Haozhe Wang

Meituan LongCat Team, Bin Xiao, Chao Wang et al.

The prevailing Next-Token Prediction (NTP) paradigm has driven the success of large language models through discrete autoregressive modeling. However, contemporary multimodal systems remain language-centric, often treating non-linguistic modalities as external attachments, leading to fragmented architectures and suboptimal integration. To transcend this limitation, we introduce Discrete Native Autoregressive (DiNA), a unified framework that represents multimodal information within a shared discrete space, enabling a consistent and principled autoregressive modeling across modalities. A key innovation is the Discrete Native Any-resolution Visual Transformer (dNaViT), which performs tokenization and de-tokenization at arbitrary resolutions, transforming continuous visual signals into hierarchical discrete tokens. Building on this foundation, we develop LongCat-Next, a native multimodal model that processes text, vision, and audio under a single autoregressive objective with minimal modality-specific design. As an industrial-strength foundation model, it excels at seeing, painting, and talking within a single framework, achieving strong performance across a wide range of multimodal benchmarks. In particular, LongCat-Next addresses the long-standing performance ceiling of discrete vision modeling on understanding tasks and provides a unified approach to effectively reconcile the conflict between understanding and generation. As an attempt toward native multimodality, we open-source the LongCat-Next and its tokenizers, hoping to foster further research and development in the community. GitHub: https://github.com/meituan-longcat/LongCat-Next

Haozhe Wang, Cong Wei, Weiming Ren et al.

Most reward models for visual generation reduce rich human judgments to a single unexplained score, discarding the reasoning that underlies preference. We show that teaching reward models to produce explicit, multi-dimensional critiques before scoring transforms them from passive evaluators into active optimization tools, improving generators in two complementary ways: at training time, structured rationales provide interpretable, fine-grained rewards for reinforcement learning; at test time, a Generate-Critique-Refine loop turns critiques into targeted prompt revisions that improve outputs without any parameter updates. To train such a reward model without costly rationale annotations, we introduce Preference-Anchored Rationalization (PARROT), a principled framework that recovers high-quality rationales from readily available preference data through anchored generation, consistency filtering, and distillation. The resulting model, RationalRewards (8B), achieves state-of-the-art preference prediction among open-source reward models, competitive with Gemini-2.5-Pro, while using 10-20x less training data than comparable baselines. As an RL reward, it consistently improves text-to-image and image-editing generators beyond scalar alternatives. Most strikingly, its test-time critique-and-refine loop matches or exceeds RL-based fine-tuning on several benchmarks, suggesting that structured reasoning can unlock latent capabilities in existing generators that suboptimal prompts fail to elicit.

Jialin Yang, Dongfu Jiang, Lipeng He et al. · amazon-science, utoronto

As Large Language Models (LLMs) become integral to software development workflows, their ability to generate structured outputs has become critically important. We introduce StructEval, a comprehensive benchmark for evaluating LLMs' capabilities in producing both non-renderable (JSON, YAML, CSV) and renderable (HTML, React, SVG) structured formats. Unlike prior benchmarks, StructEval systematically evaluates structural fidelity across diverse formats through two paradigms: 1) generation tasks, producing structured output from natural language prompts, and \textbf{2)} conversion tasks, translating between structured formats. Our benchmark encompasses 18 formats and 44 types of task, with novel metrics for format adherence and structural correctness. Results reveal significant performance gaps-even state-of-the-art models like o1-mini achieve only 75.58 average score, with open-source alternatives lagging approximately 10 points behind. We find generation tasks more challenging than conversion tasks, and producing correct visual content more difficult than generating text-only structures.

Haozhe Wang, Chao Du, Panyan Fang et al. · tsinghua

Real-Time Bidding (RTB) is an important mechanism in modern online advertising systems. Advertisers employ bidding strategies in RTB to optimize their advertising effects subject to various financial requirements, especially the return-on-investment (ROI) constraint. ROIs change non-monotonically during the sequential bidding process, and often induce a see-saw effect between constraint satisfaction and objective optimization. While some existing approaches show promising results in static or mildly changing ad markets, they fail to generalize to highly dynamic ad markets with ROI constraints, due to their inability to adaptively balance constraints and objectives amidst non-stationarity and partial observability. In this work, we specialize in ROI-Constrained Bidding in non-stationary markets. Based on a Partially Observable Constrained Markov Decision Process, our method exploits an indicator-augmented reward function free of extra trade-off parameters and develops a Curriculum-Guided Bayesian Reinforcement Learning (CBRL) framework to adaptively control the constraint-objective trade-off in non-stationary ad markets. Extensive experiments on a large-scale industrial dataset with two problem settings reveal that CBRL generalizes well in both in-distribution and out-of-distribution data regimes, and enjoys superior learning efficiency and stability.

Haozhe Wang, Chao Du, Panyan Fang et al. · tsinghua

The proliferation of the Internet has led to the emergence of online advertising, driven by the mechanics of online auctions. In these repeated auctions, software agents participate on behalf of aggregated advertisers to optimize for their long-term utility. To fulfill the diverse demands, bidding strategies are employed to optimize advertising objectives subject to different spending constraints. Existing approaches on constrained bidding typically rely on i.i.d. train and test conditions, which contradicts the adversarial nature of online ad markets where different parties possess potentially conflicting objectives. In this regard, we explore the problem of constrained bidding in adversarial bidding environments, which assumes no knowledge about the adversarial factors. Instead of relying on the i.i.d. assumption, our insight is to align the train distribution of environments with the potential test distribution meanwhile minimizing policy regret. Based on this insight, we propose a practical Minimax Regret Optimization (MiRO) approach that interleaves between a teacher finding adversarial environments for tutoring and a learner meta-learning its policy over the given distribution of environments. In addition, we pioneer to incorporate expert demonstrations for learning bidding strategies. Through a causality-aware policy design, we improve upon MiRO by distilling knowledge from the experts. Extensive experiments on both industrial data and synthetic data show that our method, MiRO with Causality-aware reinforcement Learning (MiROCL), outperforms prior methods by over 30%.

Lei Zhou, Zhiyang Liu, Runze Gan et al.

Category-level object pose estimation involves estimating the 6D pose and the 3D metric size of objects from predetermined categories. While recent approaches take categorical shape prior information as reference to improve pose estimation accuracy, the single-stage network design and training manner lead to sub-optimal performance since there are two distinct tasks in the pipeline. In this paper, the advantage of two-stage pipeline over single-stage design is discussed. To this end, we propose a two-stage deformation-and registration pipeline called DR-Pose, which consists of completion-aided deformation stage and scaled registration stage. The first stage uses a point cloud completion method to generate unseen parts of target object, guiding subsequent deformation on the shape prior. In the second stage, a novel registration network is designed to extract pose-sensitive features and predict the representation of object partial point cloud in canonical space based on the deformation results from the first stage. DR-Pose produces superior results to the state-of-the-art shape prior-based methods on both CAMERA25 and REAL275 benchmarks. Codes are available at https://github.com/Zray26/DR-Pose.git.

Yu Wu, Yana Wei, Haozhe Wang et al.

This paper introduces Grounded Image Text Matching with Mismatched Relation (GITM-MR), a novel visual-linguistic joint task that evaluates the relation understanding capabilities of transformer-based pre-trained models. GITM-MR requires a model to first determine if an expression describes an image, then localize referred objects or ground the mismatched parts of the text. We provide a benchmark for evaluating pre-trained models on this task, with a focus on the challenging settings of limited data and out-of-distribution sentence lengths. Our evaluation demonstrates that pre-trained models lack data efficiency and length generalization ability. To address this, we propose the Relation-sensitive Correspondence Reasoning Network (RCRN), which incorporates relation-aware reasoning via bi-directional message propagation guided by language structure. RCRN can be interpreted as a modular program and delivers strong performance in both length generalization and data efficiency.

Taofeng Xue, Chong Peng, Mianqiu Huang et al.

The development of native computer-use agents (CUA) represents a significant leap in multimodal AI. However, their potential is currently bottlenecked by the constraints of static data scaling. Existing paradigms relying primarily on passive imitation of static datasets struggle to capture the intricate causal dynamics inherent in long-horizon computer tasks. In this work, we introduce EvoCUA, a native computer use agentic model. Unlike static imitation, EvoCUA integrates data generation and policy optimization into a self-sustaining evolutionary cycle. To mitigate data scarcity, we develop a verifiable synthesis engine that autonomously generates diverse tasks coupled with executable validators. To enable large-scale experience acquisition, we design a scalable infrastructure orchestrating tens of thousands of asynchronous sandbox rollouts. Building on these massive trajectories, we propose an iterative evolving learning strategy to efficiently internalize this experience. This mechanism dynamically regulates policy updates by identifying capability boundaries -- reinforcing successful routines while transforming failure trajectories into rich supervision through error analysis and self-correction. Empirical evaluations on the OSWorld benchmark demonstrate that EvoCUA achieves a success rate of 56.7%, establishing a new open-source state-of-the-art. Notably, EvoCUA significantly outperforms the previous best open-source model, OpenCUA-72B (45.0%), and surpasses leading closed-weights models such as UI-TARS-2 (53.1%). Crucially, our results underscore the generalizability of this approach: the evolving paradigm driven by learning from experience yields consistent performance gains across foundation models of varying scales, establishing a robust and scalable path for advancing native agent capabilities.

Zhili Ng, Haozhe Wang, Zhengshen Zhang et al.

In this work, we present SynTable, a unified and flexible Python-based dataset generator built using NVIDIA's Isaac Sim Replicator Composer for generating high-quality synthetic datasets for unseen object amodal instance segmentation of cluttered tabletop scenes. Our dataset generation tool can render complex 3D scenes containing object meshes, materials, textures, lighting, and backgrounds. Metadata, such as modal and amodal instance segmentation masks, object amodal RGBA instances, occlusion masks, depth maps, bounding boxes, and material properties can be automatically generated to annotate the scene according to the users' requirements. Our tool eliminates the need for manual labeling in the dataset generation process while ensuring the quality and accuracy of the dataset. In this work, we discuss our design goals, framework architecture, and the performance of our tool. We demonstrate the use of a sample dataset generated using SynTable for training a state-of-the-art model, UOAIS-Net. Our state-of-the-art results show significantly improved performance in Sim-to-Real transfer when evaluated on the OSD-Amodal dataset. We offer this tool as an open-source, easy-to-use, photorealistic dataset generator for advancing research in deep learning and synthetic data generation. The links to our source code, demonstration video, and sample dataset can be found in the supplementary materials.

Haozhe Wang, Chao Qu, Zuming Huang et al.

Recently, slow-thinking systems like GPT-o1 and DeepSeek-R1 have demonstrated great potential in solving challenging problems through explicit reflection. They significantly outperform the best fast-thinking models, such as GPT-4o, on various math and science benchmarks. However, their multimodal reasoning capabilities remain on par with fast-thinking models. For instance, GPT-o1's performance on benchmarks like MathVista, MathVerse, and MathVision is similar to fast-thinking models. In this paper, we aim to enhance the slow-thinking capabilities of vision-language models using reinforcement learning (without relying on distillation) to advance the state of the art. First, we adapt the GRPO algorithm with a novel technique called Selective Sample Replay (SSR) to address the vanishing advantages problem. While this approach yields strong performance, the resulting RL-trained models exhibit limited self-reflection or self-verification. To further encourage slow-thinking, we introduce Forced Rethinking, which appends a rethinking trigger token to the end of rollouts in RL training, explicitly enforcing a self-reflection reasoning step. By combining these two techniques, our model, VL-Rethinker, advances state-of-the-art scores on MathVista, MathVerse to achieve 80.4%, 63.5% respectively. VL-Rethinker also achieves open-source SoTA on multi-disciplinary benchmarks such as MathVision, MMMU-Pro, EMMA, and MEGA-Bench, narrowing the gap with OpenAI-o1. Our empirical results show the effectiveness of our approaches.

Ziyang Yan, Wenzhen Dong, Yihua Shao et al.

End-to-end autonomous driving with vision-only is not only more cost-effective compared to LiDAR-vision fusion but also more reliable than traditional methods. To achieve a economical and robust purely visual autonomous driving system, we propose RenderWorld, a vision-only end-to-end autonomous driving framework, which generates 3D occupancy labels using a self-supervised gaussian-based Img2Occ Module, then encodes the labels by AM-VAE, and uses world model for forecasting and planning. RenderWorld employs Gaussian Splatting to represent 3D scenes and render 2D images greatly improves segmentation accuracy and reduces GPU memory consumption compared with NeRF-based methods. By applying AM-VAE to encode air and non-air separately, RenderWorld achieves more fine-grained scene element representation, leading to state-of-the-art performance in both 4D occupancy forecasting and motion planning from autoregressive world model.

Haozhe Wang, Alex Su, Weiming Ren et al.

Chain-of-thought reasoning has significantly improved the performance of Large Language Models (LLMs) across various domains. However, this reasoning process has been confined exclusively to textual space, limiting its effectiveness in visually intensive tasks. To address this limitation, we introduce the concept of reasoning in the pixel-space. Within this novel framework, Vision-Language Models (VLMs) are equipped with a suite of visual reasoning operations, such as zoom-in and select-frame. These operations enable VLMs to directly inspect, interrogate, and infer from visual evidences, thereby enhancing reasoning fidelity for visual tasks. Cultivating such pixel-space reasoning capabilities in VLMs presents notable challenges, including the model's initially imbalanced competence and its reluctance to adopt the newly introduced pixel-space operations. We address these challenges through a two-phase training approach. The first phase employs instruction tuning on synthesized reasoning traces to familiarize the model with the novel visual operations. Following this, a reinforcement learning (RL) phase leverages a curiosity-driven reward scheme to balance exploration between pixel-space reasoning and textual reasoning. With these visual operations, VLMs can interact with complex visual inputs, such as information-rich images or videos to proactively gather necessary information. We demonstrate that this approach significantly improves VLM performance across diverse visual reasoning benchmarks. Our 7B model, \model, achieves 84\% on V* bench, 74\% on TallyQA-Complex, and 84\% on InfographicsVQA, marking the highest accuracy achieved by any open-source model to date. These results highlight the importance of pixel-space reasoning and the effectiveness of our framework.

Shida Gao, Feng Xue, Xiangfeng Wang et al.

Spatio-temporal grounding and reasoning aims to locate the temporal segment and spatial region of an event in a video given a user query, while also reasoning about semantics such as causality, temporal order, and action relationships. To achieve this, current MLLMs primarily treats bounding boxes as text tokens and generates them autoregressively. However, such autoregressive spatial decoding leads to very-long output sequences, causing spatial errors to accumulated over time and the localization results to progressively drift across a video. To address this, we present a Detector-Empowered Video LLM, short for DEViL, which couples a Video LLM with an open-vocabulary detector (OVD). Specifically, the MLLM and detector are connected via a reference-semantic token (RST) that distills the user query into a rich semantic representation. Unlike tokens that merely serve as spatial prompts or segmentor switches, the RST functions as both a control signal and a replacement for the OVD's text embedding, enabling end-to-end learning of both referential understanding and spatial localization. Furthermore, we propose a tube-mined temporal regularization (TTReg) within OVD, which drives the OVD to generate temporally-consistent queries for target objects, thereby ensuring effective temporal association. Experiments demonstrate that DEViL achieves strong performance across various fine-grained video understanding tasks, particularly STVG and GroundedVQA. Code will be released on https://github.com/gaostar123/DeViL.

Yuanhao Li, Haozhe Wang, Geyong Min et al.

The integration of Generative AI models into AI-native network systems offers a transformative path toward achieving autonomous and adaptive control. However, the application of such models to continuous control tasks is impeded by intrinsic architectural limitations, including finite context windows, the lack of explicit reward signals, and the degradation of the long context. This paper posits that the key to unlocking robust continuous control is enabling agents to internalize experience by distilling it into their parameters, rather than relying on prompt-based memory. To this end, we propose a novel self-finetuning framework that enables agentic systems to learn continuously through direct interaction with the environment, bypassing the need for handcrafted rewards. Our framework implements a bi-perspective reflection mechanism that generates autonomous linguistic feedback to construct preference datasets from interaction history. A subsequent preference-based fine-tuning process distills long-horizon experiences into the model's parameters. We evaluate our approach on a dynamic Radio Access Network (RAN) slicing task, a challenging multi-objective control problem that requires the resolution of acute trade-offs between spectrum efficiency, service quality, and reconfiguration stability under volatile network conditions. Experimental results show that our framework outperforms standard Reinforcement Learning (RL) baselines and existing Large Language Model (LLM)-based agents in sample efficiency, stability, and multi-metric optimization. These findings demonstrate the potential of self-improving generative agents for continuous control tasks, paving the way for future AI-native network infrastructure.

Zhiyuan Zeng, Yichi Zhang, Yong Shan et al.

While Large Language Models (LLMs) have achieved remarkable success in code generation, they often struggle with the deep, long-horizon reasoning required for complex software engineering. We attribute this limitation to the nature of standard pre-training data: static software repositories represent only the terminal state of an intricate intellectual process, abstracting away the intermediate planning, debugging, and iterative refinement. To bridge this gap, we propose a novel paradigm: understanding via reconstruction. We hypothesize that reverse-engineering the latent agentic trajectories -- the planning, reasoning, and debugging steps -- behind static repositories provides a far richer supervision signal than raw code alone. To operationalize this, we introduce a framework that synthesizes these trajectories using a multi-agent simulation. This process is grounded in the structural realities of the source repositories (e.g., dependency graphs and file hierarchies) to ensure fidelity. Furthermore, to guarantee the logical rigor of the synthetic data, we employ a search-based optimization technique that iteratively refines the Chain-of-Thought (CoT) reasoning to maximize the likelihood of the ground-truth code. Empirical results demonstrate that continuous pre-training on these reconstructed trajectories significantly enhances Llama-3-8B's performance across diverse benchmarks, including long-context understanding, coding proficiency, and agentic capabilities.

Dongfu Jiang, Yi Lu, Zhuofeng Li et al. · utoronto

Reinforcement Learning with Verifiable Rewards (RLVR) has demonstrated success in enhancing LLM reasoning capabilities, but remains limited to single-turn interactions without tool integration. While recent Agentic Reinforcement Learning with Tool use (ARLT) approaches have emerged to address multi-turn tool interactions, existing works develop task-specific codebases that suffer from fragmentation, synchronous execution bottlenecks, and limited extensibility across domains. These inefficiencies hinder broader community adoption and algorithmic innovation. We introduce VerlTool, a unified and modular framework that addresses these limitations through systematic design principles. VerlTool provides four key contributions: (1) upstream alignment with VeRL ensuring compatibility and simplified maintenance, (2) unified tool management via standardized APIs supporting diverse modalities including code execution, search, SQL databases, and vision processing, (3) asynchronous rollout execution achieving near 2$\times$ speedup by eliminating synchronization bottlenecks, and (4) comprehensive evaluation demonstrating competitive performance across 6 ARLT domains. Our framework formalizes ARLT as multi-turn trajectories with multi-modal observation tokens (text/image/video), extending beyond single-turn RLVR paradigms. We train and evaluate models on mathematical reasoning, knowledge QA, SQL generation, visual reasoning, web search, and software engineering tasks, achieving results comparable to specialized systems while providing unified training infrastructure. The modular plugin architecture enables rapid tool integration requiring only lightweight Python definitions, significantly reducing development overhead and providing a scalable foundation for tool-augmented RL research. Our code is open-sourced at https://github.com/TIGER-AI-Lab/verl-tool.

Xiaoyu Tan, Tianchu Yao, Chao Qu et al.

The reasoning capabilities of advanced large language models (LLMs) like o1 have revolutionized artificial intelligence applications. Nevertheless, evaluating and optimizing complex reasoning processes remain significant challenges due to diverse policy distributions and the inherent limitations of human effort and accuracy. In this paper, we present AURORA, a novel automated framework for training universal process reward models (PRMs) using ensemble prompting and reverse verification. The framework employs a two-phase approach: First, it uses diverse prompting strategies and ensemble methods to perform automated annotation and evaluation of processes, ensuring robust assessments for reward learning. Second, it leverages practical reference answers for reverse verification, enhancing the model's ability to validate outputs and improving training accuracy. To assess the framework's performance, we extend beyond the existing ProcessBench benchmark by introducing UniversalBench, which evaluates reward predictions across full trajectories under diverse policy distribtion with long Chain-of-Thought (CoT) outputs. Experimental results demonstrate that AURORA enhances process evaluation accuracy, improves PRMs' accuracy for diverse policy distributions and long-CoT responses. The project will be open-sourced at https://auroraprm.github.io/. The Universal-PRM-7B is available at https://huggingface.co/infly/Universal-PRM-7B.

Songcheng Cai, Zhiheng Lyu, Yuansheng Ni et al.

Agentic repository-level code understanding is essential for automating complex software engineering tasks, yet the field lacks reliable benchmarks. Existing evaluations often overlook the long tail topics and rely on popular repositories where Large Language Models (LLMs) can cheat via memorized knowledge. To address this, we introduce SWE-QA-Pro, a benchmark constructed from diverse, long-tail repositories with executable environments. We enforce topical balance via issue-driven clustering to cover under-represented task types and apply a rigorous difficulty calibration process: questions solvable by direct-answer baselines are filtered out. This results in a dataset where agentic workflows significantly outperform direct answering (e.g., a ~13-point gap for Claude Sonnet 4.5), confirming the necessity of agentic codebase exploration. Furthermore, to tackle the scarcity of training data for such complex behaviors, we propose a scalable synthetic data pipeline that powers a two-stage training recipe: Supervised Fine-Tuning (SFT) followed by Reinforcement Learning from AI Feedback (RLAIF). This approach allows small open models to learn efficient tool usage and reasoning. Empirically, a Qwen3-8B model trained with our recipe surpasses GPT-4o by 2.3 points on SWE-QA-Pro and substantially narrows the gap to state-of-the-art proprietary models, demonstrating both the validity of our evaluation and the effectiveness of our agentic training workflow.

Yuhuan Wu, Cong Wei, Fangzhen Lin et al.

Vision-Language Models (VLMs) deployed as situated agents in high-resolution visual environments require active perception -- the ability to dynamically decide where to look through operations like zooming, cropping, and panning. However, current training paradigms produce models that mimic the surface form of such operations without functionally depending on their outputs, a phenomenon we term lazy perception. We trace this to a fundamental learning asymmetry: when coarse global views combined with language priors suffice for moderate accuracy, the model has no incentive to learn harder multi-step visual search. If a model can succeed without actively looking, it will never learn to look. This motivates Starve to Perceive, a training paradigm that constrains visual bandwidth -- restricting each observation to a tight token budget so that no single view suffices for task completion, making active perception the only viable strategy. Despite requiring no auxiliary losses, reward shaping, or architectural changes -- serving as a minimal, plug-in modification to standard post-training pipelines -- models trained under perceptual starvation achieve substantial gains of 5% average relative improvement across diverse benchmarks.

Jintao Cheng, Haozhe Wang, Weibin Li et al.

Vision-Language-Action (VLA) models have rapidly advanced embodied intelligence, enabling robots to execute complex, instruction-driven tasks. However, as model capacity and visual context length grow, the inference cost of VLA systems becomes a major bottleneck for real-world deployment on resource-constrained platforms. Existing visual token pruning methods mainly rely on semantic saliency or simple temporal cues, overlooking the continuous physical interaction, a fundamental property of VLA tasks. Consequently, current approaches often prune visually sparse yet structurally critical regions that support manipulation, leading to unstable behavior during early task phases. To overcome this, we propose a shift toward an explicit Interaction-First paradigm. Our proposed \textbf{training-free} method, VLA-IAP (Interaction-Aligned Pruning), introduces a geometric prior mechanism to preserve structural anchors and a dynamic scheduling strategy that adapts pruning intensity based on semantic-motion alignment. This enables a conservative-to-aggressive transition, ensuring robustness during early uncertainty and efficiency once interaction is locked. Extensive experiments show that VLA-IAP achieves a \textbf{97.8\% success rate} with a \textbf{$1.25\times$ speedup} on the LIBERO benchmark, and up to \textbf{$1.54\times$ speedup} while maintaining performance \textbf{comparable to the unpruned backbone}. Moreover, the method demonstrates superior and consistent performance across multiple model architectures and three different simulation environments, as well as a real robot platform, validating its strong generalization capability and practical applicability. Our project website is: \href{https://chengjt1999.github.io/VLA-IAP.github.io/}{VLA-IAP.com}.

Haozhe Wang, Haoran Que, Qixin Xu et al.

While the ``deep reasoning'' paradigm has spurred significant advances in verifiable domains like mathematics, its application to open-ended, creative generation remains a critical challenge. The two dominant methods for instilling reasoning -- reinforcement learning (RL) and instruction distillation -- falter in this area; RL struggles with the absence of clear reward signals and high-quality reward models, while distillation is prohibitively expensive and capped by the teacher model's capabilities. To overcome these limitations, we introduce REverse-Engineered Reasoning (REER), a new paradigm that fundamentally shifts the approach. Instead of building a reasoning process ``forwards'' through trial-and-error or imitation, REER works ``backwards'' from known-good solutions to computationally discover the latent, step-by-step deep reasoning process that could have produced them. Using this scalable, gradient-free approach, we curate and open-source DeepWriting-20K, a large-scale dataset of 20,000 deep reasoning trajectories for open-ended tasks. Our model, DeepWriter-8B, trained on this data, not only surpasses strong open-source baselines but also achieves performance competitive with, and at times superior to, leading proprietary models like GPT-4o and Claude 3.5.

Baode Wang, Biao Wu, Weizhen Li et al.

Document parsing from scanned images into structured formats remains a significant challenge due to its complexly intertwined elements such as text paragraphs, figures, formulas, and tables. Existing supervised fine-tuning methods often struggle to generalize across diverse document types, leading to poor performance, particularly on out-of-distribution data. This issue is further exacerbated by the limited availability of high-quality training data for layout-aware parsing tasks. To address these challenges, we introduce LayoutRL, a reinforcement learning framework that optimizes layout understanding through composite rewards integrating normalized edit distance, paragraph count accuracy, and reading order preservation. To support this training, we construct the Infinity-Doc-400K dataset, which we use to train Infinity-Parser, a vision-language model demonstrating robust generalization across various domains. Extensive evaluations on benchmarks including OmniDocBench, olmOCR-Bench, PubTabNet, and FinTabNet show that Infinity-Parser consistently achieves state-of-the-art performance across a broad range of document types, languages, and structural complexities, substantially outperforming both specialized document parsing systems and general-purpose vision-language models. We will release our code, dataset, and model to facilitate reproducible research in document parsing.

Baode Wang, Biao Wu, Weizhen Li et al.

Automated parsing of scanned documents into richly structured, machine-readable formats remains a critical bottleneck in Document AI, as traditional multi-stage pipelines suffer from error propagation and limited adaptability to diverse layouts. We introduce layoutRL, an end-to-end reinforcement learning framework that trains models to be explicitly layout-aware by optimizing a composite reward of normalized edit distance, paragraph count accuracy, and reading order preservation. Leveraging our newly released dataset, Infinity-Doc-55K, which combines 55K high-fidelity synthetic scanned document parsing data with expert-filtered real-world documents, we instantiate layoutRL in a vision-language-model-based parser called Infinity-Parser. Evaluated on English and Chinese benchmarks for OCR, table and formula extraction, and reading order detection, Infinity-Parser achieves new state-of-the-art performance in both accuracy and structural fidelity, outpacing specialist pipelines and general-purpose vision-language models. We will publicly release our code and dataset to accelerate progress in robust document understanding.

Haozhe Wang, Qixin Xu, Changpeng Wang et al.

Achieving robust perception-reasoning synergy is a central goal for advanced Vision-Language Models (VLMs). Recent advancements have pursued this goal via architectural designs or agentic workflows. However, these approaches are often limited by static textual reasoning or complicated by the significant compute and engineering burden of external agentic complexity. Worse, this heavy investment does not yield proportional gains, often witnessing a "seesaw effect" on perception and reasoning. This motivates a fundamental rethinking of the true bottleneck. In this paper, we argue that the root cause of this trade-off is an ambiguity in modality credit assignment: when a VLM fails, is it due to flawed perception ("bad seeing") or flawed logic ("bad thinking")? To resolve this, we introduce a reinforcement learning framework that improves perception-reasoning synergy by reliably rewarding the perception fidelity. We explicitly decompose the generation process into interleaved perception and reasoning steps. This decoupling enables targeted supervision on perception. Crucially, we introduce Perception Verification (PV), leveraging a "blindfolded reasoning" proxy to reward perceptual fidelity independently of reasoning outcomes. Furthermore, to scale training across free-form VL tasks, we propose Structured Verbal Verification, which replaces high-variance LLM judging with structured algorithmic execution. These techniques are integrated into a Modality-Aware Credit Assignment (MoCA) mechanism, which routes rewards to the specific source of error -- either bad seeing or bad thinking -- enabling a single VLM to achieve simultaneous performance gains across a wide task spectrum.

Asmit Kumar Singh, Haozhe Wang, Laxmi Naga Santosh Attaluri et al.

Large language models (LLMs) now sit in the critical path of search, assistance, and agentic workflows, making semantic caching essential for reducing inference cost and latency. Production deployments typically use a tiered static-dynamic design: a static cache of curated, offline vetted responses mined from logs, backed by a dynamic cache populated online. In practice, both tiers are commonly governed by a single embedding similarity threshold, which induces a hard tradeoff: conservative thresholds miss safe reuse opportunities, while aggressive thresholds risk serving semantically incorrect responses. We introduce \textbf{Krites}, an asynchronous, LLM-judged caching policy that expands static coverage without changing serving decisions. On the critical path, Krites behaves exactly like a standard static threshold policy. When the nearest static neighbor of the prompt falls just below the static threshold, Krites asynchronously invokes an LLM judge to verify whether the static response is acceptable for the new prompt. Approved matches are promoted into the dynamic cache, allowing future repeats and paraphrases to reuse curated static answers and expanding static reach over time. In trace-driven simulations on conversational and search workloads, Krites increases the fraction of requests served with curated static answers (direct static hits plus verified promotions) by up to $\textbf{3.9}$ times for conversational traffic and search-style queries relative to tuned baselines, with unchanged critical path latency.

Xiaomeng Zhu, Changwei Wang, Haozhe Wang et al.

A scene graph is a structured represention of objects and their relationships in a scene. Scene Graph Anticipation (SGA) involves predicting future scene graphs from video clips, enabling applications as intelligent surveillance and human-machine collaboration. Existing SGA approaches primarily leverage visual cues, often struggling to integrate valuable commonsense knowledge, thereby limiting long-term prediction robustness. To explicitly leverage such commonsense knowledge, we propose a new approach to better understand the objects, concepts, and relationships in a scene graph. Our approach decouples the SGA task in two steps: first a scene graph capturing model is used to convert a video clip into a sequence of scene graphs, then a pure text-based model is used to predict scene graphs in future frames. Our focus in this work is on the second step, and we call it Linguistic Scene Graph Anticipation (LSGA) and believes it should have independent interest beyond the use in SGA discussed here. For LSGA, we introduce an Object-Oriented Two-Staged Method (OOTSM) where an Large Language Model (LLM) first forecasts object appearances and disappearances before generating detailed human-object relations. We conduct extensive experiments to evaluate OOTSM in two settings. For LSGA, we evaluate our fine-tuned open-sourced LLMs against zero-shot APIs (i.e., GPT-4o, GPT-4o-mini, and DeepSeek-V3) on a benchmark constructed from Action Genome annotations. For SGA, we combine our OOTSM with STTran++ from, and our experiments demonstrate effective state-of-the-art performance: short-term mean-Recall (@10) increases by 3.4% while long-term mean-Recall (@50) improves dramatically by 21.9%. Code is available at https://github.com/ZhuXMMM/OOTSM.

Yifan Jia, Kailin Jiang, Yuyang Liang et al.

Large Multimodal Models(LMMs) face notable challenges when encountering multimodal knowledge conflicts, particularly under retrieval-augmented generation(RAG) frameworks where the contextual information from external sources may contradict the model's internal parametric knowledge, leading to unreliable outputs. However, existing benchmarks fail to reflect such realistic conflict scenarios. Most focus solely on intra-memory conflicts, while context-memory and inter-context conflicts remain largely investigated. Furthermore, commonly used factual knowledge-based evaluations are often overlooked, and existing datasets lack a thorough investigation into conflict detection capabilities. To bridge this gap, we propose MMKC-Bench, a benchmark designed to evaluate factual knowledge conflicts in both context-memory and inter-context scenarios. MMKC-Bench encompasses three types of multimodal knowledge conflicts and includes 1,573 knowledge instances and 3,381 images across 23 broad types, collected through automated pipelines with human verification. We evaluate three representative series of LMMs on both model behavior analysis and conflict detection tasks. Our findings show that while current LMMs are capable of recognizing knowledge conflicts, they tend to favor internal parametric knowledge over external evidence. We hope MMKC-Bench will foster further research in multimodal knowledge conflict and enhance the development of multimodal RAG systems. The source code is available at https://github.com/MLLMKCBENCH/MLLMKC.

Haozhe Wang, Zhiyang Liu, Lei Zhou et al.

Vision-based grasp estimation is an essential part of robotic manipulation tasks in the real world. Existing planar grasp estimation algorithms have been demonstrated to work well in relatively simple scenes. But when it comes to complex scenes, such as cluttered scenes with messy backgrounds and moving objects, the algorithms from previous works are prone to generate inaccurate and unstable grasping contact points. In this work, we first study the existing planar grasp estimation algorithms and analyze the related challenges in complex scenes. Secondly, we design a Pixel-wise Efficient Grasp Generation Network (PEGG-Net) to tackle the problem of grasping in complex scenes. PEGG-Net can achieve improved state-of-the-art performance on the Cornell dataset (98.9%) and second-best performance on the Jacquard dataset (93.8%), outperforming other existing algorithms without the introduction of complex structures. Thirdly, PEGG-Net could operate in a closed-loop manner for added robustness in dynamic environments using position-based visual servoing (PBVS). Finally, we conduct real-world experiments on static, dynamic, and cluttered objects in different complex scenes. The results show that our proposed network achieves a high success rate in grasping irregular objects, household objects, and workshop tools. To benefit the community, our trained model and supplementary materials are available at https://github.com/HZWang96/PEGG-Net.

Huaye Zeng, Dongfu Jiang, Haozhe Wang et al.

Most progress in recent coder models has been driven by supervised fine-tuning (SFT), while the potential of reinforcement learning (RL) remains largely unexplored, primarily due to the lack of reliable reward data/model in the code domain. In this paper, we address this challenge by leveraging automated large-scale test-case synthesis to enhance code model training. Specifically, we design a pipeline that generates extensive (question, test-cases) pairs from existing code data. Using these test cases, we construct preference pairs based on pass rates over sampled programs to train reward models with Bradley-Terry loss. It shows an average of 10-point improvement for Llama-3.1-8B-Ins and 5-point improvement for Qwen2.5-Coder-7B-Ins through best-of-32 sampling, making the 7B model on par with 236B DeepSeek-V2.5. Furthermore, we conduct reinforcement learning with both reward models and test-case pass rewards, leading to consistent improvements across HumanEval, MBPP, BigCodeBench, and LiveCodeBench (V4). Notably, we follow the R1-style training to start from Qwen2.5-Coder-base directly and show that our RL training can improve model on HumanEval-plus by over 25\% and MBPP-plus by 6\% for merely 80 optimization steps. We believe our results highlight the huge potential of reinforcement learning in coder models.

Haozhe Wang, Qixin Xu, Che Liu et al.

Reinforcement Learning (RL) has proven highly effective at enhancing the complex reasoning abilities of Large Language Models (LLMs), yet underlying mechanisms driving this success remain largely opaque. Our analysis reveals that puzzling phenomena like ``aha moments", ``length-scaling'' and entropy dynamics are not disparate occurrences but hallmarks of an emergent reasoning hierarchy, akin to the separation of high-level strategic planning from low-level procedural execution in human cognition. We uncover a compelling two-phase dynamic: initially, a model is constrained by procedural correctness and must improve its low-level skills. The learning bottleneck then decisively shifts, with performance gains being driven by the exploration and mastery of high-level strategic planning. This insight exposes a core inefficiency in prevailing RL algorithms like GRPO, which apply optimization pressure agnostically and dilute the learning signal across all tokens. To address this, we propose Hierarchy-Aware Credit Assignment (HICRA), an algorithm that concentrates optimization efforts on high-impact planning tokens. Our extensive experiments validate that HICRA significantly outperforms strong baselines, and offer deep insights into how reasoning advances through the lens of strategic exploration.

Che Liu, Haozhe Wang, Jiazhen Pan et al.

Improving performance on complex tasks and enabling interpretable decision making in large language models (LLMs), especially for clinical applications, requires effective reasoning. Yet this remains challenging without supervised fine-tuning (SFT) on costly chain-of-thought (CoT) data distilled from closed-source models (e.g., GPT-4o). In this work, we present AlphaMed, the first medical LLM to show that reasoning capability can emerge purely through reinforcement learning (RL), using minimalist rule-based rewards on public multiple-choice QA datasets, without relying on SFT or distilled CoT data. AlphaMed achieves state-of-the-art results on six medical QA benchmarks, outperforming models trained with conventional SFT+RL pipelines. On challenging benchmarks (e.g., MedXpert), AlphaMed even surpasses larger or closed-source models such as DeepSeek-V3-671B and Claude-3.5-Sonnet. To understand the factors behind this success, we conduct a comprehensive data-centric analysis guided by three questions: (i) Can minimalist rule-based RL incentivize reasoning without distilled CoT supervision? (ii) How do dataset quantity and diversity impact reasoning? (iii) How does question difficulty shape the emergence and generalization of reasoning? Our findings show that dataset informativeness is a key driver of reasoning performance, and that minimalist RL on informative, multiple-choice QA data is effective at inducing reasoning without CoT supervision. We also observe divergent trends across benchmarks, underscoring limitations in current evaluation and the need for more challenging, reasoning-oriented medical QA benchmarks.

Haozhe Wang, Long Li, Chao Qu et al.

Recent advances in mathematical problem-solving with language models (LMs) integrate chain-of-thought (CoT) reasoning and code execution to harness their complementary strengths. However, existing hybrid frameworks exhibit a critical limitation: they depend on externally dictated instructions or rigid code-integration templates, lacking metacognitive awareness -- the capacity to dynamically evaluate intrinsic capabilities and autonomously determine when and how to integrate tools. This rigidity motivates our study of autonomous code integration, enabling models to adapt tool-usage strategies as their reasoning abilities evolve during training. While reinforcement learning (RL) shows promise for boosting LLM reasoning at scale (e.g., DeepSeek-R1), we demonstrate its inefficiency in learning autonomous code integration due to inadequate exploration of the vast combinatorial space of CoT-code interleaving patterns. To address this challenge, we propose a novel Expectation-Maximization (EM) framework that synergizes structured exploration (E-step) with off-policy RL optimization (M-step), creating a self-reinforcing cycle between metacognitive tool-use decisions and evolving capabilities. Experiments reveal our method achieves superior results through improved exploration. Notably, our 7B model improves over 11% on MATH500 and 9.4% on AIME without o1-like CoT.

Yihua Shao, Deyang Lin, Fanhu Zeng et al.

Diffusion models have been widely adopted in image and video generation. However, their complex network architecture leads to high inference overhead for its generation process. Existing diffusion quantization methods primarily focus on the quantization of the model structure while ignoring the impact of time-steps variation during sampling. At the same time, most current approaches fail to account for significant activations that cannot be eliminated, resulting in substantial performance degradation after quantization. To address these issues, we propose Time-Rotation Diffusion Quantization (TR-DQ), a novel quantization method incorporating time-step and rotation-based optimization. TR-DQ first divides the sampling process based on time-steps and applies a rotation matrix to smooth activations and weights dynamically. For different time-steps, a dedicated hyperparameter is introduced for adaptive timing modeling, which enables dynamic quantization across different time steps. Additionally, we also explore the compression potential of Classifier-Free Guidance (CFG-wise) to establish a foundation for subsequent work. TR-DQ achieves state-of-the-art (SOTA) performance on image generation and video generation tasks and a 1.38-1.89x speedup and 1.97-2.58x memory reduction in inference compared to existing quantization methods.

Lei Zhou, Haozhe Wang, Zhengshen Zhang et al.

In the realm of robotic grasping, achieving accurate and reliable interactions with the environment is a pivotal challenge. Traditional methods of grasp planning methods utilizing partial point clouds derived from depth image often suffer from reduced scene understanding due to occlusion, ultimately impeding their grasping accuracy. Furthermore, scene reconstruction methods have primarily relied upon static techniques, which are susceptible to environment change during manipulation process limits their efficacy in real-time grasping tasks. To address these limitations, this paper introduces a novel two-stage pipeline for dynamic scene reconstruction. In the first stage, our approach takes scene scanning as input to register each target object with mesh reconstruction and novel object pose tracking. In the second stage, pose tracking is still performed to provide object poses in real-time, enabling our approach to transform the reconstructed object point clouds back into the scene. Unlike conventional methodologies, which rely on static scene snapshots, our method continuously captures the evolving scene geometry, resulting in a comprehensive and up-to-date point cloud representation. By circumventing the constraints posed by occlusion, our method enhances the overall grasp planning process and empowers state-of-the-art 6-DoF robotic grasping algorithms to exhibit markedly improved accuracy.

Che Liu, Zhongwei Wan, Haozhe Wang et al.

Medical Vision-Language Pre-training (MedVLP) has made significant progress in enabling zero-shot tasks for medical image understanding. However, training MedVLP models typically requires large-scale datasets with paired, high-quality image-text data, which are scarce in the medical domain. Recent advancements in Large Language Models (LLMs) and diffusion models have made it possible to generate large-scale synthetic image-text pairs. This raises the question: "Can MedVLP succeed using purely synthetic data?" To address this, we use off-the-shelf generative models to create synthetic radiology reports and paired Chest X-ray (CXR) images, and propose an automated pipeline to build a diverse, high-quality synthetic dataset, enabling a rigorous study that isolates model and training settings, focusing entirely from the data perspective. Our results show that MedVLP models trained exclusively on synthetic data outperform those trained on real data by 3.8% in averaged AUC on zero-shot classification. Moreover, using a combination of synthetic and real data leads to a further improvement of 9.07%. Additionally, MedVLP models trained on synthetic or mixed data consistently outperform those trained on real data in zero-shot grounding, as well as in fine-tuned classification and segmentation tasks. Our analysis suggests MedVLP trained on well-designed synthetic data can outperform models trained on real datasets, which may be limited by low-quality samples and long-tailed distributions.

Long Li, Xuzheng He, Haozhe Wang et al.

Program-of-Thought (PoT) replaces natural language-based Chain-of-Thought (CoT) as the most popular method in Large Language Models (LLMs) mathematical reasoning tasks by utilizing external tool calls to circumvent computational errors. However, our evaluation of the GPT-4 and Llama series reveals that using PoT introduces more reasoning errors, such as incorrect formulas or flawed logic, compared to CoT. To address this issue, we propose Human-Think Language (HTL), which leverages a suite of strategies that help integrate PoT and CoT, encompassing: (1) a new generation paradigm that uses full CoT reasoning to control code generation. (2) Focus Attention, that directs model attention to the CoT reasoning during PoT to generate more logical code. (3) reinforcement learning that utilizes the accuracy of both CoT and PoT responses as rewards to prevent repetitive reasoning steps in LLMs when solving difficult math problems. Our method achieves an average improvement of 6.5% on the Llama-Base model and 4.3% on the Mistral-Base model across 8 mathematical calculation datasets. It also shows significant effectiveness on five out-of-domain datasets by controlling the model's information flow, exhibiting strong transferability. Additionally, HTL shows the most significant improvement in non-mathematical natural language inference task, contributing to a unified reasoning task framework

Che Liu, Cheng Ouyang, Zhongwei Wan et al.

Recent advances in multimodal ECG representation learning center on aligning ECG signals with paired free-text reports. However, suboptimal alignment persists due to the complexity of medical language and the reliance on a full 12-lead setup, which is often unavailable in under-resourced settings. To tackle these issues, we propose **K-MERL**, a knowledge-enhanced multimodal ECG representation learning framework. **K-MERL** leverages large language models to extract structured knowledge from free-text reports and employs a lead-aware ECG encoder with dynamic lead masking to accommodate arbitrary lead inputs. Evaluations on six external ECG datasets show that **K-MERL** achieves state-of-the-art performance in zero-shot classification and linear probing tasks, while delivering an average **16%** AUC improvement over existing methods in partial-lead zero-shot classification.

Jingyun Yang, Ruoyan Avery Yin, Chi Jiang et al.

Characterization of atomic-scale materials traditionally requires human experts with months to years of specialized training. Even for trained human operators, accurate and reliable characterization remains challenging when examining newly discovered materials such as two-dimensional (2D) structures. This bottleneck drives demand for fully autonomous experimentation systems capable of comprehending research objectives without requiring large training datasets. In this work, we present ATOMIC (Autonomous Technology for Optical Microscopy & Intelligent Characterization), an end-to-end framework that integrates foundation models to enable fully autonomous, zero-shot characterization of 2D materials. Our system integrates the vision foundation model (i.e., Segment Anything Model), large language models (i.e., ChatGPT), unsupervised clustering, and topological analysis to automate microscope control, sample scanning, image segmentation, and intelligent analysis through prompt engineering, eliminating the need for additional training. When analyzing typical MoS2 samples, our approach achieves 99.7% segmentation accuracy for single layer identification, which is equivalent to that of human experts. In addition, the integrated model is able to detect grain boundary slits that are challenging to identify with human eyes. Furthermore, the system retains robust accuracy despite variable conditions including defocus, color temperature fluctuations, and exposure variations. It is applicable to a broad spectrum of common 2D materials-including graphene, MoS2, WSe2, SnSe-regardless of whether they were fabricated via chemical vapor deposition or mechanical exfoliation. This work represents the implementation of foundation models to achieve autonomous analysis, establishing a scalable and data-efficient characterization paradigm that fundamentally transforms the approach to nanoscale materials research.

Qixin Xu, Haozhe Wang, Che Liu et al.

Current document reasoning paradigms are constrained by a fundamental trade-off between scalability (processing long-context documents) and fidelity (capturing fine-grained, multimodal details). To bridge this gap, we propose CogDoc, a unified coarse-to-fine thinking framework that mimics human cognitive processes: a low-resolution "Fast Reading" phase for scalable information localization,followed by a high-resolution "Focused Thinking" phase for deep reasoning. We conduct a rigorous investigation into post-training strategies for the unified thinking framework, demonstrating that a Direct Reinforcement Learning (RL) approach outperforms RL with Supervised Fine-Tuning (SFT) initialization. Specifically, we find that direct RL avoids the "policy conflict" observed in SFT. Empirically, our 7B model achieves state-of-the-art performance within its parameter class, notably surpassing significantly larger proprietary models (e.g., GPT-4o) on challenging, visually rich document benchmarks.

Chenrui Zhang, Yiyuan Zhang, Yunfei Ye et al.

Underwater robots are widely deployed for ocean exploration and manipulation. Underactuated mechanisms are particularly advantageous in aquatic environments, as reducing actuator count lowers the risk of motor leakage while introducing inherent mechanical compliance. However, accurate modeling of underwater underactuated and soft robotic systems remains challenging because it requires identifying a high-dimensional set of internal structural and external hydrodynamic parameters. In this work, we propose a trajectory-driven global optimization framework for unified structural-hydrodynamic modeling of underwater multibody systems. Inspired by the Covariance Matrix Adaptation Evolution Strategy (CMA-ES), the proposed approach simultaneously identifies coupled internal elastic, damping, and distributed hydrodynamic parameters through trajectory-level matching between simulation and experimental motion. This enables high-fidelity reproduction of both underactuated mechanisms and compliant soft robotic systems in underwater environments. We first validate the framework on a link-by-link underactuated multibody mechanism, demonstrating accurate identification of distributed hydrodynamic coefficients, with a normalized end effector position error below 5% across multiple trajectories, varying initial conditions, and both active-passive and fully passive configurations. The identified modeling strategy is then transferred to a single octopus-inspired soft arm, showing strong real-to-sim consistency without manual retuning. Finally, eight identified arms are assembled into a swimming octopus robot, where the unified parameter set enables realistic whole body behavior without additional parameter calibration. These results demonstrate the scalability and transferability of the proposed structural-hydrodynamic modeling framework across underwater underactuated and soft robotic systems.

Changpeng Wang, Haozhe Wang, Xi Chen et al.

Recent advances in vision-language reasoning underscore the importance of thinking with images, where models actively ground their reasoning in visual evidence. Yet, prevailing frameworks treat visual actions as optional tools, boosting metrics but leaving reasoning ungrounded and crops ineffective. This gap gives rise to the illusion of thinking with images: models seem visually grounded but rely on context-agnostic actions that neither refine perception nor guide reasoning toward correct answers. We address this problem by reframing visual actions as core reasoning primitives rather than optional tools, which we term visual rationalization, the visual analogue of textual Chain-of-Thought. Building on this insight, we propose Visual Rationale Learning (ViRL), an end-to-end paradigm that grounds training in the visual rationale itself. ViRL integrates (1) Process Supervision with ground-truth rationales, (2) Objective Alignment via step-level reward shaping, and (3) Fine-Grained Credit Assignment to distinguish correct, redundant, and erroneous actions. By ensuring each action contributes meaningfully to the reasoning chain, ViRL enables models to "get the right answer for the right visual reason". Trained purely with end-to-end RL, ViRL achieves state-of-the-art results across benchmarks spanning perception, hallucination, and reasoning. This work establishes visual rationalization as a task-agnostic, process-grounded paradigm for building transparent, verifiable, and trustworthy vision-language models.

Yang Yao, Yixu Wang, Yuxuan Zhang et al. · utoronto

Artificial intelligence is undergoing the paradigm shift from closed language models to interconnected agent systems capable of external perception and information integration. As a representative embodiment, Deep Research Agents (DRAs) systematically exhibit the capabilities for task decomposition, cross-source retrieval, multi-stage reasoning, and structured output, which markedly enhance performance on complex and open-ended tasks. However, existing benchmarks remain deficient in evaluation dimensions, response formatting, and scoring mechanisms, limiting their capacity to assess such systems effectively. This paper introduces a rigorous benchmark and a multidimensional evaluation framework tailored to DRAs and report-style responses. The benchmark comprises 214 expert-curated challenging queries distributed across 10 broad thematic domains, each accompanied by manually constructed reference bundles to support composite evaluation. The framework enables comprehensive evaluation of long-form reports generated by DRAs, incorporating integrated scoring metrics for semantic quality, topical focus, and retrieval trustworthiness. Extensive experimentation confirms the superior performance of mainstream DRAs over web-search-tool-augmented reasoning models, yet reveals considerable scope for further improvement. This study provides a robust foundation for capability assessment, architectural refinement, and paradigm advancement in DRA systems.

Xuan He, Dongfu Jiang, Ping Nie et al. · utoronto

Recent advances in text-to-video generation have produced increasingly realistic and diverse content, yet evaluating such videos remains a fundamental challenge due to their multi-faceted nature encompassing visual quality, semantic alignment, and physical consistency. Existing evaluators and reward models are limited to single opaque scores, lack interpretability, or provide only coarse analysis, making them insufficient for capturing the comprehensive nature of video quality assessment. We present VideoScore2, a multi-dimensional, interpretable, and human-aligned framework that explicitly evaluates visual quality, text-to-video alignment, and physical/common-sense consistency while producing detailed chain-of-thought rationales. Our model is trained on a large-scale dataset VideoFeedback2 containing 27,168 human-annotated videos with both scores and reasoning traces across three dimensions, using a two-stage pipeline of supervised fine-tuning followed by reinforcement learning with Group Relative Policy Optimization (GRPO) to enhance analytical robustness. Extensive experiments demonstrate that VideoScore2 achieves superior performance with 44.35 (+5.94) accuracy on our in-domain benchmark VideoScore-Bench-v2 and 50.37 (+4.32) average performance across four out-of-domain benchmarks (VideoGenReward-Bench, VideoPhy2, etc), while providing interpretable assessments that bridge the gap between evaluation and controllable generation through effective reward modeling for Best-of-N sampling. Project Page: https://tiger-ai-lab.github.io/VideoScore2/

Che Liu, Zhongwei Wan, Yuqi Wang et al.

Automatic radiology report generation holds significant potential to streamline the labor-intensive process of report writing by radiologists, particularly for 3D radiographs such as CT scans. While CT scans are critical for clinical diagnostics, they remain less explored compared to 2D radiographs. To date, there has been no comprehensive benchmark for 3D radiograph report generation (3DRRG), nor sufficient investigation into the optimal training strategies for Vision Language Models (VLMs) in this context, particularly with respect to vision encoder choices, visual token compression, and model scaling. In this work, we make three key contributions. We curate **CT-3DRRG**, the largest **publicly** available 3D CT-report dataset, establishing a robust and diverse benchmark for evaluating VLM performance on 3DRRG. Furthermore, we propose a comprehensive training recipe for building high-performing VLMs for 3DRRG, exploring key factors such as vision encoder pretraining strategies, visual token compression, and the impact of data & model scale. Guided by these findings, we introduce **Argus**, a state-of-the-art family of VLMs that achieve superior performance across different model sizes and input 3D medical image resolutions, efficiently processing high-resolution 3D images up to $512 \times 512 \times 256$[^1].

Haozhe Wang, Jiale Zhou, Xuming He

Despite recent success of deep network-based Reinforcement Learning (RL), it remains elusive to achieve human-level efficiency in learning novel tasks. While previous efforts attempt to address this challenge using meta-learning strategies, they typically suffer from sampling inefficiency with on-policy RL algorithms or meta-overfitting with off-policy learning. In this work, we propose a novel meta-RL strategy to address those limitations. In particular, we decompose the meta-RL problem into three sub-tasks, task-exploration, task-inference and task-fulfillment, instantiated with two deep network agents and a task encoder. During meta-training, our method learns a task-conditioned actor network for task-fulfillment, an explorer network with a self-supervised reward shaping that encourages task-informative experiences in task-exploration, and a context-aware graph-based task encoder for task inference. We validate our approach with extensive experiments on several public benchmarks and the results show that our algorithm effectively performs exploration for task inference, improves sample efficiency during both training and testing, and mitigates the meta-overfitting problem.