In briefDarwin-Claw: An Environment-Adaptive Framework for Agents That Learn from Their Own Experience
SummaryLarge language model agents usually work in isolated sessions and struggle to learn continually from past interactions. The paper proposes Darwin-Claw, which uses a hierarchical experience consolidation system to extract fine-grained error attributions and turn them into dynamically updated prototype vectors. These are used to guide context-sensitive perception and skill refinement. The authors report substantial performance gains on WildClawBench and state that the framework is automated and compatible with mainstream agent systems.
In briefBMuon: a Muon variant that budgets channels for structured pruning
SummaryThis paper introduces BMuon, a modification of the Muon optimizer that makes hidden channels across Transformer blocks compete for a shared allocation budget. The authors observe that Muon-trained Transformers degrade sharply when MLP hidden channels are pruned after training. On ImageNet-1K with ViT-S/16 and a 1B-parameter LLaMA-style model trained on FineWeb, BMuon outperforms Muon under structured MLP pruning.
Rida Assalouh, Tom Jacobs, Rebekka Burkholz et al.
In briefA general method for building data-adaptive anytime-valid bounds for iterative algorithms
SummaryThis paper presents a general recipe for constructing anytime-valid, data-adaptive boundaries for a broad class of iterative stochastic algorithms, so they can be safely stopped at data-dependent times. The error analysis reduces to a single requirement: time-uniform concentration for predictably weighted sums of noise, formalized via a strongly adapted sub-psi condition. Ready-made envelopes are derived for settings including SGD, temporal difference learning, and Bures-Wasserstein SGD, supported by numerical experiments.
In briefFrom Adult Kidney Tumors to Pediatric Wilms Tumors: Testing Kidney Lesion MRI Segmentation Across Age Groups
SummaryResearchers trained an MRI segmentation model on adult kidney lesions and tested it on an adult test set and a pediatric Wilms tumor cohort, comparing it with a pediatric-specific nnU-Net. The adult-trained model reached a median Dice score of 0.86 on adult renal lesions and 0.85 on pediatric tumors, versus 0.90 for the pediatric model. Non-inferiority could not be shown, so the authors highlight the value of pediatric training data and suggest fine-tuning as future work.
Sarah de Boer, Myrthe A. D. Buser, Hartmut H\"antze et al.
In briefBeyond Expected Values: Distributional Tree Search for Risk-Sensitive Planning
SummaryThe paper notes that standard Monte-Carlo Tree Search keeps only mean estimates, which limits its use for tail-risk goals such as conditional value-at-risk (CVaR). It introduces two distributional variants, CATSO and PATSO, which store finite-support empirical return distributions on action edges and select actions by CVaR. The authors prove an O(n^{-1/2}) bound on root value error after n rollouts for finite-depth nested CVaR planning. Experiments show improved lower-tail returns and fewer catastrophic outcomes when risk-sensitive and mean-optimal behavior differ.
In briefLiveSync: Iterative Self-Training for Realistic Lip-Sync Video Generation
SummaryLiveSync is an iterative self-training framework that generates lip-synced video from audio while keeping the subject's identity, head pose, and surrounding face unchanged. A mask-based model is first trained on real and externally generated clips, then a frozen checkpoint creates confidence-filtered pseudo pairs that are mixed with real data to refine the model. Experiments report leading synchronization on HDTF and competitive results on the AIGC-LipSync subset, with low inference latency.
In briefBeyond Latent Personalization: LLM-Based, Interpretable HRV-Conditioned Stress Detection
SummaryThis paper proposes an LLM-enhanced personalized stress detection framework that combines heart rate variability (HRV) features with personal attributes through LLM-based semantic interpretation. It reports a 29.9% relative MCC improvement over the strongest baseline on TILES-2018. The framework also provides human-readable descriptions of individual physiological responses.
In briefSkillBrowser: Selective Skill Exposure for Scalable Skill-Using AI Agents
SummaryThis paper proposes SkillBrowser, a framework managed by a Meta-Skill that lets language model agents identify the capabilities they currently need, write targeted queries, and load skills on demand. Across six SRA-Bench subsets and five LLMs, it improves overall accuracy over static retrieval and LLM-based selection, and reduces accuracy drops on tasks the agent could already solve. The gains depend on the model and task, and the authors note under-invocation of useful skills and added latency as limitations.
In briefReason-Guided Human Reward Modeling for Preference-Based RL: A study on teaching AI agents human-aligned behavior using preference and reason feedback.
SummaryThis paper proposes a preference-based reinforcement learning framework that learns a human reward model from trajectory-level feedback. Besides preference labels, it incorporates reason tags explaining why a behavior is preferred, so the learned reward captures both human choices and their rationale. The learned human reward is combined with the environment reward during policy optimization, aiming for interpretable, human-aligned, and preference-sensitive agent behavior.
Anshul Nayak, Shahil Shaik, Jonathan Smereka et al.
In briefCausal discovery beyond conditional independence: identifiability, algorithms, and inference for causal graphs
SummaryThis review examines methods for learning causal structure from data that go beyond conditional independence, using distributional asymmetries, conditional-moment restrictions, and variation across environments or individuals. It covers discovery under latent confounding and feedback, including the extra semantics needed for cyclic models. The paper also distinguishes population identifiability, consistency, uncertainty quantification, and algorithmic guarantees, and closes with a research agenda.
In briefExecution First, Memory Second: Diagnosing Memory Failures in Video World Models
SummaryThis paper introduces MIST, an evaluation that separates setting the scene, carrying out an instructed change, and showing the result in the final frame, so a low score can be attributed to the right cause. It finds that models set the scene but fail to carry out the instructed change, a gap that also appears in general memory evaluations. The authors propose M3-Agent, which recursively revises instruction details so the model executes the requested change, more than tripling the MIST reveal score over the baseline.
In briefConvexLLM: Scaling Convex-Hull Speech Interfaces to General-Purpose Audio Intelligence
SummaryThis paper presents ConvexLLM, a general-purpose speech and audio language model whose C-Gate interface represents audio as convex combinations of frozen text embeddings. It reports a VoiceBench aggregate of 75.36 across eleven tasks and 62.50% on MMAU Test-mini. Compared with MLP and Q-Former interfaces, C-Gate gives higher spoken-language aggregate performance, with task-dependent tradeoffs in transcription and audio understanding.
In briefPanoBench: A Panoptic Occupancy Prediction Benchmark for Autonomous Driving
SummaryPanoBench is a camera-centric benchmark for panoptic occupancy prediction in autonomous driving, spanning synthetic and real-world driving scenes. It unifies PanoBench-CarlaOcc, PanoBench-nuScenes, and PanoBench-KITTI360 under shared data interfaces and evaluation protocols, with multi-resolution annotations and persistent instance identities. The paper benchmarks representative methods on semantic occupancy, panoptic occupancy, and 4D occupancy forecasting.
In briefPreciseEdit: A training-free method for stable one-step image editing
SummaryThis paper addresses one-step image editing with text-to-image models, noting that existing methods can distort structure and drift the background when strong transport is applied along unreliable directions or to unedited regions. The authors propose PreciseEdit, a training-free and inversion-free framework that applies strong semantic transport only along reliable directions and locations, using a dual-noise-scale role decomposition. Experiments on PIE-Bench are reported to show state-of-the-art performance across various editing tasks.
In briefReVision: Repurposing Image Generators for Dense Visual Perception via Instructions
SummaryThis paper proposes ReVision, a framework that keeps a multimodal understanding model and an image generator frozen while a trainable parameter generator produces transient low-rank updates from an image and a natural-language instruction. It unifies segmentation, depth estimation, and surface-normal prediction through a shared RGB output without task-specific heads. The authors report competitive benchmark performance and analyze gradient disagreement across tasks.
In briefPONDER: A Method That Learns the Best Order for Image Generation
SummaryThis paper introduces PONDER, a method that learns the order in which an image generator should fill in content and which earlier predictions to revise. A policy chooses the next positions to fill and can mask earlier positions for regeneration, a mechanism called rethinking. The policy is trained with group-relative policy gradients and a new image-level reward, FIDgrad, based on changes in Fréchet Inception Distance. With frozen generators, learned ordering lowers MAR-B's ImageNet FID from 4.09 to 2.78, and rethinking further reduces it to 2.70.
Julius Erbach, Dominik Narnhofer, Bernt Schiele et al.
In briefEduWild: How Teacher-Configured Chatbots Fail in Real Use and How to Fix Them
SummaryThis paper introduces EduWild, a dataset of 1,726 conversations with 217 teacher-configured educational chatbots, including 900 conversations from end users of 35 deployed bots. The authors build a 29-code taxonomy of factual, workflow, diagnostic, and pedagogical failures. The most common failures involve not following teacher intent, such as skipping required workflow steps. A prompt-based harness targeting the six most frequent failure modes reduced the mean number of target failure codes per conversation by 65.2% in a study with 200 college students.
In briefLearning from Examples Beyond Pretraining: Unsupported Instructions and Composed Tasks
SummaryThis paper examines whether in-context learning, where a language model adapts to a task from a few input-output examples without updating its parameters, can succeed when the instruction or target task was not covered in pretraining. The authors develop a theoretical framework that separates these two kinds of distribution shift and analyze how demonstrations identify the task as their number grows. They also study long-horizon tasks that decompose into pretraining-supported subtasks, deriving a standard accumulated-error bound.
In briefContrastive likelihood scores for calibrated factual confidence in language models
SummaryThe paper studies how to estimate language models' confidence in factual claims more reliably. It proposes Contrastive Factual Evidence (CFE) and its probabilistic form, the Contrastive Factual Posterior (CFP), which compare a claim's likelihood under factual and non-factual conditions. CFP substantially improves calibration, while discrimination remains weak across methods in lower-resource languages. The authors align target-language claims with stronger English evidence along the factual-evidence direction, improving discrimination and open-ended factual generation.
Ponhvoan Srey, Jason Lucas, Pittawat Taveekitworachai et al.
In briefInterpreting WavLM Layers for Spoof Detection
SummaryThis paper studies which layers of the self-supervised speech model WavLM capture spoofing-relevant information, using the ASVspoof 5 database. It combines per-layer linear probing, canonical correlation analysis with acoustic features, and a search over layer subsets. The authors report that five layers outperform full thirteen-layer pooling, and that voice quality measures correlate more strongly with deeper layers.
Mohamed Mallat, Michele Panariello, Massimiliano Todisco et al.
In briefTrapped in Your Story: How Narratives Sway LLM Agreement
SummaryThe paper builds a controlled benchmark of 500 debatable questions across ten domains, keeping the user's stance fixed while varying how that stance is justified. Each question is tested under 11 matched conditions, yielding 5,500 evaluation instances across seven LLMs. Narrative-induced agreement varied substantially and was not monotonic: personal experience could reduce agreement, while causal framing and worldview-based synthesis increased it.
In briefPerceive, Solve, Reason: how time series LLMs progressively build three capabilities
SummaryThe paper builds the first suite that trains and evaluates three progressive time series capabilities for LLMs: Perceive, Solve, and Reason, with 337K training and 44K evaluation samples. It finds that a dedicated encoder is essential and that learning the capabilities in order outperforms mixing the same data. Distilled into a recipe, its 4B models outperform four recent 7B time series LLMs on seven of eight out-of-domain metrics without target-domain data.
In briefAn AI framework that predicts hydrogen spread in confined spaces with uncertainty estimates to help operators stay aware of leaks
SummaryThis paper presents a two-stage AI framework for hydrogen leaks during compressed hydrogen bunkering in maritime confined spaces. It first estimates the unknown leakage rate from early sensor readings, then forecasts how the hydrogen cloud disperses, providing a concentration map with confidence bounds. The system is a decision-support tool, while an independent emergency shutdown system remains the final safety layer. On an open dataset, leakage-rate estimation had a mean absolute error of 0.052 g/s and concentration prediction had an error of about 0.018.
In briefLocalIntent: Structured prompting that helps small local AI models correctly control smart buildings
SummaryThis paper presents LocalIntent, a prompting method that routes natural-language building-control requests either to ordinary commands or to persistent trigger-action rules, followed by deterministic validation. Four locally run Gemma and Qwen models (2B to 9B parameters) were tested on 101 non-persistent and 54 persistent-rule records. For rules, intent understanding rose from 26.4% to 80.6%, and action/parameter correctness rose from 5.6% to 44.9% compared with generic two-stage prompting.
In briefAdaptive conformal prediction for Grassmann-valued regression
SummaryThis paper presents a conformal prediction framework for regression with responses on the Grassmann manifold. Prediction regions are metric balls on the manifold, using basis-invariant geometric distances as non-conformity measures, with scores scaled by a learned difficulty estimator to make regions input-dependent. Experiments on synthetic benchmarks and an airfoil inverse shape-design problem show that geometry-aware scores are advantageous. In strongly heteroscedastic settings, the method can substantially reduce region size while maintaining the target coverage.
Dimitrios Loukrezis, Zachary J. Grey, Dimitris G. Giovanis
In briefA topological signature of neural network learning with noisy labels
SummaryThe paper uses persistent homology to track loop structure in neural activations, comparing representation geometry for clean-label and corrupted-label training subsets. On CIFAR-10 with ResNet-18, corrupted-label subsets show higher persistence at epoch 50 but no detectable difference at epoch 10, even though predictions match the noisy labels for only 7-11% of corrupted probe examples. In a separate calibration comparison, mean activation norm achieved a higher in-sample R² than raw persistence (0.936 versus 0.908).
In briefMuninAI: A Self-Hosted Research Assistant Framework for Independent Groups
SummaryThis paper presents Munin, an open-source framework for a self-hosted scientific assistant that indexes a research group's own corpus and citation graph. It wraps an open-weight model in an agent harness with reading, search, and computation tools. On 199 paired literature questions, the authors report accuracy rising from 0.387 to 0.874 over the base model. They also release the evaluation suite as a single runnable command.
In briefFinding best timing for choosing which unlabeled samples to label in online learning
SummaryThis paper studies how to decide, for a stream of unlabeled data where labeling is costly, which samples to label and train on. It analyzes online linear classification in a high-dimensional Gaussian setting, comparing a generalization error objective under a labeling budget and a cumulative cost objective. The authors find that greedy myopic policies select samples with margins in shrinking intervals, numerically determine optimal threshold schedules, and test a simple policy on synthetic and medical image data. They also note that naive margin-based policies can fail to be optimal for two-layer neural networks.
In briefTAN: A Size- and Form-Adaptive Masked-Convolution Encoder for Tissue Graphs
SummaryTissue graphs represent histology images using superpixels that follow tissue boundaries, instead of cutting tissue into rectangular patches. Existing encoders pad or mask cropped superpixels, so background fill values can affect node features. The authors propose TAN, a compact convolutional encoder with about 0.41M parameters that processes each superpixel at its native shape and excludes background, making its embeddings invariant to fill. On BACH, TAN's accuracy is not significantly different from the 303M-parameter foundation model UNI, with far fewer parameters and FLOPs and no pretraining.
In briefQD-Rec: Open-ended evolution of recommendations using quality and diversity
SummaryThis paper frames recommendation as building a diverse, high-quality set of options, where each item comes with a natural-language hypothesis explaining why it fits a decision maker's needs. The proposed QD-Rec method selects items with a contextual bandit and uses a foundation model to generate hypotheses, drawing on earlier recommendations kept in an archive as stepping stones. Across M&A, customer targeting, and paper and book recommendation tasks, QD-Rec reports the highest mean quality and the largest archive in most settings.
In briefWhen Evidence Vanishes: Teaching Medical Vision-Language Models to Rely on Image Evidence
SummaryCompact medical vision-language models can answer clinical questions correctly without truly using the image. This paper proposes Evidence-Conditioned Dependency Distillation (ECDD), which transfers the teacher's drop in answer support when matched visual evidence is removed. Tested on three medical VQA settings, ECDD keeps original-image accuracy competitive while making answers more sensitive to evidence removal.
In briefS-GAP: Stealthy Graph Attack Primitives for Graph LLMs
SummaryThis paper studies vulnerabilities of Graph Large Language Models (GraphLLMs) that reason over text-attributed graphs when node text and graph structure are perturbed. It introduces the Graph Attack Primitive (GAP), a unified semantic representation that combines textual and structural perturbations under one attack intent. It also proposes S-GAP, a multi-agent framework that iteratively refines GAPs using surrogate evaluation and textual feedback. Experiments on four datasets, three GraphLLMs, and two attack settings report an average 7.02% ASR improvement over the strongest baselines.
In briefA low-dimensional dynamical basis for controlling tongue shape
SummaryThis paper proposes that holistic tongue shape modes can serve as a low-dimensional control basis for a simple dynamical law of speech articulation. The authors identify empirical tongue shape modes in a cross-speaker coordinate space and fit a model to ultrasound data from 36 British English speakers producing monophthongal vowels. A three-dimensional linear point attractor accurately recovers tongue shape dynamics, suggesting complex tongue movement arises from linearly combined modes.
In briefHydraLLM: An elastic model that adjusts its width to fit different hardware
SummaryHydraLLM is an elastic MQA-based architecture in which the number of active attention heads sets the model's width, forming nested subnetworks within one parameter set for devices with different resources. The paper describes a pipeline that converts a pre-trained checkpoint into an elastic model and extends the approach to multimodal inference. Reported results include 2.2 points higher TriviaQA accuracy than a nested baseline with 17.25% more parameters at 75% of heads active, and 75.3% higher token throughput at 25% of heads active.
In briefTriINR: Unsupervised Atmospheric Turbulence Correction Using Three Implicit Neural Representations
SummaryAtmospheric turbulence blurs and distorts images taken over long distances. This paper proposes TriINR, an unsupervised method that recovers a sharp image from multiple degraded frames. It represents the sharp image, tilt, and blur as three implicit neural representations, with tilt and blur modeled across space and time so the number of parameters does not grow with frame count. Tests on synthetic data at three turbulence strengths show it is more robust than prior methods, and it also performs well on real long-range scenes.
In briefPredictable Scaling: Predicting Large Model Performance from Small Proxies to Explore Changes Efficiently
SummaryTesting changes to recommendation models normally requires full-scale training runs that use thousands of GPU-hours. This paper presents Predictable Scaling (PS), which builds scaled-down proxy models and fits curves to project full-scale metrics with uncertainty estimates. Evaluated on two ranking models, PS correctly ranks candidate hypotheses while reducing compute by 79-88%.
Aishwarya Wesanekar, Sheng-Tao Yang, Jiawei Zang et al.
In briefGatewayBench: A Unified Benchmark for Evaluating the Capability and Security of MCP Gateways
SummaryThis paper introduces GatewayBench, a benchmark for evaluating MCP gateways, the middleware that connects AI agents to external tools. It includes 524 tasks across six domains and 11,118 real-world tools from 251 MCP servers, with irrelevant tools injected to simulate large-scale deployments. Experiments on seven widely deployed gateways show differences in task completion, tool-use efficiency, and resistance to attacks, with effects that interact with the underlying model and agent harness.
WhyFirst unified benchmark studying how MCP gateway design affects agent capability and security, with 524 tasks and experiments across seven gateways.
In briefRepairing resampling distortion in high-dimensional particle filters via regularized free deconvolution
SummaryThe paper addresses how resampling in high-dimensional particle filters distorts the ensemble covariance spectrum and proposes a regularized free-deconvolution estimator that recovers the pre-resampling spectrum from observed weights and the empirical spectrum. The repaired spectrum is used only as a preconditioner for a Metropolis-adjusted rejuvenation move, so spectral error affects mixing efficiency but not the invariant target. Pre-registered simulations report an order-of-magnitude reduction in spectral error and recovery of about 90% of the oracle's sampling efficiency in an anisotropic regime.
In brief"You" Is Not Toxic: How LLM Labelers Distort Findings About Online Conversations
SummaryPrior claims about toxicity in online threads were mostly measured with Google's Perspective API, which is shutting down. This paper finds that LLM labelers mistake second-person "you" for toxicity, creating a rising-toxicity pattern with reply depth that human crowd labels do not show. On 1.25 million news comments, Claude Haiku flagged 12% of non-toxic comments without "you" and 51% of those with four or more. The authors propose Threadsafe, a method that certifies the statistical law itself by correcting labeler predictions with human labels.
WhyShows LLM labelers mistake second-person address for toxicity, creating spurious depth laws, and proposes an unbiased estimator; supported by experiments on 1.25 million comments across seven LLMs.
In briefRECREG: Recovery-Based Shielding for Safe Reinforcement Learning with Regular Properties
SummaryThis paper introduces RecReg, a safe reinforcement learning framework that enforces non-Markovian regular safety properties through recovery-based shielding. It builds theoretical foundations in the tabular setting, providing safety and convergence guarantees. Experiments using the MASA-Safe-RL library show efficient learning consistent with the theory.
In briefScoutRoute: An efficient routing method using internal language model signals to pick experts
SummaryThis paper presents ScoutRoute, a lightweight post-hoc framework for choosing among expert language models in a Mixture-of-Experts system. It runs a single forward pass through an inexpensive scout model and uses hidden-state probes and logit-based confidence to estimate each expert's competence, without modifying the models. Experiments on five open-weight models and four benchmarks covering math, code, commonsense, and multidisciplinary knowledge show it outperforms individual routing signals and calibrates better than conventional confidence measures.
In briefShadowCast: A Context-Aware Method That Adds Synthetic Shadows to Remote Sensing Images for Better Segmentation
SummaryShadows from tall objects change how land features look in remote sensing images and blur their boundaries, which hurts semantic segmentation. Instead of detecting and removing shadows, ShadowCast synthesizes shadows and injects them into images during training. It derives object masks from segmentation labels and estimates each image's shadow direction from the surrounding brightness, without using a learned network or acquisition metadata. Experiments on the Vaihingen, Potsdam, and LoveDA benchmarks show consistent gains over the baseline across several segmentation architectures, and further improvements when combined with standard augmentation and regularization.
In briefImproving Generated Actions from Observational Data with PolicyLift
SummaryThis paper studies how to improve decision policies learned from observational data when actions are complex structures such as text or images. The authors propose PolicyLift, a three-stage framework that learns discrete action codes, a context-dependent policy over those codes, and then maps the code-level policy back to the original action space. They show theoretically that improving the code-level policy guarantees improvement in the original action space. Experiments on synthetic and real-world text generation tasks report consistent gains over the behavior policy.
Jonas Schweisthal, Dennis Frauen, Stefan Feuerriegel et al.
In briefCombining compressed communication and local steps in federated learning using a spectral method
SummaryFederated learning trains models across many edge devices without moving their data, but data heterogeneity and communication cost remain major bottlenecks. This paper proposes a new spectral federated learning method that uses both local updates and communication compression. The authors give convergence guarantees under mild assumptions and remove the bounded heterogeneity condition. Experiments on standard image classification benchmarks show the approach is efficient and robust.
Dmitriy Bystrov, George Chirkov, Daniil Medyakov et al.
In briefMomentum: Persistent Dynamics for Video World Models
SummaryThis paper introduces Momentum, a method for video world models that keeps moving objects such as vehicles and pedestrians represented after they leave the camera view, so they reappear consistently. It uses synchronized videos from multiple eight-camera reference rigs as a shared memory of scene dynamics, generated in a two-stage pipeline. Experiments on generated street scenes report a 126.2% improvement in dynamic scene consistency on unseen scenes over static image conditioning, along with better visual fidelity.
Nedko Savov, Stefan Maria Ailuro, Danda Pani Paudel et al.
In briefResidual Stochastic Concept Bottleneck Models: interpretable models that keep concept interventions working when concepts are incomplete
SummaryThe paper introduces Residual Stochastic Concept Bottleneck Models (Res-SCBMs), which jointly model concept and residual variables with a multivariate Gaussian and learn their covariance. This lets interventions on known concepts propagate to dependent residual variables through Gaussian conditioning, without requiring complete concepts during training. The authors report that across several concept-incomplete benchmarks, Res-SCBMs improve predictive performance over baselines while keeping effective test-time concept interventions.
Stephen Marcou, Mateo Espinosa Zarlenga, Sonia Laguna et al.
In briefmmPencil: Radar-based in-air handwriting recognition using vision-language models
SummaryThe paper proposes m²VLMs, an architecture linking millimeter-wave radar sensing with large vision-language models, and builds mmPencil, a 3D in-air handwriting word recognition system. It reconstructs 3D writing trajectories from mmWave signals and converts them into visual representations so the models can recognize words with less sensitivity to writing style. The authors release a dataset of 7,664 samples from 12 users, reporting 97.60% accuracy on 200 word classes for seen users and 92.50% zero-shot accuracy on 50 classes for unseen users.
In briefGeoReason: Balancing Logical Reasoning and Precise Spatial Segmentation in Remote Sensing
SummaryThis paper addresses instruction-driven reasoning segmentation for remote sensing images, where a segmentation token produced at the end of long reasoning chains loses visual grounding, causing failures on tiny targets. The authors propose GeoReason, which injects the segmentation token into the initial query, uses bounding-box-guided coarse-to-fine chain-of-thought reasoning, and adds a patch-to-token alignment loss tied to remote sensing textures. The authors report that GeoReason consistently outperforms state-of-the-art methods on the Earthreason and RRSIS-D benchmarks.
In briefUsing Saliency Maps to Evaluate Generative Image Models
SummaryThis paper proposes a human-inspired way to evaluate generative image models without relying on pretrained feature extractors built for other tasks. For each model being evaluated, the authors train a classifier to tell real images from that model's generated images, then use saliency maps to examine its decisions. The classifiers mainly learn artifacts specific to each generator rather than a universal notion of fakeness. Combining classification accuracy with saliency information yields a composite score that the authors report is aligned with human judgments.
Eleonora Ristori, Luca Bindini, Marco Lippi et al.
In briefFlash RNN-T: a memory-efficient joint network and loss for speech recognition training
SummaryThis paper presents Flash RNN-T, an implementation of the joint network and transducer loss for streaming speech recognition that combines specialized kernels with optional tiled recomputation. On an H100, the untiled version achieves 24-39% higher throughput and uses 26-56% less memory than the fastest baseline on medium and long workloads. The tiled version reduces memory further, fitting long workloads within 80 GB where the tested alternatives run out of memory.
In briefLithoWorld: A World Model for Recovering Lithography Process Windows
SummaryThis paper presents LithoWorld, an endpoint world model that uses SEM images to recover lithography process parameters when process drift pushes patterns outside their process window. It combines interior representative targets, a causal predictive branch, and a conditional flow policy that outputs a two-dimensional correction. On 200 test pairs from a 73-group subset, it achieves 88.5% qualified-region success and 46.0% interior success, exceeding the strongest comparison method by 2.5 and 6.0 points.
In briefEstimating information in discrete data with diffusion models
SummaryThis paper introduces InfoSEDD, a discrete diffusion-based method for estimating information-theoretic measures such as mutual information and computing Kullback-Leibler divergences. It avoids embedding discrete data into continuous space, is lightweight and scalable, and can integrate with pretrained models. The authors demonstrate it on motif discovery in genetic promoter data, semantic-aware model selection for text summarization, and entropy estimation in Ising models, reporting better results than methods that use the embedding trick.
Alberto Foresti, Giulio Franzese, Pietro Michiardi
In briefContract design with an agent whose environment model is inaccurate
SummaryThis paper studies contract design between a principal and an agent whose transition model of the environment may differ from the true dynamics. The principal knows the true model and is modeled as interacting repeatedly through a principal-agent Markov Decision Process. The authors characterize contract design under this model mismatch and propose a framework that lets the principal achieve desirable long-term benefits despite the agent's inaccurate beliefs.
Elif Y{\i}lmaz, Jakub T{\l}uczek, Victor Villin et al.
In briefOn the Tip of the Tongue: Testing LLM Recall Failures by Cueing the First Token
SummaryThis paper argues that many closed-book factual errors in large language models come from choosing a wrong first token, not from missing knowledge. The authors introduce First-Token Cueing, which prefills the gold answer's first token, and find that this single token corrects a substantial share of multi-token errors across 16 open models. It also improves accuracy in all seven languages of a multilingual benchmark, suggesting that recoverable errors concentrate at the first decoding step.
In briefAWAM: Active Trajectory-Conditioned View Imagination for Robust Robot Manipulation
SummaryThe paper introduces AWAM, an active world-action model that plans where to observe and imagines what will be seen along the planned trajectory before acting. It predicts task-related virtual camera motion and uses it as a query to a geometry-conditioned world model, producing a reusable spatial prior that is fused with live observations for diffusion-based action prediction. The authors report success rates of 75.9% on LIBERO-Occ, 62.4% on LIBERO-Plus, 76.5% on RLBench-OG Occlusion1, and 94.4% in real-robot trials. These results suggest improved robustness to the evaluated occlusions and distribution shifts.
In briefVoltage time-series forecasting with an adaptive-kernel WGAN that controls Lipschitz constraints
SummaryThis paper addresses the difficulty of modeling long-term dynamics in high-dimensional recursive time-series forecasting. It proposes a Wasserstein GAN with a progressive Lipschitz penalty and an adaptive kernel critic, which forecasts the next values from a data window and dynamically adjusts its kernel for each new window. Evaluated on voltage and energy time-series datasets using power ratios, correlation coefficient, and relative RMSE, the authors report that the model outperforms benchmark methods in accuracy and robustness.
In briefAssigning topics to individual sentences using word and context clues
SummaryThe paper introduces SentenceETM, a neural topic model that assigns one topic to each sentence by combining lexical evidence (word distributions defined via embeddings) and contextual evidence (Gaussian distributions over sentence embeddings). Experiments show it consistently outperforms state-of-the-art methods for sentence-level topic assignment. A case study of a presidential debate illustrates how sentence-level topics reveal how candidates address the same subjects.
Federica Granese, Serena Villata, Charles Bouveyron
In briefData-driven prediction of optimal MPI rank placement and node count for memory-bound CFD on AMD EPYC clusters
SummaryThis paper studies how to choose node count and MPI ranks per core complex when running a memory-bound OpenFOAM case (motorBike) on AMD EPYC clusters, balancing speed against node-hour cost. It draws on 560 measured configurations across five Zen 4 and Zen 5 EPYC processors, including socket power telemetry. Large meshes are fastest and cheapest at the largest node count, while only medium meshes show a real speed-versus-cost trade-off. A simple procedure reusing the optimal configuration from the nearest measured mesh is near optimal in 75 to 85% of decisions.
Marcin Lawenda, {\L}ukasz Szustak, Aleksandra Krasicka
In briefFloodVLM-Pix: Language-Guided Pixel-Level Flood Understanding
SummaryThis paper builds the first large-scale remote sensing flood multimodal benchmark, with 131,200 flood image patches and 443,100 text descriptions of flood severity and evolution. It proposes FloodVLM-Pix, which uses a flood prior module to address attentional bias and a multi-temporal pixel decoder to handle interference such as clouds and mountain shadows. Experiments report better flood description quality and pixel-level segmentation than state-of-the-art vision-language models and flood segmentation models.
In briefMOSS-Image Technical Report: An Open Image Generator for Dense Text and Complex Layouts
SummaryThis paper presents MOSS-Image, an open image generation system that combines high-fidelity visual synthesis with intent-aware layout composition. It uses a 5B single-stream diffusion transformer pretrained from scratch, plus an agentic prompt enhancement module that turns user requests into detailed content and layout specifications. The authors also introduce LongLongTextBench and RealLayoutArena for evaluating long-form text rendering and layout-rich generation. The report states that MOSS-Image reaches 96.3% average text coverage on LLTB, the best among open-weight models, and ranks second in RLA's end-to-end comparison.
SummaryThis paper proposes the Interventional Memory Evaluation Protocol (IMEP) to check whether memory in vision-language-action robot systems truly affects outcomes, rather than relying on aggregate success rates. It introduces ChainBench, where correct decisions depend on historical information, and reports that the full memory configuration raises task success from 39.1% to 74.3%. The authors conclude that verifying a memory system's benefit requires fact-specific evidence beyond overall success rates.
In briefUniPolar: Restoring Images and Polarization Data Under Combined Degradations
SummaryThis paper addresses polarimetric image restoration for color polarization filter array imaging, where device and environment degradations occur together. It proposes UniPolar, which jointly reconstructs polarization-angle images and Stokes, DoLP, and AoLP quantities using two coupled streams, cross-domain interaction modules, and physical consistency learning. On the compound-degradation benchmark, it reports average PSNR of 31.522 dB for images and 24.097 dB for DoLP, with an average AoLP error of 16.833 degrees.
In briefBenchmarking artifact segmentation methods for H&E-stained whole slide images
SummaryThis paper benchmarks six publicly available algorithms that segment technical artifacts in H&E-stained whole-slide images. The methods were compared on two public datasets and a newly introduced dataset, measuring how well they segment artifact-free tissue and specific artifact classes. Performance differed significantly across methods, with HistoKit performing best and HistoQC performing worst, and dark spots and out-of-focus regions proving hardest to segment. A simple post-processing step further improved HistoKit's results.
Julia Merta, Daniel Mendelsohn, Benedikt Schoser et al.
In briefSTC-MGNet: A multimodal graph network that predicts whether pedestrians will cross
SummaryThis paper proposes STC-MGNet, a graph network that predicts pedestrian crossing intentions to improve safety in autonomous driving. It uses spatial-temporal-channel attention to learn from pedestrian pose sequences and fuses visual and ego-vehicle motion information through cross-modal channel modulation. The authors report accuracy of 70%, 87%, and 88% on the JAADbeh, JAADall, and PIE datasets, outperforming several baselines. Ablation experiments indicate that visual and ego-vehicle motion information each contribute complementarily to performance.
In briefRTSM: Applying Recursive Models to Multivariate Time-Series Classification
SummaryThis paper introduces the Recursive Time-Series Model (RTSM), which pairs a learned convolutional tokenizer with weight-tied refinement over a single latent state and a pooled classification readout. Across all 30 UEA datasets, RTSM reports 72.81% mean accuracy, the highest observed mean among 17 evaluated deep-learning models and 1.60 points above MambaSL. Ablations indicate that temporal encoding and state organization are key design choices.
In briefClient-aware, security-adaptive early-exit Vision Transformers for heterogeneous federated learning
SummaryThis paper studies early-exit Vision Transformer inference run locally by clients in federated learning. Instead of one global threshold, it sets a client-specific base threshold from label entropy and shallow-exit confidence, then adjusts each input using entropy, confidence, class margin, and exit depth. Reported results include 88.70% CIFAR-10 accuracy with 32.0% depth saving and 73.7% ImageNet-1K top-1 accuracy with 30% saving. On six-client non-IID CIFAR-100, full-depth accuracy is 72.28%.
Sahar Refaei Ghalibafan, Samira Nazari, Christian Herglotz et al.
In briefOliO: A Linear ODE-Based Transition Operator for Self-Supervised Time Series Forecasting
SummaryThis paper proposes OliO, a plug-and-play self-supervised method for time series forecasting. It uses a transition operator derived from linear ordinary differential equations to align latent representations across arbitrary time shifts, with a strictly upper triangular constraint on the transition matrix for numerical stability. Experiments across various backbones and benchmarks report MSE reductions of up to 5.75% over the strongest grid-searched SOTA baselines.
In briefLanguage Models Learn to Improve Language Models: Training LLMs to Propose Better LM Development Changes
SummaryThe paper introduces LM2LM, which uses reinforcement learning to train LLMs to generate proposals for improving language models. To keep training affordable, it uses a frozen research world model that predicts proposal gains from experimental records, and periodically executes highly ranked proposals to keep this feedback reliable. According to the paper, across eight autoregressive and diffusion LM development environments, Qwen3.8-27B trained with LM2LM outperformed GPT-6-sol on held-out tasks in real-execution evaluations, at 33% lower data-generation cost.
In briefHow Finite-Rank Bottlenecks Shape LoRA's Optimization Landscape
SummaryThis paper offers a theoretical explanation of why low-rank adaptation (LoRA) works, without relying on linearization via the Neural Tangent Kernel. With the backbone frozen, the loss depends on each adapted weight only through a finite-rank map set by the frozen activations, so trace-norm regularization selects low-rank solutions whose rank is governed by this "bottleneck rank." The authors also give tighter rank bounds for multi-head attention, show that first-order stationary points of the perturbed factorized objective are rank-deficient, and provide generalization bounds.
WhyReplaces NTK linearization with a finite-rank bottleneck analysis, proving attention rank bounds and nonconvex rank-deficiency results, a substantive theoretical advance.
Long Nguyen-Chi, Quynh Nguyen, Thanh Nguyen-Cung et al.
In briefRobust Function Calling for LLM Agents Under Interface Drift
SummaryThis paper studies how tool-using language model agents fail when their interface text, such as tool descriptions, argument schemas, and candidate tool lists, changes after deployment. The authors propose SchemaShield-Lite, a training-free repair layer combining schema validation, candidate-set retry, and one-shot constrained repair. On Qwen3.5-9B, it raises DICE-300 accuracy from 0.717 to 0.813 and BFCL-200 accuracy from 0.580 to 0.805, recovering 19 of 22 and 33 of 45 drift-induced failures.
Aniruddha Chattopadhyay, Kanishka Halder, Partha Sarathi Roy
In briefSETAREH Transform: A Haar-wavelet method that approximates repeated inner products using fewer values
SummaryThis paper proposes the SETAREH transform, which uses Haar wavelets to sparsify high-dimensional vectors so that inner products with a fixed coefficient vector can be approximated. Both vectors undergo the same permutations and Haar transforms, which keeps the inner product exact before truncation, and only a subset of transformed coefficient-input pairs is retained. Experiments on synthetic Gaussian vectors and apparent-age estimation with a CNN on APPA-REAL show that far fewer pairs than the original dimension are needed. Practical benefit depends on approximation accuracy and the cost of storing or transmitting permutation metadata.
In briefMinimax-Optimal Scaling Laws for Ridge Regression via Transfer Learning
SummaryThis paper asks whether reusing a pretrained model through transfer learning can substantially improve performance in terms of scaling law, studied in high-dimensional ridge regression. The authors derive a deterministic equivalent for the risk of a two-step procedure, where a source ridge model is followed by a target model regularized toward the source estimator, under both model and covariance shift. They identify a regime with a shared regression function and power-law covariance spectra where the scaling law reaches the minimax-optimal exponent, and show one-pass stochastic gradient descent attains the same rate. Experiments on synthetic data and language models support the analysis.
WhyDerives deterministic risk equivalents for two-step ridge transfer, proving minimax-optimal scaling exponents, validated on synthetic and language model experiments.
In briefEKT-Router: Jointly routing reasoning experts and external evidence for embodied vision-language models
SummaryThis paper introduces EKT-Router, a framework that helps embodied vision-language models answer questions from partial observations by jointly choosing a reasoning expert, optional knowledge, and perception tools. A verifier admits only evidence predicted to improve the selected expert's answer, and the expert then responds. It achieves a mean score of 61.4 across 6 embodied QA benchmarks, the highest among compared systems, and 74.7 on 10 general vision-language benchmarks. Ablations indicate that learned routing is the main contributor and that knowledge and perception are complementary.
In briefHiPredict: Using machine learning to predict a missing Mars camera color channel
SummaryThe HiRISE camera on NASA's Mars Reconnaissance Orbiter has a failing central detector, RED4, which halves its color imaging ability. The authors present HiPredict, a U-Net-based machine learning model that reconstructs RED4 from infrared, blue-green, neighboring red observations, and metadata. On test data it achieves a 1.1% normalised mean absolute error, better than the existing synthetic red channel's 3.1%. The tool is released as open software.
In briefReliable peak-memory scheduling for inference: exact execution, proven optimality, and recomputation limits
SummaryThis paper studies how reordering operators can lower peak activation memory for real-time inference on small devices without adding latency. The authors find that handing an improved order back to a compiler and regenerating code is unreliable, so they instead model memory exactly from the generated program's statements and reorder those statements. On CPUs the realized peak matches the model with bitwise-identical results. The paper also gives computable lower bounds that certify many schedules as optimal, and analyzes selective recomputation, including its hardness and its benefit on some networks.
WhyGives an exact peak-memory reordering method with computable optimality lower bounds and NP-hardness results, substantially advancing scheduling. Validated across CPU, GPU, and public models, with mod
In briefTransNSNet: Transformer-Enhanced Multi-Scale Network for Synthesizing PET-like Images from CT
SummaryThis paper presents TransNSNet, a Transformer-enhanced multi-scale adversarial framework that generates PET-like images from CT scans. It was trained and tested on a paired CT-PET dataset from the public NSCLC Radiogenomics collection, with patient-level train-test separation. TransNSNet showed favorable image-quality trends against Pix2Pix and related models, but the patient-level statistical tests found no significant differences after correction. The authors state that the results do not establish superiority or clinical validity, and synthesized images should not replace clinically acquired PET.
In briefCatchUp: Depth-Aware Re-Admission of Visual Tokens for Faster Vision-Language Models
SummaryLarge vision-language models repeatedly process hundreds of visual tokens across decoder layers, which is costly in computation and memory. The paper proposes CatchUp, a training-free method that combines reversible visual-token routing with Local Transition Correction, which corrects the outdated hidden states of tokens re-admitted after being skipped. At a 64-token budget, CatchUp processes 88.9% fewer active visual tokens per post-routing interval, retains 92.17% of dense-model multimodal performance, and reduces end-to-end latency by 8.0%.
WhyProposes reversible visual-token routing with local transition correction to fix stale representations, validated across multiple grounding and multimodal benchmarks.
In briefLoss Function or Architecture? What Drives Competing-Risks Survival Analysis
SummaryThis paper examines which factors drive deep learning performance in competing-risks survival analysis, where several causes compete to occur first. The authors vary three training losses and four backbones independently across 20 tabular and 3 pathology-image datasets. The cumulative-incidence loss with censoring weighting performs best, bringing about twice the benefit of backbone choice on competing-risks data. Proper scoring rules reveal these differences, while the concordance index does not separate the models.
In briefAgentOnto: Ontology-Guided LLM Agents for Enterprise Data Analysis
SummaryThis paper presents AgentOnto, an LLM agent for enterprise data analytics that organizes knowledge from databases, business documents, and example tasks into an ontology. The ontology guides analysis and incorporates confirmed feedback, and an iterative preparation step refines it using self-generated analysis tasks. Evaluated on two public data sources, AgentOnto substantially outperforms representative methods, and a real-world deployment shows its practical value.
In briefJoint reconstruction in photoacoustic tomography using an untrained CNN and coefficient-space regularization
SummaryThis paper studies reconstructing the initial pressure and spatially varying sound speed together from boundary measurements in photoacoustic tomography. The unknown fields are represented in wavelet spaces, and fixed features from an untrained convolutional neural network are used to estimate pressure coefficients via Tikhonov-regularized least squares, followed by a learned coefficient-space prior for the sound-speed coefficients. Experiments show accurate results on held-out data, good generalization to unseen phantom geometries, and stability under noise up to 10%. Compared with a supervised trained CNN, it gives substantially smaller errors in about one-sixth of the computation time.
In briefFrom Memory Hits to Answerable Evidence: Query-Conditioned Memory Compilation for Long-Term Agent Memory
SummaryThis paper proposes MemCompile, a framework that compiles query-specific evidence for long-term conversational agent memory. It first locates relevant sources for a query, then allocates a bounded evidence budget across complementary packets and renders their associated claims into answer-ready context. The authors report 80.52% Judge Accuracy on LoCoMo and strong results on LongMemEval-S.
In briefSelf-Reflective Chain-of-Thought (SRCoT): a closed-loop retrieve-reflect-revise method for evidence-based text generation on local LLMs
SummaryThis paper introduces Self-Reflective Chain-of-Thought (SRCoT), a modular architecture that runs text generation as a closed Retrieve-Reflect-Revise loop, using BM25 retrieval and provenance-labeled memory stores. The authors examine it through qualitative, trace-based case studies in two fully synthetic settings and run local prototypes on an Apple M4-family laptop. The traces show source-labeled outputs and explicit gap marking, but also provenance errors and unsafe assumptions. The authors state the work does not show superiority, hallucination reduction, clinical validity, or operational safety, and that it is a research proof of concept requiring human review.
In briefGuessing a Writer's First Language from Their English Using LLM Token Probabilities
SummaryThis paper studies native language identification, which infers a non-native English speaker's first language from their English writing. The authors score each text as a continuation of a prompt naming a candidate first language and choose the candidate with the highest score. They test Llama 3.1 8B and 70B on TOEFL9, Write and Improve 2024, and BULATS-L1. The method substantially outperforms direct classification, and the 70B model reaches state-of-the-art results on all datasets.
In briefImplicit Neural Models for Direct Multi-Shell Diffusion MRI Signal Estimation
SummaryThis paper proposes using an implicit neural representation network with wavelet activation functions to represent the full diffusion MRI signal. Fourier features and spherical harmonics encode the signal space, enabling super-resolution and q-space interpolation. A group-template pretraining strategy, combined with nonlinear registration, speeds up per-subject optimisation. Tests on 10 Human Connectome Project subjects and 26 Parkinson's disease cohort subjects show high-fidelity reconstructions with much lower memory use than raw diffusion-weighted images.
Sanna Persson, Fabian Leander Sinzinger, Christoffer Olsson et al.
In briefSimOPD: Transferring Teacher Knowledge via Response Similarity in On-Policy Distillation
SummaryThis paper proposes SimOPD, a method for transferring a teacher model's internal knowledge to a student during on-policy distillation (OPD). Inspired by contrastive learning, it uses memory banks of student-generated responses to turn the teacher's representations into similarity distributions, which the student matches within its own representation space. This avoids direct hidden-state matching and cross-model projection. Experiments report average gains of up to 2.6 and 3.1 percentage points over the strongest baselines in same- and cross-architecture settings.
In briefDesigning an AI data assistant to handle vague, jargon-filled questions using knowledge graphs
SummaryEnterprise data assistants built on large language models often receive user prompts that use domain jargon and lack schema references. This study designs and compares two semantic mediation approaches: schema-based mediation using semantic similarity, and knowledge graph-based mediation built automatically from the data lake. A simulation using real-world data, with LLM-as-a-Judge and human validation, found the knowledge graph approach outperformed the schema approach in Precision, Recall, F1, and preserving user intent.
In briefLook, Reason, Draw a Region: A Recursive Model for Predicting Spatial Context
SummaryThis paper addresses spatial context prediction, which identifies vacant regions relative to objects in images, a task where large vision-language models struggle. It introduces Spatial-TRM, a compact recursive model with 11.2M trainable parameters that builds on frozen vision-language features. On the RoboSpatial-Home benchmark it reaches 29.51% context accuracy, 9.02 points above a single-step model with the same parameter count. The authors conclude that repeated refinement improves prediction at a fixed trainable parameter count.
In briefOptimist Deep Research: an open-source agent that iteratively runs autonomous literature research
SummaryThe paper presents Optimist Deep Research, an open-source agent that treats literature research as an iterative process over a mutable research state. It splits a question into subquestions, searches multiple sources, deduplicates and expands documents through citations, and builds a structured evidence store. A critic returns actions that can revise the research plan, and the final report is written section by section with every citation tied to a retrieved document. The authors also report a case study and the weaknesses it exposed.
In briefHow elastic axis position folds the transonic flutter boundary, explained by shock energy and calibrated learning
SummaryThis paper studies the transonic flutter boundary of a pitch-plunge aerofoil (Isogai's case A) in inviscid flow, which folds near the bottom of the transonic dip. The authors combine coupled unsteady Euler simulations, a Gaussian-process surrogate, and split conformal prediction to map the boundary and its variation with elastic axis position. The fold exists only for elastic axes more than about 0.4 chord ahead of the leading edge and closes between 0.39 and 0.33 chord ahead. Chordwise energy budgets are presented as the explanation for the boundary.
In briefFlashReFocus: Interactive Photo Refocusing in Seconds
SummaryFlashReFocus is a method for changing the focus and depth of field of a photo after it has been taken. It runs a diffusion model once per photograph to recover the sharp image, then renders each further edit with a lightweight network controlled by a defocus map. The authors also introduce 3CReal, a benchmark of 89 image pairs from three cameras, to test generalization. They report the best perceptual quality for zero-shot bokeh rendering, refocusing 148.3 times faster than GenRefocus, and about 462 times faster per additional edit.
Ching-Heng Cheng, Chia-Ming Lee, Ming-Ching Chang et al.
In briefOptimal ranking compromise between two classification scores
SummaryThe paper addresses how to find a single ranking that best balances two common two-class classification scores, such as accuracy, F-scores, and Cohen's kappa. It treats these scores as special cases of a parameterized family of ranking scores and defines the optimal tradeoff as the one minimizing Fréchet variance under Kendall distance. The authors give an exact analytical solution for any pair of initial scores when class priors are the same across compared performances, and experimentally examine cases where performance follows fixed-prior distributions.
S\'ebastien Pi\'erard, Adrien Deli\`ege, Marc Van Droogenbroeck
In briefNeural control boosts performance of plug-and-play networked systems while keeping stability
SummaryThis paper studies how to improve pre-stabilized distributed controllers in networks where subsystems are added or removed, using neural policies. It builds on Internal Model Control-based performance boosting and derives free parametrizations that keep a prescribed network gain bound for any parameter value, so the neural controller can be trained without stability constraints. A DC-microgrid case study shows improved voltage transients under topology changes and load variations while nominal closed-loop stability is preserved.
In briefCoTA++: Why cached diffusion multimodal models repeat words, and how to fix it
SummaryDiffusion-based multimodal LLMs generate text by iterative denoising, and approximate caching to speed them up often makes the model repeat the same token at adjacent positions. The authors call this the Repeat Curse and trace it to three disruptions in information flow within steps, across steps, and across layers. They propose CoTA++, a training-free plug-in framework with three components targeting these disruptions. Across three models and two caching backends, it cuts the repetition rate of long responses from up to 99.6% to at most 17.0% and recovers most accuracy lost to caching on nine benchmarks.
In briefPretraining Classifiers with Diffusion Models for Limited Data
SummaryThe paper proposes DiffPT, a simple diffusion-based method for pretraining a classifier backbone when domain-specific training data is limited. It is tested on time-series data, standard natural-image benchmarks, and medical imaging datasets, and compared against modern and traditional unsupervised and self-supervised methods. Contrastive pretraining performs better on CIFAR-10 and Tiny ImageNet in low-data settings, while DiffPT outperforms self-supervised baselines on medical imaging and time-series tasks. The authors conclude that DiffPT is a simple and effective approach for data-efficient pretraining across modalities.
In briefI2V-DETACH: Blocking Unauthorized Image-to-Video Generation by Detaching Source Grounding
SummaryThis paper proposes I2V-DETACH, an image immunization method that protects images from unauthorized image-to-video generation by diffusion models. Instead of degrading output quality or temporal coherence, it weakens how source-image content propagates into video latents, suppressing source-video grounding and strengthening intra-video interactions. Experiments across diverse diffusion architectures show it substantially reduces source-content preservation, outperforms prior immunization methods, and transfers well in black-box settings.
In briefAnnotrace: A tool for labelling each step of AI agent runs
SummaryThis paper presents Annotrace, a workbench for annotating execution traces of LLM agents. It reads six trace formats and converts each run into an ordered, nested list of steps, letting annotators write their own labelling codebook. A frontier model proposes an initial set of labels and answers alongside the human, making annotation a joint human-AI process. On a multi-agent question-answering corpus, the authors report reproducing 235 prior manual annotation units, with two models agreeing 91% of the time (κ = 0.82).
Chenyi Huang, Arun Josephraj Arokiaraj, Bin Wu et al.
In briefXiaomi-Lens-0 Technical Report: A Multimodal Model That Predicts Region Masks
SummaryThis report presents Xiaomi-Lens-0, a 4B multimodal large language model built on Qwen3.5 that predicts masks from its own visual representations. A lightweight decoder combines vision-encoder patch features with the language model's mask-token states, so mask supervision reaches the shared backbone without a separate segmentation encoder. Across all eight RefCOCO, RefCOCO+, and RefCOCOg splits, the model gains 5.6 points on average over Qwen3.5 in referring comprehension, and it surpasses Qwen3.5 on all perception benchmarks.
In briefExplaining Chain-of-Thought faithfulness by verbalising model computation
SummaryThis paper introduces VERITAS, which uses layer-wise Jacobian projection to describe how a model's internal reasoning evolves at each generated chain-of-thought token. A classifier trained on these structured features reveals how and where reasoning diverges toward unfaithfulness. Across four datasets from FaithCoT-Bench and BonaFide, VERITAS reports a 28.4% gain in balanced accuracy over LLM-as-Judge and a 16.0% average improvement over activation probing.
In briefDeDiCa: A scalable model for causal inference with discrete variables and hidden confounders.
SummaryDeDiCa is a framework that uses a tensorized parameterization of causal Bayesian networks to handle discrete variables with hidden confounding. It is trained once by maximum likelihood on observational data, exactly marginalizing over hidden confounders, and can then estimate any (proxy-)identifiable interventional query. The paper reports that it matches state-of-the-art neural causal generative models in ATE estimation while enabling memory-efficient CPU training and exact interventional inference in milliseconds.
Kunchangtai Liang, Adri\'an Javaloy, Antonio Vergari et al.
In briefNestedVLA: Combining Separately Trained Robot Skills Without Original Training Data
SummaryThis paper proposes NestedVLA, a method for consolidating separately fine-tuned vision-language-action (VLA) robot policies without access to their original training data. Instead of merging all updates into one checkpoint, a hypernetwork in the VLA generates action-expert weight updates for a selected task, with each task's compressed update stored separately. Evaluated on three reinforcement-learning-tuned policies across MetaWorld, ManiSkill, and CALVIN, it reaches 66.7% average success, 41.1 points above the strongest model-merging baseline. Total storage for three policies is 1.5 times that of a single full checkpoint.
