The POLAR SemEval-2026 Shared Task aims to detect online polarization and focuses on the classification and identification of multilingual, multicultural, and multi-event polarization. Accurate computational detection of online polarization is challenging due to nuanced rhetoric, implicit framing, and the high cost of human-in-the-loop annotation. Building on recent findings that contextual prompting enables large language models to function as strong polarization detectors, we present a two-stage approach for detecting polarization in social media text that combines structured supervised fine tuning with Direct Preference Optimization (DPO) refinement. We fine tune Qwen 2.5-7B-Instruct with LoRA using an interpretable slot-filling template (target, claim type, manifestation checklist, and justification). We then apply DPO with automatically generated preference pairs to reduce costly false negatives. Our submitted system achieves 0.7664 Macro-F1 on the English test set. Post-submission experiments with Mistral-Nemo-Instruct-2407 and LLM-judge-filtered preference pairs further improve to 0.8162 Macro-F1 (not submitted to CodaBench), surpassing the organiser baseline of 0.7802.

Activation steering is a popular white-box control technique that modifies model activations to elicit an abstract change in its behavior. It has also become a standard tool in interpretability (e.g., probing truthfulness, or translating activations into human-readable explanations) and safety research (e.g., jailbreakability). However, it is unclear whether steered behavior is realizable by any textual prompt. In this work, we cast this question as a surjectivity problem: for a fixed model, does every steered activation admit a preimage under the model's natural forward pass? Under practical assumptions, we prove that activation steering pushes the residual stream off the manifold of states reachable from discrete prompts. Almost surely, no prompt can reproduce the same internal behavior induced by steering. We also illustrate this finding empirically across three widely used LLMs. Our results establish a formal separation between white-box steerability and black-box prompting. We therefore caution against interpreting the ease and success of activation steering as evidence of prompt-based interpretability or vulnerability, and argue for evaluation protocols that explicitly decouple white-box and black-box interventions.

Dimensional Aspect-Based Sentiment Analysis (DimABSA) extends traditional ABSA from categorical polarity labels to continuous valence-arousal (VA) regression. This paper describes a system developed for Track A, Subtask 1 (Dimensional Aspect Sentiment Regression), aiming to predict real-valued VA scores in the [1, 9] range for each given aspect in a text. A fine-tuning approach based on XLM-RoBERTa-base is adopted, with dual regression heads with sigmoid-scaled outputs for valence and arousal prediction. Separate models are trained for each language-domain pair (English and Chinese across restaurant, laptop, and finance domains), and training and development sets are merged for final test predictions. In development experiments, the fine-tuning approach is compared against several large language models under a few-shot prompting setting, demonstrating that task-specific fine-tuning outperforms these LLM-based methods across all evaluation datasets.

Processing long-form videos with Video Large Language Models (Video-LLMs) is computationally prohibitive. Current efficiency methods often compromise fine-grained perception through irreversible information disposal or inhibit long-range temporal modeling via rigid, predefined sparse patterns. This paper introduces AdaSpark, an adaptive sparsity framework designed to address these limitations. AdaSpark first partitions video inputs into 3D spatio-temporal cubes. It then employs two co-designed, context-aware components: (1) Adaptive Cube-Selective Attention (AdaS-Attn), which adaptively selects a subset of relevant video cubes to attend for each query token, and (2) Adaptive Token-Selective FFN (AdaS-FFN), which selectively processes only the most salient tokens within each cube. An entropy-based (Top-p) selection mechanism adaptively allocates computational resources based on input complexity. Experiments demonstrate that AdaSpark significantly reduces computational load by up to 57% FLOPs while maintaining comparable performance to dense models and preserving fine-grained, long-range dependencies, as validated on challenging hour-scale video benchmarks.

Online reinforcement learning (RL) serves as an effective method for enhancing the capabilities of Android agents. However, guiding agents to learn through online interaction is prohibitively expensive due to the high latency of emulators and the sample inefficiency of existing RL algorithms. We identify a fundamental limitation in current approaches: the Single State Single Action paradigm, which updates the policy with one-to-one state-action pairs from online one-way rollouts without fully exploring each costly emulator state. In this paper, we propose Android Coach, a novel framework that shifts the training paradigm to Single State Multiple Actions, allowing the agent to sample and utilize multiple actions for a single online state. We enable this without additional emulator overhead by learning a critic that estimates action values. To ensure the critic serves as a reliable coach, we integrate a process reward model and introduce a group-wise advantage estimator based on the averaged critic outputs. Extensive experiments demonstrate the effectiveness and efficiency of Android Coach: it achieves 7.5% and 8.3% success rate improvements on AndroidLab and AndroidWorld over UI-TARS-1.5-7B, and attains 1.4x higher training efficiency than Single State Single Action methods PPO and GRPO at matched success rates.

We show that robustness to post-training quantization (PTQ) is a transferable direction in weight space. We call this direction the quantization vector: extracted from a donor task by simple weight-space arithmetic, it can be used to patch a receiver model and improve post-PTQ Top-1 accuracy by up to 60 points in a 3-bit setting, without receiver-side quantization-aware training (QAT). Because the method requires no receiver training data, it provides a zero-shot, low-cost alternative to QAT for extremely low-bit deployment. Across four ViT scales and 22 image classification tasks, donor quantization vectors often yield substantial gains even when donor and receiver tasks differ markedly. We further prove rigorously that quantization vectors are well-defined and do not suffer from reparameterization symmetries, and provide a local geometric account of their effect. Together, these results suggest that quantization robustness can be partially isolated, reused, and transferred through simple weight-space algebra.

Hadamard transforms have become a key tool for stabilizing low-precision training, but existing methods apply them uniformly across tensors and computation paths. We show that this one-size-fits-all strategy is inherently limited: Hadamard smoothing reduces quantization error only when its direction is properly aligned with the operand's outlier structure. Through a systematic study of weights, activations, and gradients in LLM training, we identify three stable outlier patterns, Row-wise, Column-wise, and None, and show that each outlier pattern pair in matrix multiplication requires a distinct transform or outlier-handling strategy. We propose AdaHOP, Adaptive Hadamard transform with Outlier-Pattern-aware strategy, which applies Inner Hadamard Transform (IHT) when inner-dimension mixing properly suppresses the operands' outliers, and selectively applies Outlier Extraction (OE) that extracts dominant outlier rows or columns into a high-precision path when it does not. With fused, hardware-aware Triton kernels, AdaHOP enables training from scratch at MXFP4 precision with BF16-level quality, while achieving up to 3.6X memory compression, 1.46X end-to-end training speedup over BF16.

Spectral graph contrastive learning often constructs low- and high-frequency views to capture complementary graph signals, but these views are commonly combined by graph-level or node-agnostic fusion rules. We show that graph-level fusion can incur irreducible regret on mixed graphs with separated node-wise spectral preferences. Motivated by this result, we propose ASPECT, a spectral graph contrastive learning method that adaptively fuses low- and high-frequency views at the node level. ASPECT learns a node-wise spectral policy and regularizes it using channel-wise contrastive evidence, enabling different nodes to use different spectral mixtures. We further introduce ASPECT-S, an optional stability-aware extension that uses generated graph-structure and feature perturbations to obtain empirical channel-wise sensitivity estimates, together with a Rayleigh-based spectral search bias for producing informative perturbations. Experiments on homophilic and heterophilic benchmarks show that ASPECT improves representation quality over competitive spectral and graph contrastive baselines, while ASPECT-S further improves performance under joint graph-structure and feature perturbations.

Morphological traits are physical characteristics of biological organisms that provide vital clues on how organisms interact with their environment. Yet extracting these traits remains a slow, expert-driven process, limiting their use in large-scale ecological studies. A major bottleneck is the absence of high-quality datasets linking biological images to trait-level annotations. In this work, we demonstrate that sparse autoencoders trained on foundation-model features yield monosemantic, spatially grounded neurons that consistently activate on meaningful morphological parts. Leveraging this property, we introduce a trait annotation pipeline that localizes salient regions and uses vision-language prompting to generate interpretable trait descriptions. Using this approach, we construct Bioscan-Traits, a dataset of 80K trait annotations spanning 19K insect images from BIOSCAN-5M. Human evaluation confirms the biological plausibility of the generated morphological descriptions. We assess design sensitivity through a comprehensive ablation study, systematically varying key design choices and measuring their impact on the quality of the resulting trait descriptions. By annotating traits with a modular pipeline rather than prohibitively expensive manual efforts, we offer a scalable way to inject biologically meaningful supervision into foundation models, enable large-scale morphological analyses, and bridge the gap between ecological relevance and machine-learning practicality.

The global Lipschitz constant of a neural network is related to robustness and generalization, yet unlike in many classical models, it is not plainly legible from the parameters. This has motivated sophisticated verification algorithms, especially semidefinite programming (SDP) based on incremental quadratic constraints on the activation functions, to improve on the fast but often loose product of layerwise Lipschitz constants (the trivial bound). We ask why Lipschitz verification is a problem in the first place. Our answer is that the difficulty is structural: estimating a network's Lipschitz constant requires knowing which hidden states are reachable, and reachability is NP-hard. If P!=NP, then reachability is a barrier to any polynomial-time algorithm. Through explicit constructions, we show that this blindness can force SDP-based bounds to inherit the same qualitative failures as the trivial bound, including but not limited to polynomial per-layer conservatism. We show that the difficulties of NP-hard questions are not isolated to worst-case computational reductions, but actually afflict every instance of the verification problem. Thus SDP is not sufficient for Lipschitz verification. We also argue that it is not necessary: several apparent failures of the trivial bound arise from removable parameterization pathologies, and can be mitigated by optimizing or regularizing the trivial bound itself. We demonstrate this claim via a "spherical cow" linear model and numerical proofs of concept. While the main contribution is theoretical and negative, we finally motivate a novel form of trigonometric layers that do not need biases for universal approximation. Combined with trivial bound regularization, they make the trivial bound provably and practically tight.

Steel surface defect analysis is critical for industrial quality control, yet existing benchmarks rely primarily on label-only annotations, limiting fine-grained semantic understanding and systematic evaluation of vision-language models. To address this gap, we introduce SteelDefectX, a vision-language dataset with multi-form textual annotations for steel surface defect analysis, comprising 7,778 images across 25 defect categories. At the class level, the dataset provides defect names, representative visual attributes, and industrial causes. At the sample level, each image is annotated with three forms of textual representations: (1) free-form natural language descriptions, (2) structured attribute annotations, and (3) template-based sentences. These annotations provide flexible textual supervision with varying levels of expressiveness and controllability. We further establish a comprehensive benchmark covering vision-language classification, segmentation, and cross-dataset transfer, along with additional evaluations such as retrieval and text-guided localization. Experimental results reveal a trade-off between structure and flexibility in textual representations. Structured attributes provide more stable semantic alignment, while natural language descriptions improve transferability and fine-grained spatial grounding. These findings highlight the critical role of textual design in industrial vision-language learning. SteelDefectX provides a new benchmark for studying semantic alignment and generalization in industrial vision-language learning. The code and dataset are available at https://github.com/Zhaosxian/SteelDefectX.

Wind disturbances remain a key barrier to reliable autonomous navigation for lightweight quadrotors, where the rapidly varying airflow can destabilize both planning and tracking. This paper introduces GustPilot, a hierarchical wind-resilient navigation stack in which a deep reinforcement learning (DRL) policy generates inertial-frame velocity reference for gate traversal. At the same time, a geometric Incremental Nonlinear Dynamic Inversion (INDI) controller provides low-level tracking with fast residual disturbance rejection. The INDI layer achieves this by providing incremental feedback on both specific linear acceleration and angular acceleration rate, using onboard sensor measurements to reject wind disturbances rapidly. Robustness is obtained through a two-level strategy, wind-aware planning learned via fan-jet domain randomization during training, and rapid execution-time disturbance rejection by the INDI tracking controller. We evaluate GustPilot in real flights on a 50g quad-copter platform against a DRL-PID baseline across four scenarios ranging from no-wind to fully dynamic conditions with a moving gate and a moving disturbance source. Despite being trained only in a minimal single-gate and single-fan setup, the policy generalizes to significantly more complex environments (up to six gates and four fans) without retraining. Across 80 experiments, DRL-INDI achieves a 94.7% versus 55.0% for DRL-PID as average Overall Success Rate (OSR), reduces tracking RMSE up to 50%, and sustains speeds up to 1.34 m/s under wind disturbances up to 3.5 m/s. These results demonstrate that combining DRL-based velocity planning with structured INDI disturbance rejection provides a practical and generalizable approach to wind-resilient autonomous flight navigation.

Large language models (LLMs) are increasingly deployed in financial research workflows, where their role is evolving from single-model assistance for human analysts toward autonomous collaboration among multiple agents. Yet real-world deployments still expose factual errors, numerical inconsistencies, and shallow analysis, which can distort assessments of corporate fundamentals and trigger severe economic losses. While existing benchmarks have begun to evaluate such failures, they score all aspects of the generated analysis in one pass, failing to distinguish whether a model fails at foundational stages like auditing and correction, or underperforms at generating research-grade insights. Consequently, it obscures capability bottlenecks and the specialized strengths essential for multi-agent role assignment. To address these gaps, we introduce FinReasoning, a hierarchical benchmark that decomposes the core capabilities of financial research into semantic consistency, data alignment, and deep insight. We further propose a fine-grained evaluation framework that strengthens hallucination-correction assessment and incorporates a 12-indicator rubric for core analytical skills. FinReasoning reveals clear capability stratification across model types. Closed-source models (like Doubao-Seed-1.8) perform strongly overall and are better suited for core reasoning agents in multi-agent financial systems; open-source general models (like Qwen3-235B) show clear capability divergence and consistently underperform on Semantic Consistency, making them less suited for quality-sensitive generation tasks; financial-domain models (like Fin-R1) generate moderate insights but lack foundational auditing skills. Our work has already been deployed in pilot tests across several real-world scenarios. The resource is available at https://github.com/TongjiFinLab/FinReasoning.

Recent video reasoning models increasingly produce spatio-temporal evidence chains that localize objects at specific timestamps. While these traces improve interpretability by grounding \emph{where} and \emph{when} evidence appears, they often leave the motion connecting observations, the \textit{how}, implicit. This makes dynamic and trajectory-dependent claims difficult to supervise, verify, or penalize when unsupported by the video. We formalize this missing component as Spatial-Temporal-Trajectory (STT) reasoning and introduce \textbf{Motion-o}, a motion-centric extension to vision-language models (VLMs) that makes trajectories explicit and verifiable. Motion-o augments evidence chains with Motion Chain of Thought (MCoT), a structured pathway that represents object motion through a discrete \texttt{<motion/>} tag summarizing direction, speed, and scale change. To supervise MCoT, we densify sparse spatio-temporal annotations into object tracks and derive motion descriptors from centroid displacement and box-area change. We then train with complementary rewards for trajectory consistency and visual grounding, including a perturbation-based signal that penalizes motion descriptions that remain unchanged when temporal evidence is removed. Across multiple video understanding benchmarks, Motion-o consistently improves trajectory-faithful reasoning without architectural modifications. These results suggest that an explicit motion interface can complement existing VLM pipelines by converting implicit dynamics into verifiable evidence. Code is available at~\href{https://github.com/ostadabbas/Motion-o}{\faGithub\ \texttt{ostadabbas/Motion-o}}.

Diffusion Transformers (DiTs) achieve strong video generation quality but suffer from high inference cost due to dense 3D attention, motivating sparse attention techniques for improving efficiency. However, existing training-free sparse attention methods for video generation still face two unresolved limitations: ignoring layer heterogeneity in attention pruning and ignoring query-key coupling in block partitioning, which hinder a better quality-speedup trade-off. In this work, we uncover a critical insight: attention sparsity is an intrinsic layer-wise property, with only minor variation across different inputs. Motivated by this observation, we propose SVOO, a training-free sparse attention framework for fast video generation via offline layer-wise sparsity profiling and online bidirectional co-clustering. Specifically, SVOO adopts a two-stage paradigm: (i) offline layer-wise sensitivity profiling to derive intrinsic per-layer pruning levels, and (ii) online block-wise sparse attention via a bidirectional co-clustering algorithm. Extensive experiments on seven widely used video generation models demonstrate that SVOO achieves a superior quality-speedup trade-off over state-of-the-art methods, delivering up to 1.93x speedup while maintaining a PSNR of up to 29 dB on Wan2.1. Code is available at: https://github.com/Mutual-Luo/SVOO.

We propose MARL-Rad, a multi-modal multi-agent reinforcement learning framework for radiology report generation that trains the entire agentic system on policy within its deployed radiology workflow. MARL-Rad addresses the limitation of post-hoc agentization, where fixed LLMs are organized into hand-designed agentic workflows without being optimized for their assigned roles. Our framework decomposes chest X-ray interpretation into region-specific agents and a global integrating agent, and jointly optimizes them using clinically verifiable rewards. Experiments on the MIMIC-CXR and IU X-ray datasets show that MARL-Rad consistently improves clinical efficacy metrics such as RadGraph, CheXbert, and GREEN scores, achieving state-of-the-art clinical efficacy performance. Further analyses show that MARL-Rad improves laterality consistency and produces more accurate and detailed reports. A blinded clinician evaluation further suggests that MARL-Rad produces reports clinically comparable to ground-truth reports.

Pneumatically-actuated anthropomorphic robots with high degrees of freedom (DOF) offer significant potential for physical human-robot interaction. However, precise control of pneumatic actuators is challenging due to their inherent nonlinearities. This paper presents the development of a compact 13-DOF upper-body humanoid robot. To assess the feasibility of an effective controller, we first investigate its key dynamic properties, such as actuation time delays, and confirm that the system exhibits highly reproducible behavior. Leveraging this reproducibility, we implement a preliminary data-driven controller for a 4-DOF arm subsystem based on a multilayer perceptron with explicit time delay compensation. The network was trained on random movement data to generate pressure commands for tracking arbitrary trajectories. Comparative evaluations with a traditional PID controller demonstrate superior trajectory tracking performance, highlighting the potential of data-driven approaches for controlling complex, high-DOF pneumatic robots.

State-of-the-art text-to-image models produce high-quality images, but inference remains expensive as generation requires several sequential ODE or denoising steps. Native one-step models aim to reduce this cost by mapping noise to an image in a single step, yet fair comparisons to multi-step systems are difficult because studies use mismatched sampling steps and different classifier-free guidance (CFG) settings, where CFG can shift FID, Inception Score, and CLIP-based alignment in opposing directions. It is also unclear how well one-step models scale to multi-step inference, and there is limited standardized out-of-distribution evaluation for label-ID-conditioned generators beyond ImageNet. To address this, we benchmark eight models spanning one-step flows (MeanFlow, Improved MeanFlow, SoFlow), multi-step baselines (RAE, Scale-RAE), and established systems (SiT, Stable Diffusion 3.5, FLUX.1) under a class-conditional protocol on ImageNet validation, ImageNetV2, and reLAIONet, our new proofread out-of-distribution dataset aligned to ImageNet label IDs. Using FID, Inception Score, CLIP Score, and Pick Score, we show that FID-focused model development and CFG selection can be misleading in few-step regimes, where guidance changes can improve FID while degrading text-image alignment and human preference signals, worsening visual quality. To make these tradeoffs explicit, we introduce CLIP-scaled and PickScore-scaled variants of FID (csFID, psFID) and Inception Score (csIS, psIS) to serve as a diagnostic for semantically aligned image generation.

Word sense plausibility rating requires predicting the human-perceived plausibility of a given word sense on a 1-5 scale in the context of short narrative stories containing ambiguous homonyms. This paper systematically compares three approaches: (1) embedding-based methods pairing sentence embeddings with standard regressors, (2) transformer fine-tuning with parameter-efficient adaptation, and (3) large language model (LLM) prompting with structured reasoning and explicit decision rules. The best-performing system employs a structured prompting strategy that decomposes evaluation into narrative components (precontext, target sentence, ending) and applies explicit decision rules for rating calibration. The analysis reveals that structured prompting with decision rules outperforms both fine-tuned models and embedding-based approaches, and that prompt design matters more than model scale for this task.

Autoregressive (AR) language models build representations incrementally via left-to-right prediction, while diffusion language models (dLLMs) are trained through full-sequence denoising. Although recent dLLMs match AR performance, whether diffusion objectives fundamentally reshape internal representations remains unclear. We perform the first layer- and token-wise representational analysis comparing native dLLMs (LLaDA), native AR models (Qwen2.5), and AR-initialized dLLMs (Dream-7B), using cosine similarity across layers and tokens alongside static inference-time layer-skipping as an analytical probe of redundancy. We find that diffusion objectives produce more global representations with substantial early-layer redundancy and reduced recency bias, while AR objectives yield tightly coupled, locally structured representations. AR-initialized dLLMs retain AR-like dynamics despite diffusion training, revealing persistent initialization bias. Leveraging this redundancy, native dLLMs absorb up to 18.75% FLOPs reduction while retaining over 90% performance on math-reasoning and coding benchmarks, whereas AR models collapse under identical skipping, revealing that diffusion objectives, rather than architecture alone, induce depth redundancy that enables principled compression.

Contact-rich dexterous manipulation with multi-finger hands remains an open challenge in robotics because task success depends on multi-point contacts that continuously evolve and are highly sensitive to object geometry, frictional transitions, and slip. Recently, tactile-informed manipulation policies have shown promise. However, most use tactile signals as additional observations rather than modeling contact state or how their action outputs interact with low-level controller dynamics. We present Contact-Grounded Policy (CGP), a visuotactile policy that grounds multi-point contacts by predicting coupled trajectories of actual robot state and tactile feedback, and using a learned contact-consistency mapping to convert these predictions into executable target robot states for a compliance controller. CGP consists of two components: (i) a conditional diffusion model that forecasts future robot state and tactile feedback in a compressed latent space, and (ii) a learned contact-consistency mapping that converts the predicted robot state-tactile pair into executable targets for a compliance controller, enabling it to realize the intended contacts. We evaluate CGP using a physical four-finger Allegro V5 hand with Digit360 fingertip tactile sensors, and a simulated five-finger Tesollo DG-5F hand with dense whole-hand tactile arrays. Across a range of dexterous tasks including in-hand manipulation, delicate grasping, and tool use, CGP outperforms visuomotor and visuotactile diffusion-policy baselines.

Imitation Learning (IL) enables robots to acquire manipulation skills from expert demonstrations. Diffusion Policy (DP) models multi-modal expert behaviors but degrades when naively increasing stacked observation horizons, limiting long-horizon manipulation. We propose Self-Evolving Gated Attention (SEGA), a temporal module that maintains a time-evolving latent state via gated attention, enabling efficient recurrent updates that accumulate long-term context into a compact latent representation while filtering irrelevant temporal information. Integrating SEGA into DP yields Self-Evolving Diffusion Policy (SeedPolicy), which resolves the temporal modeling bottleneck and extends the effective temporal horizon with moderate overhead. On the RoboTwin 2.0 benchmark with 50 manipulation tasks, SeedPolicy outperforms DP and other IL baselines. Averaged across both CNN and Transformer backbones, SeedPolicy achieves 36.8% relative improvement in clean settings and 169% relative improvement in randomized challenging settings over the DP. Compared to vision-language-action models such as RDT with 1.2B parameters, SeedPolicy achieves stronger performance in the clean setting with one to two orders of magnitude fewer parameters, demonstrating strong efficiency. These results establish SeedPolicy as a state-of-the-art imitation learning method for long-horizon robotic manipulation. Code is available at: https://anonymous.4open.science/r/SeedPolicy-64F0/.

Diffusion Transformers (DiTs) achieve state-of-the-art video generation quality, but their substantial memory and computational footprints hinder edge deployment. Quantization can reduce these costs, yet existing methods often degrade video quality due to high activation variation and the difficulty of preserving semantic and temporal coherence. We propose SemanticDialect, which advances block-wise mixed-format quantization. In this framework, each block selects an optimal format (dialect) from a candidate set (formatbook), which is augmented with lookup tables that store quantization errors and quantized indices, enabling efficient per-block format selection and quantization with minimal online overhead. We further introduce attention-guided activation decomposition, which reduces quantization error via residual quantization, and semantic-aware dialect assignment (SeDA), which reduces cross-token quantization inconsistency by enforcing format uniformity among semantically correlated tokens. Experiments demonstrate that SemanticDialect outperforms prior quantization methods and block-wise formats (MXFP4, NVFP4) while approaching FP16 quality on Open-Sora 2.0. We also validate hardware deployability through RTL design and GPU kernel implementation.

Emerging unified editing models have demonstrated strong capabilities in general object editing tasks. However, it remains a significant challenge to perform fine-grained editing in complex multi-entity scenes, particularly those where targets are not visually salient and require spatial reasoning. To this end, we propose InterCoG, a novel text-vision Interleaved Chain-of-Grounding reasoning framework for fine-grained image editing in complex real-world scenes. The key insight of InterCoG is to first perform object position reasoning solely within text that includes spatial relation details to explicitly deduce the location and identity of the edited target. It then conducts visual grounding via highlighting the editing targets with generated bounding boxes and masks in pixel space, and finally rewrites the editing description to specify the intended outcomes. To further facilitate this paradigm, we propose two auxiliary training modules: multimodal grounding reconstruction supervision and multimodal grounding reasoning alignment to enforce spatial localization accuracy and reasoning interpretability, respectively. We also construct GroundEdit-45K, a dataset comprising 45K grounding-oriented editing samples with detailed reasoning annotations, and GroundEdit-Bench for grounding-aware editing evaluation. Extensive experiments substantiate the superiority of our approach in highly precise edits under spatially intricate and multi-entity scenes.

We investigate the theoretical aspects of offline reinforcement learning (RL) under general function approximation. While prior works (e.g., Xie et al., 2021) have established the theoretical foundations of learning a good policy from offline data via pessimism, existing algorithms that are computationally tractable (often in an oracle-efficient sense), such as PSPI, only apply to finite and small action spaces. Moreover, these algorithms rely on state-wise mirror descent and require actors to be implicitly induced from the critic functions, failing to accommodate standalone policy parameterization which is ubiquitous in practice. In this work, we address these limitations and extend the theoretical guarantees to parameterized policy classes over large or continuous action spaces. When extending mirror descent to parameterized policies, we identify contextual coupling as the core difficulty, and show how connecting mirror descent to natural policy gradient leads to novel analyses, guarantees, and algorithmic insights, including a surprising unification between offline RL and imitation learning.

Moral benchmarks for LLMs typically score models on context-free prompts, implicitly treating the measured choice rate as stable. We test this assumption with a direction-flipped influence audit: for each scenario, we compare a baseline prompt with matched cues steering toward option A or option B. Across a trolley-problem-style moral triage task, BBQ, and DailyDilemmas, and across five LLM families with and without reasoning, short contextual cues shift per-condition choice rates by 12-18 percentage points on average. These shifts reveal structure that baseline scores miss: roughly 40% of baseline-neutral triage and BBQ conditions exhibit directional asymmetry under influence, and a meaningful share of significant effects backfire, moving opposite the cue's intended direction. In follow-up probes, models often recognize the cue while denying that it affected their choice. Among significant backfire trials, this stated-vs.-revealed inconsistency appears in 78% of cases. Reasoning does not eliminate contextual sensitivity but reshapes it: social-pressure cues such as user preference and emotional appeal weaken across benchmarks, while few-shot demonstrations strengthen sharply on both triage and BBQ. We recommend direction-flipped influence pairs as a standard complement to context-free moral-bias evaluation, and release the harness and data to make such audits routine.

Multimodal large language models (MLLMs) have made significant progress in mobile agent development, yet their capabilities are predominantly confined to a reactive paradigm, where they merely execute explicit user commands. The emerging paradigm of proactive intelligence, where agents autonomously anticipate needs and initiate actions, represents the next frontier for mobile agents. However, its development is critically bottlenecked by the lack of benchmarks that can address real-world complexity and enable objective, executable evaluation. To overcome these challenges, we introduce ProactiveMobile, a comprehensive benchmark designed to systematically advance research in this domain. ProactiveMobile formalizes the proactive task as inferring latent user intent across four dimensions of on-device contextual signals and generating an executable function sequence from a comprehensive function pool of 63 APIs. The benchmark features over 3,660 instances of 14 scenarios that embrace real-world complexity through multi-answer annotations. To ensure quality, a team of 30 experts conducts a final audit of the benchmark, verifying factual accuracy, logical consistency, and action feasibility, and correcting any non-compliant entries. Extensive experiments demonstrate that our fine-tuned Qwen2.5-VL-7B-Instruct achieves a success rate of 19.15%, outperforming o1 (15.71%) and GPT-5 (7.39%). This result indicates that proactivity is a critical competency widely lacking in current MLLMs, yet it is learnable, emphasizing the importance of the proposed benchmark for proactivity evaluation.

In engineering design and scientific computing, computational cost and predictive accuracy are intrinsically coupled. High-fidelity simulations provide accurate predictions but at substantial computational costs, while lower-fidelity approximations offer efficiency at the expense of accuracy. Multi-fidelity surrogate modelling addresses this trade-off by combining abundant low-fidelity data with sparse high-fidelity observations. However, existing methods rely on global correlation assumptions that can often fail in practice to capture how fidelity relationships vary across the input space, leading to poor performance, particularly under tight budget constraints. We introduce MAST, a method that blends corrected low-fidelity observations with high-fidelity predictions, trusting high-fidelity near observed samples and relying on corrected low-fidelity elsewhere. MAST achieves this through explicit discrepancy modelling and distance-based weighting with closed-form variance propagation, producing a single heteroscedastic Gaussian process. Across multi-fidelity synthetic benchmarks, MAST shows a marked improvement over the current state-of-the-art techniques. Crucially, MAST maintains robust performance across varying total budget and fidelity gaps, conditions under which competing methods exhibit significant degradation or unstable behaviour. More broadly, MAST provides a spatially adaptive framework for multi-fidelity Gaussian-process modelling, in which the contribution of low-fidelity information is governed by its proximity to high-fidelity calibration data, opening a new direction for more reliable surrogate construction under sparse and budget-constrained settings.

The core learning signal used in language model distillation is the standard Kullback-Leibler (KL) divergence between the student and teacher distributions. Traditional KL divergence tends to be dominated by the next tokens with the highest probabilities, i.e., the teacher's modes, thereby diminishing the influence of less probable yet potentially informative components of the output distribution. We propose a new tail-aware divergence that decouples the contribution of the teacher model's top-K predicted probabilities from that of lower-probability predictions, while maintaining the same computational profile as the KL Divergence. Our decoupled approach reduces the impact of the teacher modes and, consequently, increases the contribution of the tail of the distribution. Experimental results demonstrate that our modified distillation method yields competitive performance in both pre-training and supervised distillation of decoder models across various datasets. Furthermore, the distillation process is efficient and can be performed with a modest academic budget for large datasets, eliminating the need for industry-scale computing.

Chain-of-Thought (CoT) prompting significantly improves reasoning in Large Language Models, yet the temporal dynamics of the underlying representation geometry remain poorly understood. We investigate these dynamics by applying Manifold Capacity Theory (MCT) to two compositional reasoning tasks: a controlled Boolean logic tree that supports deep mechanistic analysis, and a natural-language eligibility task in which the model has to extract attributes from prose, compare them to thresholds, and compose the local decisions through a fixed evaluation tree. MCT lets us quantify the linear separability of latent representations without the confounding factors of probe training. On both tasks, and across several open-weight models, reasoning manifests as a transient geometric pulse: concept manifolds are untangled into linearly separable subspaces immediately prior to computation and rapidly compressed thereafter. This behavior diverges from standard linear probe accuracy, which remains high long after computation, suggesting a fundamental distinction between information that is merely retrievable and information that is geometrically prepared for processing. We interpret this phenomenon as Dynamic Manifold Management, a mechanism where the model dynamically modulates representational capacity to optimize the bandwidth of the residual stream throughout the reasoning chain.