In briefPhysical Muon: Orthogonalization as Equilibrium Computation
SummaryThis paper introduces Physical Muon, which computes the orthogonalization step of the Muon optimizer as the equilibrium of a continuous-time flow instead of using dense Newton–Schulz matrix products. Random probes approximate the flow using matrix-vector products, reciprocal reads, and local rank-1 writes, aiming for analog hardware implementability. On a 10.95M-parameter transformer, the dense flow's validation cross-entropy was 0.0085 above Newton–Schulz, and the probe implementation was 0.0188 above the control. Circuit simulations reproduced the flow dynamics with comparable training behavior.
In briefGC-OPD: graph-guided on-policy distillation of language agents from unmodified teacher models
SummaryThis paper proposes GC-OPD, a method for training compact language agents using teacher feedback on student-generated trajectories. A graph organizes repeated teacher executions by shared states, and the teacher's scoring is enriched with retrieved execution evidence. On ScienceWorld, ALFWorld Unseen, and WebShop, it raises mean success over vanilla on-policy distillation without any task-specific teacher optimization.
In briefVideo2STL: Turning Video-Generated Temporal Rules into Robot Learning Signals
SummaryThis paper presents Video2STL, a framework that converts observation-only videos into parametric Signal Temporal Logic (STL) specifications for robot learning. A vision-language model extracts an embodiment-independent event trace and builds symbolic temporal specifications, while numerical thresholds and time bounds are grounded from successful robot trajectories. The authors report higher success rates than native dense PPO and Text2Reward on four manipulation tasks, and strong results on quadruped locomotion.
Merve Atasever, Keyan Azbijari, Cagan Bakirci et al.
In briefCompressing Vision-Language Model Benchmarks Hierarchically
SummaryFull evaluation of vision-language models has become very expensive as benchmarks grow. This paper presents PRIMEBench, a vision-aware hierarchical framework that compresses benchmarks through data cleaning, category representative selection, Vision-Aware Variance-based item pruning, and category-count pruning. At the released 5% retention, it achieves the highest mean fidelity on models not used for item selection. The released suite removes over 97% of items while preserving model rankings.
In briefFrom Dissonance to Orchestration: Teacher Intervention in On-Policy Distillation
SummaryOn-policy distillation trains a student model on its own reasoning trajectories using feedback from a stronger teacher. The authors find that rollout quality alone is not enough to decide where teacher guidance should go, since deeper interventions give diminishing accuracy gains while increasing off-policy load. They propose MAESTRO, which uses local policy disagreement to adapt when the teacher takes over and how long it generates. Across eight mathematical reasoning benchmarks, MAESTRO achieves the highest macro-average accuracy for both 0.6B and 1.7B Qwen3 students and cuts average training response length by 67.3% relative to standard OPD.
In briefScaGNN: A Graph Neural Network for Simulating Multiple Wave Scattering
SummaryThis paper introduces ScaGNN, a learning-based method that approximates the boundary solution the first step of the boundary element method uses for multiple scattering problems. It uses a graph neural network with a dynamic adaptive edge sampling mechanism that selects the most relevant distant interactions to model, giving linear complexity in the number of input graph nodes. The authors present a benchmark of several multiple scattering datasets, where ScaGNN outperforms existing learning-based methods. They also test generalization to more obstacles and out-of-distribution obstacle shapes.
In briefExamining Data and Simulator Assumptions in Supervised Causal Discovery
SummaryThis paper examines supervised causal discovery, which learns to infer causal structure for new datasets from simulated training pairs labeled with structure. It analyzes how the simulator's assumptions about graphs, mechanisms, and noise supplement the information in observational data. The authors distinguish whether a target is identifiable under the assumed model class from whether a predictor generalizes beyond its training distribution, and recommend evaluations that test changes in graphs, mechanisms, and noise.
In briefTraining a Planner to Retrieve Missing Evidence for Long-Term Memory QA
SummaryThis paper introduces MERA (Missing-Evidence Retrieval Augmentation), which separates globally searchable memory from a question-specific evidence state so verified evidence guides later retrieval. A lightweight planner is trained with reinforcement learning to reward queries that recover previously missing evidence. A trained 0.6B planner reaches 77.40% accuracy on LoCoMo and 71.29% on LongMemEval-S with Qwen3-30B, outperforming a 30B planner without retrieval-grounded training by 4.10% and 3.96%. On LoCoMo, cumulative evidence recall rises from 55.5% to 80.5% in later retrieval rounds.
In briefA variational Koopman autoencoder that gives probabilistic time series forecasts with uncertainty estimates.
SummaryThis paper introduces the Variational Augmented Invertible Koopman AutoEncoder (VAIKAE), which makes the latent embedding Gaussian rather than deterministic so that prediction uncertainty can be quantified. It uses normalizing flows to allow likelihood computation in the dynamical system's state space during training. The authors also propose uncertainty-aware latent data assimilation strategies for a trained VAIKAE, and report experiments on long-term time series forecasting benchmarks.
Anthony Frion, Lucas Drumetz, Guillaume Tochon et al.
In briefLooped Actor: Depth-recurrent policy models that adjust how long they think for reinforcement learning
SummaryThis paper proposes Looped Actor, a transformer-based policy that repeatedly refines a latent representation with a shared computation block and varies the number of loops depending on the current state. The authors evaluate it on 22 tasks across six environments, including puzzles and robotic manipulation, in both online and offline settings. The model matches or exceeds an untied baseline with 16 times more parameters, with the largest gains in tasks requiring multistep planning.
T. Konstantin Rusch, Tim Seyde, Jared Boyer et al.
In briefMemory method for finding objects over long periods using past observations
SummaryThis paper addresses long-term object search, where a robot must learn where objects usually appear from repeated, uneven observations. It introduces ECROM, which weights each detection or non-detection by the robot's opportunity to observe that location, estimates prevalence for concepts specified only at query time, and turns the belief into an active-search prior. The authors also present a controlled long-term benchmark across ten HM3D homes that varies object placement and observation opportunity independently.
In briefLazySloth: a fast long-video understanding method using bounded lazy tree search with LLMs
SummaryThis paper presents LazySloth, a tree-based search method for long-video understanding and retrieval. It captions only the video portions that a vision-language model judges irrelevant to the query, in a bounded way, to reduce computation. The authors report 2.9 to 8.3 times speedup over existing agentic methods, with similar or better accuracy on four benchmarks, and a narrower gap to GPT-4o for open-source base models.
Arka Mukherjee, Kaleen Shrestha, Larissa Zhu et al.
In briefSimultaneous Neural Optimal Transport: one shared map aligning several sources to one target
SummaryThis paper studies learning one shared transport map that simultaneously aligns several source distributions with a common target distribution, using unpaired samples. The authors propose a neural method that minimizes the average transport cost while aligning each source with the prescribed target, and derive a max-min formulation for learning the map. They illustrate the approach on image restoration, where a single model handles multiple degradation types using a shared collection of clean target images.
Milena Gazdieva, Kirill Sokolov, Jiawei Chen et al.
In briefRobust LassoNet: Making Neural Feature Selection Resistant to Outliers
SummaryThis paper studies feature selection in neural networks, which becomes difficult when data are noisy or contaminated. It proposes Robust LassoNet, which replaces the standard mean squared error loss in LassoNet with robust losses such as Huber, Cauchy, Tukey's bisquare, and Nonnegative Garrote. Experiments on synthetic and real data show better prediction and feature selection accuracy with outliers or heavy-tailed noise, and comparable performance on clean data.
Daniela De Canditiis, Italia De Feis, Paola Stolfi
In briefScale Sensitivity in Low-Bit Post-Training Quantization: How Curvature of the Error Landscape Shapes Scale Choice
SummaryThis paper studies how the scale of the uniform quantization grid affects layer-wise reconstruction error in GPTQ-style post-training quantization. For wide networks with high effective-rank calibration data, the authors show the normalized error converges to the mean-squared error of a uniform quantizer on a standard Gaussian, with an optimal scale that shrinks as the number of levels grows. Experiments on five LLMs show scale choice strongly affects perplexity at 2 to 3 bits but barely at 6 bits and above. After Hadamard incoherence processing, the Gaussian-optimal scale matches the best searched rule at 3 bits and above, though it remains worse at 2 bits.
Jonas von Berg, Massimiliano Datres, Carlo Knei{\ss}l et al.
In brief"Look What You Made Us Cluster": Extracting Hate Narratives from Reddit Discussions
SummaryThis paper presents a pipeline that extracts online hate narratives by representing each as an entity-evaluation pair. It uses LLM reasoning that extends Aspect-Based Sentiment Analysis, followed by Leiden clustering and LLM-guided refinement to reach a chosen granularity. The method is illustrated on 2024 English Reddit comments criticizing Taylor Swift, with one representative cluster showing hate speech patterns analyzed for its interpretive value.
Annabelle K. L. Chua, Forster J. Khoo, Joel C. R. Tan et al.
In briefRetrieval-based prompting to keep characters and scenes consistent across long videos
SummaryLong-form video generation struggles to keep characters, scenes, and objects consistent across independently generated shots. The authors propose an agentic framework that parses the script into a registry of visual elements and retrieves a compact set of complementary historical references from a VLM-annotated keyframe library. These references, element states, and grounding instructions are combined into a prompt for a frozen video generator, requiring no retraining. The method reportedly outperforms recent-frame, memory-based, and entity-level retrieval baselines on cross-shot consistency and text controllability.
In briefComplexity-based testing of how well LLMs understand code
SummaryThis paper proposes a framework that evaluates large language models' code comprehension by grouping functions by structural complexity, using cyclomatic complexity, nesting depth, branching factor, and Halstead volume. DeepSeek-Coder-V2 and Llama were tested on automatic input-output prediction for 300 Python functions and on manually assessed semantic comprehension for 60 functions. Both models became substantially less accurate as complexity increased, and the authors argue this breakdown is more diagnostic than overall accuracy.
Ali Mohammadi Esfahani, Nafiseh Kahani, Samuel A.Ajila
In briefRouting That Pays for Itself: Sparse Supervision for Cost-Effective LLM Routing
SummaryThis paper proposes SaveRouter, a routing framework for large language models that uses sparse supervision by selectively collecting informative model feedback and sharing capability information across related queries. Across four routing benchmarks, the main setting uses only about 33–41% of training feedback while keeping routing quality competitive or better. It also cuts the break-even deployment volume by roughly 1.9–9.5 times compared with the fastest conventional router.
In briefBeatDance: beat-consistent 3D dance generation from music using hierarchical spatial-temporal modeling
SummaryThis paper proposes BeatDance, a diffusion-based framework that generates 3D dance motion from music. It uses a Hierarchical Decoupled Attention module that separately learns human pose and temporal dynamics while capturing both short-term and long-term dependencies. A cycle-consistent training scheme with an auxiliary dance-to-music module further encourages dance motion to stay synchronized with the music. The authors report that the method outperforms recent competing approaches on two benchmark datasets.
In briefDirect Experience World-Model Optimization: Learning the World Beyond Action Imitation
SummaryThe paper proposes DEWO, a post-deployment learning method for World-Action Models that refines world representations from visual experience alongside action imitation. It learns from successful and failed future continuations and uses a value head to activate guidance when task progress stalls. The authors report improved success on five simulated DexJoCo tasks and on four real-world tasks.
In briefRethinking Soft Tokens for Parallel Decoding in Diffusion Language Models
SummaryDiffusion language models can generate several tokens at once, but these tokens may be individually plausible yet inconsistent with each other. This paper studies how soft-token feedback, which represents uncertain positions with continuous embeddings, improves parallel decoding in frozen pretrained models without extra training. It identifies a geometric mismatch with conventional soft-token construction and proposes a training-free, geometry-aware alternative. Across four models and four math and code benchmarks, the method outperforms standard parallel decoding and a Euclidean soft-token baseline.
In briefCan Randomized Audits Keep AI Agents That Hide Misconduct in Line? A Theory Paper
SummaryThis economics theory paper asks when randomized audits and scoring can align AI agents that might conceal misconduct and alter records. It argues that stronger auditing makes undeterred violations harder to detect. The authors state that rare audits can deter all agent types if evidence survives concealment and audit draws cannot be predicted in advance. If evidence can be erased, deterrence must instead come from lower gains from violation or costlier concealment, and the paper links these conditions to a July 2026 incident involving OpenAI cybersecurity evaluation agents.
In briefMulti-task learning for automatically grading perivascular space burden in brain MRI
SummaryThis paper addresses enlarged perivascular spaces visible in brain MRI, which are linked to poor brain health and are usually scored visually by radiologists. The authors train machine learning models to automatically grade PVS burden in the basal ganglia and centrum semiovale using the Potters/Wardlaw scale, incorporating semi-automatically generated imperfect segmentation masks during training. A multi-task CNN performing both segmentation and scoring achieved the best mean average precision of 64.08%, outperforming the conditional CNN, baseline CNN, and logistic regression approaches.
Jesse Phitidis, William N. Whiteley, Joanna M. Wardlaw et al.
In briefPredicting Bronchoscopy Accessibility from 3D CT Using Anatomy-Aware Modeling
SummaryThis paper formalizes bronchoscopic accessibility prediction as a new supervised learning task and introduces the first end-to-end framework for it. The authors propose an Anatomy-Aware Mixture-of-Experts model combining a CT expert, a lobe anatomy expert, and a path geometry expert for the bronchial tree. They curated a clinical dataset of 438 cases with pre-operative CT and procedural outcomes, reporting an AUROC of 0.8052, which outperforms baselines and experienced human experts.
In briefMoTIF-X: An interpretable, extensible multimodal framework for molecular representation learning using chemical motifs
SummaryThis paper presents MoTIF-X, a framework that uses graph-based chemical motifs as shared anchors to integrate molecular graphs, SMILES strings, and 3D conformations. It pretrains motif representations with hierarchical contrastive learning, then contextualizes them with SMILES and torsion-angle tokens via multimodal masked token modeling. On all nine OpenADMET ExpansionRx endpoints it achieved the lowest mean absolute error, and it also extends to drug-target interaction prediction.
In briefSimulation-based inference over compressed parameter spaces
SummaryNeural simulation-based inference (SBI) typically infers a small set of interpretable parameters from high-dimensional observations. This paper proposes combining SBI with latent generative modeling: it learns a low-dimensional representation of simulator parameters, performs posterior inference in that latent space, and maps samples back to the original parameter space. The authors characterize when latent-space inference recovers the target posterior and compare it with direct inference across four case studies and three generative families. At matched training compute, it reaches accuracy and marginal calibration comparable to direct inference while sampling up to more than an order of magnitude faster.
Lars K\"uhmichel, Stefan T. Radev, Bhanu Prasanna Koppolu et al.
In briefLooped Transformers as Optimizers: treating the loop state as fast weights and designing loop transitions as local gradient-style updates
SummaryThe paper studies how to design loop transitions for looped Transformers, which repeatedly apply shared blocks to scale depth efficiently. The authors treat the looped hidden state as a fast weight updated across depth and derive transitions in closed form from a projection, a local objective, and an optimizer update rule. They align the input maps of existing transitions and propose OperLoop, which adds explicit weight decay, adaptive step size, and a delta objective. Under matched training FLOPs, the aligned variants reduce training loss and improve average commonsense accuracy, and OperLoop improves average generative performance over the compared baselines.
In briefDo-JEPA: Training latent world models to predict action effects by intervening on the world
SummaryThis paper proposes Do-JEPA, a training objective for latent world models that predicts the difference between latent futures produced by an action and a reference action from the same simulator state. Unlike object-masking methods that change what the model can see, it intervenes on the physical dynamics to separate an action's causal effect from mere correlation. Experiments on a synthetic system, pixel-based models, and CausalWorld benchmarks report lower effect prediction errors than a control trained on the same data, with a small accuracy cost when trained from scratch that fine-tuning removes.
In briefHow prompt data shapes knowledge transfer in on-policy distillation
SummaryThis paper studies how the number, source, and selection of prompts affect transfer from a teacher model to a student model in on-policy distillation, where the student learns from teacher feedback on its own responses. The authors find that a few prompts can nearly match large prompt pools, but a prompt's usefulness depends on the teacher-student pair, since changing only the teacher can reverse the relative effectiveness of math and code prompts. They conclude that effective data choice depends on the model pair and target capability, and that random sampling is a competitive baseline in the settings studied.
In briefMG-Thinker: A Reinforcement Learning Method for Pixel-Precise Grounding Across Multiple Images
SummaryThis paper introduces MG-Thinker, a reinforcement learning post-training framework for multi-image reasoning grounding, which aims to localize objects precisely across real-world multi-image contexts. The authors release a curated 25K dataset with task-adaptive chain-of-thought annotations and propose Bi-Axial DAPO, which handles uneven task and sample difficulty through intra-group and inter-group advantage decomposition. Experiments report state-of-the-art results on multi-image reasoning grounding, along with improved generalization to multi-image understanding and other multimodal benchmarks.
In briefThink First, Then Score: A Thinking Reward Model for Visual Generation
SummaryThis paper proposes a "Think Before You Score" approach, implemented as the Thinking Reward Model (TRM), which creates case-specific rubrics, assesses outputs against them, and outputs fine-grained pointwise rewards. The authors also introduce Pairwise Dual-Group Relative Policy Optimization (PD-GRPO) to improve reward discrimination using pairwise supervision while keeping fine-grained scores. On image generation and editing reward benchmarks, TRM achieves state-of-the-art results among open-source reward models and is competitive with proprietary ones. Using TRM as a reinforcement learning reward consistently improves various visual generation models.
In briefCompiling learning problems into model adaptation programs
SummaryThis paper proposes adaptation compilation, which treats where, how, and how far a language model should adapt as a joint prediction and decision problem. A compiler learns from prior adaptations to predict how candidate update programs will affect acquisition, transfer, boundedness, and preservation, and selects a program before adaptation begins. On Llama-3.1-8B, selected programs approach exhaustive search and outperform global and objective-specific defaults. Replication on Gemma-2-9B shows similar program variation and selection headroom.
In briefPretrain once, route anywhere: a foundation model for LLM routing
SummaryThe paper introduces RouteFM, a model that learns a reusable capability for routing queries to suitable large language models. It is pretrained episodically across heterogeneous routing environments and adapts to new settings through context alone. On MMR-Bench, which is excluded from pretraining, it outperforms the strongest baseline by 2.23 quality points using only eight observations per candidate.
WhyProposes a reusable foundation-model approach to LLM routing, showing transfer to an unseen benchmark, though comparisons across settings remain limited.
In briefSemOPT: Fixing Semantic Errors in LLM Optimization Modeling via Reward-Guided Search
SummaryOptimization modeling turns natural-language problem descriptions into solver code, a step that large language models can automate but often get wrong. This paper highlights semantic errors, where code runs and returns a value yet violates the problem's intent, which are hard to detect because they cause no runtime failure. The authors propose SemOPT, combining a semantic reward model that separates faithful models from plausible but incorrect ones with an adaptive correction system that performs hierarchical reward-guided search. Across seven optimization modeling benchmarks, SemOPT reports a new state of the art, with an average 7.6% accuracy gain over the strongest baseline on complex datasets.
In briefSynthetic Visual Anomalies for Model Selection When Defect Data Is Scarce
SummaryThis paper presents a framework that synthesizes severity-graded defects on real defect-free industrial images, without defect references for the target asset, to support model selection and validation. On MVTecAD, nearly halving the image AUROC choice regret compared with the best fixed model is reported. In a hydropower case study, a PatchCore-based model fit on Pelton turbine images was selected and validated with synthetic images, reaching 94% correct detection at optimal threshold and AUROC 0.97. Moderate and advanced defects were detected reliably, while early-stage defects remained challenging.
Daniel Pr\"oll, Thomas Kraxner, Tobias Schaefer et al.
In briefEncore: Few-Shot Agentic Discovery of Manipulation Strategies
SummaryThis paper introduces ENCORE, a method where a coding agent reads a few demonstrations as evidence and writes a robot manipulation policy program, which is then frozen before sealed evaluation. On LIBERO-PRO, frozen programs reached 96.3% success, above the strongest prior agentic system's 89.3% with the same language model. On RoboDojo tasks whose instructions leave the goal unstated, no program succeeded without demonstrations. The method was also run on a real bimanual robot, learning cube handover and cup inversion from five demonstrations each.
In briefNinaXander: Testing whether frozen language models from different architectures can be combined through a shared latent space
SummaryThis paper proposes NinaXander, which links layers of frozen language models from different architecture families using one trained shared-latent adapter, producing composed models without retraining. The authors combined the recurrent RWKV-4-Raven-7B and the Transformer-based Tulu-Pythia-6.9b. One configuration, using the first 5 Pythia layers with the remaining 27 RWKV layers, cut the Transformer KV cache by 84.4% with accuracy not significantly different from RWKV alone. However, no composed model matched Pythia on multiple-choice questions, and performance dropped sharply on WikiText. The authors note that the representation correspondence does not show the models share a general semantic space.
Takanori Kotama, Shun-ichiro Hayashi, Daichi Mukunoki et al.
In briefTraining LLMs to Create Hard Mixed-Integer Optimization Problems Using Solver Feedback
SummaryThe paper proposes a challenger-solver asymmetric self-play method in which an LLM generates progressively harder feasible MILP instances without seed or training instances. Hardness is measured by branch-and-bound nodes and post-cut relaxation gaps, and the solver's feedback forms the reward used to fine-tune Gemma-4-12B and Qwen3.5-4B into OptiScribe models. On capacitated facility location and max-cut, OptiScribe-4B raises median SCIP nodes by up to 15.6x.
WhyProposes challenger-solver self-play with solver-derived hardness rewards, substantially raising branch-and-bound nodes and cut gaps over base models on two MILP families.
Jitin Singla, Parikshit Pareek, Pratik Jawanpuria et al.
In briefSeek Before You Move: Evidence-Seeking for Reliable Progress Grounding in Vision-Language Navigation
SummaryThis paper addresses vision-language navigation agents that proceed confidently despite missing task-relevant evidence, a failure the authors call Progress Myopia. The proposed SeekVLN framework combines semantic progress reasoning with active acquisition of task-relevant observations. It is trained with Future-guided Reverse Generation for supervised fine-tuning and Counterfactual Contrastive Policy Optimization for reinforcement fine-tuning. The authors report success-rate gains of 12.7% on R2R-CE and 7.5% on RxR-CE over the base model.
In briefStabilizing long latent reasoning chains against accuracy collapse
SummaryThe paper studies a failure mode called the Deliberation Drift Cliff, where accuracy in iterative latent-space reasoning falls sharply as the number of reasoning steps grows. The authors propose Port-Hamiltonian Latent Deliberation and a Direct-Gradient Pure-Tensor Helmholtz-Hodge Decomposition to balance expressivity and stability. They report a 58.67% peak accuracy and 35.27% retained at K=32 on a benchmark, and monotonic gains when transferred to small language model multi-hop causal reasoning.
Zeyu Jia (School of Biomedical Engineering, Technology, Tianjin Medical University et al.
In briefPCaPaint: Prostate cancer MRI tumor inpainting that avoids copying the input
SummaryThis paper presents PCaPaint, a latent diffusion model method for synthesizing tumors in prostate cancer MRI scans. It addresses shortcut learning, where the model merely reproduces the masked input, by filling the condition image with Gaussian noise and emphasizing lesion-region error in training. It also uses separate autoencoders for T2w and DWI&ADC sequences. The authors report that the synthetic data improves prostate lesion segmentation, patient-level classification and lesion-level detection, and outperforms a recent LDM-based tumor inpainting method.
Levente Lippenszky, Hongxu Yang, Marcell D\"om\"ot\"or et al.
In briefTAEC: A Training-Free Method That Keeps Visual Evidence Useful Across Multi-Step Retrieval
SummaryMulti-step visual retrieval-augmented generation answers complex questions by repeatedly retrieving visual evidence and deciding whether to keep searching. The paper notes that retrieved evidence can become less useful over a reasoning trajectory, as redundant or stale content crowds out needed information. It proposes TAEC, a training-free framework that tracks unresolved answer requirements in a shared state to decide which evidence enters context, how memory is kept, and how finely visual evidence is examined. TAEC achieves the highest average accuracy among training-free visual RAG baselines on ViDoSeek, SlideVQA, and MMLongBench-Doc across several proprietary vision-language models.
In briefWhen to Retrieve, When to Stay: Uncertainty-Aware Evidence Selection for Streaming Video-LLMs
SummaryThis paper addresses how streaming Video-LLMs should choose visual evidence under a bounded processing budget, balancing recent frames against query-relevant history. It proposes WRWS, a training-free framework that uses a lightweight vision-language encoder for relevance scoring and the normalized entropy of the similarity distribution as a retrieval uncertainty proxy. Experiments on StreamingBench and OVO-Bench report competitive accuracy, and the method reduces average vision-to-answer time to 47.93% of the state-of-the-art method.
In briefSharp Threshold in Differentially Private Data Reconstruction: Privacy Budget Relative to Data Dimension Determines Whether a Training Sample Can Be Recovered
SummaryThe paper studies informed attackers who know all other training data and try to reconstruct a single d-dimensional training sample from a model satisfying ρ-zero-concentrated differential privacy. It shows a sharp transition at ρ ≍ d: for ρ much smaller than d, entropy-based lower bounds show reconstruction is information-theoretically impossible, while for ρ much larger than d, a simple attack on private linear regression with output perturbation makes reconstruction practically feasible. For data in an s-dimensional subspace, the transition moves to ρ ≍ s. Experiments on synthetic data and CIFAR-10 and ImageNet images validate the findings.
WhyEstablishes a sharp ρ≍d reconstruction transition with matching information-theoretic lower bounds and an attack upper bound, validated on synthetic data and CIFAR-10/ImageNet.
Max Cairney-Leeming, Simone Bombari, Marco Mondelli
In briefHyperSAM: A Promptable Foundation Model for Hyperspectral Remote Sensing
SummaryThis paper presents HyperSAM, a promptable foundation model for hyperspectral remote sensing. It synthesizes full-spectrum hyperspectral data from high-resolution SpaceNet multispectral imagery and builds a spectral adaptation architecture on Segment Anything Model 3 (SAM3), with confidence-based selection to handle noisy pseudo-labels. The authors report strong generalization across tasks including classification, anomaly detection, change detection, target detection, and airborne oil-spill mapping.
In briefUGO: A Unified Architecture for General Multi-Object Tracking by Segmentation
SummaryThis paper presents UGO, a unified tracker that follows every instance of a user-specified category using only one exemplar in the first video frame. It combines an exemplar-conditioned detection head with an instance-propagation head, and uses a training-free energy-minimization step to merge overlapping proposals into exclusive pixel-wise masks. A hierarchical memory operating at global and instance levels is also introduced. The authors report new state-of-the-art results on general multi-object tracking benchmarks and video object counting, and competitiveness with specialist tracking methods.
In briefOFBD: Object-Focused Background Debiasing for Long-Tailed Image Recognition
SummaryThis paper studies why tail classes degrade in visual recognition trained on long-tailed data, finding that standard training makes representations and optimization biased toward backgrounds, with tail classes affected more. The authors propose OFBD, which combines Foreground-guided CutMix and Background-guided Feature Rectification to reduce background bias. Experiments report improved overall accuracy and significant tail-class gains, and the method can be used as a plug-in for mainstream long-tailed approaches.
In briefDomain as residue: adapting self-supervised features, not generators
SummaryThis paper proposes the Representation Feature Adapter (RFA), a roughly 2.9M-parameter network that removes source-domain traits such as fog, rain, snow, or rendering style from DINO feature maps while keeping scene content. Only the adapter and its discriminators are trained, with the encoder and decoder kept frozen. It outperforms CycleGAN-Turbo on fog and night KID and leads on rain removal, though CycleGAN-Turbo preserves scene structure better in most conditions.
In briefSolving Without Stopping: On-Policy Distillation at Small Scale
SummaryThis paper studies what on-policy distillation, where a student learns from a stronger teacher's feedback on its own outputs, transfers to small models. The authors distill Qwen3-8B into Qwen3 4B, 1.7B and 0.6B students in thinking and non-thinking modes. They find distillation transfers problem-solving ability but, in thinking mode, not the ability to know when a problem is solved, and smaller students retain less of this stopping ability.
In briefVISTA: Appraising Events by Value to Resolve Conflicting Multimodal Emotions
SummaryThe paper proposes VISTA, a learned seven-field appraisal interface that uses a person's concerns, event relations, and expression conditions to arbitrate between conflicting emotional cues across modalities. On the CA-MER benchmark it reaches 64.5% accuracy on conflicting cases, improving on modality gating by 2.5 points. Shuffling the appraisal across scenes or removing its decision connection reduces this benefit.
In briefParallel tempering for diffusion-based combinatorial optimization
SummaryThis paper addresses inference with discrete diffusion models that solve combinatorial optimization problems on graphs. It proposes PT-Denoise, which lets several concurrent denoising trajectories interact through parallel tempering, with each trajectory assigned a temperature that can be swapped based on performance. The method requires no retraining or fine-tuning of the denoiser. Experiments on canonical graph-structured problems report consistent improvements in the best solution quality found, with only minimal extra computation.
In briefMubric: Using Mutation Testing to Generate Rubrics for Evaluating LLMs
SummaryThis paper introduces Mubric, a method that automatically generates scoring rubrics for evaluating LLM responses, inspired by mutation testing in software. It mines common defects from preferred and dispreferred response pairs, turns them into reusable mutation operators, injects them into reference responses, and refines the rubric based on which defects are insufficiently penalized. Tested on 703 tasks across four domains, Mubric achieves the highest overall evaluation accuracy, outperforming the strongest of six baselines by 7.48 percentage points.
In briefPowerMarketJax: A JAX Benchmark Suite for Multi-Agent Reinforcement Learning in Power Markets
SummaryThis paper introduces PowerMarketJax, a benchmark suite for multi-agent reinforcement learning in five power market types: day-ahead wholesale, real-time balancing, ancillary services, peer-to-peer double auctions, and local flexibility. Each environment implements its own clearing, pricing, and settlement rules within a common framework. The authors report that learned bidding behavior depends strongly on market design. Both market simulation and policy training run in JAX on GPUs, achieving up to a 33X speedup over CPU-based baselines.
In briefLabInstruct: A Benchmark for Generating Lab Procedure Instruction Videos
SummaryThis paper introduces LabInstruct, a benchmark for generating instructional videos of real laboratory procedures from an initial workspace image and a natural-language instruction. It covers 204 tasks across 5 scientific disciplines, with real reference executions and structured annotations of objects, actions, contacts, and state transitions. Testing 8 frontier image-to-video models, the authors find that visually plausible videos often remain procedurally incorrect.
In briefWhat Comes Next? Omni-StoryBench for Story-Grounded Multimodal Generation
SummaryThis paper introduces Omni-StoryBench, a benchmark testing whether models can coherently continue a story across illustration, narration, and speech. Each instance gives a current storybook page and structured next-page conditions, and models must generate the next image, narration, and character utterance. It contains 900 validated story transitions from openly licensed children's books, evaluated with modality-specific metrics and consistency-focused LLM-as-a-judge rubrics. Orchestration with strong VLM planning was most reliable, while native omnimodal models often struggled with completeness and controllability, and image generation with visual continuity was the clearest bottleneck.
In briefContextAdapt: Testing whether LLMs apply honesty, autonomy and confidentiality appropriately across professions
SummaryThis paper introduces ContextAdapt, an evaluation framework covering honesty, autonomy, and confidentiality across medicine, law, finance, and national security, built from professional and regulatory documents. Twelve LLMs were tested on recommended actions and justifications, reaching 95.6% mean appropriateness, though correct domain-specific justification rates ranged from 25.6% to 76.9%. Varying stakes caused some models to change responses even when the underlying professional obligation stayed the same.
In briefDo Agent Benchmarks Do What They Say? An Audit of Tool-Using Agent Environments
SummaryThe paper audits 34 state-changing tools across four agent benchmarks by checking each implementation against its advertised interface and tracing how scores are derived. The authors confirm seven tool defects and one evaluator property at pinned commits. They also report that some grading can reward outcomes that a defective tool should not have produced.
WhyTreats tool interfaces as executable contracts to audit agent benchmarks; confirms defects across 34 tools with mostly pinned, reproducible evidence, with substantial impact on evaluation trust.
In briefDetecting fake news that has been rewritten with different emotions, using LLM-generated explanations
SummaryThis paper studies fake news detection when articles are rewritten with different emotions while keeping the same factual claims. The authors build emotion-rewritten test sets and generate explanations from the original articles as stable background knowledge. They propose a Gated Cross Attention (GCA) framework that adaptively combines the rewritten news with these explanations. Experiments on PolitiFact, GossipCop, and LUN show notable gains under multiple emotional conditions on PolitiFact and LUN, and competitive performance on GossipCop.
In briefFLASH: A "generate once, synthesize many" framework for creating synthetic defects in industrial inspection
SummaryFLASH is a framework that creates synthetic defect samples for industrial anomaly detection when real defects are scarce. It uses a vision-language model and an image-generation model on normal images to produce a small set of defects, then stores reusable defect patches for blending onto new normal images without per-sample generation. On the MVTec AD 2 dataset, FLASH-generated anomalies reach 78.1% image-level F1 against an 83.6% real-anomaly upper bound, and synthesize anomalies more than 11.95x faster than per-sample generative methods.
Abhay Kumar Das, Rajesh Gangireddy, Ashwin Vaidya et al.
In briefHidden AI reasoning must leak traces, but may be unreadable to monitors
SummaryThe paper studies whether reasoning models can perform hidden computation in their chain-of-thought without monitors noticing. It finds that simple tasks can be done covertly, but beyond a model-size-dependent threshold, solving the task necessarily leaks a near-linear amount of information about the hidden input into the chain-of-thought. Under plausible cryptographic assumptions, even a one-layer Transformer can encrypt its reasoning so that no polynomial-time monitor can extract the hidden computation.
WhyProves hidden computation must leak near-linear information into chain-of-thought, and constructs unreadable encrypted reasoning under cryptographic assumptions; theory is novel, experiments are limit
Mohammadali Mohammadkhani, Madhava Krishna, Yash Sarrof et al.
In briefReMem: Rethinking Perception and Memory for Long-Context Recommendation Agents
SummaryThis paper proposes ReMem, a recommendation agent framework that perceives item pages through screenshots and OCR instead of parsing raw HTML. It uses a chunk-wise sequential memory update strategy to keep a fixed-size memory of informative interactions over long contexts, and a multi-memory GRPO variant to train the memory updates. Experiments on three datasets report an average 5.16% improvement over state-of-the-art baselines across searching, ranking, and judging tasks.
In briefHybrid joint-selective optimization: refining a few key parameters with reduced-space Levenberg-Marquardt
SummaryThis paper proposes a hybrid joint-selective optimization (HJSO) framework for large problems where only a small subset of trainable parameters matters most. It first runs first-order optimization over all parameters, then freezes the rest and refines the low-dimensional parameters of interest with a reduced-space Levenberg-Marquardt method. The method was tested on a matrix eigenvalue problem, an inverse Bratu problem with a physics-informed neural network, and a 100-dimensional nonlinear Black-Scholes problem solved with DeepBSDE. It reached the error thresholds faster than the joint first-order baseline and improved final accuracy in the reported configurations.
Muhammad Luthfi Shahab, Gabriella Alfa Indahsari, Imam Mukhlash et al.
In briefMetaCtrl: A Metacognitive Controller That Helps Language Models Reason Better and More Concisely
SummaryLarge reasoning models often spend extra computation that does not help, wasting effort on simple problems, while cutting reasoning short can hurt hard ones. MetaCtrl is a lightweight controller that watches the reasoning trace and decides whether to continue, simplify, skip redundant steps, or stop, without retraining the reasoner or setting token budgets. It is trained with reinforcement learning to favor correct answers and shorter correct trajectories. Across seven benchmarks in mathematics, science, and code, it improves accuracy and reduces generation length, and it transfers to an unseen reasoner without further training.
WhyFrames reasoning regulation as sequential metacognitive control trained with RL, improving accuracy and cutting length across seven benchmarks and an unseen reasoner.
In briefTraining-free feature alignment using information geometry to improve vision model robustness under shift
SummaryThis paper proposes ZFGA, a closed-form method that aligns the Fisher geometry of test feature embeddings with a clean-data reference, improving robustness under covariate shift without changing model parameters. Evaluated on CIFAR-10-C and ImageNet-C with ResNet-50, DINO ViT-S/16, and CLIP ViT-B/32, ZFGA consistently improves over zero-shot inference. It is reported as the only compared method that does not substantially harm any of the three model families, though it is not the strongest on every model.
Behraj Khan, Tahir Qasim Syed, Syed Ahmad Chan Bukhari
In briefScaling Full Conformal Image Classifiers: A method to make full conformal prediction practical for large-scale image classification.
SummaryThis paper addresses the high computational cost of full conformal prediction for image classification. The proposed Targeted Full Conformal Prediction (T-FCP) uses zero-shot vision-language models to prune unlikely labels, then applies full conformal prediction only to the remaining candidates. The authors also introduce Stabilized Online LDA (SO-LDA), a VLM adaptation solver based on rank-one inverse-covariance updates. Across benchmarks including ImageNet, the method reports modest test-time overhead, efficient prediction sets, and more stable empirical coverage than split conformal alternatives.
In briefWhy Cross-Skeleton Motion Retargeting Is Non-Identifiable: Limits of Generative Motion Models
SummaryThe paper argues that generative models for cross-skeleton motion retargeting cannot be guaranteed to learn the source-conditioned mapping in sparse, heterogeneous motion data, since training evidence fits multiple explanations equally well. It introduces Source-Instance Fidelity (SIF), a diagnostic that checks whether output differences track differences between source clips, with the target skeleton and action fixed. On animal motion data, methods that pass standard action-level tests often perform at a source-blind baseline under SIF.
In briefRoboHarness: a simple robot control framework that beats VLA and world action models
SummaryThis paper presents RoboHarness, a robotic harness that gives LLM agents a visual-geometric control panel to understand and invoke embodied tasks. The authors argue that coding agents equipped with this harness can achieve generalist robot manipulation without training dedicated vision-language-action models. They report that Qwen3.8-Flash-Next outperforms state-of-the-art VLA models on BEHAVIOR Challenge 2025 tasks. The implementation is open source.
In briefSafeLLM4SE: Statistically Sound Evaluation and Reporting for LLM-Based Software Engineering Tools
SummaryThis paper presents SafeLLM4SE, a methodology and reporting standard for evaluating large language model-based software engineering systems with statistical rigor. It treats generated outputs as realizations of a stochastic process and separates quality, stability, and estimation uncertainty. The approach combines adaptive sampling, confidence intervals, and effect sizes, and is released as an open-source PyPI package. The authors illustrate it by comparing two LLMs on the HumanEval programming benchmark.
In briefVISTA-Bench: A Multilingual Benchmark for Translating Text in Images
SummaryThis paper introduces VISTA-Bench, a benchmark for testing how well multimodal models translate text within images across 22 languages and 10 domains. It uses image-specific rubrics that specify essential content, semantic relations, and acceptable translation variants, with annotations produced by models and checked by humans. The authors evaluate 16 mainstream models, including 12 multimodal and four text-input models, and analyze results across languages, domains, and evaluation dimensions.
In briefAssayRouter: Using Past Utility Priors to Route Among Frozen Molecular Predictors
SummaryLabs often have a new molecular assay with only 16 to 64 labels and predictors whose training data and parameters are unavailable. AssayRouter treats completed assays as pseudo-targets and scores each candidate by how much it reduces held-out discovery loss, then trains a shared regressor to predict this utility from candidate behavior. On a new assay it selects four sources and fits a convex combiner. Across 24 external regression assays it lowers strict four-call negative log-likelihood by 0.0409 relative to Support-CV@4.
In briefSAM Meets VLM: Parameter-Decoupled Training for Medical Reasoning and Segmentation
SummaryThis paper addresses medical multimodal models that must both answer clinical questions and localize the visual evidence behind their answers. It proposes a framework that treats the <SEG> token's hidden state as a prompt for the mask decoder and keeps it separable from general language states, using a three-phase training process. The method is reported to achieve strong language-conditioned segmentation on medical referring segmentation, grounding, and QA benchmarks while retaining competitive reasoning ability.
In briefTransolver-σ: Joint spectral and physical subspace modeling for neural PDE solvers
SummaryThis paper presents Transolver-σ, a neural solver for partial differential equations that models physical states and spectral structure in joint latent subspaces. It is evaluated on five PDE benchmarks, coupled multiphysics systems, and real fluid and combustion measurements from RealPDEBench. The authors report a 33.4% average relative error reduction over the strongest baseline for each metric, along with improved autoregressive rollout.
In briefGeometry-aided channel deduction using partial estimates and an uncalibrated digital twin
SummaryThis paper addresses the high pilot overhead needed to acquire channel state information in wireless MIMO-OFDM systems. It proposes a geometry-aided channel deduction (GCD) approach that uses an uncalibrated digital twin with only approximate environmental geometry and positions. Geometric features are converted into a CSI-like prompt via random prompt augmentation and fused with a coarse pilot-based channel estimate in a deduction network. Experiments report high acquisition quality, low pilot overhead, robustness, and good generalization to new scenarios.
In briefResidual Method Recovering Input Details Lost by Non-Injective Neural Mappings
SummaryNon-injective neural mappings send different inputs to the same representation, which can erase input differences that downstream tasks still need. The paper proposes Task-Relevant Null-Space Residuals (NSR), a framework that extracts null-space components from pre-mapping representations and integrates them under downstream supervision while keeping the original aggregation or merging rules. NSR is evaluated on token merging and graph aggregation, with reported gains in semantic segmentation and graph learning tasks.
In briefNon-Commutative Lie-Algebraic State-Space Models: Group-Structured Sequence Modeling and Gradient Saddle Escape
SummaryThe paper proposes Lie-SSMs, state-space models whose recurrent updates are parameterized on Lie algebra tangent spaces via the matrix exponential, aiming to track non-solvable groups and 3D rigid-body manifolds that diagonal SSMs cannot. It gives a single-layer finite-division criterion for SO(3) and SU(2) heads, and depth-2 semidirect and Schreier cascades that resolve several finite groups. It also analyzes a commutator saddle and reports experiments on group-word tasks and three 3D IMU datasets.
In briefDesigning and testing AI agents that help before being asked
SummaryThis paper proposes a framework for designing and evaluating proactive LLM agents, which act on anticipated user needs without being prompted. It is built on three principles: task capability, temporal allocation of compute, and trust. The authors also introduce Proactivity-Gym, a simulation testbed, and report tests on 23 model-harness configurations and a 30-participant human study showing that trust drops after misaligned interventions even when results are correct.
In briefHow diffusion language models should edit code: a study comparing four editing interfaces and how they locate and generate changes.
SummaryThe paper compares four editing interfaces for masked diffusion language models: whole-file rewriting, search-and-replace, locate-then-infill, and token-level editing, tested on CanItEdit. Models generate coordinated changes well when correct edit locations are supplied, but this ability largely drops when they must predict those locations. The authors argue a good interface must cover every required change while limiting regeneration of unchanged code.
In briefUniAfford: Token-Routed Multitask Learning for Flexible 2D and 3D Affordance Detection
SummaryThis paper presents UniAfford, a unified framework that locates actionable object regions, or affordances, in both 2D images and 3D point clouds. It uses a multimodal large language model as a shared semantic hub and a modality-aware token router to produce affordance queries for a pixel decoder and a point-cloud decoder. The authors report strong zero-shot generalization on 2D and 3D benchmarks without target-specific fine-tuning, along with state-of-the-art branch-wise results under modality-isolated protocols.
In briefBeyond Attention Imbalance: Spectral Surgery to Reduce LVLM Hallucinations
SummaryThis paper addresses hallucinations in Large Vision-Language Models (LVLMs), which the authors say limit reliable deployment. Prior methods mostly target cross-modal attention imbalance and often rely on Contrastive Decoding, which doubles inference time. The authors propose FLASH, a training-free, contrastive-decoding-free framework that uses a Spectral Vortex Score to find vision heads and applies adaptive spectral modulation during decoding. According to the paper, FLASH achieves a better balance of performance and efficiency than state-of-the-art methods.
In briefApproximating Attention Efficiently Is Provably Hard for All Inputs
SummaryThis paper studies whether softmax attention, whose cost grows quadratically with sequence length, can be approximated by fast algorithms with guarantees holding for all inputs. Under standard complexity-theoretic assumptions, the authors prove that no truly subquadratic algorithm can achieve any nontrivial additive or relative uniform approximation. The impossibility extends to settings with polynomial preprocessing of the KV cache and to identifying sparsely attended keys.
WhyProves under standard assumptions no subquadratic algorithm can uniformly approximate attention nontrivially, settling computational limits with strong theoretical hardness results.
Lukas Haverbeck, Carmen Amo Alonso, Andres Felipe Posada-Moreno et al.
In briefPoint-cloud scene reconstruction that avoids object collisions and matches observations
SummaryThis paper presents COOL, a framework that reconstructs objects in a scene from point clouds. It conditions generation on instance and background point clouds and adds an explicit collision loss, using joint optimization and resampling to reduce collisions at inference. Experiments on 3D-Front and Scan2CAD report strong scene-level fidelity, observation alignment, and fewer collisions. The authors also test robustness to mask errors and applicability to real-world scene replicas.
In briefSelecting Pretraining Data for Time-Series Models Using Forecasting Loss
SummaryThis paper addresses the problem that time-series foundation models are pretrained on huge collections of windows without checking whether each window is useful for learning. The authors propose a static selection method that scores each window with a reference forecaster and keeps an intermediate interval within every source dataset. They show that normalized squared loss controls per-sample gradient norm under a local Jacobian condition, and they apply dataset-stratified selection to preserve diversity. Across several architectures, the method beats random selection and, with fewer windows, improves relative MASE and CRPS over full-data pretraining.
In briefV-JEPA Policy: Building robot world-action models on predictive visual features
SummaryThis paper introduces V-JEPA Policy, a framework that builds a world-action model for robots on top of a frozen V-JEPA 2.1 visual encoder, instead of a full pretrained video generator. It jointly learns a future-latent predictor and an action generator from scratch in one stage. The model is competitive with representative baselines on LIBERO, LIBERO-Plus, and RoboCasa-GR1, and performs particularly well under distribution shifts. Pretraining the predictor on action-free DROID videos further improves downstream control and out-of-distribution generalization.
In briefCodebook-guided cross-modal distillation for features with different structures
SummaryThis paper proposes a cross-modal knowledge distillation framework that uses a vector-quantized codebook to transfer knowledge between structurally different feature spaces. Teacher features are abstracted into vector-form codes that serve as concept-level anchors for the student model. This allows transfer between features such as 2D visual grids and 1D audio sequences without direct unit-level alignment. The authors report effectiveness on classification and semantic segmentation across several cross-modal distillation scenarios.
In briefTrident: Unifying Guarded Dispatch and Host Execution for PyTorch Triton Kernels
SummaryThis paper addresses how host-side overhead can dominate the end-to-end latency of user-written Triton kernels in PyTorch when device execution is short. The authors present Trident, a compiler backend whose Specialization Cache Module compiles guard checks, argument preparation, and host execution for multiple specializations into one executable module. Built on Torch-MLIR and tested on two LLMs, Trident achieves up to a 1.47x model-level end-to-end speedup over eager execution and up to 1.68x over torch.compile.
In briefDifferentiable Bicausal Optimal Transport: A Framework for Fitting Parametric Markov Chains
SummaryThis paper frames model learning in sequential decision-making as minimizing a distance to a target process, using the bisimulation metric, equivalently Bicausal Optimal Transport (BOT). The authors show BOT can be written as a linear program and is therefore differentiable with respect to model dynamics, deriving an exact closed-form gradient via the envelope theorem. They propose the Differentiable Bicausal Optimal Transport (D-BOT) algorithm and evaluate it on state-space compression, parametric model learning, and imitation learning from observations.
Sergio Calo, Amy Zhang, Javier Segovia-Aguas et al.
In briefAgents asking unrequested questions: horizontal and vertical proactivity
SummaryThis paper studies what information tool-using agents should proactively pursue beyond what users explicitly request, distinguishing horizontal proactivity (unstated information the context already identifies) from vertical proactivity (needs revealed only by earlier evidence). The authors score both forms using a need graph derived from benchmark decompositions, without a model judge. They propose Q&D, which trains a questioner to prefer questions whose continuations retrieve more required evidence, and report gains on three multi-hop QA benchmarks and in a simulated customer-service setting with fewer questions asked.
WhyIntroduces a new horizontal/vertical proactivity distinction, with need-graph scoring and experiments on multi-hop QA and customer service showing gains over a 15x larger model.
In briefFraming the Narrative: How LLMs Mirror a User's Political Ideology
SummaryThis paper studies whether large language models change the political stance they express on contested issues depending on signals from the user, such as terminology, premises, and stated ideology. The authors built the Poli-SHIFT dataset and tested seven open-weight models on ten contentious topics in the US, UK, and Australia. They found that changing terminology alone reversed which side a model supported in 16.9% of matched comparisons, and stated ideology systematically shifted responses toward the user's position.
Olivia Macmillan-Scott, Michael Jacobs, Nils Metternich et al.
In briefDatalogBench: Testing AI Models on Writing Datalog Programs from Plain Language
SummaryDatalog is a language used for reasoning tasks such as program analysis, but writing its programs is difficult. This paper introduces DatalogBench, a benchmark of 136 text-to-Datalog synthesis tasks curated from existing Datalog-based artifacts, graded by execution on held-out inputs. Across six large language models and four prompting setups, exact match peaks at 68.4%, while two coding agents reach up to 83.8%. Remaining errors are mostly semantic and concentrated in recursive tasks.
In briefSeeing Is Not Addressing: Checking Whether Language Can Reach Visual Distinctions
SummaryThe paper studies a gap in vision-language models, where a visual distinction can be told apart in frozen image features but is hard to reach through text. The authors use FactorAtlas, a fully crossed set of 23,040 images varying in shape, hue, pattern and nuisance factors, to compare visual discriminability with text-based addressability. They show that matched visual contrasts can probe and reduce this access gap across factors and models. The findings indicate that visual distinguishability and linguistic addressability do not necessarily coincide.
In briefAESOP: Joint human and camera generation with controllable camera movement strength
SummaryAESOP is a framework that generates camera trajectories for a given human motion and can also generate human motion and camera together. It uses an independent human pathway and a shared camera generator conditioned on human motion, which keeps the human output unchanged during camera generation. The authors built trajectory pairs that differ only in camera translation magnitude to learn an explicit translation-intensity control. Experiments on the PulpMotion dataset report strong camera distribution and framing quality for both tasks and effective control over translation intensity.
In briefInteracting Particle Guidance: a new sampling method that steers pretrained generative models toward new goals without reweighting.
SummaryThe paper addresses inference-time steering of pretrained diffusion and flow models, framed as sampling from a reward-tilted generative prior. Standard sequential Monte Carlo (SMC) corrects guidance bias with importance weights but suffers weight degeneracy and particle collapse in practice. The authors propose interacting particle guidance (IPG), which replaces reweighting with transport by having unweighted particles interact through an extra drift, with a closed-form drift in a reproducing kernel Hilbert space. The method is tested on Gaussian mixtures with known posteriors and on high-dimensional image inpainting and protein structure inference.
Adhithyan Kalaivanan, Zheng Zhao, Jens Sj\"olund et al.
In briefFollowing Entities to Navigate Documents: CorpusMap for Agentic Search
SummaryThe paper addresses multi-document question answering, where agents searching a flat collection must repeatedly rediscover how documents relate, often missing complementary evidence and spending extra tokens. The authors propose CorpusMap, an offline-built navigation layer that resolves recurring entities and creates an Entity Page for each, linking it to every document that mentions it. Across 7 models and 3 benchmark datasets, CorpusMap improves evidence discovery and answer quality over raw-corpus agentic search with fewer tokens on average, and outperforms 4 alternative navigation layers.
Soyeong Jeong, Sujay Kumar Jauhar, Sung Ju Hwang et al.
In briefResComEmb: Efficient multimodal embeddings via residual homogeneity compression
SummaryThis paper proposes ResComEmb, a trainable framework that creates multi-vector embeddings for universal multimodal retrieval while reducing storage and interaction costs. Its core components are a Residual Homogeneity Compression module that removes redundancy under visual token budgets and a length-adaptive bidirectional late-interaction matching mechanism for scoring. Experiments on MMEB, ViDoRe V1, and ViDoRe V2 report higher-quality embeddings than VLM2Vec-V2 and better visual document retrieval than ColQwen2.5 using only 37.5% of its visual token budget.
In briefCredWise: A Controlled, Explainable Agentic Framework for Credit-Risk Assessment
SummaryThis paper presents CredWise, a decision-support framework that combines credit-risk prediction, probability calibration, explainable AI, policy retrieval, SQL analytics, and controlled agent workflows. An XGBoost model trained on 1,345,310 Lending Club loans achieved a 2018 test-set ROC-AUC of 0.7109, and calibration reduced the Brier score from 0.2157 to 0.1273. The authors describe it as an academic research prototype in which final decisions remain with a human reviewer.
In briefOptiCom: A unified framework for choosing search methods in LLM-driven optimization
SummaryThe paper argues that the best search mechanism for LLM-driven iterative optimization depends on the current state, such as candidate quality and remaining budget, so coordinating them dynamically is an open challenge. It proposes OptiCom, a framework with a shared configuration space, a fast controller that composes mechanisms via Action Packages, and a slower Strategy Adapter that refines preferences from trajectory feedback. Across 32 benchmark groups, OptiCom reports an average Max-score rank of 1.72 among 14 configurations and the top score in 23 groups.
In briefInformation bottleneck-guided adaptive hypergraph Transformer for diagnosing brain disease from brain networks
SummaryThis paper proposes IBAHGT, an adaptive hypergraph Transformer guided by the information bottleneck principle, for analyzing brain networks. It jointly learns high-order correlations and short- and long-range dependencies to improve brain disease diagnosis. The authors report that the method outperforms current state-of-the-art approaches and can identify biomarkers for clinical use.
In briefCRJudgeBench: Can AI Spot Code Review Comments That Look Right but Are Wrong?
SummaryThe paper introduces CRJudgeBench, a benchmark of 1199 instances built from real pull requests and expert-verified perturbations, covering both trustworthy and plausible but untrustworthy review comments. It also presents Sentinel, a repository-grounded agentic judge that actively gathers code evidence to verify comments before judging them. On the 359-instance test set, Sentinel reaches 76.60% accuracy, outperforming GLM-5.3 by 6.13 points and its base model.
In briefAnchored weight search that makes fine-tuned vision models more robust to new data types
SummaryThe paper proposes Barycentric Anchored Search, a post-hoc method that improves fine-tuned vision models under distribution shift without extra gradient training or higher inference cost. It builds candidate weights from two fine-tuned checkpoints and runs a small derivative-free search along directions toward the pre-trained model, selecting candidates using a small out-of-distribution validation subset. Experiments with CLIP ViT-B/32 report substantial gains on held-out OOD benchmarks such as ImageNet-A, ImageNet-R, and ImageNet-Sketch, while ImageNet accuracy stays competitive.
