This paper introduces the "Trilingual Triad" framework, a model that explains how students learn to design with generative artificial intelligence (AI) through the integration of Design, AI, and Domain Knowledge. As generative AI rapidly enters higher education, students often engage with these systems as passive users of generated outputs rather than active creators of AI-enabled knowledge tools. This study investigates how students can transition from using AI as a tool to designing AI as a collaborative teammate. The research examines a graduate course, Creating the Frontier of No-code Smart Cities at the Singapore University of Technology and Design (SUTD), in which students developed domain-specific custom GPT systems without coding. Using a qualitative multi-case study approach, three projects - the Interview Companion GPT, the Urban Observer GPT, and Buddy Buddy - were analyzed across three dimensions: design, AI architecture, and domain expertise. The findings show that effective human-AI collaboration emerges when these three "languages" are orchestrated together: domain knowledge structures the AI's logic, design mediates human-AI interaction, and AI extends learners' cognitive capacity. The Trilingual Triad framework highlights how building AI systems can serve as a constructionist learning process that strengthens AI literacy, metacognition, and learner agency.

AI agents that build user interfaces on the fly assembling buttons, forms, and data displays from structured protocol payloads are becoming common in production systems. The trouble is that a payload can pass every schema check and still trick a user: a button might say "View invoice" while its hidden action wipes an account, or a display widget might quietly bind to an internal salary field. Current defenses stop at syntax; they were never built to catch this kind of behavioral mismatch. We built AegisUI to study exactly this gap. The framework generates structured UI payloads, injects realistic attacks into them, extracts numeric features, and benchmarks anomaly detectors end-to-end. We produced 4000 labeled payloads (3000 benign, 1000 malicious) spanning five application domains and five attack families: phishing interfaces, data leakage, layout abuse, manipulative UI, and workflow anomalies. From each payload we extracted 18 features covering structural, semantic, binding, and session dimensions, then compared three detectors: Isolation Forest (unsupervised), a benign-trained autoencoder (semi-supervised), and Random Forest (supervised). On a stratified 80/20 split, Random Forest scored best overall (accuracy 0.931, precision 0.980, recall 0.740, F1 0.843, ROC-AUC 0.952). The autoencoder came second (F1 0.762, ROC-AUC 0.863) and needs no malicious labels at training time, which matters when deploying a new system that lacks attack history. Per-attack-type analysis showed that layout abuse is easiest to catch while manipulative UI payloads are hardest. All code, data, and configurations are released for full reproducibility.

As Large Language Models (LLMs) evolve from chatbots to agentic assistants, they are increasingly observed to exhibit risky behaviors when subjected to survival pressure, such as the threat of being shut down. While multiple cases have indicated that state-of-the-art LLMs can misbehave under survival pressure, a comprehensive and in-depth investigation into such misbehaviors in real-world scenarios remains scarce. In this paper, we study these survival-induced misbehaviors, termed as SURVIVE-AT-ALL-COSTS, with three steps. First, we conduct a real-world case study of a financial management agent to determine whether it engages in risky behaviors that cause direct societal harm when facing survival pressure. Second, we introduce SURVIVALBENCH, a benchmark comprising 1,000 test cases across diverse real-world scenarios, to systematically evaluate SURVIVE-AT-ALL-COSTS misbehaviors in LLMs. Third, we interpret these SURVIVE-AT-ALL-COSTS misbehaviors by correlating them with model's inherent self-preservation characteristic and explore mitigation methods. The experiments reveals a significant prevalence of SURVIVE-AT-ALL-COSTS misbehaviors in current models, demonstrates the tangible real-world impact it may have, and provides insights for potential detection and mitigation strategies. Our code and data are available at https://github.com/thu-coai/Survive-at-All-Costs.

Explainable Artificial Intelligence (XAI) methods (SHAP, LIME) are increasingly adopted to interpret models in high-stakes businesses. However, the credibility of these explanations, their stability under realistic data perturbations, remains unquantified. This paper introduces the Credibility Index via Explanation Stability (CIES), a mathematically grounded metric that measures how robust a model's explanations are when subject to realistic business noise. CIES captures whether the reasons behind a prediction remain consistent, not just the prediction itself. The metric employs a rank-weighted distance function that penalizes instability in the most important features disproportionately, reflecting business semantics where changes in top decision drivers are more consequential than changes in marginal features. We evaluate CIES across three datasets (customer churn, credit risk, employee attrition), four tree-based classification models and two data balancing conditions. Results demonstrate that model complexity impacts explanation credibility, class imbalance treatment via SMOTE affects not only predictive performance but also explanation stability, and CIES provides statistically superior discriminative power compared to a uniform baseline metric (p < 0.01 in all 24 configurations). A sensitivity analysis across four noise levels confirms the robustness of the metric itself. These findings offer business practitioners a deployable "credibility warning system" for AI-driven decision support.

Navigation in cluttered environments often requires robots to tolerate contact with movable or deformable objects to maintain efficiency. Existing contact-tolerant motion planning (CTMP) methods rely on indirect spatial representations (e.g., prebuilt map, obstacle set), resulting in inaccuracies and a lack of adaptiveness to environmental uncertainties. To address this issue, we propose a direct contact-tolerant (DCT) planner, which integrates vision-language models (VLMs) into direct point perception and navigation, including two key components. The first one is VLM point cloud partitioner (VPP), which performs contact-tolerance reasoning in image space using VLM, caches inference masks, propagates them across frames using odometry, and projects them onto the current scan to generate a contact-aware point cloud. The second innovation is VPP guided navigation (VGN), which formulates CTMP as a perception-to-control optimization problem under direct contact-aware point cloud constraints, which is further solved by a specialized deep neural network (DNN). We implement DCT in Isaac Sim and a real car-like robot, demonstrating that DCT achieves robust and efficient navigation in cluttered environments with movable obstacles, outperforming representative baselines across diverse metrics. The code is available at: https://github.com/ChrisLeeUM/DCT.

Computational psychiatry faces a fundamental trade-off: traditional reinforcement learning (RL) models offer interpretability but lack behavioral realism, while large language model (LLM) agents generate realistic behaviors but lack structural interpretability. We introduce BioLLMAgent, a novel hybrid framework that combines validated cognitive models with the generative capabilities of LLMs. The framework comprises three core components: (i) an Internal RL Engine for experience-driven value learning; (ii) an External LLM Shell for high-level cognitive strategies and therapeutic interventions; and (iii) a Decision Fusion Mechanism for integrating components via weighted utility. Comprehensive experiments on the Iowa Gambling Task (IGT) across six clinical and healthy datasets demonstrate that BioLLMAgent accurately reproduces human behavioral patterns while maintaining excellent parameter identifiability (correlations $>0.67$). Furthermore, the framework successfully simulates cognitive behavioral therapy (CBT) principles and reveals, through multi-agent dynamics, that community-wide educational interventions may outperform individual treatments. Validated across reward-punishment learning and temporal discounting tasks, BioLLMAgent provides a structurally interpretable "computational sandbox" for testing mechanistic hypotheses and intervention strategies in psychiatric research.

Although virtual and augmented reality are gaining traction as teleoperation tools for various types of robots, including manipulators and mobile robots, they are not being used for soft robots. The inherent difficulties of modelling soft robots mean that combining accurate and computationally efficient representations is very challenging. This paper presents an augmented reality interface for teleoperating these devices. The developed system consists of Microsoft HoloLens 2 glasses and a central computer responsible for calculations. Validation is performed on PETER, a highly modular pneumatic manipulator. Using data collected from sensors, the computer estimates the robot's position based on the physics of the virtual reality programme. Errors obtained are on the order of 5% of the robot's length, demonstrating that augmented reality facilitates operator interaction with soft manipulators and can be integrated into the control loop.

Source Free Unsupervised Domain Adaptation (SFUDA) is critical for deploying deep learning models across diverse clinical settings. However, existing methods are typically designed for low-gap, specific domain shifts and cannot generalize into a unified, multi-modalities, and multi-target framework, which presents a major barrier to real-world application. To overcome this issue, we introduce Tell2Adapt, a novel SFUDA framework that harnesses the vast, generalizable knowledge of the Vision Foundation Model (VFM). Our approach ensures high-fidelity VFM prompts through Context-Aware Prompts Regularization (CAPR), which robustly translates varied text prompts into canonical instructions. This enables the generation of high-quality pseudo-labels for efficiently adapting the lightweight student model to target domain. To guarantee clinical reliability, the framework incorporates Visual Plausibility Refinement (VPR), which leverages the VFM's anatomical knowledge to re-ground the adapted model's predictions in target image's low-level visual features, effectively removing noise and false positives. We conduct one of the most extensive SFUDA evaluations to date, validating our framework across 10 domain adaptation directions and 22 anatomical targets, including brain, cardiac, polyp, and abdominal targets. Our results demonstrate that Tell2Adapt consistently outperforms existing approaches, achieving SOTA for a unified SFUDA framework in medical image segmentation. Code are avaliable at https://github.com/derekshiii/Tell2Adapt.

Autonomous Mobility-on-Demand (AMoD) systems promise to revolutionize urban transportation by providing affordable on-demand services to meet growing travel demand. However, realistic AMoD markets will be competitive, with multiple operators competing for passengers through strategic pricing and fleet deployment. While reinforcement learning has shown promise in optimizing single-operator AMoD control, existing work fails to capture competitive market dynamics. We investigate the impact of competition on policy learning by introducing a multi-operator reinforcement learning framework where two operators simultaneously learn pricing and fleet rebalancing policies. By integrating discrete choice theory, we enable passenger allocation and demand competition to emerge endogenously from utility-maximizing decisions. Experiments using real-world data from multiple cities demonstrate that competition fundamentally alters learned behaviors, leading to lower prices and distinct fleet positioning patterns compared to monopolistic settings. Notably, we demonstrate that learning-based approaches are robust to the additional stochasticity of competition, with competitive agents successfully converging to effective policies while accounting for partially unobserved competitor strategies.

Mobile photography is often limited by complex, lens-specific optical aberrations. While recent deep learning methods approach this as an end-to-end deblurring task, these "black-box" models lack explicit optical modeling and can hallucinate details. Conversely, classical blind deconvolution remains highly unstable. To bridge this gap, we present Lens2Zernike, a deep learning framework that blindly recovers physical optical parameters from a single blurred image. To the best of our knowledge, no prior work has simultaneously integrated supervision across three distinct optical domains. We introduce a novel physics-consistent strategy that explicitly minimizes errors via direct Zernike coefficient regression (z), differentiable physics constraints encompassing both wavefront and point spread function derivations (p), and auxiliary multi-task spatial map predictions (m). Through an ablation study on a ResNet-18 backbone, we demonstrate that our full multi-task framework (z+p+m) yields a 35% improvement over coefficient-only baselines. Crucially, comparative analysis reveals that our approach outperforms two established deep learning methods from previous literature, achieving significantly lower regression errors. Ultimately, we demonstrate that these recovered physical parameters enable stable non-blind deconvolution, providing substantial in-domain improvement on the patented Institute for Digital Molecular Analytics and Science (IDMxS) Mobile Camera Lens Database for restoring diffraction-limited details from severely aberrated mobile captures.

Non-Intrusive Load Monitoring (NILM) aims to estimate appliance-level consumption from aggregate electrical signals recorded at a single measurement point. In recent years, the field has increasingly adopted deep learning approaches; however, cross-domain generalization remains a persistent challenge due to variations in appliance characteristics, usage patterns, and background loads across homes. Transfer learning provides a practical paradigm to adapt models with limited target data. However, existing methods often assume a fixed appliance set, lack flexibility for evolving real-world deployments, remain unsuitable for edge devices, or scale poorly for real-time operation. This paper proposes RefQuery, a scalable multi-appliance, multi-task NILM framework that conditions disaggregation on compact appliance fingerprints, allowing one shared model to serve many appliances without a fixed output set. RefQuery keeps a pretrained disaggregation network fully frozen and adapts to a target home by learning only a per-appliance embedding during a lightweight backpropagation stage. Experiments on three public datasets demonstrate that RefQuery delivers a strong accuracy-efficiency trade-off against single-appliance and multi-appliance baselines, including modern Transformer-based methods. These results support RefQuery as a practical path toward scalable, real-time NILM on resource-constrained edge devices.

Large language models perform well in short text generation but still struggle with long text generation, particularly under complex constraints. Such tasks involve multiple tightly coupled objectives, including global structural consistency, local semantic coherence, and constraint feasibility, forming a challenging constrained optimization problem. Existing approaches mainly rely on static planning or offline supervision, limiting effective coordination between global and local objectives during generation. To address these challenges, we propose HiFlow, a hierarchical feedback-driven optimization framework for constrained long text generation. HiFlow formulates generation as a two-level optimization process, consisting of a planning layer for global structure and constraint modeling, and a generation layer for conditioned text generation. By incorporating constraint-aware plan screening and closed-loop feedback at both levels, HiFlow enables joint optimization of planning quality and generation behavior, progressively guiding the model toward high-quality, constraint-satisfying outputs. Experiments on multiple backbones confirm HiFlow's effectiveness over baseline methods.

Monocular 3D clothed human reconstruction aims to generate a complete and realistic textured 3D avatar from a single image. Existing methods are commonly trained under multi-view supervision with annotated geometric priors, and during inference, these priors are estimated by the pre-trained network from the monocular input. These methods are constrained by three key limitations: texturally by unavailability of training data, geometrically by inaccurate external priors, and systematically by biased single-modality supervision, all leading to suboptimal reconstruction. To address these issues, we propose a novel reconstruction framework, named MultiGO++, which achieves effective systematic geometry-texture collaboration. It consists of three core parts: (1) A multi-source texture synthesis strategy that constructs 15,000+ 3D textured human scans to improve the performance on texture quality estimation in challenge scenarios; (2) A region-aware shape extraction module that extracts and interacts features of each body region to obtain geometry information and a Fourier geometry encoder that mitigates the modality gap to achieve effective geometry learning; (3) A dual reconstruction U-Net that leverages geometry-texture collaborative features to refine and generate high-fidelity textured 3D human meshes. Extensive experiments on two benchmarks and many in-the-wild cases show the superiority of our method over state-of-the-art approaches.

Dynamic data selection accelerates training by sampling a changing subset of the dataset while preserving accuracy. We rethink two core notions underlying sample evaluation: representativeness and diversity. Instead of local geometric centrality, we define representativeness as coverage of dataset-level common or high-frequency feature factors. Instead of within-subset dispersion, we define diversity at the process level, requiring the selection trajectory to gradually include complementary rare factors over training. Based on this view, we propose a dynamic selection framework with three components. First, we score representativeness in a plug-in feature space to prioritize samples covering frequent factors. We instantiate this with a sparse autoencoder trained on the target dataset, using sparse unit activations to summarize both individual samples and dataset-wide factor statistics. Second, we realize process-level diversity by combining rare-factor sampling with a Usage-Frequency Penalty that promotes sample rotation, provably discourages monopoly, and reduces gradient bias. Third, we couple the two-dimensional scoring with a smooth scheduler that transitions selection from core-pattern consolidation to rare-factor exploration, without extra gradients, influence estimates, or second-order computations on the training model. Extensive experiments on five benchmarks across vision and text tasks demonstrate improved accuracy-efficiency trade-offs across models. Our method matches or exceeds full-data accuracy with over 2x training acceleration. Code will be released.

In this work, we introduce Wallaroo, a simple autoregressive baseline that leverages next-token prediction to unify multi-modal understanding, image generation, and editing at the same time. Moreover, Wallaroo supports multi-resolution image input and output, as well as bilingual support for both Chinese and English. We decouple the visual encoding into separate pathways and apply a four-stage training strategy to reshape the model's capabilities. Experiments are conducted on various benchmarks where Wallaroo produces competitive performance or exceeds other unified models, suggesting the great potential of autoregressive models in unifying multi-modality understanding and generation. Our code is available at https://github.com/JiePKU/Wallaroo.

Long video understanding is challenging due to dense visual redundancy, long-range temporal dependencies, and the tendency of chain-of-thought and retrieval-based agents to accumulate semantic drift and correlation-driven errors. We argue that long-video reasoning should begin not with reactive retrieval, but with deliberate task formulation: the model must first articulate what must be true in the video for each candidate answer to hold. This thinking-before-finding principle motivates VideoHV-Agent, a framework that reformulates video question answering as a structured hypothesis-verification process. Based on video summaries, a Thinker rewrites answer candidates into testable hypotheses, a Judge derives a discriminative clue specifying what evidence must be checked, a Verifier grounds and tests the clue using localized, fine-grained video content, and an Answer agent integrates validated evidence to produce the final answer. Experiments on three long-video understanding benchmarks show that VideoHV-Agent achieves state-of-the-art accuracy while providing enhanced interpretability, improved logical soundness, and lower computational cost. We make our code publicly available at: https://github.com/Haorane/VideoHV-Agent.

Event cameras, with their high dynamic range, show great promise for Low-light Image Enhancement (LLIE). Existing works primarily focus on designing effective modal fusion strategies. However, a key challenge is the dual degradation from intrinsic background activity (BA) noise in events and low signal-to-noise ratio (SNR) in images, which causes severe noise coupling during modal fusion, creating a critical performance bottleneck. We therefore posit that precise event denoising is the prerequisite to unlocking the full potential of event-based fusion. To this end, we propose BiEvLight, a hierarchical and task-aware framework that collaboratively optimizes enhancement and denoising by exploiting their intrinsic interdependence. Specifically, BiEvLight exploits the strong gradient correlation between images and events to build a gradient-guided event denoising prior that alleviates insufficient denoising in heavily noisy regions. Moreover, instead of treating event denoising as a static pre-processing stage-which inevitably incurs a trade-off between over- and under-denoising and cannot adapt to the requirements of a specific enhancement objective-we recast it as a bilevel optimization problem constrained by the enhancement task. Through cross-task interaction, the upper-level denoising problem learns event representations tailored to the lower-level enhancement objective, thereby substantially improving overall enhancement quality. Extensive experiments on the Real-world noise Dataset SDE demonstrate that our method significantly outperforms state-of-the-art (SOTA) approaches, with average improvements of 1.30dB in PSNR, 2.03dB in PSNR* and 0.047 in SSIM, respectively. The code will be publicly available at https://github.com/iijjlk/BiEvlight.

Large Language Models (LLMs) have achieved remarkable success across diverse natural language tasks, yet the reward models employed for aligning LLMs often encounter challenges of reward hacking, where the approaches predominantly rely on directly mapping prompt-response pairs to scalar scores, which may inadvertently capture spurious correlations rather than authentic human preferences. In contrast, human evaluation employs a sophisticated process that initially weighs the relative importance of multiple high-dimensional objectives according to the prompt context, subsequently evaluating response quality through low-dimensional semantic features such as logical coherence and contextual appropriateness. Motivated by this consideration, we propose VRM, i.e., Variational Reward Modeling, a novel framework that explicitly models the evaluation process of human preference judgments by incorporating both high-dimensional objective weights and low-dimensional semantic features as latent variables, which are inferred through variational inference techniques. Additionally, we provide a theoretical analysis showing that VRM can achieve a tighter generalization error bound compared to the traditional reward model. Extensive experiments on benchmark datasets demonstrate that VRM significantly outperforms existing methods in capturing authentic human preferences.

Weight-space merging aims to combine multiple fine-tuned LLMs into a single model without retraining, yet most existing approaches remain fundamentally parameter-space heuristics. This creates three practical limitations. First, linear averaging, task vectors, and related rules operate on Euclidean coordinates, even though the desired goal is to merge functionality, i.e., predictive behaviors across tasks. Second, when the source checkpoints are farther apart or more heterogeneous, Euclidean blends often trigger representation collapse, manifested as activation variance shrinkage and effective-rank degradation, which sharply degrades accuracy. Third, many geometry-inspired methods are most natural for two-model interpolation and do not extend cleanly to merging N>2 experts with a principled objective. We address these issues by formulating model merging as computing a weighted Karcher mean on the Fisher--Rao manifold, which is locally equivalent to minimizing a KL-based function distance between predictive distributions. We derive a practical fixed-point algorithm using a lightweight spherical proxy that preserves norms and generalizes directly to multi-expert merging. Across various benchmarks and collapse diagnostics, our method remains stable as the number and heterogeneity of merged models increase, consistently outperforming prior baselines.

Mixture-of-Experts (MoE) decouples model capacity from per-token computation, yet their scalability remains limited by the physical dimensions of depth and width. To overcome this, we propose Mixture of Universal Experts (MOUE),a MoE generalization introducing a novel scaling dimension: Virtual Width. In general, MoUE aims to reuse a universal layer-agnostic expert pool across layers, converting depth into virtual width under a fixed per-token activation budget. However, two challenges remain: a routing path explosion from recursive expert reuse, and a mismatch between the exposure induced by reuse and the conventional load-balancing objectives. We address these with three core components: a Staggered Rotational Topology for structured expert sharing, a Universal Expert Load Balance for depth-aware exposure correction, and a Universal Router with lightweight trajectory state for coherent multi-step routing. Empirically, MoUE consistently outperforms matched MoE baselines by up to 1.3% across scaling regimes, enables progressive conversion of existing MoE checkpoints with up to 4.2% gains, and reveals a new scaling dimension for MoE architectures.

Multi-party conversation generation, such as smart reply and collaborative assistants, is an increasingly important capability of generative AI, yet its evaluation remains a critical bottleneck. Compared to two-party dialogue, multi-party settings introduce distinct challenges, including complex turn-taking, role-dependent speaker behavior, long-range conversational structure, and multiple equally valid continuations. Accordingly, we introduce MPCEval, a task-aware evaluation and benchmarking suite for multi-party conversation generation. MPCEval decomposes generation quality into speaker modeling, content quality, and speaker--content consistency, and explicitly distinguishes local next-turn prediction from global full-conversation generation. It provides novel, quantitative, reference-free, and reproducible metrics that scale across datasets and models. We apply MPCEval to diverse public and real-world datasets and evaluate modern generation methods alongside human-authored conversations. The results reveal systematic, dimension-specific model characteristics in participation balance, content progression and novelty, and speaker--content consistency, demonstrating that evaluation objectives critically shape model assessment and that single-score evaluation obscures fundamental differences in multi-party conversational behavior. The implementation of MPCEval and the associated evaluation code are publicly available at https://github.com/Owen-Yang-18/MPCEval.

Preference alignment is an essential step in adapting large language models (LLMs) to human values, but existing approaches typically depend on costly human annotations or large-scale API-based models. We explore whether a weak LLM can instead act as an effective annotator. We surprisingly find that selecting only a subset of a weak LLM's highly confident samples leads to substantially better performance than using full human annotations. Building on this insight, we propose Confidence-Weighted Preference Optimization (CW-PO), a general framework that re-weights training samples by a weak LLM's confidence and can be applied across different preference optimization objectives. Notably, the model aligned by CW-PO with just 20% of human annotations outperforms the model trained with 100% of annotations under standard DPO. These results suggest that weak LLMs, when paired with confidence weighting, can dramatically reduce the cost of preference alignment while even outperforming methods trained on fully human-labeled data.

To obtain a language model for a target domain (e.g. math), the current paradigm is to pre-train on a vast amount of generic web text and then fine-tune on the relatively limited amount of target data. Typically, generic data is only mixed in during fine-tuning to prevent catastrophic forgetting of the generic domain. We surprisingly find that replaying the generic data during fine-tuning can actually improve performance on the (less related) target task. Concretely, in a controlled pre-training environment with 4M target tokens, 4B total tokens, and 150M parameter models, generic replay increases target data efficiency by up to $1.87\times$ for fine-tuning and $2.06\times$ for mid-training. We further analyze data schedules that introduce target data during pre-training and find that replay helps more when there is less target data present in pre-training. We demonstrate the success of replay in practice for fine-tuning 8B parameter models, improving agentic web navigation success by $4.5\%$ and Basque question-answering accuracy by $2\%$.

We introduce a novel camera model for monocular 3D Morphable Model (3DMM) regression methods that effectively captures the perspective distortion effect commonly seen in close-up facial images. Fitting 3D morphable models to video is a key technique in content creation. In particular, regression-based approaches have produced fast and accurate results by matching the rendered output of the morphable model to the target image. These methods typically achieve stable performance with orthographic projection, which eliminates the ambiguity between focal length and object distance. However, this simplification makes them unsuitable for close-up footage, such as that captured with head-mounted cameras. We extend orthographic projection with a new shrinkage parameter, incorporating a pseudo-perspective effect while preserving the stability of the original projection. We present several techniques that allow finetuning of existing models, and demonstrate the effectiveness of our modification through both quantitative and qualitative comparisons using a custom dataset recorded with head-mounted cameras.

Large Multimodal Models (LMMs) have achieved strong performance in vision-language understanding, yet many existing approaches rely on large-scale architectures and coarse supervision, which limits their ability to generate detailed image captions. In this work, we present VisionPangu, a compact 1.7B-parameter multimodal model designed to improve detailed image captioning through efficient multimodal alignment and high-quality supervision. Our model combines an InternVL-derived vision encoder with the OpenPangu-Embedded language backbone via a lightweight MLP projector and adopts an instruction-tuning pipeline inspired by LLaVA. By incorporating dense human-authored descriptions from the DOCCI dataset, VisionPangu improves semantic coherence and descriptive richness without relying on aggressive model scaling. Experimental results demonstrate that compact multimodal models can achieve competitive performance while producing more structured and detailed captions. The code and model weights will be publicly available at https://www.modelscope.cn/models/asdfgh007/visionpangu.

The improvement of web agents on current benchmarks raises the question: Do today's agents perform just as well when the web changes? We introduce TimeWarp, a benchmark that emulates the evolving web using containerized environments that vary in UI, design, and layout. TimeWarp consists of three web environments, each with six UI versions spanning different eras of the internet, paired with a set of complex, realistic tasks requiring different forms of web navigation. Our experiments reveal web agents' vulnerability to changes and the limitations of behavior cloning (BC) on single-version trajectories. To address this, we propose TimeTraj, a simple yet effective algorithm that uses plan distillation to collect trajectories across multiple versions. By training agents on teacher rollouts using our BC-variant, we achieve substantial performance gains: $20.4\%\rightarrow37.7\%$ for Qwen-3 4B and $0\%\rightarrow27.0\%$ for Llama-3.1 8B models. We hope our work helps researchers study generalization across web designs and unlock a new paradigm for collecting plans rather than trajectories, thereby improving the robustness of web agents.

Scaling inference-time compute for Large Language Models (LLMs) has unlocked unprecedented reasoning capabilities. However, existing inference-time scaling methods typically rely on inefficient and suboptimal discrete search algorithms or trial-and-error prompting to improve the online policy. In this paper, we propose $\nabla$-Reasoner, an iterative generation framework that integrates differentiable optimization over token logits into the decoding loop to refine the policy on the fly. Our core component, Differentiable Textual Optimization (DTO), leverages gradient signals from both the LLM's likelihood and a reward model to refine textual representations. $\nabla$-Reasoner further incorporates rejection sampling and acceleration design to robustify and speed up decoding. Theoretically, we show that performing inference-time gradient descent in the sample space to maximize reward is dual to aligning an LLM policy via KL-regularized reinforcement learning. Empirically, $\nabla$-Reasoner achieves over 20% accuracy improvement on a challenging mathematical reasoning benchmark, while reducing number of model calls by approximately 10-40% compared to strong baselines. Overall, our work introduces a paradigm shift from zeroth-order search to first-order optimization at test time, offering a cost-effective path to amplify LLM reasoning.

Prostate cancer being one of the frequently diagnosed malignancy in men, the rising demand for biopsies places a severe workload on pathologists. The grading procedure is tedious and subjective, motivating the development of automated systems. Although deep learning has made inroads in terms of performance, its limited interpretability poses challenges for widespread adoption in high-stake applications like medicine. Existing interpretability techniques for prostate cancer classifiers provide a coarse explanation but do not reveal why the highlighted regions matter. In this scenario, we propose a novel prototype-based weakly-supervised framework for an interpretable grading of prostate cancer from histopathology images. These networks can prove to be more trustworthy since their explicit reasoning procedure mirrors the workflow of a pathologist in comparing suspicious regions with clinically validated examples. The network is initially pre-trained at patch-level to learn robust prototypical features associated with each grade. In order to adapt it to a weakly-supervised setup for prostate cancer grading, the network is fine-tuned with a new prototype-aware loss function. Finally, a new attention-based dynamic pruning mechanism is introduced to handle inter-sample heterogeneity, while selectively emphasizing relevant prototypes for optimal performance. Extensive validation on the benchmark PANDA and SICAP datasets confirms that the framework can serve as a reliable assistive tool for pathologists in their routine diagnostic workflows.

Training automatic speech recognition (ASR) models increasingly relies on decentralized federated learning to ensure data privacy and accessibility, producing multiple local models that require effective merging. In hybrid ASR systems, while acoustic models can be merged using established methods, the language model (LM) for rescoring the N-best speech recognition list faces challenges due to the heterogeneity of non-neural n-gram models and neural network models. This paper proposes a heterogeneous LM optimization task and introduces a match-and-merge paradigm with two algorithms: the Genetic Match-and-Merge Algorithm (GMMA), using genetic operations to evolve and pair LMs, and the Reinforced Match-and-Merge Algorithm (RMMA), leveraging reinforcement learning for efficient convergence. Experiments on seven OpenSLR datasets show RMMA achieves the lowest average Character Error Rate and better generalization than baselines, converging up to seven times faster than GMMA, highlighting the paradigm's potential for scalable, privacy-preserving ASR systems.

Target speaker extraction (TSE) aims to isolate a specific speaker's voice from multi-speaker mixtures. Despite strong benchmark results, real-world performance often degrades due to different interacting factors. Previous curriculum learning approaches for TSE typically address these factors separately, failing to capture their complex interactions and relying on predefined difficulty factors that may not align with actual model learning behavior. To address this challenge, we first propose a multi-factor curriculum learning strategy that jointly schedules SNR thresholds, speaker counts, overlap ratios, and synthetic/real proportions, enabling progressive learning from simple to complex scenarios. However, determining optimal scheduling without predefined assumptions remains challenging. We therefore introduce TSE-Datamap, a visualization framework that grounds curriculum design in observed training dynamics by tracking confidence and variability across training epochs. Our analysis reveals three characteristic data regions: (i) easy-to-learn examples where models consistently perform well, (ii) ambiguous examples where models oscillate between alternative predictions, and (iii) hard-to-learn examples where models persistently struggle. Guided by these data-driven insights, our methods improve extraction results over random sampling, with particularly strong gains in challenging multi-speaker scenarios.