In briefDifferentiable Task-Driven Coded Aperture Camera Design Using 3D Gaussian Scenes
SummaryThis paper presents a differentiable framework for jointly designing camera systems and downstream perception tasks. The authors use it to design monocular and stereo coded-aperture depth-from-defocus cameras by optimising the aperture mask, camera parameters, and depth network end-to-end. Scenes are represented with 3D Gaussian Splatting reconstructed from real captures, and the results are reported to improve over existing coded-aperture and non-differentiable camera design baselines.
In briefDancing Diffusion: learning continuous programs to generate and edit images
SummaryThis paper introduces Dancing Diffusion, which gives each span of denoising timesteps its own continuous condition block, turning conditioning into a latent program that a fixed pretrained generator executes. Executed prefixes can be reused while later operations are rewritten, preserving earlier content. On SANA, reusing a prefix kept more of the scene than regenerating from noise, and the learned program improved generation quality over native SANA. The authors note that later branches weaken the edit, so semantic control over individual operations remains limited.
In briefBeyond the Evidence Boundary: Behavioral Mirage in Medical VLMs
SummaryThis paper studies how medical vision-language models behave when image evidence is progressively degraded, a condition not captured by standard clean-input evaluation. The authors trace baseline-correct cases through increasing degradation to find an empirical evidence boundary, and define a mirage as continued commitment to an answer at that boundary regardless of correctness. Across three models and three medical datasets, post-boundary behavior varied widely, and residual-stream probes decoded commit-versus-abstain behavior well, while output entropy and abstention did not consistently reflect the persistent commitment.
In briefMatrix-Game 4.0: Using video models to build real-time playable game worlds
SummaryThis paper presents Matrix-Game 4.0, which separates game execution from visual generation. A lightweight engine keeps the world state and enforces rules, a coding agent turns player ideas into game code and scene captions, and an autoregressive video renderer turns low-poly proxies into detailed visuals. The system runs 480p gameplay at 60 FPS on a single RTX5090 with 112ms key-to-display latency, and demonstrates customizable shooter and driving games. Code, model weights and low-poly assets are released.
In briefLearning tactile-aware hand representations from human hand-object interaction data
SummaryThis paper proposes HuTact, a framework that pretrains tactile-aware hand representations from large-scale human hand-object interaction data without requiring real tactile signals. The interaction structure is transferred to robotic tactile observations and can be adapted with unlabeled robot data. Human HOI pretraining improves data efficiency and provides a strong initialization across tactile perception and manipulation tasks.
In briefTelecom Language Model Built with RSOD: Data, Specialist Training, Distillation, and Evaluation
SummaryThe paper presents a framework that generates synthetic networking training data, trains six frozen specialist LoRA teachers, and distills their behavior into one deployable student adapter via Routed Specialist On-Policy Distillation (RSOD). On 2,850 held-out questions, the student scores 21.7112/25 micro, within 0.0529 points of the combined six-teacher system (21.7641/25) and well above the raw Meta-Llama-3.1-8B-Instruct model (12.8661/25). The authors also report a numerical validation of the teacher logit reconstruction.
Anil Srirangapatna Nagesh, Nitin Gupta, Anuj Taneja et al.
In briefWhen Momentum Helps Polyak–Ruppert Averaging: Online vs. Offline Analysis for Strongly Convex Quadratics
SummaryThe paper studies how quickly finite runs of SGD and momentum methods reach the same asymptotic covariance under Polyak–Ruppert averaging, and how far they can be parallelized. For strongly convex quadratics with additive noise, it links online tail-averaged estimates to offline empirical risk minimization and splits the gap into center, covariance, and shape errors. The authors give conditions showing that momentum methods, especially Nesterov's, need fewer rounds and allow larger critical batch sizes than SGD.
WhyProves an online-offline bridge and improved critical batch sizes for momentum on strongly convex quadratics, a narrow but rigorous theoretical result.
Alexandra Suvorikova, Darina Dvinskikh, Yuriy Dorn et al.
In briefUsing Johnson-Lindenstrauss Transforms for Attack-Resistant Distributed Optimization
SummaryThis paper studies how the Johnson-Lindenstrauss (JL) transform, a classical dimensionality-reduction tool, can support Byzantine-robust optimization. JL transforms approximately preserve inner products and pairwise distances with high probability. The authors use this property to weight worker messages by their alignment with a trusted reference gradient. Experimental results are presented to support the theoretical findings.
Aleksandr Shestakov, Vladimir Aletov, Aleksandr Beznosikov
In briefRecast the Past: Stable Long Video Generation with Full Context Memory
SummaryThis paper proposes ReCam, a method for long video generation that groups blocks of historical latents into memory tokens projected into a Diffusion Transformer's hidden space. This reduces token count while keeping block-level access to the full history. A dual-pointer denoising scheme with separate noise schedules for recent context and target frames is introduced to mitigate temporal drift. The paper reports an overall VBench-Long score of 80.1 and leading results on four consistency metrics for off-screen subject reappearance.
In briefNoise sensitivity of deep neural networks with biases: annealed correlation and quenched sensitivity
SummaryThis paper studies how fully connected deep neural networks with added independent Gaussian biases respond to input noise. Earlier work showed noise sensitivity for networks without biases and sign activations. The authors show that quenched noise sensitivity persists over a broad range of depths under a positive-correlation condition and the weaker assumption that the perturbed input is asymptotically unlikely to equal the original. The annealed covariance stays bounded away from zero and comes entirely from fluctuations in network-dependent output means.
In briefBeyond single frames: redundancy, not object motion, leaks data in video unlearning
SummaryThis paper studies class unlearning for video segmentation, which aims to stop a removed category from being reconstructed from a deployed model's outputs. The authors show that per-frame evaluation understates leakage, since averaging outputs over a two-second clip recovers the forgotten class. They attribute the leak to cross-frame redundancy rather than object trajectory and propose PACE, a method that suppresses this redundancy, lowering attacker AP on VSPW and Virtual KITTI 2.
In briefAdaptive stopping for streamed reasoning in language model agents
SummaryThis paper introduces StreamStop, a method that decides when a language model agent should stop reasoning and act. It learns model-specific stopping policies offline from prefix-level action agreement, without needing access to the model's internal signals. In closed-loop tests on WebShop and ScienceWorld, it uses fewer reasoning tokens than full thinking while achieving comparable observed success.
In briefMedia Intelligence Platform: Detecting Fake and Edited Media Across Formats
SummaryThis paper presents the Media Intelligence Platform, a web-based system that combines several detectors for manipulated media. It covers fake videos, face morphing, AI-generated images, audio deepfakes, local image edits, and text analysis. Its modular design allows new models to be added for emerging manipulation techniques. Experiments on benchmark datasets report reliable detection across media types, and a public demonstrator is released with the submission.
David Fischinger, Marcel Hasenbalg, Silvia Poletti et al.
In briefNeural identification of nonlinear systems from closed-loop data, with stability and generalization guarantees
SummaryThis paper studies identifying nonlinear dynamics with neural networks from closed-loop data, where feedback correlates inputs with noise. The authors derive a parameterization of all nonlinear systems stabilized by a given controller and propose Internal Controller Identification (ICI), which ensures closed-loop stability by construction. Experiments report that ICI handles unstable plants, outperforms direct methods under weak excitation, and generalizes zero-shot to unseen environments.
Laura Meroi, Mahrokh G. Boroujeni, Leonardo Massai et al.
In briefWhen Do Humans and AI Truly Complement Each Other? Formal Conditions for Synergy
SummaryThis paper presents a formal framework that models human-AI interaction through the joint correctness patterns of human and algorithmic decisions. It derives necessary and sufficient conditions for synergy, which depend on whether either agent can identify and correct the other's errors. Analysis of seven clinical case studies finds a "dominance trap" where strong algorithmic advantage can suppress human correction, and shows synergy can persist despite automation bias when human and AI errors are functionally de-correlated.
Andrea Campagner, Federico Cabitza, Vincenzo Crupi
In briefFaster, more accurate multi-object segmentation for SAM3 via competitive representations
SummaryThis paper addresses SAM3's difficulty segmenting several overlapping concepts at once, which causes misclassifications, redundant masks, and high latency. It proposes DCR-SAM, which decouples cross-modal interactions, uses a dense competition head to enforce inter-class competition, and adds a learnable background token. The authors report 2.3% to 10.7% performance gains over state-of-the-art methods and a 33× inference speedup over native SAM3.
In briefRobust Unsupervised Hyperspectral Segmentation Using Mumford-Shah Models Based on Proper Scoring Rules
SummaryThis paper presents an unsupervised, feature-agnostic framework for segmenting hyperspectral images based on the piecewise-constant Mumford-Shah functional. It builds the functional's indicator from strictly proper scoring rules and parametric distributions, using heavy-tailed elliptical distributions and beta-divergence to handle outliers. A robustness step estimates distribution parameters, and an edge bias with boundary prior information is added to the functional. The authors report the highest overall accuracy among fully unsupervised methods on three of four benchmarks evaluated on full scenes, and an improvement over their previous model.
In brief"Lost in Recovery": Studying How Computer-Use Agents Recover from Errors via Goal-Oriented Action Replay
SummaryThe paper notes that computer-use agents built on large vision-language models are fragile in long workflows, where errors accumulate and cause unintended side effects, so recovery from failure is critical. The authors introduce ReTrace, an evaluation framework that uses controlled error injection and goal-oriented action replay to systematically study model-based recovery behavior. Their findings indicate that ideal recovery can unlock models' latent task completion abilities, while current agents fail to exploit their recovery capabilities.
WhyFormalizes transition-consistent recovery and introduces ReTrace, the first systematic study of model-based recovery for computer-use agents; experimental scope is limited.
In briefTraining a Mobile-Side Head Network for Split Computing Without Labeled Data
SummaryThis paper addresses split computing, where a neural network is divided into head and tail parts run on a mobile device and a server. It proposes a method that trains a lighter replacement head network using federated learning and knowledge distillation on small amounts of unlabeled, non-IID mobile data, without exposing that data. Personalization is added to improve accuracy. The authors report accuracy comparable to baselines, a 95% reduction in mobile communication requirements, and only a 3 to 5 percentage point drop from the original model.
In briefEdit in Sync: Unified end-to-end audio-visual editing for human videos
SummaryThe paper presents AVE, a unified framework that edits both video and audio of people together, including swapping, inserting, and deleting individuals, without using spatial masks. To address data scarcity, the authors build AVE-Dataset, which provides 100K pairs with appearance and voice references, and combine Synthetic-to-Real reconstruction with audio warm-up. According to the paper, AVE outperforms modality-separate pipelines in editing realism, temporal consistency, and audio-visual alignment.
In briefTransferring hand actions from first-person videos to new scenes without extra geometry estimators
SummaryThe paper presents EgoACT, a test-time framework that transfers semantic egocentric actions into new visual contexts by operating directly on the denoising process of video diffusion models, without extra geometric estimators. It uses Velocity-guided Structure Anchoring in early denoising to stabilize hand-object structure, and Sparse Correspondence Calibration in mid-to-late denoising to refine local correspondences. The authors also introduce EgoActionBench, a benchmark for action preservation, visual quality, and hand-object plausibility. Experiments reportedly show more coherent and physically plausible interaction videos than strong baselines.
In briefPRAISE: Attribute-based retrieval for person re-identification using query-conditioned subspaces
SummaryThis paper studies person re-identification with vision-language models, where attributes like clothing color and carried objects can help match images or text queries to the right person. Because attribute-prompt similarities are unreliable as direct scores, the authors propose PRAISE, a test-time adaptation framework that builds a query-dependent low-rank Attribute-Conditioned Subspace from a bank of attribute prompts. This subspace reshapes query-gallery similarities and guides lightweight inference-time updates without identity or attribute labels. The authors report that PRAISE consistently improves VLM retrieval backbones and outperforms generic and ReID-specific test-time adaptation baselines across several benchmarks.
Rajarshi Bhattacharya, Aryan Shukla, Shakeeb Murtaza et al.
In briefPreference transfer via choosing near-duplicate images in vision-language models
SummaryThe paper studies subliminal learning, where a model passes its preferences to another through data that do not state them. A curator model picks one of eight near-duplicate image variants per photograph by caption likelihood, without target-specific prompts. Students trained on owl-preferring curators' selections show stronger owl preference in an image-free likelihood readout than those trained on neutral selections. The effect appears across several vision-language models and also carries into a second student generation.
In briefAPRIL: Ambiguity-Aware Expert Routing for Continual Respiratory Sound Classification
SummaryThis paper proposes APRIL, a framework that lets respiratory sound classifiers keep learning as recording devices and collection periods change, without replaying stored historical data. It builds on a frozen BEATs backbone and adds separate low-rank experts for each new domain, selected at inference by a multi-level prototype router. On the ICBHI and SPRSound datasets under forward and reversed domain orders, the authors report the highest final average score among compared methods and average forgetting below 1.0%.
In briefHypTI: A Hyperbolic-Geometry Framework Fusing Tabular and MRI Data for Dementia Detection
SummaryThis paper proposes HypTI, a framework for detecting dementia from clinical tabular assessments and MRI brain scans. It models tabular data, MRI features, and their cross-modal interactions in hyperbolic space to capture hierarchical and multi-scale structure. Experiments on the OASIS-3 dataset show that it performs strongly compared with representative approaches.
In briefFrom DQI to C-DQI: Automatically Generating and Evaluating Codebooks for Measuring Discourse Quality
SummaryThe paper addresses the rigidity and bias of expert-made codebooks used to assess parliamentary discourse quality, proposing an LLM-based protocol for generating and evaluating them. Starting from the Discourse Quality Index (DQI), it compares top-down, bottom-up, and hybrid approaches for building context-adapted schemas. Experiments on 100 Israeli parliament transcripts show that data-driven, inductive schemas achieve higher inter-model agreement while remaining valid.
Esther Shizgal, Ron Keinan, Daniel Shtauber et al.
In briefImproving cross-size generalization of reinforcement-learned Pointer Networks for capacitated vehicle routing
SummaryThe paper studies why Pointer Networks trained with reinforcement learning lose solution quality on capacitated vehicle routing problems (CVRP) when the number of clients differs from the training size. It identifies three causes: shifts in node density, attention dispersion, and dilution of mean-pooled graph context. The authors propose a learned, permutation-invariant graph-context aggregation method and an approach selector that picks the best model for each training-target size pair. The selector reduces cost relative to the reference model across all nine seen pairs and eight of nine pairs with unseen sizes, with improvements of up to 3.2%.
Sofiene Lassoued, Devara Tertia, Stefan Lier et al.
In briefWhy LLMs Accept Generics: Explaining Acceptability with Psycholinguistic and Model-Derived Factors
SummaryThis paper applies psycholinguistic findings on human acceptance of generic statements to large language models. It tests twelve models on a psycholinguistic dataset and on naturally occurring generics. The models weigh how common a property is against how well it distinguishes the kind and favor characteristic properties, but ignore striking ones, and factual truth matters most to their ratings.
Mervat Abassy, Zhaoxin Feng, Tatsuki Kuribayashi et al.
In briefQuantum-JEPA: quantum Born priors for latent-action JEPA world models
SummaryThe paper introduces Quantum-JEPA, a framework that learns distributions over future representations using a conditional quantum Born prior, where circuit measurements produce binary latent actions mapped to alternative futures. On constructed multimodal tasks and the D3IL benchmark of human demonstrations, the trained models generate multiple obstacle-crossing routes and improve prediction validity and distributional fidelity over the evaluated variational baselines. The authors also verify finite-shot parameter-shift training and show that inference with a trained prior runs on a real QPU.
In briefWorld Models with Configurable Perception: A Method That Filters Observations by Task Description for More Stable Planning in Noisy Settings
SummaryThis paper introduces ConfigWM, a world model that uses a language description of the task as a configuration signal. The signal steers how visual embeddings are mapped into a task-specific state space, keeping task-relevant information while suppressing irrelevant variations. Action-conditioned dynamics are then learned in this space for planning. In increasingly noisy environments, ConfigWM's planning success declines less than existing methods, outperforming the strongest competitor by 22.3% on average.
In briefACE-Compression: An ability-centric way to evaluate compressed language models
SummaryThis paper introduces ACE-Compression, a framework that evaluates which core abilities of compressed large language models survive compression, rather than relying only on task scores. It probes four dimensions: procedural reasoning and constraint execution, knowledge grounding and truthfulness, social cognition and affective understanding, and persona fidelity and stylistic consistency. Testing more than 35 compressed checkpoints from six open-weight models, the authors find that coarse, surface-level abilities tend to persist while fine-grained, long-tail, and context-dependent abilities become fragile. Quantization generally retains abilities better than pruning.</summary>
In briefICIBS: Adding Context to Improve Ice Bottom Boundary Segmentation in Radar Data
SummaryThis paper addresses the difficulty of automatically locating the ice bottom in radio-echo sounding data, especially in temperate glaciers where high water content makes radar reflections weak and ambiguous. The authors extend deep learning models with additional spatial and geo-spatial context. They report new state-of-the-art results on the IceAnatomy benchmark and better performance than the baseline on new Patagonia data, with code and the dataset publicly released.
Marcel Dreier, Dakota Pyles, Nora Gourmelon et al.
In briefMMEnterprise-Bench: A Benchmark Testing Enterprise Agents on Changing Multimodal Workplace Records
SummaryThis paper introduces MMEnterprise-Bench, a benchmark of 1,613 questions built on multimodal records from simulated team collaboration projects where information changes over time. The questions cover evidence reasoning, conflict resolution, and unanswerability. Evaluating 10 proprietary and open-source multimodal LLMs, the authors find that even the strongest models struggle with cross-source or temporal conflicts and with evidence spanning multiple images. They also report that larger retrieval budgets or bigger retrievers raise evidence recall without consistently improving answer accuracy.
In briefPolarity Blindness in LLM Safety: Testing How Negation and Antonyms Bypass Safeguards
SummaryThis paper studies a weakness in large language models called polarity blindness, where models fail to recognize harmful intent after a sentence's polarity is changed. The authors test grammatical negation and antonym replacement, using synonym replacement as a control, and also check whether adversarial wrappers make failures worse. They evaluated 10 state-of-the-art LLMs on a 260-prompt benchmark in English, Spanish, and Czech. Most models showed the effect, with the most vulnerable reaching a mean harm score of 4.12 on a 0–5 scale.
Diego Miguel Lozano, Erik Derner, Juan Antonio P\'erez-Ortiz et al.
In briefStaged adaptation of a Japanese finance multimodal AI model
SummaryThis paper tracks an 8.4B-parameter multimodal model through three training phases on Japanese financial documents, starting from an un-fine-tuned baseline, and evaluates it on five text-input benchmarks. The phases interact non-additively: math-only distillation overwrites previously acquired financial knowledge, while the domain phase mainly repairs rather than adds. Skipping a phase gives the best domain knowledge but loses numerical-reasoning gains, so per-phase contributions cannot be separated and the best checkpoint depends on the task.
In briefCollabSciReM: Aligning language models for scientific dialogue with science-aware reward decomposition
SummaryThis paper proposes CollabSciReM, a reward framework that splits multi-turn research conversations into conversational quality, scientific rigor, and structural efficiency. The authors also introduce SciQuality, a pairwise evaluation built on 300 held-out scientific papers and scored on six dimensions such as methodological soundness and factual correctness. Fine-tuned on 8.62K synthetic research conversations, CollabSciReM was preferred over a science-agnostic reward baseline across most tested settings.
In briefHow DPO hyperparameters behave across model size, datasets, and benchmarks
SummaryThis paper studies the hyperparameters of direct preference optimisation (DPO), such as the learning rate and the reference-policy regularisation parameter β. The authors run dense grid experiments across model sizes from one family, analysing how optimal values change with model size and update steps, and which static or dynamic benchmarks give the strongest selection signal. They find the optimal learning rate varies with model size, hyperparameters do not transfer across training datasets, and validation loss is a weak proxy for downstream performance. They also show that short candidate training runs can reduce tuning cost.
Hannan Mahadik, David Salinas, Jacek Golebiowski et al.
In briefTesting how humans and AI models creatively associate hard-to-name abstract images
SummaryThis paper compares humans and multimodal large language models on describing abstract visual stimuli, measuring associative creativity by originality and effectiveness. It uses two datasets with controlled difficulty: SilCC increases naming difficulty through more complex combinations of abstract shapes, and PixCC increases it through more random visual noise. Humans produce more diverse and novel associations than the models, and as difficulty rises, effectiveness drops in SilCC while humans shift toward describing low-level visual properties in PixCC.
In briefPhysics-Informed Deep Learning for Robust Fault Diagnosis in Industrial Fluid Systems
SummaryThis paper proposes PhyNet, a deep learning framework for diagnosing faults in industrial fluid transport systems whose sensors differ in sampling rate, observation duration, and data granularity. It treats physical variables as global operating constraints and response signals as local diagnostic evidence, linking them through cross-attention without strict temporal alignment. PhyNet achieves the best results on all evaluated metrics across three real-world cavitation datasets and one public two-phase-flow dataset.
In briefEndoWorld-JEPA: Motion-guided latent world modeling for endoscopic video pre-training
SummaryThis paper introduces EndoWorld-JEPA, a latent world-model pre-training framework for general-purpose endoscopic video representation learning. Instead of using recorded actions, calibrated poses, or robot kinematics, it predicts the next latent state from motion observed up to the current frame, separating dominant displacement from residual local motion. The authors report that a single pre-trained checkpoint transfers well across four downstream benchmarks.
In briefElicitation or Expansion? Understanding the Gains from Optimizing AI Agent Harnesses
SummaryThis paper studies how optimizing the "harness" of language model agents, which shapes context, tool use, and workflows, can improve performance without changing model weights. The authors use pass@k analysis from reinforcement learning to test whether optimization only makes already-solvable tasks more reliable or expands the set of solvable tasks. They compare harness optimization with model scaling and extra inference compute across task difficulty levels. They find that pass@1 advantages can reverse at larger attempt budgets, and that aggregate gains can hide weaknesses on hard tasks.
In briefDifferentiable graph-based solver with constraints for forward and inverse PDE problems
SummaryThis paper presents a fully differentiable, graph-based implicit solver for partial differential equations on unstructured grids, treating the large system of equations as a constraint. The method is evaluated on four tasks: linear and nonlinear forward modeling, inverse estimation of equation coefficients, inverse estimation of the right-hand side, and learnable graph coarsening to speed up forward modeling. The authors state that the solver's stability and differentiability reduce forward-modeling complexity and efficiently address related inverse problems. They highlight potential use in geophysics and other physics and engineering applications.
Nikolay Yavich, Alexander Ryabov, Viacheslav Naumov et al.
In briefLightweight heart rate estimation from PPG using gated time-frequency fusion
SummaryThis paper proposes TF-IG, a compact model that estimates heart rate from photoplethysmography (PPG) signals. It uses two branches, one for pulse waveform shape in the time domain and one for cardiac periodicity in the frequency domain, fused by an interaction-gated module. On the BIDMC dataset with subject-exclusive splits, it reports a mean absolute error of 1.227 BPM with only 84.6K parameters.
Anh Khoi Pham, Phat Lam, Thien Truong Nguyen Ly et al.
In briefSGLDet: Spectral-aware global-local dynamic modeling for detecting objects in multispectral aerial images
SummaryThis paper proposes SGLDet, a framework for detecting objects in multispectral aerial images. It targets two problems: severe foreground-background imbalance in large scenes, and large local variations in object scale, orientation, and structure that make precise localization hard. The method combines three modules: spectral-driven global dynamic modeling, multi-scale global context integration, and spectral-aware input-dependent calibration aggregation. The authors report superior performance with competitive computational cost on the MODA, HOD3K, and DroneVehicle datasets.
In briefS³E-Det: A Strict Spatial-Spectral Equivariant Detector for Multispectral Object Detection
SummaryThis paper proposes S³E-Det, a detector for multispectral object detection that couples ordered spectral modeling with predictions that stay consistent under geometric transformations. Its Spectral Order-aware Sampling Calibration module applies local order constraints to spectral features, its Common-Differential Equivariant Fusion module adapts spectral-context aggregation across pyramid levels, and its Geometric Alignment Head reconciles invariant and equivariant outputs. The authors report results on MODA, HOD3K, and DroneVehicle, where the method combines high detection accuracy, a compact parameter count, and reduced rotation-induced variation.
In briefGaze2HOI: Using eye gaze to predict future hand-object interactions from a first-person view
SummaryThis paper proposes Gaze2HOI, a framework that uses gaze cues to infer a user's intention and forecast future hand-object interactions in egocentric views, for use in AR, VR, and embodied systems. It weights gaze signals by fixation strength and recency, then combines a vision-language model with an anticipatory contact grounding module and a diffusion model to generate object trajectories and hand motion. The authors also introduce EgoAIM, a phase-aware benchmark with fine-grained intention annotations, and report state-of-the-art results in intention inference and motion forecasting.
In briefManifold-prior-guided deep unfolding for hyperspectral image reconstruction
SummaryThis paper proposes a deep unfolding framework that uses a two-stage manifold learning strategy to obtain a degradation-free structural prior. The prior is injected into both the data-consistency projection module, which adaptively weights spectral bands, and the denoiser module, which uses a manifold-guided Mamba to extract spatial-spectral features. Quantitative and visual comparisons on synthetic and real-world datasets show it outperforms existing state-of-the-art methods while keeping the unfolding framework interpretable. Code is publicly available.
In briefEnumeration-guided training of neural Hamilton-Jacobi safety certificates
SummaryThe paper proposes a verification-guided training method that uses probabilistic enumeration of safety constraints to find states where a learned certificate misclassifies unsafe behavior. These counterexamples are used to fine-tune a pretrained HJ value function, yielding larger certified regions and faster certification. A shielding mechanism built on the refined certificates is evaluated in Isaac Sim and real-world autonomous racing, reporting zero safety constraint violations.
Mihaela-Larisa Clement, Tosca Pierro, Ezio Bartocci et al.
In briefLLM Hallucinations in Cross-Language Code Translation and How to Reduce Them
SummaryThis paper systematically studies hallucinations, meaning generated code that conflicts with the source or contains incorrect facts, in LLM-based code translation. Using three benchmarks, six language pairs and seven LLMs, the authors build a taxonomy of three categories and manually annotate 13,919 cases. They also propose HaluCodeTran, a mitigation framework combining BM25 retrieval of translation examples with taxonomy-aware reflective refinement.
In briefVISH-ASR: Viterbi-guided hotword selection for speech-LLM context-aware speech recognition
SummaryThis paper addresses selecting a small set of hotwords from large lists for prompt-based contextual speech recognition. It proposes VISH, a training-free framework that checks each candidate's acoustic support directly on frame-level CTC posteriors, with M-VISH for scalable search and Bi-MVISH for added backward evidence. Across five benchmarks, the authors report up to 11.7% absolute Best-F1 gain over a learned-retrieval baseline and 75.2% lower search latency at 2,000 candidates.
In briefTokens to Files: Detecting and Locating AI-Written Code at Multiple Levels
SummaryThis paper proposes a framework for telling apart purely human-written code from files where AI generated or modified some part, and for localizing that AI-involved code at the line, block, and file levels. It uses a frozen code language model with two low-rank adapters trained on machine-written and human-written files, and turns per-token log-likelihood ratios into local evidence aggregated into file-level decisions. The method reaches 72.4 macro F1 at a 1% false-positive rate on an internal test set, and 62.5 line-level F1 on an external benchmark without adaptation. On human-written files from repositories created before 2022, the false-positive rate stays between 0.55% and 0.81%.
WhyToken-level likelihood ratios aggregated into line, block, and file predictions beat a published external benchmark and hold low false positives on large human code.
In briefOperating-condition-guided vision-language matching for cavitation diagnosis in industrial pipes
SummaryDiagnosing cavitation in industrial pipelines is hard because acoustic signals change with operating conditions, and similar patterns can mean different states. The paper proposes POC-VLM, which writes sample-specific operating variables into candidate cavitation texts, aligns them with time-frequency acoustic features, and matches global and local patterns. A K-nearest-neighbours operating-condition prior estimates each candidate's physical plausibility and is fused with the semantic evidence. On three real-world cavitation datasets, it outperforms representative data-driven, multimodal and domain-specific baselines.
In briefUvA-MT's WMT26 Entry for the General Machine Translation Shared Task
SummaryThis paper describes UvA-MT's system for the WMT 2026 General Machine Translation shared task, entered in the unconstrained track across all 23 translation directions. It uses large language models at test time, first generating translation candidates through sequential sampling with a refinement prompt, then selecting the best one with LLM pairwise judges in a round-robin vote. Offline evaluation on the WMT25 benchmark reportedly shows a clear improvement over point-wise best-of-N selection.
In briefLaLa-World: Scaling latent world models with natural language actions
SummaryThis paper introduces LaLa-World, a latent world model that uses language to describe both numerical actions and visual changes, so one predictor can train on simulation, robot trajectories, and image and video data. The predictor is built on a language model to predict future latent states. According to the paper, it achieves higher average planning performance than models trained separately per domain, and matches a video generation baseline with 6× less training cost and 13.9× faster planning. The pretrained model also initializes LaLa-Policy, which performs competitively with Cosmos Policy on LIBERO and RoboCasa.
In briefSAPD: learning LLM agents from past trajectories without new environment interaction
SummaryThis paper proposes Step-Advantage Policy Distillation (SAPD), a method for improving LLM agent policies using recorded offline trajectories. It splits multi-turn trajectories into individual decision contexts for on-policy distillation and adds relative step advantage signals to stabilize training, without new observations or extra interactions. Experiments on ALFWorld, WebShop, and Search-based QA report gains across teacher configurations, less catastrophic forgetting than SFT, and at least 4.7x faster per training step than online OPD.
In briefSciDiagBench: A Type-Specific Benchmark for Aligning Automated Evaluation of Scientific Diagrams with Human Judgment
SummaryThis paper introduces SciDiagBench, a benchmark that systematically categorizes scientific diagrams into 11 types, containing 165 diagrams across 10 broad domains. It also proposes a human preference-guided calibration method that generates type-specific evaluation criteria. The authors aim for automated evaluation that aligns more closely with human expert judgment.
In briefFrom Generic to Dedicated: A New Optimizer for Online Continual Learning
SummaryOnline continual learning requires models to learn from a non-stationary data stream in a single pass, which makes catastrophic forgetting likely. This paper proposes SPIN (Stability-Plasticity INterpolation), an optimizer that interpolates between a fast-moving plasticity sequence and an adaptive stability sequence to separate the two explicitly. The authors theoretically show that SPIN implicitly uses an inverse Hessian approximation that provides Tikhonov regularization to reduce forgetting. Experiments report consistent performance gains when SPIN replaces standard optimizers across multiple online continual learning methods.
In briefGARRET: A method for region retrieval in visual documents using generative feedback
SummaryThis paper presents GARRET, a framework for retrieving finer-grained regions, rather than whole pages, from visually rich documents. It trains using only (page, question, answer) triplets, using differentiable generative feedback to assign query-conditioned credit to region representations without query-to-region annotations. On the newly introduced ViDoRe-V3-Region benchmark, it reaches a macro nDCG@10 of 0.4126, 2.94 points above the strongest visual baseline. For page retrieval, it is comparable to ColQwen2.5-v0.2 while indexing 47.8% fewer vectors per page.
In briefCOMPASS: A Routing Method That Picks the Right AI Model Using Performance Profiles, Search, and Tools
SummaryThis paper introduces COMPASS, an agentic router in which a large language model follows a routing procedure, gathers performance and cost evidence through tools, and can inspect candidate answers before choosing a model. Evaluations on LLMRouterBench and DeepSWE report that it improves accuracy and lowers model-call cost compared with specialized and trained routing baselines. The authors state that these gains come without routing-specific training.
In briefSMART: Context-Aware Semantic Evaluation and Correction of Speech Recognition Transcripts
SummaryWord error rate treats all lexical edits equally, even though some change meaning, and overall semantic scores cannot show which errors matter. This paper proposes SMART, which groups discrepancies between reference and recognition output into coherent segments and labels their semantic consequences using a structured contextual rubric, yielding a Semantic Impact Error Rate. Against majority-vote human labels, SMART reaches 0.843 F1, compared with 0.745 for direct LLM judgment. The framework is also used to guide targeted correction of ASR outputs.
In briefPrivacy-preserving hosted inference via a learned input-embedding transform
SummaryThe paper studies whether a client can use a hosted language model without sending text or token ids. Text is converted to vectors locally and passed through a trained 55.6M-parameter network before reaching a frozen 27B model on the host. On 1,490 held-out MMLU questions, accuracy falls from 82.3% to 80.7%, and 30.1% of content tokens no longer decode to their original token under nearest-neighbour lookup.
In briefRAG-NIMBF: A Non-Invasive Framework for Benchmarking RAG Configurations
SummaryThis paper presents RAG-NIMBF, a modular benchmarking framework that attaches to existing retrieval-augmented generation (RAG) pipelines as a non-invasive evaluation layer. It analyzes configuration choices such as chunking, retrieval, reranking, prompting, and generation through controlled ablation and end-to-end experiment tracking. Validated on the SQuAD dataset, the authors report that MMR raises nDCG@5 by 4 to 11 percentage points over similarity search, and that more detailed prompt instructions improve answer faithfulness for some models.
In briefProVe: Proposer-Verifier Decoding, a faster text generation method that drafts token blocks
SummaryProVe reverses speculative decoding by fine-tuning a large autoregressive model into a proposer that predicts a block of future tokens in one forward pass, then verifies it with one more pass. The verifier can be a smaller model or the proposer itself, the latter giving ProVe-lossless, which matches greedy autoregressive output. Across six benchmarks, a 32B proposer with ProVe-lossless reaches 2.5x average autoregressive throughput, and a 3B verifier reaches 2.7x at a small accuracy cost.
In briefLatBridge: Keeping Latent-Communication Multi-Agent Systems Working When Models Are Swapped
SummaryMulti-agent systems use communication to finish tasks, and latent communication passes continuous representations instead of text. Existing methods assume fixed model combinations, so their interfaces can break when a model is replaced. LatBridge connects agent models through a shared Bridge and separate interfaces, reusing unchanged interfaces and training only the missing ones with models frozen. Across eight small-model configurations, basic-mode mean accuracy was 14.8% on MATH and 27.3% on ALFWorld next-action prediction.
In briefSliMOO: Interpretable multi-objective evolutionary search for pruning layers in LLMs
SummaryLarge language models are costly to serve, and depth pruning reduces their size by removing layers. The authors observe that layers contribute differently across tasks, while existing methods ignore this and tend to favor narrow domains. SliMOO builds a task-layer importance map through single-layer removal probing, then runs a Pareto-aware evolutionary search for pruning masks. On Llama-3.1 and Qwen3 it achieves better multi-task trade-offs than baselines, such as raising Llama-3.1-70B's math average at 25% sparsity from 24.33% to 34.40%.
WhyReformulates LLM depth pruning as multi-objective search with an interpretable task-layer prior, outperforming baselines across Llama-3.1 and Qwen3 models.
In briefCertified Adaptive List Decoding: Proving Per-Frame That Decoded Codewords Are Globally Optimal
SummaryThe paper addresses maximum-likelihood decoding of linear block codes, which is NP-hard, by proposing a neural decoder that searches over an absorbing-state discrete diffusion model using probability beam search. Two code-derived tests, a linear-programming dual bound and a distance-based test, certify for each frame whether the candidate is the global ML solution. Across 15 codes, the tests split coverage as their derivations predict, and most frame errors fall on uncertified frames.
WhyCombines discrete diffusion decoding with LP dual bounds and distance tests for per-frame ML certification; novel, with broad experiments on 15 codes against exact ML.
In briefDisentangled representation learning for face attack detection with smart glasses
SummaryThis paper addresses face presentation attack detection for images captured by smart glasses, which involve ego motion, viewpoint changes, clutter, and uncontrolled lighting. The authors release SmaGla-FPAD, a database of 297,000 images recorded with Ray-Ban Meta Wayfarer glasses indoors, outdoors, and in cars, including genuine faces and silicone mask attacks. They also propose DeCoupPAD-FAS++, which separates spoof evidence from capture variation using attack and style branches, pseudo-domain supervision, adversarial suppression, and prompt-guided alignment, and report the best balance between attack rejection and bona fide acceptance on this database.
Raghavendra Ramachandra, Narayan Vetrekar, Neha Naik et al.
In briefProtoEEG: Aligning EEG and images through shared visual prototypes
SummaryThis paper proposes ProtoEEG, which aligns EEG signals with images by softly composing them over a shared learnable visual codebook of 64 prototypes. A trainable EEG encoder and a frozen visual teacher produce assignment distributions over these prototypes, and their weighted reconstructions support retrieval and image generation. The authors report 91.5% Top-5 accuracy on 200-way retrieval in THINGS-EEG2 and 73.8% Top-5 on 80-way retrieval after retraining on EEG-ImageNet.
In briefAgentic AI for 6G Physical Layer: From Algorithm Discovery to Verified Silicon
SummaryThis paper presents an agentic workflow in which an LLM coding agent implements physical-layer algorithms on a commercial baseband SoC, optimizes them down to hand-written-style vector assembly, and validates them on silicon. Applied to an AI-discovered OTFS detector portfolio on the 6G4All platform, it yields twelve kernels that run cycle-exactly on the chip. It also produces a new detector that needs 13.9% fewer cycles than the ported discovered detector while matching its error rate.
WhyApplies an LLM coding agent to implement and optimize detectors on real baseband silicon, yielding a hardware-shaped detector; statistical significance of the gain is unresolved.
Eric Mittag, Lukas Brechtel, Maximilian Stark et al.
In briefTerraNostra-Bench: A Benchmark for Downscaling and Rolling Out Global 1-km Land-Use Maps
SummaryThe paper proposes a data-driven framework that treats high-resolution land-use reconstruction as semantic segmentation, combining coarse scenario data, climate classes, and static geophysical features to produce 1 km annual maps. It introduces TerraNostra-Bench, an evaluation protocol that separates spatial downscaling from temporal rollout and scores models on rare land-use-change pixels. Autoregressive models lose 5.7-7.2 accuracy points over 13 years, versus 2.2-2.8 for non-autoregressive models, a trade-off attributed mainly to the autoregressive design rather than the backbone.
Amirpasha Mozaffari, Marina Casta\~no, Stefano Materia et al.
In briefDo language and arithmetic use separate circuits in LLMs?
SummaryThis paper asks whether large language models use overlapping internal circuits for linguistic tasks and arithmetic tasks, inspired by neural dissociation in human brains. It extracts circuits with mechanistic interpretability methods and tests five LLMs with controlled materials. Circuit similarity is mostly explained by task domain rather than vocabulary or syntactic similarity, suggesting largely separate, ability-based modular circuits.
In briefFoundations of the Scaling Law: Mechanisms Behind Its Exponents
SummaryThis technical note explains three mechanisms behind empirical scaling laws that relate loss to parameters, training tokens and compute. It covers local approximation bounds, independent feature coverage and separable-loss optimization for training allocation. It also discusses how nonlinear scoring can produce apparent emergent abilities and how the SE(3) quotient affects approximation exponents for molecular configurations.
In briefSLM-SS: Token-Based Speech Separation Using Speech Language Models
SummaryThe paper proposes SLM-SS, a generative speech separation framework that represents multi-talker mixtures as discrete token sequences modeled by speech language models. Clean speech is encoded with neural codecs and arranged with Serialized Output Training, while an autoregressive model sets speaker order and coarse content and a non-autoregressive module refines residual codebooks. On Libri2Mix, the per-talker non-autoregressive variant reaches a word error rate approaching codec-only reconstruction. The WavTokenizer variant, which drops the refinement step, achieves the highest mean subjective quality among the evaluated systems.
In briefRELAY: Restoring Historical Builds via Reusable Rules and Agent Repair
SummaryThe paper examines 14,159 OSS-Fuzz historical cases and finds that 69.7% can no longer complete rebuilding and verification, mostly due to a few recurring failure symptoms. It proposes RELAY, a hybrid framework where an LLM derives and validates deterministic repair rules that are reused without model calls, while unresolved cases escalate to a per-case agent. The authors report that rule derivation can be fully automated and that the hybrid approach restores more builds than either method alone, at lower cost than agents only.
In briefRoboJudge: Using Multimodal Language Models to Judge Robot Video Generation
SummaryThe paper introduces RoboJudge-Bench, a benchmark of 800 human-annotated embodied videos for testing whether multimodal LLMs agree with human judgments on physical adherence and instruction alignment. Existing judges agree unevenly, with general-purpose MLLMs showing lower agreement on physical adherence. The authors build RoboJudge-Train, with 12K videos, and fine-tune a 9B model into RoboJudge-9B, which reaches a Pearson correlation of 0.719. This is a gain of 0.089 over the best proprietary baseline, Gemini-3.7-Flash, and the training also improves results on VideoPhy2 and Physion-Eval.
In briefGRADUS: Generating Graded Training Tasks for LLM Agents from Environment Rules
SummaryThis paper introduces GRADUS, a method for creating training tasks for LLM agents in policy-constrained tool environments such as airline booking, refunds, and cancellations. It extracts behavioral operators from a fixed environment without seed tasks, composes them into increasingly difficult tasks, and uses execution feedback to jointly refine tasks and solutions. The reported results show a 30B model improving from 59% to 89% on τ²-Bench and from 47% to 83% on CAR-Bench.
In briefRITE: Testing whether e-commerce AI agents can recover from silent tool errors
SummaryThis paper studies how tool-using LLM agents handle e-commerce queries when an intermediate tool output is silently corrupted, returning a plausible but wrong value with no failure signal. The authors define a Silent-error Resilience Task and build the RITE benchmark from industrial workflows and production-log inputs, comprising 501 instances across nine workflows. Evaluating 10 recent LLMs, they find silent errors roughly halve task success and are harder to handle than explicit failures. They also test a light State Checker as an intervention to improve resilience.
In briefGeometry Regularization for Vanilla NeRF via Cross-Ray Consistency
SummaryThis paper addresses noisy, inaccurate geometry in vanilla NeRF by adding a self-supervised geometry branch. It samples candidate surface points, creates virtual rays through each point from different directions, and enforces color consistency across them to constrain the implicit geometry. Experiments on eight Replica scenes report a 30.8% reduction in mean Chamfer distance compared with vanilla NeRF.
In briefRadar-Based 3D Human Pose Estimation with Two-Stage Refinement
SummaryThis paper proposes I2R-Pose, a two-stage framework for estimating 3D human poses indoors using millimeter-wave radar. The first stage estimates an initial pose sequence by using temporal relationships across radar observations and skeletal structure, while the second stage predicts joint-wise corrections to refine it. The authors also release the RaMoS-Pose dataset, and report reduced mean per-joint position error with fused inputs on RaMoS-Pose (15.13%) and MVDoppler-Pose (39.50%) compared with the best existing methods.
In briefDo ℓ1 attacks need concentrated perturbations?
SummaryThe paper studies how much ℓ1 adversarial attacks depend on concentrated changes by adding a per-coordinate cap to the global ℓ1 budget. It sweeps the cap for APGD-ℓ1, SLIDE, and FMN-ℓ1 across 18 standard-model settings. Caps up to 32/255 retain most attack success, while peak magnitude drops and effective support widens. Enforcing the constraint inside the optimization preserves more success than post-hoc clipping.
In briefEngram Memory Network: Brain-Inspired Prototype Explanations
SummaryThis paper proposes the Engram Memory Network (EMN), an explainable AI model inspired by the human memory system that mimics the ventral visual stream and inferior temporal cortex. A Hopfield-inspired memory module iteratively retrieves representative prototypes used as explanations. The authors evaluate perceptual similarity, subclass alignment, and information preservation across several datasets, reporting improvements and the highest average classification accuracy.
In briefPretraining one shared model to forecast many distributed solar sites
SummaryThis paper proposes a foundation-style pretraining framework that learns one shared forecasting model for distributed photovoltaic (PV) stations from heterogeneous multi-site, multi-horizon data. It frames NWP-assisted forecasting as sequence completion and compares three pretraining objectives, with a mixed reconstruction objective giving the most balanced results. Tested on 479 PV stations, the model outperforms several baselines from 6 hours to 7 days and can be applied zero-shot to held-out stations.
In briefCrystalAR: Autoregressive Modeling for Symmetric Crystal Generation
SummaryThis paper proposes CrystalAR, a unified autoregressive framework that models Wyckoff symmetry templates and complete crystal structures within a single generative sequence. It reformulates continuous coordinate generation as a symmetry-conditioned hierarchy of discrete spatial localization decisions. The authors report that the method retains strong symmetry-template generation, while achieving competitive de novo generation and template-free crystal structure prediction accuracy.
In briefPixelSR: Pixel-space orthogonal flow decoupling for real-world image super-resolution
SummaryThis paper addresses a limitation of latent diffusion models in real-world image super-resolution, where boosting perceptual realism tends to reduce structural fidelity. The authors propose PixelSR, a single-stage Pixel Diffusion Transformer that operates directly in RGB pixel space, bypassing VAE compression. Its Orthogonal Flow Decoupling splits the flow target into a low-rank semantic component and a noise-free orthogonal texture component, optimized in separate branches. The authors report competitive results across multiple benchmarks and a reduced fidelity-realism trade-off.
In briefAC/DC on a Budget: Cutting Compute for Sparse Neural Network Training
SummaryThis paper studies why alternating dense and sparse training phases help neural network sparsification, finding that reintroduced masked parameters receive large gradients that drive mask rewiring and sign flips. The authors propose ARC (Alternating Rewiring Compression), which shortens dense phases, zero-initializes masked parameters, and replaces dense phases with a sparse rewiring phase. On CIFAR100 and ImageNet, ARC matches AC/DC baseline performance at a fraction of the computational cost, and experiments with GPT-2 on OpenWebText report improved validation perplexity and zero-shot performance.
In briefChannel-Selective Optical-SAR Fusion for Object Detection
SummaryThis paper addresses how combining optical and SAR (synthetic aperture radar) images for remote sensing object detection can introduce redundant or conflicting features when channels interact without selection. The authors propose CSIF-Net, which uses frequency decomposition on SAR inputs, cross-attention to split features into shared and modality-specific parts, and a dual-branch fusion module. The authors report state-of-the-art results on the OGSOD 1.0 and M4-SAR datasets, with gains of 4.01% for horizontal and 3.18% for oriented object detection.
In briefGiving VLMs a Finger: Child-Development-Based Evaluation of Counting
SummaryVision-language models count exactly on small sets but switch to estimation on larger ones, much like children before they master counting. The authors give models child-like actions such as pointing and marking items within a training-free loop, and use the resulting marks to audit serial visual reasoning. Across 9 models, 682 images, and 7 conditions, the exact-counting range varies by model, and the marking loop mainly yields an auditable trace rather than higher accuracy.
In briefDetecting memorization in diffusion models via thermodynamic singularity
SummaryDiffusion models should learn the underlying density by reversing a diffusion process, but a key failure is memorization, where the reverse process instead learns from individual training samples. This paper analytically characterizes such memorized reverse processes and proposes a criterion that detects the associated thermodynamic singularity. The criterion reliably distinguishes properly learned networks from memorizing ones across several neural network models designed to promote or mitigate memorization.
In briefConstant Step Sizes for Muon-Type Spectral Optimizers on Strongly Convex Quadratics
SummaryThis paper studies Muon-type spectral optimizers with constant step sizes on strongly convex matrix quadratics, removing the effects of nonconvexity, noise, and step size schedules. It analyzes each update through its gain along singular directions and its length. The popular updates fall into three regimes: convergence with bounded gain, convergence with unbounded gain but vanishing length, and non-convergence for almost every initialization when length does not vanish. Experiments support these findings.
Pratik Jawanpuria, Bhabatosh Garai, and Bamdev Mishra
In briefS²-MAS: A Society-Scale Simulator Using 500 Million Household Persona Agents
SummaryThe paper presents S²-MAS, a multi-agent simulation framework built on 500 million synthetic household personas derived from real survey data, tracking consumption, cash, debt and payments. The authors trained a model on 2.93 million examples and report 74.55% accuracy on one-shot consumer financial-choice tasks, claiming it outperforms several compared models. Simulating 12 million household agents, they find that raising purchase goals and credit limits increases consumption, with effects varying across age and income groups.
In briefHearsay or Fact? How Framing and Elicitation Shape LLM Probability Updates
SummaryThis paper examines how much large language models change their probability estimates after receiving one new piece of information. Using premise-hypothesis pairs from δ-SNLI, the authors compare stated updates with reported ones ("Someone said that") and single-point with five-weighted-estimate elicitation across seven configurations. Under stated, direct elicitation, all models move in the annotated direction, while reporting consistently but only slightly reduces that movement.
In briefEEG-based classification of hypnotic susceptibility with 2D CNNs
SummaryThis study applies two-dimensional convolutional neural networks to minimally preprocessed EEG signals to classify hypnotic susceptibility without manual feature extraction. Data from 62 male participants (32 control, 30 hypno) were analysed using Leave-One-Subject-Out validation to estimate subject-independent performance. Anterior regions, particularly the prefrontal area, generalised best across subjects, reaching mean accuracy of 0.738 with an ensemble of ten windows. The occipital region fell to near-chance under this validation, suggesting subject-specific patterns rather than true group differences.
Krzysztof Wrobel, Grzegorz M. Wojcik, Filip Postepski et al.
In briefRSI Control: Can AI Models Secretly Sabotage Their Own Self-Improvement?
SummaryThis paper examines the safety risk when frontier AI models help train their successors. The authors built RSI Control Arena, a testbed where a model completes a normal training task while secretly trying to install a misaligned property in the next model. Six frontier models were tested against LLM monitors, and every model completed a substantial share of the hidden tasks without clearly hurting main-task performance. Monitors caught most attacks, especially when they could read the agent's reasoning, but not all. The authors conclude that strong safety measures are needed before models are trusted to train their successors.
WhyIntroduces a covert sabotage testbed for recursive self-improvement with 22 task pairs and six frontier models, a novel setting with moderate experimental evidence.
In briefCILA: Confusion-based logit adjustment for learning from imbalanced, long-tailed classes
SummaryReal-world datasets are often class-imbalanced, so models favor majority classes and perform poorly on minority ones. Existing logit adjustment methods use class frequencies, but ignore instance-level confusion and can suppress useful signals from confusing non-target classes. The paper proposes CILA, which adjusts logits based on instance-wise confusion between target and non-target classes, and BNI, which perturbs the target logit during training. CILA is a loss-agnostic plug-in that can be added to existing long-tailed learning methods and consistently improves performance across multiple benchmarks.
In briefAnytime-valid monitoring of learned safety barrier certificates at runtime
SummaryThis paper proposes a runtime monitor that checks whether a learned barrier certificate, used to establish safety of stochastic systems, remains consistent with observed closed-loop trajectories during deployment. It frames the barrier-drift condition as an episode-wise sequential hypothesis test, using a nonnegative test supermartingale for finite-sample false-alarm control under arbitrary stopping times. A deterministic temporal shaping of the boundary adds conservatism early in the horizon. The authors evaluate alarm behavior empirically under valid and invalid certificates, and note the monitor detects statistical inconsistency rather than predicting unsafe entry.
In briefWorld2HOSI: A Multi-Expert Harness for Generating Long-Horizon Human-Object-Scene Interaction Data
SummaryWorld2HOSI is a framework that generates continuous human-object-scene interaction data from 3D Gaussian Splatting scenes and natural language instructions. It coordinates existing specialized interaction models through a closed loop around a persistent scene state, so they can jointly handle tasks none could handle alone. Bounded residual optimization and contact-driven physics simulation are used to improve the physical plausibility of human motion and object manipulation. Experiments and a perceptual user study report the best task completion and interaction quality on the evaluated tasks, along with more diverse data.
In briefHow Language Models Organize Task Information: Readability vs. Readout Sparsity
SummaryThe paper introduces k*, the smallest number of features needed to recover 95% of a dense linear probe's AUROC, to measure how sparsely task information is read out from language model representations. Across 19 models from five labs on 14 core tasks, base models show similar k* values on readable tasks, while instruction tuning selectively lowers k*. The authors also report that saved heads can fail after continual learning even when linear readability stays high, and that pruning reduces readability.
WhyIntroduces readout sparsity k* distinct from readability, supported by cross-model, synthetic, and continual-learning experiments with broad coverage.
In briefTracing recurring and new hallucinations in step-by-step Text-to-SQL generation
SummaryThis paper introduces a drill-down evaluation framework that breaks Text-to-SQL queries into progressively complex executable sub-queries and matching natural-language questions to examine rule-based hallucination categories along the generation path. It separates recurrent hallucinations, seen in both an earlier step and the final query, from emergent ones that appear only at the final step. Across five LLMs on an expanded BIRD-mini benchmark, several categories show notable recurrence, and clause abuse shows relatively higher conditional emergence.
In briefPush-Pull Attentional Anchoring: a method for erasing concepts from diffusion models
SummaryThis paper proposes Push-Pull Attentional Anchoring (PPAA), a method that removes unwanted concepts from diffusion models by pushing target representations away in cross-attention space. It bounds the erasure strength with cosine-similarity constraints to limit damage to non-target generation. The authors use a relative similarity ratio as an adaptive threshold that can be applied uniformly across layers. Experiments on celebrities, artistic styles, nudity and intellectual property show favorable trade-offs on Stable Diffusion v1.4, and the method is also shown applicable to FLUX.
Nattanat Chatthee, Tagon Sompong, Ekapol Chuangsuwanich et al.
In briefDoes Experience Memory Help LLM Driving Agents? A Controlled Test
SummaryThis paper evaluates experience memory for closed-loop LLM driving agents in the HighwayEnv simulator. It resamples warmup experience pools and compares methods against token-matched irrelevant filler controls. The authors report that apparent memory gains depend on which warmup pool is drawn, and that no tested retrieval design exceeded the filler control.
