To address this gap, we conduct a Virtual Reality experiment in which pedestrians interact with automated shuttles under varying approach angles (45°, 90°, 135°) and continuous-traffic conditions (single shuttle, two shuttles with 3 or 5-second gaps), collecting synchronized motion, eye gaze, and head orientation data. To investigate to what extent, under what conditions, and in what form fine-grained eye gaze is informative for pedestrian motion prediction, we develop a multi-modal prediction model that fuses these signals through modality-specific encoders, and systematically ablate gaze representations against head orientation and situational context. We report three main results. First, the predictive value of eye gaze is angle-dependent and tightly coupled with eye-head-body coordination: at acute angles where pedestrians actively redirect gaze to acquire the shuttle, eye gaze carries information that head orientation alone misses. Second, continuous gaze orientation outperforms categorical semantic fixation labels, with the optimal encoding frame (global or body-relative) depending on whether gaze is used alone or jointly with context. Third, eye gaze and situational context provide complementary predictive information: their combination reduces final displacement error (FDE) by 8.47%, close to the sum of their individual contributions. Together, these findings highlight the value of incorporating human perceptual signals into pedestrian behavior prediction and motivate a human-centered complement to vehicle-centric modeling approaches. Our code is available at https://github.com/danyayay/GazeX.git.

Reinforcement learning has emerged as a powerful paradigm for improving large language model (LLM) reasoning, where rollouts are sampled from the policy and reward signals computed on those rollouts are used to update the policy. However, in data-scarce scenarios, obtaining ground-truth labels to verify rollouts at scale often requires expensive human annotation or labor-intensive expert verification. For instance, evaluating mathematical proofs demands expert review, and open-ended question answering lacks definitive ground truth. When ground-truth labels are scarce, the effectiveness of reinforcement learning fine-tuning is constrained. Inspired by the success of semi-supervised learning in propagating labels from labeled to unlabeled samples, we propose MemReward, a graph-based experience memory framework that integrates reward propagation directly into online policy optimization. MemReward stores rollouts (thinking processes and final answers) from an initial LLM policy as nodes in a heterogeneous graph connected by similarity and structural edges, over which a GNN propagates rewards from labeled to unlabeled rollouts. To train such a framework, we first warm up the GNN on labeled rollouts to predict rewards via heterogeneous aggregation over query, thinking, and answer nodes. During online RL fine-tuning, unlabeled rollouts are attached to the graph by query similarity, and the GNN predicts their rewards, yielding a hybrid reward acquisition strategy that combines ground-truth and GNN-predicted rewards. Experiments on Qwen2.5-1.5B and 3B in mathematics, question answering, and code generation demonstrate that MemReward, with ground-truth rewards on only 20% of rollouts, achieves 96.6% of Oracle performance on 1.5B and 97.3% on 3B, and closely approaches Oracle on out-of-domain tasks.

We evaluate Large Language Models (LLMs) in repeated game-theoretic settings to assess whether strategic performance reflects genuine reasoning or reliance on memorized patterns. We consider two canonical games, Prisoner's Dilemma (PD) and Rock-Paper-Scissors (RPS), upon which we introduce counterfactual variants that alter payoff structures and action labels, breaking familiar symmetries and dominance relations. Our multi-metric evaluation framework compares default and counterfactual instantiations, showcasing LLM limitations in incentive sensitivity, structural generalization and strategic reasoning within counterfactual environments.

This work presents a multi-label temporal event detection framework for video capsule endoscopy (VCE) that addresses the extreme class imbalance inherent in the Galar dataset by combining two principal contributions: an Angular Separation Loss on class prototypes and a Biological State Machine temporal decoder. The backbone remains BiomedCLIP, a biomedical vision-language foundation model. Three consecutive frames are fused through a Local Differencing Attention module that amplifies transient pathological signals by suppressing static temporal redundancy. An Anatomy Context Head then conditions pathological predictions on soft anatomical activations, exploiting the known spatial co-occurrence structure of GI findings. Learnable text-feature prompts and prototype-based logit augmentation are trained alongside an Angular Separation Loss that penalizes off-diagonal cosine similarity between class prototypes, preventing the prototype collapse that afflicts rare classes under extreme imbalance. To counteract the skewed label distribution, the training regime combines asymmetric focal loss, inverse-frequency weighted sampling, temporal Mixup, Exponential Moving Average, and per-class threshold calibration. The Biological State Machine decoder replaces naive gap merging with a physiologically grounded forward-only state transition over anatomy labels, eliminating the fragmentation artefact that produced hundreds of spurious anatomy events per video in the prior approach and reducing per-video anatomy output to 2--3 clinically realistic events. On the held-out RARE-VISION test set comprising three NaviCam examinations (161,025 frames), the updated pipeline achieves an overall temporal mAP@0.5 of 0.3597 and mAP@0.95 of 0.3399, representing a relative improvement of 46% and 44% respectively over the prior submission, with total inference completed in approximately 21 minutes on a single GPU.

Large Reasoning Models (LRMs) exhibit backtracking and self-verification mechanisms that enable them to revise intermediate steps and reach correct solutions, yielding strong performance on complex logical benchmarks. We hypothesize that such behaviors are beneficial only when the model has sufficiently strong ``critique'' ability to detect its own mistakes. This work systematically investigates how current LRMs recover from errors by inserting arithmetic mistakes in their intermediate reasoning steps. Notably, we discover a peculiar yet important phenomenon: despite the error propagating throughout the entire chain-of-thought (CoT) without any verbalized correction, the model still reaches the correct final answer after the thinking process finishes. This recovery implies the existence of an internal mechanism helping the model to detect errors and trigger self-correction, which we refer to as the \textit{hidden critique ability}. Building on feature space analysis, we identify a highly interpretable \textit{critique vector} representing this behavior. Extensive experiments across multiple model scales and families demonstrate that steering latent representations with this vector improves the model's error detection capability and enhances the performance of test-time scaling at no extra training cost. Our findings provide a valuable understanding of LRMs' critique behavior, suggesting a promising direction to control and improve their self-verification mechanism. Our code is available at: https://github.com/mail-research/lrm-critique-vectors.

Political speakers often avoid answering questions directly while maintaining the appearance of responsiveness. Despite its importance for public discourse, such strategic evasion remains underexplored in Natural Language Processing. We introduce SemEval-2026 Task 6, CLARITY, a shared task on political question evasion consisting of two subtasks: (i) clarity-level classification into Clear Reply, Ambivalent, and Clear Non-Reply, and (ii) evasion-level classification into nine fine-grained evasion strategies. The benchmark is constructed from U.S. presidential interviews and follows an expert-grounded taxonomy of response clarity and evasion. The task attracted 124 registered teams, who submitted 946 valid runs for clarity-level classification and 539 for evasion-level classification. Results show a substantial gap in difficulty between the two subtasks: the best system achieved 0.89 macro-F1 on clarity classification, surpassing the strongest baseline by a large margin, while the top evasion-level system reached 0.68 macro-F1, matching the best baseline. Overall, large language model prompting and hierarchical exploitation of the taxonomy emerged as the most effective strategies, with top systems consistently outperforming those that treated the two subtasks independently. CLARITY establishes political response evasion as a challenging benchmark for computational discourse analysis and highlights the difficulty of modeling strategic ambiguity in political language.

Autonomous vehicles rely on map information to understand the world around them. However, the creation and maintenance of offline high-definition (HD) maps remains costly. A more scalable alternative lies in online HD map construction, which only requires map annotations at training time. To further reduce the need for annotating vast training labels, self-supervised training provides an alternative. This work focuses on improving the latent birds-eye-view (BEV) feature grid representation within a vectorized online HD map construction model by enforcing geospatial consistency between overlapping BEV feature grids as part of a contrastive loss function. To ensure geospatial overlap for contrastive pairs, we introduce an approach to analyze the overlap between traversals within a given dataset and generate subsidiary dataset splits following adjustable multi-traversal requirements. We train the same model supervised using a reduced set of single-traversal labeled data and self-supervised on a broader unlabeled set of data following our multi-traversal requirements, effectively implementing a semi-supervised approach. Our approach outperforms the supervised baseline across the board, both quantitatively in terms of the downstream tasks vectorized map perception performance and qualitatively in terms of segmentation in the principal component analysis (PCA) visualization of the BEV feature space.

Muon has recently shown promising results in LLM training. In this work, we study how to further improve Muon. We argue that Muon's orthogonalized update rule suppresses the emergence of heavy-tailed weight spectra and over-emphasizes the training along noise-dominated directions. Motivated by the Heavy-Tailed Self-Regularization (HT-SR) theory, we propose HTMuon. HTMuon preserves Muon's ability to capture parameter interdependencies while producing heavier-tailed updates and inducing heavier-tailed weight spectra. Experiments on LLM pretraining and image classification show that HTMuon consistently improves performance over state-of-the-art baselines and can also serve as a plug-in on top of existing Muon variants. For example, on LLaMA pretraining on the C4 dataset, HTMuon reduces perplexity by up to $0.98$ compared to Muon. We further theoretically show that HTMuon corresponds to steepest descent under the Schatten-$q$ norm constraint and provide convergence analysis in smooth non-convex settings. The implementation of HTMuon is available at https://github.com/TDCSZ327/HTmuon.

Large language model (LLM) post-training enhances latent skills, unlocks value alignment, improves performance, and enables domain adaptation. Unfortunately, post-training is known to induce forgetting, especially in the ubiquitous use-case of leveraging third-party pre-trained models, which is typically understood as a loss of parametric or factual knowledge. We argue that this accuracy-centric view is insufficient for modern foundation models and instead define forgetting as systematic model drift that degrades behavior and user experience. In this context, we introduce CapTrack, a capability-centric framework for analyzing forgetting in LLMs that combines a behavioral taxonomy with an evaluation suite centered on capability-specific metrics. Using CapTrack, we conduct a large-scale empirical study across post-training algorithms, domains, and model families, including models up to 80B parameters. We find that forgetting extends beyond parametric knowledge, with pronounced drift in robustness and default behaviors. Instruction fine-tuning induces the strongest relative drift, while preference optimization is more conservative and can partially recover lost capabilities. Differences across model families persist, and no universal mitigation emerges.

Recent advances in generative image compression (GIC) have delivered remarkable improvements in perceptual quality. However, many GICs rely on large-scale and rigid models, which severely constrain their utility for flexible transmission and practical deployment in low-bitrate scenarios. To address these issues, we propose Progressive Generative Image Compression (ProGIC), a compact codec built on residual vector quantization (RVQ). In RVQ, a sequence of vector quantizers encodes the residuals stage by stage, each with its own codebook. The resulting codewords sum to a coarse-to-fine reconstruction and a progressive bitstream, enabling previews from partial data. We pair this with a lightweight backbone based on depthwise-separable convolutions and small attention blocks, enabling practical deployment on both GPUs and CPU-only devices. Experimental results show that ProGIC attains comparable compression performance compared with previous methods. It achieves bitrate savings of up to 57.57% on DISTS and 58.83% on LPIPS compared to MS-ILLM on the Kodak dataset. Beyond perceptual quality, ProGIC enables progressive transmission for flexibility, and also delivers over 10 times faster encoding and decoding compared with MS-ILLM on GPUs for efficiency.

As artificial intelligence systems move toward clinical deployment, ensuring reliable prediction behavior is fundamental for safety-critical decision-making tasks. One proposed safeguard is selective prediction, where models can defer uncertain predictions to human experts for review. In this work, we empirically evaluate the reliability of uncertainty-based selective prediction in multilabel clinical condition classification using multimodal ICU data. Across a range of state-of-the-art unimodal and multimodal models, we find that selective prediction can substantially degrade performance despite strong standard evaluation metrics. This failure is driven by severe class-dependent miscalibration, whereby models assign high uncertainty to correct predictions and low uncertainty to incorrect ones, particularly for underrepresented clinical conditions. Our results show that commonly used aggregate metrics can obscure these effects, limiting their ability to assess selective prediction behavior in this setting. Taken together, our findings characterize a task-specific failure mode of selective prediction in multimodal clinical condition classification and highlight the need for calibration-aware evaluation to provide strong guarantees of safety and robustness in clinical AI.

Ray tracing has become a standard for accurate radio propagation modeling, but suffers from exponential computational complexity, as the number of candidate paths scales with the number of objects raised to the interaction order. This bottleneck limits its use in large-scale or real-time applications, forcing traditional tools to rely on heuristics that reduce path candidates at the cost of potentially reduced accuracy. To overcome this limitation, we propose a machine-learning-assisted framework that replaces exhaustive path searching with intelligent sampling via Generative Flow Networks. Applying these generative models to this domain presents challenges, particularly sparse rewards due to the rarity of valid paths, which can lead to convergence failures and trivial solutions when evaluating high-order interactions in complex environments. To ensure robust learning and efficient exploration, our framework incorporates three key components. First, an \emph{experience replay buffer} captures and retains rare valid paths. Second, a uniform exploratory policy improves generalization and prevents overfitting to simple geometries. Third, a physics-based action masking strategy filters out physically impossible paths before the model considers them. Validated on idealized street-canyon scenarios, our model achieves substantial speedups over exhaustive search -- up to $10\times$ faster on GPU and $100\times$ faster on CPU -- while maintaining high coverage accuracy and successfully uncovering complex propagation paths. However, out-of-distribution evaluations on real-world Manhattan street geometries reveal that generalizing to substantially different urban morphologies requires further advancement in model capacity or alternative training strategies. Source code, tests, and a tutorial are available at https://github.com/jeertmans/sampling-paths.

Natural language processing for the Turkic language family, spoken by over 200 million people across Eurasia, remains fragmented, with most languages lacking unified tooling and resources. We present TurkicNLP, an open-source Python library providing a single, consistent NLP pipeline for Turkic languages across four script families: Latin, Cyrillic, Perso-Arabic, and Old Turkic Runic. The library covers tokenization, morphological analysis, part-of-speech tagging, dependency parsing, named entity recognition, bidirectional script transliteration, cross-lingual sentence embeddings, and machine translation through one language-agnostic API. A modular multi-backend architecture integrates rule-based finite-state transducers and neural models transparently, with automatic script detection and routing between script variants. Outputs follow the CoNLL-U standard for full interoperability and extension. Code and documentation are hosted at https://github.com/turkic-nlp/turkicnlp .

Masked Diffusion Models (MDMs) enable flexible decoding orders, yet existing samplers remain largely greedy, selecting locally certain tokens without accounting for their downstream effects. We show that this myopia can increase cumulative uncertainty and lead to suboptimal generation. To address this, we propose the **Info-Gain Sampler**, a training-free decoding method that uses the bidirectional structure of MDMs to balance immediate uncertainty with the information gained over remaining masked positions. Across reasoning, coding, creative writing, and image generation tasks, Info-Gain Sampler consistently outperforms existing MDM samplers, improving average reasoning accuracy by 2.9--11.6 percentage points and achieving a 62.8% average win rate in creative writing. The code is available at https://github.com/yks23/Information-Gain-Sampler.

Recent work has shown that language models (LMs) trained on synthetic corpora can exhibit typological preferences that resemble cross-linguistic regularities in human languages, particularly for syntactic phenomena such as word order. In this paper, we extend this paradigm to differential argument marking (DAM), a semantic licensing system in which morphological marking depends on semantic prominence. Using a controlled synthetic learning method, we train GPT-2 models on 18 corpora implementing distinct DAM systems and evaluate their generalization using minimal pairs. Our results reveal a dissociation between two typological dimensions of DAM. Models reliably exhibit human-like preferences for natural markedness direction, favoring systems in which overt marking targets semantically atypical arguments. In contrast, models do not reproduce the strong object preference in human languages, in which overt marking in DAM more often targets objects rather than subjects. These findings suggest that different typological tendencies may arise from distinct underlying sources.

Remote Operation is touted as being key to the rapid deployment of automated vehicles. Streaming imagery to control connected vehicles remotely currently requires a reliable, high throughput network connection, which can be limited in real-world remote operation deployments relying on public network infrastructure. This paper investigates how the application of computer vision assisted semantic communication can be used to circumvent data loss and corruption associated with traditional image compression techniques. By encoding the segmentations of detected road users into colour coded highlights within low resolution greyscale imagery, the required data rate can be reduced by 50% compared with conventional techniques, while maintaining visual clarity. This enables a median glass-to-glass latency of below 200 ms even when the network data rate is below 500 kbit/s, while clearly outlining salient road users to enhance situational awareness of the remote operator. The approach is demonstrated in an area of variable 4G mobile connectivity using an automated last-mile delivery vehicle. Results indicate that large-scale deployment of remotely operated automated vehicles could be possible even on the often constrained public 4G/5G mobile network, providing the potential to expedite the nationwide roll-out of automated vehicles.

Although foundation models have demonstrated remarkable success in general domains, the application of these models to electroencephalography (EEG) analysis is constrained by substantial data requirements and high parameterization. These factors incur prohibitive computational costs, thereby impeding deployment in resource-constrained clinical environments. Conversely, general-purpose automated machine learning frameworks are often ill-suited for this domain, as exploration within an unbounded programmatic space fails to incorporate essential neurophysiological priors and frequently yields solutions that lack scientific plausibility. To address these limitations, we propose NeuroWeaver, a unified autonomous evolutionary agent designed to generalize across diverse EEG datasets and tasks by reformulating pipeline engineering as a discrete constrained optimization problem. Specifically, we employ a Domain-Informed Subspace Initialization to confine the search to neuroscientifically plausible manifolds, coupled with a Multi-Objective Evolutionary Optimization that dynamically balances performance, novelty, and efficiency via self-reflective refinement. Empirical evaluations across five heterogeneous benchmarks demonstrate that NeuroWeaver synthesizes lightweight solutions that consistently outperform state-of-the-art task-specific methods and achieve performance comparable to large-scale foundation models, despite utilizing significantly fewer parameters.

Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as a dominant paradigm for enhancing Large Language Models (LLMs) reasoning, yet its reliance on external verifiers limits its scalability. Recent findings suggest that RLVR primarily functions by eliciting latent capabilities, motivating the development of verifier-free algorithms. However, in such settings, standard methods like Group Relative Policy Optimization face a critical challenge: destructive gradient variance that often leads to training collapse. To address this issue, we introduce Verifier-Independent Curriculum Reinforcement Learning (VI-CuRL), a framework that leverages the model's intrinsic confidence to construct a curriculum independent from external verifiers. By prioritizing high-confidence samples, VI-CuRL effectively manages the bias-variance trade-off, specifically targeting the reduction of action and problem variance. We provide a rigorous theoretical analysis, proving that our estimator guarantees asymptotic unbiasedness. Empirically, VI-CuRL promotes stability and consistently outperforms verifier-dependent/independent baselines across math and general reasoning benchmarks with/without verifiers.

Frontier AI systems are increasingly capable and deployed in high-stakes multi-agent environments. However, existing AI safety benchmarks largely evaluate single agents, leaving multi-agent risks such as coordination failure and conflict poorly understood. We introduce GT-HarmBench, a benchmark of 1,535 high-stakes scenarios spanning game-theoretic structures such as the Prisoner's Dilemma, Stag Hunt and Chicken. Scenarios are drawn from realistic AI risk contexts in the MIT AI Risk Repository. Across 15 frontier models, agents fail to choose socially beneficial actions in 38% of high-stakes cases, such as military escalation, election manipulation, and medical malpractice. We measure sensitivity to game-theoretic prompt framing and ordering, and analyze reasoning patterns driving failures. We further show that game-theoretic interventions improve socially beneficial outcomes by up to 18%. Our results highlight substantial reliability gaps and provide a broad standardized testbed for studying alignment in multi-agent environments. The benchmark and code are available at https://github.com/causalNLP/gt-harmbench.

Graph Prompt Learning (GPL) has recently emerged as a promising paradigm for downstream adaptation of pre-trained graph models, mitigating the misalignment between pre-training objectives and downstream tasks. Recently, the focus of GPL has shifted from in-domain to cross-domain scenarios, which is closer to the real world applications, where the pre-training source and downstream target often differ substantially in data distribution. However, why GPLs remain effective under such domain shifts is still unexplored. Empirically, we observe that representative GPL methods are competitive with two simple baselines in cross-domain settings: full fine-tuning (FT) and linear probing (LP), motivating us to explore a deeper understanding of the prompting mechanism. We provide a theoretical analysis demonstrating that jointly leveraging these two complementary branches yields a smaller estimation error than using either branch alone, formally proving that cross-domain GPL benefits from the integration between pre-trained knowledge and task-specific adaptation. Based on this insight, we propose GP2F, a dual-branch GPL method that explicitly instantiates the two extremes: (1) a frozen branch that retains pre-trained knowledge, and (2) an adapted branch with lightweight adapters for task-specific adaptation. We then perform adaptive fusion under topology constraints via a contrastive loss and a topology-consistent loss. Extensive experiments on cross-domain few-shot node and graph classification demonstrate that our method outperforms existing methods.

Modern LLM-based agents and chat assistants rely on long-term memory frameworks to store reusable knowledge, recall user preferences, and augment reasoning. As researchers create more complex memory architectures, it becomes increasingly difficult to analyze their capabilities and guide future memory designs. Most long-term memory benchmarks focus on simple fact retention, multi-hop recall, and time-based changes. While undoubtedly important, these capabilities can often be achieved with simple retrieval-augmented LLMs and do not test complex memory hierarchies. To bridge this gap, we propose StructMemEval - a benchmark that tests the agent's ability to organize its long-term memory, not just factual recall. We gather a suite of tasks that humans solve by organizing their knowledge in a specific structure: transaction ledgers, to-do lists, trees and others. Our initial experiments show that simple retrieval-augmented LLMs struggle with these tasks, whereas memory agents can reliably solve them if prompted how to organize their memory. However, we also find that modern LLMs do not always recognize the memory structure when not prompted to do so. This highlights an important direction for future improvements in both LLM training and memory frameworks.

Diffusion large language models (dLLMs) have recently gained significant attention due to their inherent support for parallel decoding. Building on this paradigm, Mixture-of-Experts (MoE) dLLMs with autoregressive (AR) initialization have further demonstrated strong performance competitive with mainstream AR models. However, we identify a fundamental mismatch between MoE architectures and diffusion-based decoding. Specifically, a large number of experts are activated at each denoising step, while only a small subset of tokens is ultimately accepted, resulting in substantial inference overhead and limiting their deployment in latency-sensitive applications. In this work, we propose TEAM, a plug-and-play framework that accelerates MoE dLLMs by enabling more accepted tokens with fewer activated experts. TEAM is motivated by the observation that expert routing decisions exhibit strong temporal consistency across denoising levels as well as spatial consistency across token positions. Leveraging these properties, TEAM employs three complementary expert activation and decoding strategies, conservatively selecting necessary experts for decoded and masked tokens and simultaneously performing aggressive speculative exploration across multiple candidates. Experimental results demonstrate that TEAM achieves up to 2.2x speedup over vanilla MoE dLLM, with negligible performance degradation. Code is released at https://github.com/PKU-SEC-Lab/TEAM-MoE-dLLM.

In long-video understanding, conventional uniform frame sampling often fails to capture key visual evidence, leading to degraded performance and increased hallucinations. To address this, recent agentic thinking-with-videos paradigms have emerged, adopting a localize-clip-answer pipeline in which the model actively identifies relevant video segments, performs dense sampling within those clips, and then produces answers. However, existing methods remain inefficient, suffer from weak localization, and adhere to rigid workflows. To solve these issues, we propose VideoTemp-o3, a unified agentic thinking-with-videos framework that jointly models video grounding and question answering. VideoTemp-o3 exhibits strong localization capability, supports on-demand clipping, and can refine inaccurate localizations. Specifically, in the supervised fine-tuning stage, we design a unified masking mechanism that encourages exploration while preventing noise. For reinforcement learning, we introduce dedicated rewards to mitigate reward hacking. Besides, from the data perspective, we develop an effective pipeline to construct high-quality long video grounded QA data, along with a corresponding benchmark for systematic evaluation across various video durations. Experimental results demonstrate that our method achieves remarkable performance on both long video understanding and grounding.

Predictive coding (PC) is a biologically plausible alternative to standard backpropagation (BP) that minimises an energy function with respect to network activities before updating weights. Recent work has improved the training stability of deep PC networks (PCNs) by leveraging some BP-inspired reparameterisations. However, the full scalability and theoretical basis of these methods remain unclear. To address this gap, we study the infinite width and depth limits of PCNs. For linear residual networks, we show that the set of width- and depth-stable feature-learning parameterisations for PC is exactly the same as for BP. Moreover, under any of these parameterisations, the PC energy with equilibrated activities converges to the quadratic BP loss when the model width is much larger than the depth, resulting in PC computing the same gradients as BP. Experiments show that, as long as an activity equilibrium is reached, convergence to BP holds for nonlinear models including convolutional networks and transformers. Overall, this work constrains the types of parameterisation that are scalable with PC, while showing a way in which BP can be effectively implemented with only local updates in much wider than deep networks like the brain.

Vision--Language--Action (VLA) models bridge multimodal reasoning with physical control, but adapting them to new tasks with scarce demonstrations remains unreliable. While fine-tuned VLA policies often produce semantically plausible trajectories, failures often arise from unresolved geometric ambiguities, where near-miss actions lead to divergent execution outcomes under limited supervision. We study few-shot VLA adaptation from a \emph{generation--selection} perspective and propose a novel framework \textbf{VGAS} (\textbf{V}alue-\textbf{G}uided \textbf{A}ction-chunk \textbf{S}election). It performs inference-time best-of-$N$ selection to identify action chunks that are both semantically faithful and geometrically precise. Specifically, \textbf{VGAS} employs a finetuned VLA as a high-recall proposal generator and introduces the \textrm{Q-Chunk-Former}, a geometrically grounded Transformer critic to resolve fine-grained geometric ambiguities. In addition, we propose \textit{Explicit Geometric Regularization} (\texttt{EGR}), which shapes a discriminative value landscape to preserve action ranking resolution among near-miss candidates while mitigating value instability under scarce supervision. Experiments and theoretical analysis demonstrate that \textbf{VGAS} consistently improves success rates and robustness under limited demonstrations and distribution shifts. Our code is available at https://github.com/Jyugo-15/VGAS.

Watermarking is an important tool for promoting the responsible use of large language models (LLMs). Existing watermarks insert a signal into generated tokens that either flags LLM-generated text (zero-bit watermarking) or encodes more complex messages (multi-bit watermarking). Though a number of recent approaches insert multiple bits into text without perturbing average next-token predictions, they largely extend design principles from the zero-bit setting, such as encoding a single bit per token. In contrast, a watermarker capable of embedding multiple bytes into the text would dramatically increase the potential applications, by embedding information such as the ID of the user who submitted the prompt, the precise model version that was used, or even the prompt itself. We address this problem by introducing ArcMark: a new watermark construction based on coding and information-theoretic principles that is capable of reliably embedding multiple bytes of information into just a few hundred tokens, without any distortion of the underlying LLM next-token distribution. We derive ArcMark by formulating the distortion-free watermarking problem as a channel coding problem, and deriving an information-theoretic channel capacity that establishes the fundamental limit of embedding information in LLM output in a distortion-free manner. This capacity formulation informs the design of ArcMark. In practice, ArcMark outperforms competing multi-bit distortion-free watermarks in terms of reconstruction accuracy, including in the face of attacks that alter a subset of the LLM text. ArcMark output is also shown to be indistinguishable from unwatermarked text in terms of perplexity, and in downstream task quality.

The quest for expert-level reasoning in Large Language Models (LLMs) has been hampered by a persistent \textit{reward bottleneck}: traditional reinforcement learning (RL) relies on scalar rewards that are \textbf{costly} to scale, \textbf{brittle} across domains, and \textbf{blind} to the underlying logic of a solution. This reliance on external, impoverished signals prevents models from developing a deep, self-contained understanding of reasoning principles. We introduce \textbf{ALIVE} (\emph{Adversarial Learning with Instructive Verbal Evaluation}), a hands-free alignment framework that moves beyond scalar reward optimization toward intrinsic reasoning acquisition. Grounded in the principle of \emph{Cognitive Synergy}, ALIVE unifies problem posing, solving, and judging within a single policy model to internalize the logic of correctness. By coupling adversarial learning with instructive verbal feedback, ALIVE enables models to internalize evaluative criteria directly from raw corpora, effectively transforming external critiques into an endogenous reasoning faculty. Empirical evaluations across mathematical reasoning, code generation, and general logical inference benchmarks demonstrate that ALIVE consistently mitigates reward signal limitations. With identical data and compute, it achieves accuracy gains, markedly improved cross-domain generalization, and higher self-correction rates. These results indicate that the reasoning trinity fosters a self-sustaining trajectory of capability growth, positioning ALIVE as a scalable foundation for general-purpose reasoning alignment without human-in-the-loop supervision.

Aligning video generative models with human preferences remains challenging: current approaches rely on Vision-Language Models (VLMs) for reward modeling, but these models struggle to capture subtle temporal dynamics. We propose a fundamentally different approach: repurposing video generative models, which are inherently designed to model temporal structure, as reward models. We present the Generative-Transformer-based Self-Supervised Video Judge (\modelname), a novel evaluation model that transforms state-of-the-art video generation models into powerful temporally-aware reward models. Our key insight is that generative models can be reformulated as energy-based models (EBMs) that assign low energy to high-quality videos and high energy to degraded ones, enabling them to discriminate video quality with remarkable precision when trained via contrastive objectives. To prevent the model from exploiting superficial differences between real and generated videos, we design challenging synthetic negative videos through controlled latent-space perturbations: temporal slicing, feature swapping, and frame shuffling, which simulate realistic but subtle visual degradations. This forces the model to learn meaningful spatiotemporal features rather than trivial artifacts. \modelname achieves state-of-the-art performance on GenAI-Bench and MonteBench using only 30K human-annotations: $6\times$ to $65\times$ fewer than existing VLM-based approaches.

Human papillomavirus (HPV) status is a critical determinant of prognosis and treatment response in head and neck and cervical cancers. Although attention-based multiple instance learning (MIL) achieves strong slide-level prediction for HPV-related whole-slide histopathology, it provides limited morphologic interpretability. To address this limitation, we introduce Concept-Level Explainable Attention-guided Representation for HPV (CLEAR-HPV), a framework that restructures the MIL latent space using attention to enable concept discovery without requiring concept labels during training. Operating in an attention-weighted latent space, CLEAR-HPV automatically discovers keratinizing, basaloid, and stromal morphologic concepts, generates spatial concept maps, and represents each slide using a compact concept-fraction vector. CLEAR-HPV's concept-fraction vectors preserve the predictive information of the original MIL embeddings while reducing the high-dimensional feature space (e.g., 1536 dimensions) to only 10 interpretable concepts. CLEAR-HPV generalizes consistently across TCGA-HNSCC, TCGA-CESC, and CPTAC-HNSCC, providing compact, concept-level interpretability through a general, backbone-agnostic framework for attention-based MIL models of whole-slide histopathology.

Large language models (LLMs) are enabling reasoning over 2D and 3D structures, yet existing methods remain modality-specific and typically compress structural inputs through sequence-based tokenization or fixed-length query connectors. Such architectures either omit the geometric grounding requisite for mitigating structural hallucinations, or impose inflexible modality fusion bottlenecks that concurrently over-compress and suboptimally allocate structural tokens, thereby impeding the realization of generalized all-atom reasoning. We introduce Cuttlefish, a unified multimodal LLM that grounds language reasoning in geometric cues while scaling modality tokens with structural complexity. First, Scaling-Aware Patching leverages an instruction-conditioned gating mechanism to generate variable-size patches over structural graphs, adaptively scaling the query token budget with structural complexity to mitigate fixed-length connector bottlenecks. Second, Geometry Grounding Adapter refines these adaptive tokens via cross-attention to modality embeddings and injects the resulting modality tokens into the LLM, exposing explicit geometric cues to reduce structural hallucination. Experiments across interdisciplinary all-atom benchmarks demonstrate that Cuttlefish achieves superior performance in heterogeneous structure-grounded reasoning. Code: github.com/zihao-jing/Cuttlefish.