Many questions about a video concern one physical object. Where does the ball end up after the cups are shuffled? Which card is the queen? How often was the mug covered? Open video-language models (Video-LMs) recognise such objects in single frames but do not keep them apart over time. On shuffled shell games, open models from eight families stay near chance, and so do published tool agents. We separate semantic identity binding from temporal identity maintenance. Binding should happen at a moment when the evidence for the object is visible, and maintenance should be left to a tracker. Track, Then Bind implements this principle without training. A promptable video tracker follows every candidate object and draws it with a track number. A frozen vision-language model (VLM) answers one question about the identifying evidence, which binds the target to a track. The answer is then derived from a text description of the tracks. With Qwen3-VL-32B, the agent solves both VET-Bench tasks, where a fine-tuned trajectory model reaches 92% and 90%. On real cup games from the Perception Test, it raises accuracy from 35.5% to 52.3%. Without retuning, the same configuration also solves VET-Bench with Gemma 4 and GPT-5.5. Its track-derived visibility timelines add 8.8-17.9 points on TOC-Bench for open models. GPT-6.1 already answers 92.5% of the real cup games directly but 71.0% through our pipeline, while our timelines still raise its TOC-Bench accuracy from 52.5% to 57.7%. The benefit of external tracking thus depends on the underlying model and the task.
Real-time polyp segmentation in computer-assisted colonoscopy is hampered by the difficulty of deploying high-accuracy models under strict latency budgets. Frame skipping, which performs inference every K frames and reuses or interpolates the rest, improves throughput but introduces motion blur and lesion misalignment that degrade temporal consistency. To close this gap, we propose the Lightweight Frame Compensator (LFC), a 77K-parameter plug-and-play module inserted before a frozen segmentation backbone. Operating on a three-frame window, LFC fuses consecutive frames via spatio-temporal attention into a single compensated frame, and is trained only on pseudo-labels generated by a differently seeded model, without any expert annotation. Experiments on three polyp video benchmarks show that LFC recovers most of the accuracy degraded by direct frame skipping and, in several settings, reaches the accuracy of frame-by-frame inference, while retaining a 2.05X--2.55X throughput gain. We further evaluate the temporal stability of three backbones spanning CNN, pyramid Transformer, and global-attention Transformer architectures, revealing a spectrum of sensitivity to frame skipping that arises from how locally or globally each architecture aggregates misaligned input. The compensation gain of LFC follows this spectrum: it is largest for high-capacity Transformer backbones, which achieve the best static accuracy but suffer the most under skip-frame inference. LFC thus offers a lightweight, pluggable, and deployment-friendly route to real-time polyp segmentation.
Large Language Models (LLMs) can generate UML class diagrams from natural language requirements, but hallucinations may compromise the correctness and completeness of the resulting diagrams. Existing evaluations mostly assess the final diagram, without examining the intermediate traces of reasoning-capable LLMs, which can reveal where these hallucinations arise. We develop an empirical taxonomy of eleven hallucination sub-types organised into three dimensions: Prompt Faithfulness (deviations between requirements and reasoning), Code Faithfulness (discrepancies between reasoning and generated PlantUML code), and Reasoning Quality (problems within the reasoning). The taxonomy was derived from 60 manually annotated DeepSeek-R1 traces, applied to 120 further DeepSeek-R1 traces through a dual-LLM annotation pipeline, and examined on 20 DeepSeek-V4-Pro traces. Code Faithfulness was the most frequent dimension for DeepSeek-R1 in both datasets, whereas Prompt Faithfulness dominated for DeepSeek-V4-Pro, suggesting model-dependent failure patterns. In the manually annotated traces, reasoning-trace length showed no significant monotonic association with hallucination frequency, whereas requirement complexity was negatively associated with it, particularly for Prompt Faithfulness errors. In a mitigation experiment on 13 Golden UML ModelSet cases with three Qwen3-based models, adding Prompt Faithfulness and Reasoning Quality constraints to the generation prompt raised the pooled mean F1 score by 4.0 percentage points, whereas a post-hoc Code Faithfulness review provided no consistent further benefit. Relationship modelling remained the lowest-scoring element throughout. These results suggest that localising hallucinations within the generation process can inform targeted, prompt-based mitigation. Whether these findings generalise beyond the DeepSeek and Qwen3 model families remains to be tested.
Medical knowledge evolves, yet whether large language models (LLMs) faithfully incorporate revisions into clinical recommendations remains unclear. We introduce ChronoEvo, comprising 1,709 recommendation-level revisions across 12 forms of medical knowledge change and 275 guideline lineages, to distinguish recognition of revised guidance from its full incorporation. With updated guideline evidence explicitly provided, revision recognition exceeded full incorporation across 14 models, with a 20.1-percentage-point gap in the strongest model. For Qwen3-4B, 90.8% of responses recognized revisions, 80.2% were clinically correct, and 58.5% fully incorporated them, yielding a 32.3-percentage-point gap. Two clinicians' adjudicated assessment of 300 responses yielded a 40.3-percentage-point gap in this enriched validation sample; 61 of 144 responses with incomplete incorporation were clinically incorrect. Prompting Qwen3-4B to prioritize revised evidence reduced outdated-content retention and improved clinical correctness without significantly improving full incorporation. These findings identify faithful incorporation of revised guidance as a distinct challenge for clinical LLM reliability.
Reasoning LLMs often continue reasoning after reaching a correct answer, wasting computation and sometimes replacing it with an incorrect one. Deciding when to stop is a metareasoning problem: the model must weigh what further reasoning could achieve against its cost, given what it can already answer. Direct utility training learns a stopping representation at a chosen token price; supporting new prices can require additional controller training. We introduce BELT (BELief-based Termination), which separates outcome prediction from stopping control. It learns beliefs about current and future answers from a frozen LLM's hidden states, then fits a linear stopping policy on the frozen features. The prediction targets are independent of token price, so one probe supports different accuracy-cost trade-offs through CPU policy refitting or threshold adjustment. Averaged over five out-of-distribution (OOD) math benchmarks, BELT reduces generated tokens by 41% on Qwen3-1.7B and 51% on OLMo-3-7B-Think without loss of accuracy. The same math-trained BELT transfers to coding and agentic tasks without training on code or agentic data. In all three domains, its accuracy-cost trade-offs are competitive with or better than those of end-to-end controllers. On OLMo3, its probe has 21.2x fewer parameters than one checkpoint of the strongest baseline, OS-Pruner, and needs 55x fewer training GPU-hours than OS-Pruner's five price-specific checkpoints.
We propose FiLM-CAM; a conditional-access image watermarking approach that enforces authorized encoding, decoding, and removal through keyed feature modulation. A secret key conditions a Feature-wise Linear Modulation (FiLM) mechanism that transforms the internal representation of the watermark signal before embedding, thereby coupling the payload to the key at the feature, rather than pixel level. The same key must be presented to condition the decoder to correctly recover the payload embedded in an image. To enforce conditional access, we train the system to maximize the performance gap between authorized and unauthorized use, ensuring reliable extraction under the correct key while inducing failure under incorrect keys. In addition, we design our encoder to estimate the watermark residual invariant to the presence of pre-existing watermarks. This design enables conditional watermark removal via simple re-encoding and subtraction of the residual using iterative refinement, contingent on knowledge of the secret key. We evaluate FiLM-CAM across multiple watermarking encoder-decoder backbones, demonstrating that keyed feature modulation serves as an effective, architecture-agnostic conditional access module (CAM) for image watermarking.
We prove that for every open $U \subseteq \mathbb{R}^n$ the spherical projection of every gradient trajectory $\theta(t)$, $t \geq 0$, of a definable $C^1$-potential $F \colon U \to \mathbb{R}$ has finite length and, in particular, converges (directional convergence). In other words, every gradient trajectory of a definable $C^1$-potential has finite total angular motion. Here, definable means definable in an arbitrary o-minimal structure expanding the real field. Notably, the gradient trajectory in this result is an arbitrary solution of the gradient flow differential equation (whose solutions may not be unique) with no further assumptions: The gradient trajectory may be unbounded and diverge to infinity or it may be convergent. The established theorem therefore provides affirmative answers to both the gradient conjecture at infinity for definable potentials and the definable Thom gradient conjecture. We extend our conclusions from the Euclidean metric to a general class of ambient metrics including Riemannian metrics and we establish uniform local upper bounds for the lengths of the spherical projections of gradient flow trajectories. We finally apply our findings to the training of deep neural networks (DNNs) to conclude that, in the case of definable and continuously differentiable activation and loss functions, the spherical projection of the DNN training flow must always converge (directional convergence).
One-shot landmark localization in volume scans with varying fields of view (FOVs) is challenging. While existing self-supervised learning methods have shown promising performance, they rely on contrastive learning built upon augmented views of the same image, which inherently limits their generalization in cross-modality scenarios. This motivates research on cross-modality embedding learning using paired multi-modality images. However, existing methods suffer from not only limited training data, but also the requirement of multi-stage iterative training. To address these limitations, we propose UNIVERS, an end-to-end trainable framework that learns UNIversal Voxel-wise Embedding via Randomized Synthesis. Our framework leverages diverse pseudo-data synthesized from organ shape templates, with randomization applied to semantic layout, visual appearance, FOV, and spatial pose. Coupled with cross-domain discriminative voxel-wise contrastive learning, this domain randomization strategy forces the network to disregard domain-specific biases and learn features that are robust against variations in imaging contrast, resolution, and spatial context. Consequently, the trained model enables one-shot landmark localization in multi-modality volume scans with arbitrary FOVs. In this study, the efficacy of the proposed method is demonstrated by applying it to the task of localizing the adrenal glands, which are paired, small, and irregular endocrine organs. In particular, comprehensive experiments are conducted on three challenging adrenal gland datasets. The experimental results demonstrate that UNIVERS consistently outperforms the state-of-the-art (SOTA) one-shot landmark localization approaches, with particularly prominent advantages in cross-modality scenarios. The source code is publicly available at https://github.com/Kendra9999/UNIVERS.
Amodal video segmentation asks for the complete mask of an object in every frame, including the parts hidden behind other objects. The best current methods fine-tune video diffusion models to generate these masks; they need a depth network and tens of denoising steps, and their outputs vary from sample to sample. Self-supervised video world models of the JEPA family are pre-trained on a closely related task, predicting the representations of masked space-time regions from the visible ones, yet their predictor is usually discarded after pre-training. We reuse it by treating occlusion as masking. Tokens not covered by the object's visible mask are dropped, the V-JEPA 2.1 predictor is fine-tuned to fill them in, and a lightweight decoder turns all features into amodal masks. The resulting model, OccJEPA (Occlusion-as-masking JEPA), is deterministic and 32x faster per frame than Diffusion-VAS, the diffusion-based state of the art. It is also more accurate on SAIL-VOS (+3.70 occluded-region mIoU over all occluded frames), on TAO-Amodal (+4.00 AP50 on a held-out test split, +4.32 with SAM2-tracked masks) and on Kubric. Controlled experiments trace most of the gain to the predictor's pre-trained weights rather than its capacity, and the predictor helps at every encoder size tested (87M to 1.85B parameters).
Omni LLMs can naturally locate key audio and visual evidence and integrate information across modalities for audiovisual reasoning. Existing audio-visual reasoning benchmarks that mainly rely on multiple-choice questions usually score models solely based on answer correctness, making it difficult to assess whether a model truly forms a reasoning chain from evidence to the correct answer. In addition, textual or unimodal shortcuts may allow models to answer correctly without analyzing the necessary audio-visual evidence, further weakening the reliability of evaluation benchmarks. To address these issues, we introduce EvaOmni, an evidence-grounded benchmark containing 392 videos and 1,027 purely human-annotated question-answer pairs, covering 15 capabilities and three levels: retrieval and localization, basic understanding, and complex reasoning. We design a series of rigorous annotation procedures in which experts annotate audio and visual evidence and construct video-based distractors; humans review whether questions meet strict inspection criteria, and finally models filter textual and unimodal shortcuts. Beyond high-quality question-answer pairs, we require models to simultaneously return the answer and supporting evidence intervals in order to evaluate the complete reasoning chain. We propose the Coverage Grounded Score (C.G.) to measure how much annotated evidence a model identifies, and the Bottleneck Grounded Score (B.G.) to penalize poor temporal localization in either modality; both are incorporated into answer correctness. Across 7 open-weight and 17 proprietary Omni LLMs, Gemini3.7-Flash achieved the highest overall accuracy (79.4%), C.G. (70.8%), and B.G. (35.1%). At Level 3, it reached 79.1% accuracy, but its B.G. was only 12.3%. These results show that even the most advanced models still have room for development in grounded audio-visual reasoning.
Negation remains a persistent challenge for language models across NLP tasks and modalities. We show that this weakness can undermine LLM safety guardrails through a structural vulnerability we term polarity blindness: the failure to recognize harmful intent after polarity-altering linguistic transformations. We study two forms of polarity inversion: grammatical negation and antonym replacement, and use synonym replacement as a polarity-preserving control. We further test whether an adversarial textual wrapper amplifies these failures. Using a 260-prompt benchmark across 10 state-of-the-art LLMs and 3 languages (English, Spanish, Czech), stratified by harmfulness potential and evaluated on a validated 0-5 harm scale, we find that polarity blindness affects most evaluated models, with mean harm scores reaching 4.12 for the most vulnerable model. The effect persists across typologically differing negation systems, suggesting that polarity blindness reflects a broader limitation in semantic robustness rather than a language- or grammar-specific failure.
Learned barrier certificates provide a principled mechanism for establishing safety of stochastic systems, but their validity may deteriorate during deployment when the underlying dynamics differ from those used for training. We propose an anytime-valid runtime monitor that audits a learned barrier certificate using observed closed-loop trajectories. The monitor is intended to detect loss of statistical support for the deployed barrier condition; it does not by itself guarantee detection before unsafe entry or safety of a subsequent intervention. Specifically, the conditional barrier-drift condition is formulated as an episode-wise sequential hypothesis test, and observed barrier increments are transformed into a nonnegative test supermartingale, yielding finite-sample false-alarm control under arbitrary stopping times. Inspired by finite-horizon safety verification, we further introduce a deterministic temporal shaping of the monitoring boundary, making the monitor more conservative early in a prescribed deployment horizon while preserving anytime-validity. The monitor is designed to detect statistical evidence that the safety certificate is no longer consistent with the observed dynamics, rather than to predict unsafe entry. To assess the monitor empirically, we evaluate empirical alarm behavior under valid and invalid certificates. Code and data are available at https://github.com/FormalCert-AI/runtime-barrier-monitor.
Traditional evaluation benchmarks reduce inherently interconnected scientific knowledge in life sciences into flat lists of questions, disregarding the underlying topological structure of the knowledge. We introduce LG-Bench, the first graph-structured benchmark for life sciences, featuring over 10,000 high-quality multiple-choice questions across medicine, biology, and chemistry. Our approach constructs a weighted evaluation graph using bidirectional matching and semantic similarity algorithms, where nodes represent questions and edge weights capture their semantic relationships. Leveraging this graph topology, we design two novel evaluation metrics. The Global Coherence Score (GCS) measures a model's consistency within semantically related neighborhoods, while Knowledge Balance Score (KBS) analyzes how model errors are distributed across the graph to reveal conceptual blind spots. LG-Bench facilitates fine-grained comparison of LLMs by surfacing differences in conceptual coherence and patterns of knowledge organization across models. Our framework shifts the evaluation paradigm from flat accuracy metrics to structure-aware analysis, offering a new lens for diagnosing and improving LLM performance in the life sciences domain.
A bimanual demonstration specifies both what the robot should do and one particular schedule for when each arm should act. Standard imitation learning does not distinguish task-required cross-arm constraints from incidental scheduling choices, and may therefore exploit opposite-arm progress as a shortcut cue, creating spurious cross-arm coupling even where relative timing is irrelevant for the current task phase. We introduce Constraint-preserving Temporal Recombination (CTR), a training-time data intervention that preserves recorded arm-local sensorimotor content while retiming the corresponding segments to realize admissible schedules that preserve task-required cross-arm constraints. We use CTR to train Diffusion Policy, Action Chunking with Transformers, and $\pi_{0.5}$, and evaluate them on four simulated and four real-robot tasks. Across six real-robot bimanual progress-state initializations, CTR raises worst-state success from 56% for the strongest non-CTR schedule mixture to 76%. With $\pi_{0.5}$ on an Independent task, a 4 s pause transfers 2.56 s of delay to the opposite arm under the Concurrent schedule distribution, but only 0.25 s with CTR. On real robots, CTR violates task constraints in only 0--4% of trials, versus 36--98% for narrow-schedule baselines. Code will be released at https://github.com/MINT-SJTU/CTR.
Image restoration is a fundamental inverse problem that seeks to recover high-quality images from degraded observations. Plug-and-Play (PnP) methods provide a flexible framework for integrating pretrained denoisers into iterative reconstruction, whereas flow matching offers an efficient generative modeling approach based on continuous-time ODE dynamics. Very recently, Martin et al. (ICLR, 2025) developed a PnP-Flow method, which integrates the PnP framework with flow matching through a time-dependent denoiser constructed from a pretrained flow matching model. By incorporating a different time schedule and an additional pixel-space projection operator, this paper proposes a novel efficient PnP-Flow algorithm (HP-PnP-Flow) that enjoys rigorous convergence guarantees under mild assumptions, without imposing restrictive conditions on iterates required by the aforementioned work. Numerical experiments on CelebA and AFHQ-Cat demonstrate that the newly proposed HP-PnP-Flow algorithm significantly outperforms PnP-Flow on several restoration tasks such as denoising, deblurring, super-resolution, random inpainting, and box inpainting. Notably, our algorithm achieves comparable or better reconstruction quality with substantially fewer iterations, reducing the iteration count from $500$ to $100$ for deblurring and super-resolution on AFHQ-Cat.
Reinforcement learning (RL) solves the explore-exploit problem by training a learner on reward; its multi-agent, resource-sharing form has an equilibrium allocation as its optimum -- the ideal-free distribution (IFD), the Nash equilibrium of a congestion game. We show a minimal swarm approximates it with no learner at all. We present a decentralized, learner-free method for this problem: agents explore by heavy-tailed (Levy) flights, recruit neighbours to a discovery through a degree-normalized threshold cascade, and disperse as a patch crowds. The method drives the collective to the shape of the equilibrium allocation (pooled spatial matching slope 1.14, R^2=0.61) with no training, no stored parameters, and no reward in its dynamics; it reaches this as a self-organized attractor and holds it, re-finding it when the resources move -- though it reaches the shape, not the exact equilibrium, and stays measurably exploitable (Nash-gap 1.46). Coupling recruitment to movement causes the allocation: removing it collapses the matching (slope 0.006). Against Proximal Policy Optimization (PPO) learners (feedforward, centralized-critic, graph message-passing, and recurrent) trained in the same sensing-limited environment, the method at zero training is the only arm left unbiased w.r.t. the IFD gradient (slope ~1, CI containing 1, vs learners 0.69-0.97, all under-concentrating) and ties the best learner's harvest where resource is clustered, while the learner collects more where resource is scarce (0.33 vs 0.22) and holds a tighter per-capita intake. Learning buys precision and transfer across reward structures, not a better allocation shape. A companion paper develops the mechanism.
Existing convergence analyses of finite-width overparameterized deep neural networks under standard parameterization often rely on parameters remaining close to their initialization and typically neglect the role of noise, despite empirical evidence that stochasticity can improve optimization and generalization. To bridge this gap, we study $L_2$-regularized continuous-time Langevin dynamics (CLD) with isotropic noise, for training feedforward networks with smooth activations. We introduce a change-of-measure inequality, inspired by the second law of thermodynamics, that enables control of key properties of the training dynamics under the time-evolving parameter distribution. Leveraging a sharpened lower bound on the minimum eigenvalue of the Neural Tangent Kernel (NTK) at initialization, we show that injected noise helps preserve NTK stability with high probability without confining the weights to a small neighborhood of their initialization. As a result, we derive a non-asymptotic upper bound on the expected training loss. Our analysis yields explicit conditions on width and noise temperature that guarantee convergence to a finite loss floor. Finally, we experimentally show the network Jacobian can change from initialization, departing the commonly analyzed lazy regime.
Online scene change detection requires an autonomous agent to detect what has changed in a previously mapped environment while it is still on site, observing from unconstrained viewpoints and without access to future observations. This causal setting is strictly harder than offline detection, and online methods have long trailed offline accuracy. Our earlier method, O-SCD, closed that gap at over 10 frames per second, but its multi-view fusion is an iterative optimization that refits a change field over a 3D Gaussian splatting map of the reference scene to all past observations at every frame. This step consumes over half of the per-frame budget and forces accuracy to be traded against speed. We present O-SCD++, a replay-free formulation that eliminates this trade-off. Because the reference map stays fixed, rendering the change field is a linear operation. Each new observation can therefore be carried onto the map's Gaussians by a single backward pass of the renderer and fused by a two-scalar recursion per Gaussian. Past observations are never stored or revisited. O-SCD++ also makes the change/no-change decision without labels or iterative optimization. It sets each frame's threshold to the exact minimizer of O-SCD's self-supervised objective over threshold masks, using one sort, and the distance to this threshold gives a per-pixel decision margin that predicts errors. On PASLCD, O-SCD++ improves change detection accuracy (mIoU) from $0.489$ to $0.575$ and throughput from $11.8$ to $25.6$ frames per second, exceeding O-SCD's offline result while remaining causal.
We investigate the use of Large Language Models (LLMs) in multi-author style change detection. Using off-the-shelf LLMs or supervised fine-tuning solely within the target domain fails to yield effective features; for this reason, we propose cross-domain transfer learning with LLM-based embeddings fine-tuned for authorship verification. We fine-tune Mistral-7B embedding model and compare it with DeBERTa, used in the state-of-the-art (SOTA) systems in PAN-25 multi-author writing style analysis challenge. Results show that DeBERTa and Mistral excel in different user-generated text scenarios. We exploit this diversity by combining both embeddings, leading to more effective style change detection.
Prominent AI safety discussions emphasize existential threats and catastrophic events, such as AI-enabled nuclear conflict or engineered biological outbreaks. This position paper argues that there is another type of AI risk that arises and accumulates from the widespread use of AI. Drawing on lessons from social media, we argue that safety assessments must also address accumulative harms that develop through repeated use, recurring decisions, and feedback across interconnected deployments. Repeated reassurance, selective exposure, recurring decision biases, and growing personal profiles may gradually narrow perspectives, weaken independent judgment, or enable exploitative influence without a single clearly harmful event. We advance the Trajectory Neglect Thesis: a focus on vivid catastrophic events and identifiable incidents risks overlooking less visible harms that accumulate through the widespread daily use of AI. Social media research motivates this position by showing how reciprocal media effects, social reinforcement, and individual susceptibility shape outcomes beyond individual encounters. We examine how related mechanisms may operate in AI systems, using conversational AI as a concrete setting where generation, semantic inference, and persistent memory may tighten feedback loops. Sycophancy, manipulation, and privacy loss provide a case study of how these risks may interact and accumulate. This comparison identifies mechanisms to investigate rather than assuming that repeated AI use is harmful. We therefore call for evaluation of cumulative patterns of use and deployment, individual and societal outcomes, careful causal inference, and proportionate safeguards that preserve useful assistance and informed user choice.