In briefLoopQuant: Accurate, efficient post-training quantization for looped Transformers
SummaryLooped language models reuse shared Transformer modules recursively to improve parameter efficiency, but this creates challenges for low-bit deployment. The paper proposes LoopQuant, a loop-aware post-training quantization framework with three components: cross-loop distribution canonicalization, interaction-aware loop-wise precision allocation, and loop-conditioned task calibration. On Ouro-2.6B, W4A4 LoopQuant surpasses all W4A8 baselines, reduces memory by 65.3%, and achieves 1.7 times decoding speedup over BF16.
In briefWorld Models Need a Physical Registry: Learning Readable Internal Predictive States for Physically Plausible Video Generation
SummaryThis paper proposes the Physical Registry, a set of temporally grounded slots inside a video diffusion transformer that co-evolve with video tokens and are supervised to hold readable predictions of scene dynamics. On Physics-IQ Verified, it improves vanilla fine-tuning by 9.9 points on I2V and 8.7 points on V2V, and it raises scores across Cosmos and WAN backbones. The authors also introduce the Physical Registry Selector, which ranks candidate futures by how far their intermediate registry readouts depart from the shared forecast, improving I2V scores by 5.6% to 15.9% across four backbones.
In briefSpeech-Driven Visual Grounding in Satellite Imagery: Analyzing Semantic Supervision
SummaryThis paper studies speech-driven visual grounding in remote sensing imagery, where spoken descriptions are used to locate targets without typing. The authors graft a frozen Whisper speech encoder onto a grounding model and compare three ways of injecting semantic knowledge, using linear probes to measure what is learned. They also release two speech datasets, Audio-DIOR-R and Audio-DIOR-Comp. The results indicate that whether direct speech alignment outperforms a speech-recognition cascade depends on recognizer quality.
In briefHexaBee-4B: A six-language translation model trained with multilingual answering and bidirectional supervision.
SummaryThe paper presents HexaBee-4B, a translation model adapted from Gemma 3 4B IT for six languages: Arabic, Chinese, English, Hebrew, Persian, and Spanish. It is trained in three stages: Wikipedia continued pretraining, joint multilingual answering and answer-text translation, and bidirectional translation specialization. On 36,572 synthetic test examples across 30 directions, it reaches a direction-macro chrF++ of 73.75, versus 57.11 for Gemma 3 4B IT and 66.83 for Gemma 3 12B IT.
Iman Barati, Mohammadmostafa Rostamkhani, Mohammad Badzohreh
In briefVideoSlate: A stateful framework for selectively revising video segments in AI video reasoning
SummaryThe paper introduces VideoSlate, a framework for agentic video reasoning systems that answer questions by selecting relevant video spans. A shared evidence blackboard records candidate spans, their provenance, and verifier observations, and a certifier judges whether a proposed revision is supported relative to the current span. The controller then either adopts the revision or keeps the current span, logging each decision and its evidence. The authors report that evaluations on video grounding and grounded question-answering benchmarks show improved temporal localization through selective revision.
In briefnaspaclust: R package hybridising spatial fuzzy clustering with nature-inspired optimisation
SummaryThis paper introduces the R package naspaclust, which improves spatial fuzzy clustering by combining Fuzzy Geographically Weighted Clustering with nature-inspired optimisation algorithms. It provides seven optimisation algorithms, including modified versions with altered randomisation terms and updating rules. Additional distance metric options are offered via the rdist package. The paper demonstrates the package in various cases and compares the performance of the basic algorithms.
Bahrul Ilmi Nasution, Rezzy Eko Caraka, Robert Kurniawan
In briefSelf-distillation with verbal prompts that keep models rethinking
SummaryThe paper studies on-policy self-distillation (OPSD), where a language model acts as its own teacher, conditioned on privileged information such as verified solutions. The authors observe that this privileged information can push the model toward shortcut reasoning, with less verification and a sharp drop in output entropy. They propose VRSD, which uses system prompts to regulate how the teacher uses this information to encourage rethinking. Across three models and four tasks, VRSD improves Pass@1 over OPSD in 11 of 12 settings, by 3.9 points on average, while generally keeping token entropy higher.
In briefLoongSage: A Production-Grade, Efficient Agentic RL Framework for Frontier LLMs
SummaryThis paper presents LoongSage, a fully asynchronous reinforcement learning framework for post-training large language models on agentic tasks such as tool use and multi-turn interaction. Its core is about 14K lines of code, built on Megatron for training and SGLang for inference, and it includes AgentFlow, sliding-window staleness control, TransferMesh weight synchronization, and a Teacher Manager for multi-teacher on-policy distillation. The authors report that task rewards improved steadily across four complete training runs, the train-inference gap stayed under control, and asynchronous execution removed synchronous idle time while sustaining hardware utilization.
LoongSage Team: Bowei Song, Caiwei Zhao, Chuwen Liu et al.
In briefRIFT: Token Influence Follows the Path of Least Resistance
SummaryThe paper proposes RIFT, which treats a Transformer's layers as a causal network with attention weights as conductances and scores each context token by the series resistance of its least-resistance path to the answer. The resistance is computed exactly with a shortest-path dynamic program. On synthetic retrieval tasks and extractive question answering over Wikipedia paragraphs, RIFT predicts the effect of token replacement with Spearman correlations of 0.52 to 0.68, without gradients or extra forward passes.
In briefSex classification of young adults from DWI connectomes using a supervised variational autoencoder
SummaryThis paper classifies sex in 787 young adults aged 15 to 25 using diffusion-weighted imaging structural connectomes across 82 Desikan-Killiany atlas regions. The authors apply a supervised variational autoencoder to extract compact latent features, which are then fed into several machine learning classifiers. The multi-layer perceptron achieved the highest accuracy, at 80%.
Aditya Kothapalli, Aaryan Gera, Sathvik Reddy V et al.
In briefRescuing Points Stuck in Wrong Decision Regions: A Target for Evaluating Forward-Forward Learning
SummaryForward-Forward learning replaces backpropagation with local, layer-wise objectives and is usually judged by linear-probe accuracy. The authors show that a probe on an untrained convolutional network already performs well above chance, so accuracy alone can overstate what learning contributes. They introduce a Boundary Proximity Index analysis and apply it to seven Forward-Forward and related methods, finding that initially misclassified points near the decision boundary are more often rescued, while rescue rates for deeply embedded points vary widely across methods.
Alireza Rahmansetayesh, Amirhossein Yousefiramandi, Maxwell Spiridonov et al.
In briefChronoScribe: Splitting Language Understanding and Numerical Forecasting for Instruction-Driven Time Series Prediction
SummaryThe paper presents ChronoScribe, a system in which a LoRA-adapted LLM interprets natural-language instructions and selects an execution path and context length, while a frozen time series foundation model (TSFM) produces quantile forecasts for the requested horizon. The LLM does not generate or revise predicted values. Across 23 evaluation dimensions and 30 public datasets, ChronoScribe reports the lowest error on sMAPE, MASE, WAPE, RMSE, and MAE among all compared systems, and a Task-Level Score of 65.5, 10.6 points above the strongest direct-LLM baseline.
In briefEfficient multi-agent reasoning through confidence-guided adaptive debate (paper)
SummaryThis paper examines when multi-agent debate actually helps large language models reason, arguing that its benefits depend on timing and structure. The authors propose LASE, a leader-centric framework with an asymmetric leader-supporter design that engages in debate only in non-neutral regimes and otherwise uses simple aggregation. Experiments on diverse reasoning benchmarks report that LASE reaches multi-agent-level performance at near single-agent token cost, improving efficiency over static debate and voting baselines.
In briefSyncReward: A Human-Aligned Reward Model for Audio-Visual Sync
SummaryThis paper addresses the problem of audio-visual synchronization in generated videos, where speech and sounds often fail to line up with visual events. The authors release SyncReward-Data-80K, a dataset of 80K professional synchronization ratings, and SyncReward-Bench, a 1K-clip benchmark, showing that existing metrics and omni-modal models disagree substantially with human judgments. They propose SyncReward, a reward model that achieves the strongest agreement with human ratings and can improve synchronization via best-of-N selection and reinforcement learning post-training.
WhyIntroduces an 80K human-rated dataset and 1K benchmark, plus a reward model combining contrastive pretraining and human supervision, with strong human-agreement results.
In briefWorldRoPE: A Probabilistic World-Centric Positional Encoding for Multi-View Transformer Attention
SummaryThis paper proposes WorldRoPE, a probabilistic rotary positional embedding for attention in multi-view Transformers. Each patch token is back-projected into volumetric Gaussian primitives in a shared world frame to represent depth uncertainty and spatial extent. The method is evaluated on novel view synthesis, depth estimation, and 3D geometry reconstruction, where the authors report its effectiveness.
In briefTRAP: Testing LLM agents for tool results reused in the wrong context
SummaryThis paper describes tool-result misbinding, where a real tool output is reused in an execution context it does not match, potentially causing an AI agent to take an unsafe action. The authors introduce ReplayBench, a benchmark covering 19 scenarios, 4 failure properties and 3 replay mechanisms, tested on eight LLMs and three agent frameworks. Under cross-task replay, the mean Final-State Error Rate was 0.649. The authors propose runtime validity mechanisms that eliminated final-state errors in their evaluated settings and call for explicit contextual binding in tool-using agent runtimes.
In briefBAT3R: A Bayesian method for more stable online 3D scene reconstruction from video
SummaryBAT3R is a training-free framework for building 3D scenes from long video streams one frame at a time. It combines State Innovation Gating, which keeps the internal state from drifting away from its initialization, with Bayesian-guided Adaptive State Updating, which uses a Kalman-filter-style update to weigh observation reliability and spatial relevance. According to the paper, it outperforms existing online 3D reconstruction methods on standard benchmarks including 7-Scenes, TUM, ScanNet, and KITTI, with the largest gains on long sequences and complex scenes.
In briefExplicitly aligned multi-view breast cancer risk prediction using variable-length screening histories
SummaryThe paper proposes FlexLMV-Net, a longitudinal framework that models both CC and MLO mammography views for each breast across screenings, accepting a variable number of prior examinations. It explicitly aligns longitudinal feature maps and fuses views through cross-view temporal fusion and a dual-stream attention block. Evaluated on EMBED and CSAW-CC, it reportedly outperforms existing methods across breast density, cancer subtypes, and demographic subgroups. The authors also release BCRBench, an open-source benchmark for standardized, reproducible comparison of breast cancer risk models.
WhyExtends explicit longitudinal alignment to jointly model CC and MLO views with variable prior exams. Strong comparisons on EMBED and CSAW-CC; impact is confined to breast cancer risk prediction.
Solveig Thrun, Zijun Sun, Suaiba A. Salahuddin et al.
In briefFast Generative Simulation of Bubble Flows Using Spherical Harmonics
SummaryThis paper proposes a neural surrogate that jointly predicts bubble trajectories and interface deformation in a compact shared representation, replacing costly mesh-resolved simulation with fast generative inference. It treats forecasting of nonlinear, chaotic two-phase flow as a conditional generative problem, predicting whole future windows in one pass. The model is trained and evaluated on BubbleSH, a Front-Tracking dataset of rising deformable bubble swarms. The authors report that it reproduces key flow statistics while running several orders of magnitude faster than the underlying DNS solver.
Kiet Bennema ten Brinke, Rachna Ramesh, Douwe Orij et al.
In briefWhen a Conceptor Complement Becomes Gain: Operator Geometry in Language Models
SummaryThis paper examines a target-centered steering method for language models by removing its fitted covariance filter while keeping the same target center and gain. Across three models and seven targets, all 21 fitted maps expanded every direction at the generation layers. On a four-target panel, retaining the matrix did not produce a consistent extra suppression benefit. The authors conclude that whole-intervention comparisons cannot attribute generated behavior to the fitted matrix, and that the full transformation and its output effects must be analyzed directly.
In briefTraining-Free Video Anomaly Analysis via Anomaly-Guided Observation and Two-Level Semantic Feedback
SummaryThis paper proposes a training-free framework that uses frozen vision-language and language models to detect and describe abnormal events in videos. It selects frames guided by an upstream anomaly-score curve, generates one open-ended anomaly description, and feeds that description back to refine temporal detection at both video and local levels. The authors report improved description quality on HIVAU evaluations and detection gains, with AUC rising from 84.36% to 86.25% on UCF-Crime and AP from 68.07% to 72.58% on XD-Violence.
In briefBehaviorFormer: Finding Player Behavior Patterns Through Clusterable Next-Token Prediction
SummaryThis paper tackles discovering behavior patterns in open-ended games, which lack clear goals or labels. It extends a next-token prediction Transformer with the DCSS deep clustering algorithm to learn well-separated clusters. In a controlled Minecraft session with 26 players, the method finds 23 behavior patterns, grouped into seven categories: stationary, roaming, travel, mining, base work, logging, and farming.
Tobias Schneider, Danial Fahrner, Soren Antebi et al.
In briefEstimating building age from widely available multi-source geospatial data
SummaryThis paper proposes the first deep-learning framework that estimates building age at both pixel and instance levels using multi-source geospatial data, such as satellite imagery, Digital Surface Models, and building footprints. It combines a pretrained backbone, mixture-of-experts fusion modules, and a loss that enforces the temporal ordering of predicted ages. The authors also build a building age benchmark for Hong Kong, and the source code and datasets are publicly available.
In briefHyWe: A Framework Unifying Search and Operation for Real-World Hybrid Web Tasks
SummaryReal-world web tasks often require both information gathering and page interaction, but these abilities have usually been developed and evaluated separately. The paper introduces HyWe, which unifies search and operation within one model and trajectory using dedicated modes, explicit switching, and shared history. The authors train HyWe-35B on automatically generated hybrid demonstrations and report 47.0% success on their 100-task HyWeBench, a 19.0-point gain over the untrained base model. The model also reports strong results on Online-Mind2Web and WebVoyager.
In briefA controllable way for LLMs to abstain from answering, without training them to abstain
SummaryThe paper compares two ways for large language models to decline questions they would likely get wrong: selective answering with an external confidence threshold, and abstention training inside the model. It splits error into answering ability and distinguishing ability, finding that abstention training mainly wins by answering better. The authors propose TAPS, which trains for accuracy and then selects answers post hoc, keeping coverage controllable. Across three models and three factual QA benchmarks, TAPS with a probing-based confidence estimator shows 1.5 to 9.4 points lower mean risk at matched coverage than three representative abstention training methods.
In briefReTAC: Residual test-time agentic computation for video spatial reasoning
SummaryThis paper proposes ReTAC, a training-free multiagent framework for video spatial reasoning. It allocates extra computation only to requirements the base multimodal model leaves unresolved, and uses stage-aligned verification to localize failures and recompute only affected states. Experiments report that ReTAC achieves the highest overall score among compared methods across three mainstream MLLMs and two benchmarks.
In briefNarrative paradoxes in LLM reasoning: a method that steers models toward harmful outputs
SummaryThis paper examines whether the multi-step reasoning process of large language models can itself be unsafe, even when each step looks coherent. The authors propose PDRA, an attack that pairs a malicious actor with a benign task and forces the model through a three-stage analysis to produce harmful content. According to the paper, PDRA achieves state-of-the-art attack success rates across diverse LLMs.
In briefRAR: Rate-Adaptive Rollout for Consistent Interactive World Models
SummaryThis paper addresses a limitation of video-based interactive world models, which are usually trained and rolled out at a fixed temporal rate, so generation speed mismatches make the world evolve too fast or too slow. It proposes Rate-Adaptive Rollout (RAR), whose core component, Stride-RoPE, aligns latent frames sampled at different rates on a shared physical-time coordinate. According to the paper, RAR achieves temporally consistent multi-rate rollout and online rate switching in action- and camera-controlled environments while preserving high-rate quality and extending physical horizons under matched training budgets.
In briefFON: A two-organ language model that drafts and revises without quadratic attention
SummaryThe paper proposes FON (Fusion-Organ Network), which splits one weight stack into a drafting organ that generates text with learnable delays and recursion, and a revision organ that reprocesses the written prefix without generating. At 30M parameters under matched parameter and token budgets, FON reports a WikiText-103 perplexity of 131.03, lower than Transformer (197.73) and Mamba2 (152.53). The authors also report limitations, including weaker long-range retrieval, weaker out-of-domain code infilling, and a 6.5 to 12.3 times wall-clock slowdown.
Helong Hu (1), Hongdan Pan (2), Shuiqing Hu (3) ((1) Zhanjiang Campus et al.
In briefLearning Which Video Moments to Use for Explaining Anomalies
SummaryThis paper studies video anomaly understanding, which generates language explanations of anomalous events in long videos. The authors propose a semantic-conditioned Temporal Sampling Q-Former that learns to select temporal visual evidence through the generation objective, conditioned on CLIP-embedded object tags extracted from the video. On HIVAU-70K, the model achieves the best performance among the compared fine-tuned models, and ablations show it outperforms fixed anomaly-focused sampling.
In briefInteger-only training of deep networks to cut memory use
SummaryThis paper presents a training algorithm in which weights, gradients, and activations are all stored as INT8 and every training operation uses integer arithmetic. Tensors are represented as integer mantissas with power-of-two exponents, and a per-group velocity mechanism preserves gradient precision. The authors report competitive accuracy on CIFAR-10, CIFAR-100, and ImageNet with ResNet, MobileNetV2, and Vision Transformer models, at roughly four times lower training memory than the FP32 baseline.
Hazem Hesham Yousef Shalby, Fabrizio Pittorino, Manuel Roveri
In briefSLDBench v2: Active Scaling Law Discovery Benchmark
SummaryThis paper introduces SLDBench v2, testing whether language-model agents can discover scaling laws, expanding scenarios from 8 to 127 and selecting 30 challenging ones. It adds an active setting where agents choose experiments, pay their costs, and revise their laws as results arrive. Across 8 frontier agent systems, median R² falls from 0.72–0.75 in the passive setting to 0.24–0.57 in the active one, and agents tend to favor cheap experiments.
In briefClothAtlas: Identifying Garments and Estimating Their Shape Despite Self-Occlusion
SummaryThe paper introduces ClothAtlas, a method that jointly identifies a garment and estimates its complete configuration from partial, self-occluded point clouds and a library of canonical meshes. A shared graph-conditioned flow-matching model samples configurations for each candidate mesh and ranks candidates by how well they fit the observation, and new garments can be added without retraining. On held-out DexGarmentLab garments, the correct mesh ranks first among 221 candidates for all 22 garments from a flat observation, and identifies 15 of 22 garments from a self-occluded observation.
In briefSemantic-guided 3D Gaussian Splatting for detailed urban scene reconstruction
SummaryThis paper presents SG-GS, a method that improves 3D Gaussian Splatting for reconstructing fine structures in complex urban scenes. It uses semantic segmentation from the Segment Anything Model (SAM) to guide both adaptive density control, including boundary-enhanced densification, and a multi-feature loss function trained in progressive phases. Experiments on three datasets report PSNR gains of about 2.7% to 5.2% over the original 3DGS, with reduced artifacts on linear structures and less boundary blurring.
In briefFaster structured transforms that compress LLM attention memory while preserving inner products
SummaryThe paper studies Quantized Johnson-Lindenstrauss (QJL), a 1-bit quantization method that compresses KV caches in large language models while approximately preserving attention inner products. The authors propose two constructions built on compositions of randomized Hadamard transforms, using a Gaussian diagonal in the second stage, with added Gaussian dithering and a random permutation respectively. They prove embedding-dimension and running-time guarantees analogous to those of Fast JL in the non-quantized setting. Both constructions can be applied in O(d log d) time.
In briefREVEAL: Reverse-engineering agent latents from frontier-model traces for training-free steering of smaller agents
SummaryThe paper proposes REVEAL, a training-free framework that transfers experience from frontier LLM agents to frozen smaller open-source agents using textual trajectories. It matches local events from successful and failed rollouts of the same task and derives a target-specific steering direction from response-transition contrasts, applied before each response. On Terminal-Bench 2.1 and tool-enabled Humanity's Last Exam, the authors report higher accuracy than unmodified agents across three target models.
In briefProGAN-based synthetic images to augment inspection data for metal parts
SummaryThis paper studies generating synthetic images with Progressive Growing GANs (ProGAN) to address the scarcity and low diversity of training data for defect detection in metal parts. The authors build a ProGAN pipeline using Wasserstein loss with gradient penalty and evaluate key hyperparameters on the public MVTec AD (metal nut) and MPDD (tube) datasets. They report that training remained stable and produced realistic images that can augment training data.
Hossein Omid Beiki, Sepehr Asgari, Oliver Petrovic et al.
In briefMoment-based anomaly detection method for high-dimensional mixed-type data
SummaryThe paper proposes an anomaly detection method for high-dimensional data with mixed feature types, using standardized moment-based transformations to produce interpretable feature-level scores. These scores are aggregated with a dependence-aware step into a Mixed-data Influence (MDI) score measuring how far an observation departs from the reference distribution. Simulations and two real-data applications, to multi-omics and ESG data, assess finite-sample calibration, detection performance, and interpretability.
In briefRelLedger: An audit framework for relational information governance in LLM agent organizations
SummaryThis paper introduces RelLedger, a framework for auditing how shared records in LLM agent organizations shape later decisions. It holds the event, evidence, risk and message fixed while varying the sender's relationship histories, and tracks delivered information into recipients' downstream decisions. Across five models, Claude sent the same factual notice 62.22 percentage points more often with conflict histories than with friendly ones.
In briefagent-risk: A Method for Assessing Risk of Changes to AI Agents
SummaryThis paper introduces agent-risk, a method for assessing the risk of changes to deployed AI agents, such as stricter safety gates or new models, before approval. It replays logged episodes under the assumption that nothing else changes, combining logs, declared stop values, saved-state reruns, and bounds. In a bash sandbox experiment, the log alone could not decide two retry gates, but rerunning only the blocked episodes from saved states reached both decisions using about half the agent calls of a live test. The authors conclude that approval requires evidence designed in advance, such as specific log fields, saved states, and targeted reruns.
In briefDepthTile: Depth-guided self-tiling of existing images with input-side tileability assessment
SummaryThe paper presents DepthTile, a method that turns an existing image into a seamlessly repeatable tile while keeping its original content and local details, without retraining the diffusion backbone. It converts monocular depth into a spatial preservation prior and combines it with the timestep-dependent masking of Tiled Diffusion and a differential diffusion mask. The authors also introduce the Self-Tiling Ability Score (STAS), a lightweight input-side predictor of whether an image is suitable for self-tiling. On a 335-image benchmark, DepthTile reduces MAE, LPIPS, and Tiling Score by 12.8%, 26.7%, and 8.4%, respectively, relative to the default Tiled Diffusion pipeline.
In briefFlowAvatar: Real-time full-body avatars on consumer XR headsets from sparse sensor inputs
SummaryThis paper presents FlowAvatar, a system that creates personalized look-alike avatars in real time on a consumer headset (Meta Quest Pro) using sparse egocentric inputs. It tracks full-body pose, hand articulation, facial expression, and eye gaze. The authors report competitive spatial accuracy and state-of-the-art temporal consistency on the AMASS benchmark. Two user studies (N=12 and N=10) suggest that tracking completeness and personalized appearance are complementary, with only their combination yielding positive total embodiment.
In briefA Survey of Flow Matching in Reinforcement Learning: Reviewing Policy Generation via Deterministic Probability Transport
SummaryThis paper is a survey of flow matching methods in reinforcement learning. Flow matching generates policies by sampling from learned velocity fields through ordinary differential equations, which the authors say can reduce inference latency compared with diffusion policies that rely on stochastic denoising chains. The survey organizes the literature by the target distribution being modeled and by how RL signals are integrated, and it covers discrete and non-Euclidean action spaces, comparisons with Gaussian and diffusion baselines, and open problems.
Nabuat Zaman Nahim, Fairoz Nower Khan, and Peizhong Ju
In briefDeep speech models encode word timing as a circular path in a low-dimensional subspace
SummaryThe paper studies how speech models represent word timing, using representational geometry. It finds that many deep speech models, both supervised and self-supervised, form near-circular trajectories from word onset to offset in a two-dimensional temporal subspace. The authors propose Rotary Segmentation (RoSeg), which detects word boundaries when these cycles complete, reaching a 64.3% boundary F-score on Librispeech with HuBERT.
In briefGraph3R: Joint View Graph Setup and Feed-Forward 3D Reconstruction
SummaryFeed-forward reconstruction estimates camera poses and dense geometry from image sets, but uncurated collections often mix views from different scenes, which can corrupt joint reconstruction. Graph3R links view graph initialization with feed-forward reconstruction through a shared geometric representation. A lightweight head predicts pairwise covisibility from multi-view features to build a graph that groups views by scene, and each group is then decoded separately. The method improves recall on three benchmarks and reconstruction on mixed-scene inputs, and it runs about 35.9 times faster than the baseline when reconstructing the query group.
In briefDeep learning for selecting gamma events in TAIGA-IACT stereo image analysis
SummaryThe TAIGA observatory has four Cherenkov telescopes spaced 300-500 m apart, so air showers can be recorded by several telescopes at once. The paper proposes and studies neural network methods for selecting gamma-ray events from joint two-telescope images. One method performs single-stage analysis of the two images, while the other uses an autoencoder to extract features encoding shower physical parameters, including the primary particle type.
In briefPrediction-error correction in trained TTT layers
SummaryThis paper examines how test-time-training (TTT) layers update fast-weight memory while processing a sequence, focusing on prediction-derived update components in released implementations. The authors derive the inner gradients for dot-product objectives and the LayerNorm-based squared-error objective used by TTT-Linear and TTT-MLP. On three trained checkpoints, removing the full normalized prediction contribution raised perplexity by 2.8-2.9% on 100,303 tokens of War and Peace for TTT-Linear models, while freezing all fast-weight updates raised it by 4.2-4.5 times.
In briefFrequency Limits of Unsupervised EEG Representation Learning with Deep Networks
SummaryThis paper studies unsupervised time-domain feature extraction from EEG signals, comparing three autoencoders (LSTMVAE, VAEEG, Transformer) and a state-of-the-art EEG foundation model. The autoencoders capture large-scale changes such as alpha slowing but are biased against high frequencies under 1/f-decaying spectra, due to the Mean Squared Error loss. The authors argue for adding spectral-domain loss terms and gradient-balancing techniques to preserve high-frequency features important for diagnosing brain disorders.
David Aquilu\'e-Llorens, Abel Puigseslloses, Angel Blanco et al.
In briefHardware-friendly token-group activation quantization for low-bit Mamba super-resolution
SummaryThis paper addresses the difficulty of low-bit post-training quantization for Mamba-based super-resolution models, where activation ranges vary across spatial tokens. It proposes Token-Group Quantization, in which a lightweight predictor assigns tokens to groups from pre-scan features and selects offline-calibrated clipping bounds at inference. A Frequency-Preserving Refinement objective aligns Fourier amplitude and phase to retain high-frequency detail. The authors report consistent gains over prior baselines and a 2.23x speedup on Jetson Orin Nano relative to FP16.
WhyToken-group clipping for Mamba's selective scan plus a Fourier-based calibration objective substantially improves low-bit super-resolution, validated across protocols, backbones, and edge-device speed
In briefScreenASR-Bench: A Tiered Benchmark for Visually Aided Speech Recognition
SummaryThis paper introduces ScreenASR-Bench, a speech recognition benchmark with three tiers of increasing difficulty: direct on-screen text, visible non-text content, and scene-level implicit semantics. Experiments on ASR models, large audio-language models, and omni models show that adding visual context improves recall over audio-only input in all three families. Gains are largest for direct on-screen text and smallest for implicit semantic cues. The benchmark is publicly released.
In briefEgoWholeView: A camera-only head-mounted system capturing full-body and two-hand motion
SummaryThe paper presents EgoWholeView, a head-mounted rig of six synchronized wide-angle RGB cameras that captures body and bimanual hand motion. Outward cameras localize the rig in a world frame, while a downward camera pair predicts SMPL-X body and MANO hand parameters, which are fused using reliability cues. Compared with monocular prediction, the authors report PA-MPJPE reductions of 3.7% for the body and 11.2% for the hands, and temporal refinement lowers angular-acceleration magnitude by 32.2% and 31.0%. The system uses only RGB cameras, offering a low-cost, portable approach to egocentric interaction data.
In briefMFlowHOI: Whole-Body Human-Object Interaction Generation via Part-Aware MeanFlow
SummaryThis paper presents MFlowHOI, a framework that generates whole-body human-object interactions, covering body and finger motion, object trajectories, and contacts, in one step. A Contact-Aware Multipart RVQVAE builds a structured latent space, and an ST-DiT generates the full interaction in a single MeanFlow evaluation. The authors report better motion quality and text alignment than existing base generators, with much lower inference time.
In briefAvoiding What Your Teacher Already Knows: Teacher-Specific Entity Familiarity for Training Deep-Search Agents
SummaryThe paper studies why some questions give weak training signal when distilling search agents from teacher models: teachers familiar with key entities need little search. It proposes Evidence-Verified Entity Familiarity (EVF), an accurate but costly measure, and Self-Reported Entity Familiarity (SEF), a cheap signal where the teacher states whether it knows an entity. Using SEF-gated graphs raised Qwen3.8-27B's BrowseComp subset accuracy from 44.10% to 66.41%.
In briefReflected proximal dynamics: a faster, accurate way to sample smooth log-concave distributions
SummaryThis paper studies sampling from a distribution proportional to exp(−V(x)), where V is strongly convex and smooth. It introduces reflected proximal dynamics, which combines proximal sampling with an exact reflection of an auxiliary variable, kinetic motion, and controlled refreshment to retain directional information. Two realizations are developed; from a warm start with constant χ² divergence, both achieve total variation error ε with Õ(√κ d^{1/4}) expected oracle queries. Including an initialization stage that removes the warm-start assumption, the total cost is Õ(κ^{6/7} d^{1/4}), with polylog(1/ε) dependence on accuracy.
In briefFrom Simulated Fluids to Living Tissue: Transferring a Foundation Model Zero-Shot to Perturbed Embryos
SummaryThe paper finetunes Walrus, a continuum dynamics foundation model pretrained on diverse simulated continuous systems, using tissue velocity fields from only 12 wild-type Drosophila blastoderm videos. It evaluates autoregressive prediction in-distribution and zero-shot under unseen temperature and genetic perturbations. Walrus achieves the best CRPS under the most severe distribution shift, and its predictions capture the mean tissue velocity trajectory, including the peak timing at Germ Band Extension onset.
Louisa Cornelis, Haewon Jeong, Irina Espejo et al.
In briefSeg2Click: Pixel-level masks for locating where GUI agents can click
SummarySeg2Click reframes GUI visual grounding as predicting the region where a click can validly carry out an instruction, rather than a single point. It is trained with bounding-box supervision to output an executable-region mask, and the final click point is computed by a deterministic rule. The authors report 92.06%, 94.10%, and 55.22% on ScreenSpot, ScreenSpot-v2, and ScreenSpot-Pro for Seg2Click-9B, and say it outperforms coordinate-generation and patch-based methods on matched backbones.
WhyReframes GUI grounding as pixel-level executable-region mask prediction, with matched-backbone experiments showing gains over coordinate and patch methods.
In briefAutoNovo: Using AI Agents to Discover Nucleic Acid Design Algorithms That Transfer Across Models
SummaryThis paper uses agentic program search to repeatedly propose, run, evaluate, and revise code for designing nucleic acids. The authors report that AutoNovo designers achieved the best observed mean among evaluated methods on 3 DNA-design tasks in-silico and on 6 transfer tasks across independent models. The designers form a family of discovered methods rather than one universal best designer, and the authors plan to release them.
In briefLearning to GUIDE the Group: Self-Learning LLMs That Facilitate Conversations Across Many Kinds of Group Tasks
SummaryThis paper introduces GUIDE, a framework for building general-purpose language model facilitators that steer conversations among multiple people across different tasks. The model learns a strategy playbook from 432 human group conversations and refines it using feedback from an LLM judge aligned with human annotators. In held-out tests it won 61.4% of comparisons against three task-specific expert facilitators, and in a live study it kept group productivity while improving speaking comfort and reducing friction.
In briefNetRL: A Reinforcement Learning Framework for Network Decisions Using NetGRPO
SummaryThis paper presents NetRL, a two-stage framework using large language models for network decision tasks such as adaptive bitrate streaming, load balancing, and congestion control. It combines supervised fine-tuning on demonstration trajectories with online policy optimization, where the core NetGRPO method evaluates candidate actions over multiple future trajectories with confidence-aware budgeting. The authors report 23.6% higher QoE than NetLLM, 12.2% lower job completion time than Genet, and 24.1% higher network utility than Orca.
In briefLook Back to Move Forward: Learning When to Revisit Earlier Images in Agentic Image Generation
SummaryThe paper introduces GenRevisit, an agentic framework that jointly selects which candidate image to continue editing from and how to refine it, rather than always using the latest image. It is trained with supervised fine-tuning followed by agentic reinforcement learning, along with Normalized Action Credit Assignment (NACA) to distribute advantage across individual actions. On GenEval2, GenRevisit reaches GM scores of 84.3 and 74.4 with Qwen-Image-2512 and Z-Image-Turbo, outperforming the strongest same-generator baselines by 12.7 and 10.9 points. It also reports state-of-the-art results among baselines on GenEval, GenEval++, and GenEval2.
In briefWhy participant-disjoint validation alone may not show generalizable Parkinson's gait classification
SummaryThis paper classifies Parkinson's disease using 159 pressure-walkway summary features from 133 participants. Acquisition site alone reached an AUC of 0.955, higher than the gait-feature model at 0.882, and adding gait features gave no measurable gain. The authors argue that participant-disjoint pooled validation cannot separate disease-related signal from site-associated variation. They recommend within-site or independent validation before claiming generalization across institutions.
In briefOneShot: Label-efficient weighting of detectors for time series anomaly detection
SummaryNo single detector works best for all time series, so the right choice depends on each series. This paper proposes OneShot, which uses one confirmed anomaly to score each detector in a pool and turns those scores into soft mixture weights. Across 152 univariate and 73 multivariate series, OneShot outperforms the strongest single detector, averaging the whole pool, and selecting the best detector on the confirmed event.
In briefSafeDSE: Jointly Choosing Safety Patterns, Mapping, and Scheduling for Dataflow Systems
SummaryThis paper presents SafeDSE, a design space exploration framework for synchronous dataflow applications. A single declarative constraint model jointly decides each actor's safety architecture pattern, fault-model-required placement, static-order schedule, and platform, and prices safety integrity level promotion. On nine benchmark instances, joint exploration was always feasible and never more expensive than rules that fix patterns before mapping. In the generated safety arguments, 49 to 62% of leaves cite verifier checks.
WhyJointly exploring safety patterns, mapping, and scheduling with generated safety arguments is novel. Nine benchmarks and proven optimality on synthetic graphs give solid evidence. Impact is useful wit
In briefMP-MNCA: A Tissue-Morphology-Guided Method for Faster Spatial Domain Identification
SummaryThis paper introduces MP-MNCA, a method for identifying spatial domains in spatial transcriptomics data. It keeps the full 3,000-dimensional highly variable gene representation and uses histological similarity as an attention prior, trained with a self-supervised masked neighbor-contrastive objective. Tested on 12 human dorsolateral prefrontal cortex sections, it reaches coherent spatial structure in fewer training epochs. Its clustering quality is comparable to the strongest baseline, STAIG, with a slightly higher mean refined NMI.
In briefVGGA: Geometry-Grounded Video Autoencoder for Better Generation
SummaryThis paper proposes VGGA, a video autoencoder that adds geometric constraints to reconstruction, aiming to improve latent spaces for video generation. It grounds latent representations in structural depth, motion correspondence, and projective warping, optimized jointly with RGB reconstruction through a shared upsampling trunk. The authors report that gFVD drops by 28.1% on RealEstate10K and 37.7% on Taichi-HD while reconstruction quality remains comparable.
In briefLüshi-1840: A Classical Chinese Model Trained on Pre-1840 Texts to Probe Its Knowledge Cutoff
SummaryThe authors train a 983M-parameter Classical Chinese language model from scratch on 12,262 books, about 1.28B characters, all composed before 1840. Their first corpus, dated by a common heuristic, effectively ended at 1644 and contained twentieth-century material, so they re-dated and cleaned the data. Probing the model on post-1840 annals shows higher surprisal, and the probe reflects source coverage rather than one exact cutoff year.
In briefMSC: A caster-aware control framework combining uncertainty-aware NMPC and an RL execution layer
SummaryThis paper addresses caster misalignment in industrial AMRs and SDVs, which causes a gap between commanded and actual motion during low-speed or high-curvature maneuvers. The proposed Multi-Stage Caster-aware Control (MSC) framework has two parts: an uncertainty-aware NMPC that schedules caster-aware costs using predictive uncertainty, and a reinforcement learning execution layer that estimates caster states and converts velocity commands into executable wheel speeds. The authors report validation in Isaac Sim and on real robots, claiming better caster-state estimation, less misalignment, and more robust tracking than existing decoupled compensation approaches.
In briefCorrect Choices, Failed Sequences: A Benchmark for Whether Models Keep Correct Answers When Built Step by Step
SummaryThe paper introduces Choice2Seq, a diagnostic benchmark that compares whether a model keeps a correct answer when it selects the complete answer at once versus when it builds the answer through sequential local decisions. Some models reach over 97% selection accuracy on arithmetic and lexical retrieval but retain only a small share of those correct answers under sequential construction. The authors also test interventions such as restricting local options to candidate characters and adding explicit position indexing.
In briefExploration strategies for Gaussian process bandit optimization with guaranteed inference regret
SummaryThe paper argues that existing cumulative regret guarantees for Bayesian optimization do not characterize the quality of the final returned output, which matters under noisy evaluations. The authors propose exploration strategies that occasionally replace acquisition queries with exploration queries and can be added to any Bayesian optimization procedure. They bound the Gaussian process posterior standard deviation using the fill distance of the queries, and use this to derive upper bounds on inference regret. Experiments on synthetic benchmarks, hyperparameter optimization, and a control task show improved final recommendations for existing methods.
In briefLabel-centric knowledge retrieval for zero-shot text classification with lightweight models
SummaryThis paper proposes SLKR, a framework that uses a flagship LLM offline to generate label definitions, retrieval keywords, and boundary rules for confusable label pairs, forming a reusable knowledge store. At inference time, hybrid retrieval selects candidate labels and relevant rules, and a lightweight LLM performs classification in a single call. Across three public benchmarks and a multilingual book-classification dataset, the method achieves strong results without task-specific labeled examples.
In briefDIGS: A training-free method for open-vocabulary 3D segmentation using Gaussian splatting
SummaryThis paper presents DIGS, a training-free framework for open-vocabulary 3D segmentation built on Gaussian splatting. It separates class-agnostic instance discovery from linguistic naming and keeps a label distribution per Gaussian primitive, so boundary primitives can preserve multiple instance labels. On LERF-OVS, DIGS reports 71.95% mean mIoU, outperforming the training-free SFS and training-based LaGa baselines.
WhyTraining-free distribution-informed Gaussian labels resolve boundary ambiguity; consistent gains across several benchmarks suggest a useful advance for open-vocabulary 3D segmentation.
In briefEgoAlign: Agent-Centric 3D Understanding Without Text Situation Descriptions
SummaryThis paper addresses situated 3D question answering, where models must interpret a 3D scene relative to an agent's pose. Existing methods often rely on natural-language situation descriptions, which may be hard to obtain in real deployments. The authors propose EgoAlign, a simple 3D language model that represents scenes in an agent-centric frame and adds an egocentric alignment stage with four tasks. EgoAlign performs best on SQA3D, Real-3DQA, and MSQA, and removing text descriptions keeps performance comparable or better, while shuffling its numerical pose degrades results substantially.
In briefProCARE: Automating Real-World Studies Using Data-Profile-Based Evidence Contracts
SummaryReal-world studies analyze routinely collected healthcare data, such as electronic health records and insurance claims, to estimate treatment effects and monitor safety. The paper presents ProCARE, a framework that treats study automation as constrained compilation, converting data profiles into Evidence Contracts and study objectives into a typed intermediate representation. Deterministic validation and localized repair check the generated workflows. On 36 EHR-based benchmark tasks, ProCARE outperforms strong LLM-agent baselines in code executability, semantic consistency, and conclusion traceability.
In briefEventLens: Event-structure reinforcement learning for video understanding
SummaryThe paper notes that multimodal large language models often fail to recover a video's temporal structure, such as event boundaries, single-event progression, and relations among events. It proposes EventLens, a vision-centric reinforcement learning framework that trains models to recover event structure under controlled perturbations, using three verifiable task families with deterministic rewards. Multi-teacher on-policy distillation then consolidates the specialized policies into one model. Experiments on Qwen3-VL backbones report gains on temporal grounding and event-sensitive reasoning while preserving general video question answering performance.
In briefDeepThink: A Conceptual Model for Knowledge-Addressable Recurrent Language Modeling
SummaryThis paper proposes DeepThink, a framework that stores typed knowledge units in a Knowledge Bank, selects them by activating units from the current execution state, assembles them into an executable configuration, and updates the state with a shared thinking operator before generation. It formalizes these interfaces and their execution semantics. It also gives an exact correspondence to an existing fixed-depth model and two initialization methods for moving from pretrained models to dynamically assembled recurrent computation.
In briefGeometric Parrots: Rule Absence vs. Rule Fragility in Vision-Language Models
SummaryThis paper introduces GRIP, a benchmark of 100,000 procedurally generated images and 600,000 questions across 34 reasoning domains, testing vision-language models on geometric, spatial, visual, and physical reasoning. It proposes Dual-Loop Evaluation, where a closed loop uses a five-step question ladder to locate reasoning failures and an open loop re-asks questions under challenges such as peer-model answers and unsupported user assertions. Because ground truth is known exactly, answer changes are classified as corrections or capitulations, separating rule absence from rule fragility.
Bilal Jawaid, Hamza Hanif, Muhammad Maaz Uddin et al.
In briefPredicting how moving solids make liquids flow in 4D
SummaryThis paper introduces Interactive 4D Volumetric Liquid Forecasting, which predicts how a dense 3D liquid velocity field evolves over time given how a solid object moves through it. The authors propose TIDE/TIDES, a two-level recurrent framework, where TIDE handles localized interface-driven forecasting and TIDES adds transport-aware coupling and stabilization to reduce rollout error. Experiments report that TIDES achieves the best forecasting performance, and ablations identify localized interaction, transport coupling, and stabilization as key contributors.
In briefEsports World Model Dataset: Multi-View Human Gameplay from StarCraft II, Dota 2 and CS2
SummaryThe paper introduces the Esports World Model Dataset (EWM), a collection of human gameplay from StarCraft II, Dota 2 and Counter-Strike 2, with 3,053 games and maps and about 4,880 player-view hours. Replay-clock alignment links rendered video with game-native actions, structured state and events from the same match. The dataset covers the differing observation conditions of each game, such as fog of war, shared team vision and first-person views. It is intended to support research on jointly controlled multiplayer world models learned from human play.
In briefThink Before You Speak: Training-Free Latent Communication Between Language Models
SummaryThis paper presents CacheFuse, a training-free method that lets a downstream language model use an upstream model's key-value (KV) cache through linear interpolation. The authors evaluate it on four reasoning and knowledge benchmarks. They report accuracy gains of up to 8.6 percentage points over single-model baselines and time-to-first-token speedups of up to 12.2x over text-based communication in multi-model inference.
In briefFisher Information and the Geometry of Memorization in Neural Networks
SummaryThis paper studies how neural networks separate generalizable structure from noisy or idiosyncratic features in training data, using Fisher information as a diagnostic. In vision models trained on CIFAR-10 with controlled noise that can only be learned by memorization, projecting weights onto the top Fisher eigenvectors separates general task performance from noisy-sample memorization, with differences across layers. The same effects appear in small transformers on a synthetic in-context learning task and are reproduced in closed form in an analytically solvable linear model.
Daniel T. Bernstein, Chase Goddard, Luca Di Carlo et al.
In briefCIPHEUR: LLM-Guided Search Control for Optimizing Conflict Graphs
SummaryThis paper proposes CIPHEUR, a framework that uses a large language model to link conflict evidence to structured intervention units that guide search past stagnation in conflict-graph optimization. Tested on satellite-ground scheduling, its online variant achieves the highest mean objective among 11 configurations and increases mean scheduled contact duration by 14.84% over the greedy baseline GWMIN. The authors also report that a one-shot LLM-generated controller improves on CHILS without runtime model calls.
In briefStress-testing LLM agents with realistic, task-valid user personas
SummaryStandard LLM-based user simulators tend to be overly cooperative, which can overstate how robust interactive agents really are. The paper introduces TIPER, which evolves user personas through behavioral mutations and Pareto selection while keeping the task valid and solvable. Across four interactive-agent benchmarks, TIPER lowered held-out avg@3 scores in all 12 GPT-5.4 settings and in all 28 agent-domain pairs across seven agents. The authors also provide a taxonomy linking discovered user behaviors to recurring agent failure modes.
WhyProposes persona evolution under task-integrity constraints, beyond predefined user behaviors; consistent score drops across four benchmarks and seven agents, with impact mainly on agent robustness ev
In briefWildTrace: Benchmarking Training Data Detection in Realistic Large-Scale LLM Post-Training
SummaryThis paper examines training data detection for large language models after post-training, motivated by concerns about benchmark contamination and copyright. It introduces WildTrace, a benchmark covering models from SFT, DPO, and RL stages, with distribution-matched members and non-members across four tasks. Even the best method reaches only 0.529 average AUC, close to random guessing.
In briefLearning equilibria in unbounded random games with unknown Gaussian transitions
SummaryThe paper studies discounted general-sum stochastic games with unbounded state space and an unknown nonlinear drift observed under Gaussian noise. It proposes learning from a single trajectory, using Gaussian-process models of the drift and solving successively refined finite games. It proves that stable equilibrium selection and uniform KL accuracy of the transition models imply regional uniform convergence of values and approximate Nash behavior.
In briefBARTER: Compact Dynamic Sparse Training Using Standard Dense Kernels
SummaryThis paper presents BARTER, a framework for dynamic sparse training that keeps training tensors physically compact by pruning and cloning attention heads and feed-forward channels, instead of relying on sparse kernels. Experiments with Transformer-based vision models on two image classification benchmarks show Top-1 accuracy competitive with full-tensor baselines. On an NVIDIA L40, it reports training speedups of up to 6.29 times and peak GPU memory reductions of up to 81.3%.
In briefUnravel: Rebuilding a Movable 3D Room from One Photo
SummaryUnravel is a training-free method that turns a single image into a 3D scene in which each object can be moved independently. It works by peeling the scene apart: a vision-language model picks fully visible objects, and an image editor removes them and fills in hidden areas, with each edit verified before reconstruction. The scene is then rebuilt in reverse removal order, physically settled, and audited to fix placement errors. According to the paper, experiments on diverse indoor scenes show it recovers more objects and produces more physically plausible scenes than existing methods.
In briefPollard: Governed Execution Trees for Controlling and Replaying AI Agents Across Frameworks
SummaryThe paper presents Pollard, a framework-independent runtime that records execution, governance, and replay in a single governed semantic execution tree. Experiments include replicated state-model evaluation, concurrent resource reservation tests on PostgreSQL, strict replay that avoids external calls, and detection of semantic mutations. Multi-framework and cost-capped hosted-model studies found every completed trial produced a verified tree, supporting the core mechanisms under stated lease, boundary, and seal assumptions.
In briefHashrope: a persistent data structure for managing LLM context
SummaryThis paper proposes Hashrope, a persistent weight-balanced rope in which each node stores a polynomial hash of its subtree over a Mersenne prime. The authors argue it can handle four LLM context-management mechanisms: incremental edits, branching and snapshots, KV-cache prefix reuse, and repeated-segment encoding. They compare it against Ropey, a PagedAttention block table, and the SGLang radix matcher using pre-registered, replicated experiments. The reported results include a speedup over Ropey that grows with size, and exact identification of a one-million-token shared prefix on a live vLLM deployment.
Muntaser Syed (Florida Institute of Technology), Ameerah Alsulami (Florida Institute of Technology), Marius Silaghi (Florida Institute of Technology)
In briefAll-St3R SLAM: Real-time RGB SLAM that removes moving pixels
SummaryThis paper presents a real-time RGB SLAM system for dynamic environments that removes moving objects at the pixel level without predefined labels or segmentation models. It compares pixel motion predicted from ego-motion using DUSt3R with point tracking from All-Tracker, removing only pixels that actually move. The system runs at about 10 FPS by subsampling every third frame of a 30 FPS stream, and the authors report this outperforms prior dense dynamic SLAM methods running at 0.5 to 3 FPS.
In briefDistributional regression for arbitrary responses using ReLU networks
SummaryThis paper proposes a unified framework for distributional regression that handles arbitrary response variables by modeling conditional event probabilities indexed by measurable sets. Two estimators are studied: a preliminary one based on separately regressing event indicators, and a constrained one that enforces inclusion-consistent predictions via a directed acyclic graph over a finite set family. General error bounds are established, and for dense ReLU networks the authors derive explicit convergence rates that overcome the curse of dimensionality. Simulations and real-data experiments demonstrate practical effectiveness.
Junpeng Ren, Carlos Misael Madrid Padilla, Oscar Hernan Madrid Padilla
In briefPhantom Merge: LLM Agents Pick One Entity but Describe Another
SummaryThe paper identifies Phantom Merge, a failure where an LLM agent attributes unsupported claims to the entity it has committed to, even when the task is completed correctly. The authors measure its prevalence across four agent domains and four model families. They propose a two-stage method: Anchor-Grounded Risk (AGR) estimates claim-level risk from hidden states and Jacobian information, and Fixed-Anchor Correction (FAC) revises or removes flagged claims using observed evidence. On the Shopping detection cohort, AGR-slot reaches an AUROC of 96.2%, and FAC reduces trajectory-level Phantom Merge from 51.4% to 6.2%.
WhyFormalizes a new entity-evidence binding failure and reduces trajectory-level PM from 51.4% to 6.2% on Shopping, with AUROC 96.2%, supported by experiments.
In briefLLM-Assisted Transportation Planning with a Differentiable User Equilibrium Layer
SummaryThis paper proposes a human-AI transportation planning framework combining a large language model, a model context protocol (MCP) interface, the open-source planning engine AequilibraE, and a differentiable user equilibrium assignment layer. The UE layer is built by directly differentiating the KKT conditions of the Beckmann traffic assignment problem. The framework lets non-expert users run network analyses through natural language. It is demonstrated on calibrating travel time functions from observed traffic counts and on network design under budget constraints.
In briefOmni-modal LLMs know more than they can express
SummaryOmni-modal large language models process vision, audio, and IMU signals alongside text, but the paper finds they often fail to express non-textual information even when they perceive it correctly, which the authors call the knows-but-cannot-say gap. The authors introduce Token-level Modality Alignment (TMA) and show that non-textual representations align less with response tokens and receive less attention than textual ones. They propose OmniCatalyst, a perceive-then-infer strategy that supplements inputs with self-generated text, which improves performance by 7.29% across seven models and yields more balanced modality preferences.
In briefProMoDD: Shrinking 3D Human Motion Datasets with Prototype-Guided Diffusion
SummaryThis paper proposes ProMoDD, a dataset distillation method for skeleton-based action recognition that aims to keep a large dataset's training value in a much smaller set of examples. It conditions a motion diffusion model on class-wise prototypes of temporal recognition features and combines real and generated sequences through budget-aware hybrid selection. On single-person classes of NTU RGB+D 60, it reaches 40.32% top-1 accuracy on Full49 with three sequences per class, 4.57 points above class-wise k-means.
In briefSolving Puzzles by Asking: A Benchmark for Multi-Turn Knowledge Reasoning
SummaryThis paper introduces the Active Probing Benchmark (APB), which frames concept identification as an interactive puzzle in which models must gather information through multi-turn questioning and inference. Models must pinpoint 960 specialized target concepts across eight academic domains. The evaluation finds that most open-source reasoning models perform significantly worse than their standard instruction-tuned counterparts. The authors also provide a scalable pipeline for generating benchmarks that test knowledge reasoning in interactive settings.
In briefXForecast: A KDD 2026 Challenge for Forecasting Stock Moves Using News
SummaryThis paper introduces XForecast, a benchmark and competition that asks participants to predict stock price movements four weeks ahead using historical prices and paired news. It covers 100 large-cap stocks, training on 2019–2022 and evaluating on 2023. The authors use a weighted hit rate metric that gives more weight to cases where news is expected to have stronger influence. The challenge is hosted on Kaggle as part of KDD 2026 Finance Day.
In briefPID-gated continual prompt adaptation for LLM agents facing noisy feedback
SummaryThe paper addresses how LLM agents in changing environments should update their instructions when feedback is noisy, since rewriting after every mistake can lock in errors. It proposes lightweight gates inspired by proportional-integral-derivative control that respond to current error, accumulate evidence over time, and roll back harmful edits. Across drifting tasks, model sizes, noise levels and seeds, the gates reportedly give a better trade-off between task performance and editor-call budget than baselines including TextGrad and GEPA.
In briefA Concept-informed Interpretable AI Framework for Breast Ultrasound Diagnosis
SummaryThis paper presents IBUS, an AI framework for breast ultrasound that incorporates BI-RADS structured descriptors to better match radiologists' reasoning. It was built on a multicenter dataset of 9,941 patients across 20 datasets and tested on 14 independent datasets. In a reader study with eight radiologists, IBUS assistance increased diagnostic sensitivity for both junior and senior readers.
In briefA benchmark testing whether AI models can interpret Hong Kong fire safety codes
SummaryThis paper evaluates how well large language models interpret Hong Kong fire safety codes, a task relevant to regulatory compliance in construction. The authors, working with domain experts, built a localized benchmark with multiple-choice and open-ended questions and tested 12 proprietary and open-source models. Top proprietary models reached 80-90% accuracy on multiple-choice questions, while expert assessments of open-ended answers found only 50-70% accuracy.
Le Qiu, Zhuan Anne Zhang, Hok Ming Adeline Chan et al.
In briefImitation as Evolution: Local Repair and Verified Skill Learning for Agents
SummaryThe paper proposes passive imitation (PI), a loop that finds the violated part of an agent's experience, derives a local goal, tests a small patch, and keeps evidence only after full-task verification. In ten simulated tool environments, PI solved all 600 held-out composition tasks with three small trials, versus 6.17% for a whole-trajectory baseline. On an unseen tool revision it reached 70.0% success, and the authors stress this is a limited correction-and-retention mechanism, not general intelligence.
