are partly missing. It models each intersection's observations and actions separately to handle differing road layouts, and adds a lightweight one-hop neighbor communication module to supply spatial context. A sequential policy update scheme is used to improve coordination among the heterogeneous agents. The authors report that on two real-world road networks, HALight outperforms representative baselines in traffic efficiency and convergence.
In briefSPARC: Accurate Photonic Matrix Multiplication via Representative Vectors
SummaryThis paper addresses how low-bit quantization and analog noise reduce the inference accuracy of photonic neural network accelerators. It groups similar activation vectors in deep neural networks and computes a representative vector for each group digitally at high precision. The remaining residual inner products, which have smaller norms, are offloaded to the photonic accelerator, reducing quantization and noise effects. The approach keeps the photonic inner-product pipeline intact, so it requires no hardware changes, and experiments on six representative DNN models show accuracy gains across different architectures and domains.
In briefWhen and how SignSGD beats SGD: a functional scaling law view
SummaryThis paper studies how the sign operator interacts with gradient noise to shape the learning dynamics and scaling behavior of SignSGD. Using power-law kernel regression with diagonal features, the authors derive a functional scaling law for SignSGD under general learning rate schedules, describing learning as signal learning and noise forgetting. They argue that gradient noise smooths the sign operator, yielding an effective preconditioned SGD, and that SignSGD learns signal faster but forgets noise more slowly than SGD. The relative performance of the two optimizers depends on task difficulty and batch size.
In briefOptimization methods that use node similarity to detect approximate network symmetry
SummaryThis paper studies how node-level similarity, with network centralities as one example, can guide search toward plausible vertex correspondences when detecting approximate symmetry in networks. The authors adapt several optimization algorithms, including simulated annealing, genetic algorithms, and tabu search, to use similarity guidance, and introduce BASIN, which learns a soft assignment matrix via bilinear embeddings. Experiments on synthetic graphs, brain connectomes, and empirical networks with known automorphism groups show that the effect of guidance depends on the regime, and tabu search performs best overall.
David Hartman, Jaroslav Hlinka, Dominik Klepl et al.
In briefPl@ntNet machine-to-machine API for automated, large-scale plant identification in biodiversity monitoring
SummaryThis paper presents Pl@ntNet's machine-to-machine plant identification service, which gives programmatic access to its AI plant recognition models through an API so identification can run without human intervention. The service covers 85,000 species, reports a species prediction accuracy of 0.89, and processes images up to 666 times faster than human-in-the-loop workflows. The authors also show a reproducible example that builds a dataset of 708 plant-pollinator interactions from 2,424 iNaturalist insect photos in about 30 minutes on a standard laptop.
Benjamin Bourel (IROKO), Jean-Christophe Lombardo (IROKO), Alexis Joly (IROKO) et al.
In briefContext-predictive low-rank adaptation for face spoof detection (CP-JEPA)
SummaryThis paper presents CP-JEPA, a framework for face presentation attack detection, which separates live faces from spoofs. It applies low-rank tuning to a pretrained CLIP image encoder and adds a joint-embedding predictive objective, in which a student predicts teacher features at masked spatial locations. Only the LoRA-adapted visual encoder and binary classifier are used at inference. On the OCIM benchmark it reports 3.74% mean HTER and 98.73% mean AUC, and adding CelebA-Spoof training reduces mean HTER from 3.45% to 2.35%.
Aminu Yusuf, Shakeel Muhammad Ibrahim, Sulaimon Oyeniyi Adebayo et al.
In briefWhether AI-written math proofs have a distinct style compared with human proofs
SummaryThe paper asks three large language models to reprove all 465 propositions of Euclid's Elements and compares their citation choices with human proofs. AI-generated proofs are more similar to each other than to human proofs, and this difference is stable across runs. The AI relies more on foundational elements such as definitions, common notions and postulates, which the authors call "building from the ground up".
In briefLost in Delegation: Studying Whether LLMs Benefit from Delegating Tasks in Generative Search
SummaryThis paper examines how well large language models perform when they delegate subtasks to external tools or other LLMs. The authors use purpose-built datasets split into pre- and post-cutoff information to compare several delegation modes. They find that delegation does not consistently improve accuracy, with results varying across models and tasks. When tasks require knowledge the orchestrating model lacks, delegation can significantly reduce accuracy, whereas direct tool use by the same model performs better.
Tamara Stojanovska, Elisabeth Kirsten, Leon Swazinna et al.
In briefZONOS3-Mini Technical Report: An Ultra-Compact Zero-Shot Voice-Cloning Speech Model
SummaryZONOS3-Mini is a compact, fast text-to-speech model that clones a voice from a short sample without additional training and can run in real time on a laptop or smartphone. It uses a latent mixture-of-experts backbone with 886M total parameters (100M active per token) and a new audio codec, Zyphra Audio Codec 1, which represents 44.1 kHz audio at a 15 Hz frame rate. On the new ZTTS2-Eval benchmark, it reaches a 6.71% macro-average word error rate across 55 language sets, which the authors describe as competitive with larger, slower models. The model weights, inference code, and training code are released.
Gabriel Clark, Sofian Mejjoute, George Close et al.
In briefCombining an LLM and an optimizer to schedule drones for victim search
SummaryThis paper proposes an agent-aided drone scheduling system for searching for victims in wilderness areas. A ground control station agent coordinates multiple drones across successive search cycles, using a large language model to reason about where to search and an optimizer to decide how to dispatch drones under energy constraints. In simulations, the proposed system consistently completed the search mission, while none of six non-LLM baselines did, and direct LLM scheduling was considerably less reliable.
In briefRESYNC: a method for deadline-bound streaming prediction using synchronized latent states
SummaryThis paper proposes RESYNC, a method for making per-observation predictions on high-rate data streams within a strict deadline. A large model runs asynchronously, while a fast path updates its latest but stale latent state using observations received since that state was computed, then applies the unchanged prediction head. The updater learns from the frozen model's representations without task labels. In live replay, RESYNC served at least 99.4% of arrivals on time with 0.35–0.62 ms median latency on one CPU thread.
In briefPISim: A unified simulator for testing photonic AI chip designs before manufacturing
SummaryPISim is a unified photonic inference simulator that combines pretrained deep neural network execution with physics-based models of accelerator architectures, modulators, light sources, and readout schemes. It models impairments such as quantization, shot noise, thermal drift, and crosstalk, along with an interface timing model to reveal electro-optical bottlenecks. Simulations on MNIST multilayer perceptron benchmarks show optimal precision of 6 bits for Mach–Zehnder and microring modulators under multi-impairment stress, and a 52.5 percentage-point accuracy gap between VCSEL-array and parallel-fanout laser architectures.
In briefUniSpark: A unified AI model that both understands and creates images and text
SummaryThis paper presents UniSpark-9B, a unified multimodal model that combines full attention with GatedDeltaNet linear attention on a Mixture-of-Transformer-Experts backbone. It supports visual question answering, text-to-image generation, instruction-guided editing, and interleaved text-image prediction. The model scores 2470.3 on MME-S and 0.83 on GenEval. The authors conclude that hybrid attention is an effective backbone for unified multimodal understanding and generation.
In briefPiLLVD: A Real-World Low-Light Video Dataset of Dynamic Scenes Filmed with Off-the-Shelf Cameras
SummaryThis paper introduces PiLLVD, a paired low-light video dataset captured with a Raspberry Pi-based multi-camera system using off-the-shelf cameras. It provides aligned low-light and normal-light reference sequences of dynamic indoor and outdoor scenes with human and object motion under varying illumination. The authors benchmark pre-trained low-light video enhancement methods on the dataset in a zero-shot setting. The results show that existing methods do not generalise uniformly across the sequences.
In briefLearning time series dynamics with derivative representations for forecasting
SummaryThis paper introduces DeltaTSF, a forecasting framework that uses discrete derivative representations of time series to capture system dynamics. It jointly learns the series and its discrete temporal derivatives across input and forecast horizons in a model-agnostic way. Experiments across diverse benchmarks and backbone architectures report improved forecasting performance and more dynamically coherent predictions than strong state-of-the-art methods.
In briefSignRot: Sign-adjusted rotation for quantizing large language models with massive outliers
SummaryThis paper addresses how massive outlier activations in large language models hinder low-bit quantization, and how rotation methods based on Hadamard transforms can fragment activation distributions. The authors propose SignRot, which identifies massive outlier channels and builds a sign adjustment vector that aligns the signs of the corresponding Hadamard row, turning a bimodal distribution into a unimodal one, and pairs it with asymmetric activation quantization. It needs only one calibration forward pass and adds only element-wise sign flips at inference. Applied to QuaRot, it reports a WikiText2 perplexity of 5.88 and average zero-shot accuracy of 72.73 on LLaMA3-70B, surpassing SpinQuant.
In briefEvenVLM: A Dataset for the Endangered Even Language with Vision-Language Model Fine-Tuning
SummaryThis paper introduces EvenVLM, the first text and multimodal (visual question answering) dataset for Even (Lamut), an endangered Northern Tungusic language spoken in Siberia. The dataset contains 50,521 Even-Russian entries extracted from dictionary and phrasebook sources, from which 107,543 training examples were synthesized across translation, grammar, reading comprehension, and semantic classification. The authors also propose a tokenizer expansion method that reduces tokens per word by 27%, and fine-tune Qwen3-VL-4B with LoRA, reaching 65% accuracy on visual question answering.
In briefRecalibrating Laya's Confidence Scores: A Hands-On Tutorial
SummaryThis tutorial gives a five-step recipe (measure, diagnose, fit, validate, deploy) for recalibrating the confidence scores of Laya, a typed-decision model that assigns probabilities to caller-defined options, using about 130 lines of numpy with standard methods. On 1,065 synthetic, LLM-labelled items, grouping by question type and option count gives a bucket-macro ECE of 0.143, which a cross-fitted temperature per bucket reduces to 0.073. A single global temperature makes calibration worse, and the findings apply only to these checkpoints and this data.
In briefHow much LLM-labelled data a decision encoder needs, and how wrong its labels can be
SummaryThe paper studies how many LLM-written training rows a decision encoder such as Laya needs and how much label error it tolerates. The authors ran 19 fine-tunes on nested subsets of a 2,818-row pool and tested on 1,065 in-distribution held-out items. Accuracy rises with more rows but the curve flattens, and 20% symmetric label noise costs 3.4 points while 40% costs 8.7 points.
In briefReA-Bench: A benchmark for auditing whether AI agents behave responsibly
SummaryThis paper introduces ReA-Bench, a diagnostic benchmark for auditing the behavior of autonomous agents. It organizes auditing challenges into six tiers, including internal safety knowledge, contradictions within trajectories, and missing information. The benchmark includes benign cases that resemble risky actions but are authorized and consistent with the task. Most samples come from real production trajectories, with paired versions to test auditor reliability.
In briefTime and Space: An evidence-grounded context framework for personal AI agents
SummaryThe paper argues that personal agents are limited by how context is represented rather than by model capacity. It proposes spatio-temporal context (STC), organized along a temporal axis and a relational semantic space, with every relation anchored to immutable source records. The authors formalise six invariants, a context operator, conflict detection, and a four-regime consumption taxonomy. They report measurements from a local SQLite-based retrieval prototype over 341 documents.
In briefHeMeR: reusing cached model states to keep robot agents' mixed memories efficient
SummaryHeMeR is an inference framework for embodied agents that combines declarative, procedural, and working memory. It keeps key-value states for memory contents that have not changed and recomputes only those affected by updates. In real-world robotic deployments, it reports 83.20% task success and cuts per-step inference time from 2.94 to 1.87 seconds compared with MemRL.
In briefContain, Conform, Compare (Extended): A Local Test Kit for Small Coding Agents and Its Full Experiment Record
SummaryThe paper presents drydock-eval, a local evaluation kit for the small coding agent Gemma 4 E4B, with a container-isolated runner that passed 200 scripted escape tests, eight checks aligning the local setup with the hosted scorer, and paired statistics that respect repository clusters. It publicly releases the full record of its experiments, including 9,660 agent runs and 2,494 grading controls. Three prompt-driven agent changes that looked better on screening tasks were not confirmed on new tasks. Prompts and agent bundles are withheld.
In briefOne Model, Two Precisions: Can Quantized Copies Disagree to Flag Small Model Errors?
SummaryThis paper asks whether disagreement between an 8-bit and a lower-precision quantized copy of the same model can serve as a per-question error score for abstention, escalation, or re-asking. The authors test Qwen3-0.6B and Qwen3-1.7B on 8,398 multiple-choice questions from ARC, MMLU and HellaSwag subsets. They find the disagreement score adds little over entropy, and that comparable gains appear with Gaussian weight noise, suggesting the signal is not specific to quantization.
In briefWhere a one-pass decision model looks in long inputs
SummaryThe paper studies Laya, a small model that answers multiple-choice questions about a text in one pass within a 512-token window. Using 177 items, five tasks, three checkpoints and 93,275 scored requests, the authors show that when the evidence falls beyond the window, accuracy drops to about 26.4%, near chance. Model confidence does not reliably separate cut requests from read ones.
In briefRobots That Talk Too Much and Don't Listen Enough: A Look at Turn-Taking Failures in Social Robots
SummaryThis short paper reviews earlier methods for improving speech recognition and turn-taking in interactions with social robots. It notes that speech recognition has improved considerably, but turn-taking failures and other misunderstandings remain too frequent. The authors argue that multimodal bidirectional streaming is needed to reduce how often turn-taking failures occur.
In briefHow should planning agents reshape their search space when users revise requests?
SummaryPlanning agents and benchmarks usually treat a user's initial request as fixed, even though users often discover acceptable trade-offs only after seeing options. The paper formalizes this as the ReSHAPE task, judging a revised request by whether its search space contains acceptable plans and a planner can find one. Analyzing 5,800 agent-revised specifications in travel and shopping, the authors find that one correct trade-off can raise a planner's success rate up to 4.6 times. Current agents admit at most 32% of acceptable plans, and showing concrete plans to users improves results over asking about preferences.
In briefFrom State of the Art to Human Prediction Algorithms: Comparing Interactive Art and Human Trajectory Prediction
SummaryThis paper examines human trajectory prediction algorithms, noting that as quantitative performance improves, consideration of human behavior, future intentions, and social interaction tends to decline. Using interactive art and prediction algorithms as examples, the authors analyze their structural and developmental commonalities through perception, interaction, and timing. They also report that some art-inspired prediction algorithms can derive intriguing patterns of human behavior from data.
