Hao Liang

Aditi, Niket Agarwal, Arslan Ali et al.

We introduce Cosmos 3, a family of omnimodal world models designed to jointly process and generate language, image, video, audio, and action sequences within a unified mixture-of-transformers architecture. By supporting highly flexible input-output configurations, Cosmos 3 seamlessly unifies critical modalities for Physical AI -- effectively subsuming vision-language models, video generators, world simulators, and world-action models into a single framework. Our evaluation demonstrates that Cosmos 3 establishes a new state-of-the-art across a diverse suite of understanding and generation tasks, demonstrating omnimodal world models as scalable, general-purpose backbones for embodied agents. Our post-trained Cosmos 3 models were ranked as the best open-source Text-to-Image and Image-to-Video models by Artificial Analysis, and the best policy model by RoboArena at the time the technical report was written. To accelerate open research and deployment in Physical AI, we make our code, model checkpoints, curated synthetic datasets, and evaluation benchmark available under the Linux Foundation's OpenMDW-1.1 https://openmdw.ai/license/1-1/ License at https://github.com/nvidia/cosmos}{github.com/nvidia/cosmos and https://huggingface.co/collections/nvidia/cosmos3 . The project website is available at https://research.nvidia.com/labs/cosmos-lab/cosmos3 .

Zhengyang Zhao, Shengjie Ye, Lu Ma et al.

AI agents are increasingly being tasked with automating AI research itself, particularly the critical post-training phase that transforms base LLMs into aligned assistants. However, recent evaluations reveal that even frontier agents struggle to perform this task. While the success of post-training fundamentally relies on acquiring high-quality data, relying on agents to autonomously curate targeted training datasets from the open web introduces severe challenges. Executing the long-horizon tasks of searching, filtering, and balancing data within noisy web environments frequently overwhelms an agent's limited context, ultimately leading to degraded dataset quality and suboptimal downstream training performance. To bridge this gap, we introduce Andes (Agent Native Data Evolving Synthesis), a framework that reimagines data generation as a plug-and-play \emph{agent skill}. Rather than forcing agents to devise complex data-gathering strategies from scratch, \textsc{Andes} provides an intelligent abstraction layer. By leveraging a self-evolving World Tree routing mechanism and actionable diagnostic reports, it allows trainer agents to dynamically steer data synthesis through an interactive, closed-loop interface. We demonstrate that under strict compute constraints, equipping foundationally weaker agents with Andes improves automated alignment, securing state-of-the-art performance on PostTrainBench and robust cross-task generalization. Our project is available at https://github.com/zzy1127/ANDES.

Zinan Lin, Hao Liang, Giulia Fanti et al. · microsoft-research

We study the problem of learning generative adversarial networks (GANs) for a rare class of an unlabeled dataset subject to a labeling budget. This problem is motivated from practical applications in domains including security (e.g., synthesizing packets for DNS amplification attacks), systems and networking (e.g., synthesizing workloads that trigger high resource usage), and machine learning (e.g., generating images from a rare class). Existing approaches are unsuitable, either requiring fully-labeled datasets or sacrificing the fidelity of the rare class for that of the common classes. We propose RareGAN, a novel synthesis of three key ideas: (1) extending conditional GANs to use labelled and unlabelled data for better generalization; (2) an active learning approach that requests the most useful labels; and (3) a weighted loss function to favor learning the rare class. We show that RareGAN achieves a better fidelity-diversity tradeoff on the rare class than prior work across different applications, budgets, rare class fractions, GAN losses, and architectures.

Tao Zhang, Chenglin Zhu, Yanjun Shen et al.

The adeptness of Large Language Models (LLMs) in comprehending and following natural language instructions is critical for their deployment in sophisticated real-world applications. Existing evaluations mainly focus on fragmented constraints or narrow scenarios, but they overlook the comprehensiveness and authenticity of constraints from the user's perspective. To bridge this gap, we propose CFBench, a large-scale Comprehensive Constraints Following Benchmark for LLMs, featuring 1,000 curated samples that cover more than 200 real-life scenarios and over 50 NLP tasks. CFBench meticulously compiles constraints from real-world instructions and constructs an innovative systematic framework for constraint types, which includes 10 primary categories and over 25 subcategories, and ensures each constraint is seamlessly integrated within the instructions. To make certain that the evaluation of LLM outputs aligns with user perceptions, we propose an advanced methodology that integrates multi-dimensional assessment criteria with requirement prioritization, covering various perspectives of constraints, instructions, and requirement fulfillment. Evaluating current leading LLMs on CFBench reveals substantial room for improvement in constraints following, and we further investigate influencing factors and enhancement strategies. The data and code are publicly available at https://github.com/PKU-Baichuan-MLSystemLab/CFBench

Hao Liang, Pietro Perona, Guha Balakrishnan

We propose an experimental method for measuring bias in face recognition systems. Existing methods to measure bias depend on benchmark datasets that are collected in the wild and annotated for protected (e.g., race, gender) and non-protected (e.g., pose, lighting) attributes. Such observational datasets only permit correlational conclusions, e.g., "Algorithm A's accuracy is different on female and male faces in dataset X.". By contrast, experimental methods manipulate attributes individually and thus permit causal conclusions, e.g., "Algorithm A's accuracy is affected by gender and skin color." Our method is based on generating synthetic faces using a neural face generator, where each attribute of interest is modified independently while leaving all other attributes constant. Human observers crucially provide the ground truth on perceptual identity similarity between synthetic image pairs. We validate our method quantitatively by evaluating race and gender biases of three research-grade face recognition models. Our synthetic pipeline reveals that for these algorithms, accuracy is lower for Black and East Asian population subgroups. Our method can also quantify how perceptual changes in attributes affect face identity distances reported by these models. Our large synthetic dataset, consisting of 48,000 synthetic face image pairs (10,200 unique synthetic faces) and 555,000 human annotations (individual attributes and pairwise identity comparisons) is available to researchers in this important area.

Hao Liang, Zhixuan Ge, Soumendu Majee et al.

Reconstructing a photorealistic 3D face avatar from a single unconstrained photograph is challenging: feed-forward 3D Gaussian Splatting (3DGS) models degrade on out-of-distribution inputs, while pretrained diffusion models produce high-fidelity images but lack multi-view consistency. We observe that these paradigms are fundamentally complementary: explicit 3D representations guarantee geometric consistency, whereas 2D diffusion priors ensure photorealism. Building on this, we propose SplatShot, a training-free framework that couples these representations directly within the denoising process. Given a base 3DGS face model and a single reference image, we jointly denoise all target views using a per-step 3D feedback loop. At each timestep, we predict clean images from the noisy latents, refit the 3DGS to these multi-view predictions, and back-propagate the photometric discrepancy between the 3DGS re-renderings and 2D predictions into the noise estimate. This steers the sampling trajectory toward strictly 3D-coherent, identity-faithful outputs. Experiments on diverse in-the-wild images demonstrate that SplatShot produces 3D avatars with superior identity preservation, photorealism, and multi-view consistency.

Zhengren Wang, Qinhan Yu, Shida Wei et al.

Modern QA systems entail retrieval-augmented generation (RAG) for accurate and trustworthy responses. However, the inherent gap between user queries and relevant documents hinders precise matching. We introduce QAEncoder, a training-free approach to bridge this gap. Specifically, QAEncoder estimates the expectation of potential queries in the embedding space as a robust surrogate for the document embedding, and attaches document fingerprints to effectively distinguish these embeddings. Extensive experiments across diverse datasets, languages, and embedding models confirmed QAEncoder's alignment capability, which offers a simple-yet-effective solution with zero additional index storage, retrieval latency, training costs, or catastrophic forgetting and hallucination issues. The repository is publicly available at https://github.com/IAAR-Shanghai/QAEncoder.

Yifan Dai, Zhenhua Wu, Bohan Zeng et al.

Joint audio-visual reasoning is essential for omnimodal understanding, yet current multimodal large language models (MLLMs) still struggle when reasoning requires fine-grained evidence from both modalities. A central limitation is that explicit text-based chain-of-thought (CoT) compresses continuous audio-visual signals into discrete tokens, weakening temporal grounding and shifting intermediate reasoning toward language priors. We argue that a unified latent space is a better medium for such reasoning because it preserves dense sensory information while remaining compatible with autoregressive generation. Based on this insight, we propose \textbf{LatentOmni}, a cross-modal reasoning framework that interleaves textual reasoning with audio-visual latent states. LatentOmni introduces feature-level supervision to align latent reasoning states with task-relevant sensory features and uses Omni-Sync Position Embedding (OSPE) to maintain temporal consistency between latent audio and visual states. We further construct \textbf{LatentOmni-Instruct-35K}, a dataset of audio-visual interleaved reasoning trajectories for supervising latent-space reasoning. Comprehensive evaluation across multiple audio-visual reasoning benchmarks demonstrates that LatentOmni achieves the best performance among the evaluated open-source models and consistently outperforms the Explicit Text CoT baseline, supporting latent-space joint reasoning as a promising path toward stronger omnimodal understanding.

Yu Gao, Lutong Su, Hao Liang et al.

Neural Radiance Fields (NeRF) use multi-view images for 3D scene representation, demonstrating remarkable performance. As one of the primary sources of multi-view images, multi-camera systems encounter challenges such as varying intrinsic parameters and frequent pose changes. Most previous NeRF-based methods assume a unique camera and rarely consider multi-camera scenarios. Besides, some NeRF methods that can optimize intrinsic and extrinsic parameters still remain susceptible to suboptimal solutions when these parameters are poor initialized. In this paper, we propose MC-NeRF, a method that enables joint optimization of both intrinsic and extrinsic parameters alongside NeRF. The method also supports each image corresponding to independent camera parameters. First, we tackle coupling issue and the degenerate case that arise from the joint optimization between intrinsic and extrinsic parameters. Second, based on the proposed solutions, we introduce an efficient calibration image acquisition scheme for multi-camera systems, including the design of calibration object. Finally, we present an end-to-end network with training sequence that enables the estimation of intrinsic and extrinsic parameters, along with the rendering network. Furthermore, recognizing that most existing datasets are designed for a unique camera, we construct a real multi-camera image acquisition system and create a corresponding new dataset, which includes both simulated data and real-world captured images. Experiments confirm the effectiveness of our method when each image corresponds to different camera parameters. Specifically, we use multi-cameras, each with different intrinsic and extrinsic parameters in real-world system, to achieve 3D scene representation without providing initial poses.

Chengyu Shen, Yanheng Hou, Minghui Pan et al.

Reliable evaluation is essential for developing and deploying large language models, yet in practice it often requires substantial manual effort: practitioners must identify appropriate benchmarks, reproduce heterogeneous evaluation codebases, configure dataset schema mappings, and interpret aggregated metrics. To address these challenges, we present One-Eval, an agentic evaluation system that converts natural-language evaluation requests into executable, traceable, and customizable evaluation workflows. One-Eval integrates (i) NL2Bench for intent structuring and personalized benchmark planning, (ii) BenchResolve for benchmark resolution, automatic dataset acquisition, and schema normalization to ensure executability, and (iii) Metrics \& Reporting for task-aware metric selection and decision-oriented reporting beyond scalar scores. The system further incorporates human-in-the-loop checkpoints for review, editing, and rollback, while preserving sample evidence trails for debugging and auditability. Experiments show that One-Eval can execute end-to-end evaluations from diverse natural-language requests with minimal user effort, supporting more efficient and reproducible evaluation in industrial settings. Our framework is publicly available at https://github.com/OpenDCAI/One-Eval.

Linzhuang Sun, Tianyu Guo, Hao Liang et al.

Recent advances in Text-to-SQL have achieved strong results in static, single-turn tasks, where models generate SQL queries from natural language questions. However, these systems fall short in real-world interactive scenarios, where user intents evolve and queries must be refined over multiple turns. In applications such as finance and business analytics, users iteratively adjust query constraints or dimensions based on intermediate results. To evaluate such dynamic capabilities, we introduce DySQL-Bench, a benchmark assessing model performance under evolving user interactions. Unlike previous manually curated datasets, DySQL-Bench is built through an automated two-stage pipeline of task synthesis and verification. Structured tree representations derived from raw database tables guide LLM-based task generation, followed by interaction-oriented filtering and expert validation. Human evaluation confirms 100% correctness of the synthesized data. We further propose a multi-turn evaluation framework simulating realistic interactions among an LLM-simulated user, the model under test, and an executable database. The model must adapt its reasoning and SQL generation as user intents change. DySQL-Bench covers 13 domains across BIRD and Spider 2 databases, totaling 1,072 tasks. Even GPT-4o attains only 58.34% overall accuracy and 23.81% on the Pass@5 metric, underscoring the benchmark's difficulty. All code and data are released at https://github.com/Aurora-slz/Real-World-SQL-Bench .

Hao Liang, Linzhuang Sun, Minxuan Zhou et al.

With the rapid progress of Multimodal LLMs, evaluating their mathematical reasoning capabilities has become an increasingly important research direction. In particular, visual-textual mathematical reasoning serves as a key indicator of an MLLM's ability to comprehend and solve complex, multi-step quantitative problems. While existing benchmarks such as MathVista and MathVerse have advanced the evaluation of multimodal math proficiency, they primarily rely on digitally rendered content and fall short in capturing the complexity of real-world scenarios. To bridge this gap, we introduce MathScape, a novel benchmark focused on assessing MLLMs' reasoning ability in realistic mathematical contexts. MathScape comprises 1,369 high-quality math problems paired with human-captured real-world images, closely reflecting the challenges encountered in practical educational settings. We conduct a thorough multi-dimensional evaluation across nine leading closed-source MLLMs, three open-source MLLMs with over 20 billion parameters, and seven smaller-scale MLLMs. Our results show that even state-of-the-art models struggle with real-world math tasks, lagging behind human performance, highlighting critical limitations in current model capabilities. Moreover, we find that strong performance on synthetic or digitally rendered images does not guarantee similar effectiveness on real-world tasks. This underscores the necessity of MathScape in the next stage of multimodal mathematical reasoning.

Qifeng Cai, Hao Liang, Chang Xu et al.

The data-centric paradigm has become pivotal in AI, especially for Text-to-SQL, where performance is limited by scarce, simplistic, and low-diversity datasets. To address this, we propose Text2SQL-Flow, a SQL-aware data augmentation framework that generates large-scale, semantically valid, and structurally diverse Text-to-SQL pairs from minimal seed data. It operates across six augmentation dimensions and integrates an end-to-end pipeline featuring SQL execution verification, natural language question generation, chain-of-thought reasoning traces, and data classification. A modular Database Manager ensures cross-database compatibility and scalability. Using this framework, we build SQLFlow, a high-quality dataset of 89,544 annotated examples. We evaluate SQLFlow in two settings: (1) For open-source LLMs, fine-tuning on SQLFlow consistently improves performance across benchmarks under the same data budget. (2) For closed-source LLMs, we introduce a masked alignment retrieval method that treats SQLFlow as both knowledge base and training data for the retriever. This enables structure-aware example matching by modeling fine-grained alignments between questions and SQL queries. Experiments show our retrieval strategy outperforms existing methods, underscoring the value of SQLFlow's high-fidelity data and our novel technique. Our work establishes a scalable, data-centric foundation for advancing Text-to-SQL systems and highlights the critical role of high-quality structured data in modern AI.

Zijun Shen, Sihan Yang, Ruichuan An et al.

Unified Multimodal Models (UMMs) excel in general tasks but struggle to bridge the gap between personalized understanding and generation. Prior works largely rely on implicit token-level alignment via supervised fine-tuning, which fails to fully capture the potential synergy between comprehension and creation. In this work, we propose Sync-R1, an end-to-end reinforcement learning framework that jointly optimizes personalized understanding and generation within a single, explicit reasoning loop. Through this unified feedback process, Sync-R1 enables personalized comprehension to guide content creation, while the resulting generation quality reciprocally refines understanding within an integrated reward landscape. To efficiently orchestrate this dual-task synergy, we introduce Sync-GRPO, a reinforcement learning method utilizing an ensemble reward system. Furthermore, we propose Dynamic Group Scaling (DGS), which adaptively filters low-potential trajectories to reduce gradient variance and accelerate convergence. To better reflect real-world complexity, we introduce UnifyBench++, featuring denser textual descriptions and richer user contexts. Experimental results demonstrate that Sync-R1 achieves state-of-the-art performance, showcasing superior cross-task reasoning and robust personalization without requiring complex cold-start procedures. The code and the UnifyBench++ dataset will be released at: https://github.com/arctanxarc/UniCTokens.

Hao Liang, Qihan Lin, Zhaoyang Han et al.

Large language models (LLMs) are increasingly used in K-12 education, yet existing benchmarks such as C-Eval, CMMLU, GaokaoBench, and EduEval mainly evaluate factual recall through exam-style question answering. Effective educational AI additionally requires curriculum cognition: understanding how knowledge is structured through prerequisite chains, concept taxonomies, experiment-concept links, and pedagogical sequencing. To address this gap, we introduce K12-KGraph, a curriculum-aligned knowledge graph extracted from official People's Education Press textbooks across mathematics, physics, chemistry, and biology from primary to high school. The graph contains seven node types (Concept, Skill, Experiment, Exercise, Section, Chapter, Book) and nine relation types covering taxonomy, prerequisite, association, verification, assessment, location, and order. Based on this graph, we construct two resources: (1) K12-Bench, a 23,640-question multi-select benchmark spanning five graph-derived task families (Ground, Prereq, Neighbor, Evidence, and Locate); and (2) K12-Train, a KG-guided supervised fine-tuning corpus of approximately 2,300 QA pairs synthesized from graph structure and node attributes. Experiments reveal substantial deficiencies in curriculum cognition: on K12-Bench, Gemini-3-Flash achieves only 57% exact match, while the best open-source model, Gemma-4-31B-IT, reaches 46%. Under a strictly matched 2,300-sample SFT budget on Qwen3-4B-Base and Llama-3.1-8B-Base, K12-Train consistently outperforms equally sized subsets from eight mainstream instruction-tuning corpora on both GaokaoBench and EduEval, demonstrating that curriculum-structured supervision is highly sample-efficient for educational tuning. We release the graph, benchmark, training data, and full construction pipeline.

Hao Liang, Jiapeng Li, Tianyi Bai et al.

Recently, with the rise of web videos, managing and understanding large-scale video datasets has become increasingly important. Video Large Language Models (VideoLLMs) have emerged in recent years due to their strong video understanding capabilities. However, training and inference processes for VideoLLMs demand vast amounts of data, presenting significant challenges to data management, particularly regarding efficiency, robustness, and effectiveness. In this work, we present KeyVideoLLM, a text-video frame similarity-based keyframe selection method designed to manage VideoLLM data efficiently, robustly, and effectively. Specifically, KeyVideoLLM achieves a remarkable data compression rate of up to 60.9 times, substantially lowering disk space requirements, which proves its high efficiency. Additionally, it maintains a 100% selection success rate across all video formats and scales, enhances processing speed by up to 200 times compared to existing keyframe selection methods, and does not require hyperparameter tuning. Beyond its outstanding efficiency and robustness, KeyVideoLLM further improves model performance in video question-answering tasks during both training and inference stages. Notably, it consistently achieved the state-of-the-art (SoTA) experimental results on diverse datasets.

Linzhuang Sun, Hao Liang, Jingxuan Wei et al.

Large Language Models (LLMs) have exhibited exceptional performance across a broad range of tasks and domains. However, they still encounter difficulties in solving mathematical problems due to the rigorous and logical nature of mathematics. Previous studies have employed techniques such as supervised fine-tuning (SFT), prompt engineering, and search-based methods to improve the mathematical problem-solving abilities of LLMs. Despite these efforts, their performance remains suboptimal and demands substantial computational resources. To address this issue, we propose a novel approach, BEATS, to enhance mathematical problem-solving abilities. Our method leverages newly designed prompts that guide the model to iteratively rewrite, advance by one step, and generate answers based on previous steps. Additionally, we introduce a new back-verification technique that uses LLMs to validate the correctness of the generated answers. Furthermore, we employ a pruning tree search to optimize search time while achieving strong performance. Notably, our method improves Qwen2-7b-Instruct's score from 36.94 to 61.52, outperforming GPT4's 42.5 on the MATH benchmark.

Hengyi Feng, Hao Liang, Mingrui Chen et al.

Real-world audio-visual understanding requires chaining evidence that is sparse, temporally dispersed, and split across the visual and auditory streams, whereas existing benchmarks largely fail to evaluate this capability. They restrict videos to short clips, isolate modalities, or reduce questions to one-hop perception. We introduce TraceAV-Bench, the first benchmark to jointly evaluate multi-hop reasoning over long audio-visual trajectories and multimodal hallucination robustness. TraceAV-Bench comprises 2,200 rigorously validated multiple-choice questions over 578 long videos, totaling 339.5 hours, spanning 4 evaluation dimensions and 15 sub-tasks. Each question is grounded in an explicit reasoning chain that averages 3.68 hops across a 15.1-minute temporal span. The dataset is built by a three-step semi-automated pipeline followed by a strict quality assurance process. Evaluation of multiple representative OmniLLMs on TraceAV-Bench reveals that the benchmark poses a persistent challenge across all models, with the strongest closed-source model (Gemini 3.1 Pro) reaching only 68.29% on general tasks, and the best open-source model (Ming-Flash-Omni-2.0) reaching 51.70%, leaving substantial headroom. Moreover, we find that robustness to multimodal hallucination is largely decoupled from general multimodal reasoning performance. We anticipate that TraceAV-Bench will stimulate further research toward OmniLLMs that can reason coherently and faithfully over long-form audio-visual content.

Zelin Tao, Zeran Su, Peiran Liu et al.

Achieving general-purpose humanoid control requires a delicate balance between the precise execution of commanded motions and the flexible, anthropomorphic adaptability needed to recover from unpredictable environmental perturbations. Current general controllers predominantly formulate motion control as a rigid reference-tracking problem. While effective in nominal conditions, these trackers often exhibit brittle, non-anthropomorphic failure modes under severe disturbances, lacking the generative adaptability inherent to human motor control. To overcome this limitation, we propose Heracles, a novel state-conditioned diffusion middleware that bridges precise motion tracking and generative synthesis. Rather than relying on rigid tracking paradigms or complex explicit mode-switching, Heracles operates as an intermediary layer between high-level reference motions and low-level physics trackers. By conditioning on the robot's real-time state, the diffusion model implicitly adapts its behavior: it approximates an identity map when the state closely aligns with the reference, preserving zero-shot tracking fidelity. Conversely, when encountering significant state deviations, it seamlessly transitions into a generative synthesizer to produce natural, anthropomorphic recovery trajectories. Our framework demonstrates that integrating generative priors into the control loop not only significantly enhances robustness against extreme perturbations but also elevates humanoid control from a rigid tracking paradigm to an open-ended, generative general-purpose architecture.

Miao Zheng, Hao Liang, Fan Yang et al.

In recent years, the rise of Large Language Models (LLMs) has spurred a growing demand for plug-and-play AI systems. Among the various AI techniques, prompt engineering stands out as particularly significant. However, users often face challenges in writing prompts due to the steep learning curve and significant time investment, and existing automatic prompt engineering (APE) models can be difficult to use. To address this issue, we propose PAS, an LLM-based plug-and-play APE system. PAS utilizes LLMs trained on high-quality, automatically generated prompt complementary datasets, resulting in exceptional performance. In comprehensive benchmarks, PAS achieves state-of-the-art (SoTA) results compared to previous APE models, with an average improvement of 6.09 points. Moreover, PAS is highly efficient, achieving SoTA performance with only 9000 data points. Additionally, PAS can autonomously generate prompt augmentation data without requiring additional human labor. Its flexibility also allows it to be compatible with all existing LLMs and applicable to a wide range of tasks. PAS excels in human evaluations, underscoring its suitability as a plug-in for users. This combination of high performance, efficiency, and flexibility makes PAS a valuable system for enhancing the usability and effectiveness of LLMs through improved prompt engineering.

Hao Liang, Zhengyang Zhao, Meiyi Qiang et al.

Data-centric training has emerged as a promising direction for improving large language models (LLMs) by optimizing not only model parameters but also the selection, composition, and weighting of training data during optimization. However, existing approaches to data selection, data mixture optimization, and data reweighting are often developed in isolated codebases with inconsistent interfaces, hindering reproducibility, fair comparison, and practical integration. In this paper, we present DataFlex, a unified data-centric dynamic training framework built upon LLaMA-Factory. DataFlex supports three major paradigms of dynamic data optimization: sample selection, domain mixture adjustment, and sample reweighting, while remaining fully compatible with the original training workflow. It provides extensible trainer abstractions and modular components, enabling a drop-in replacement for standard LLM training, and unifies key model-dependent operations such as embedding extraction, inference, and gradient computation, with support for large-scale settings including DeepSpeed ZeRO-3. We conduct comprehensive experiments across multiple data-centric methods. Dynamic data selection consistently outperforms static full-data training on MMLU across both Mistral-7B and Llama-3.2-3B. For data mixture, DoReMi and ODM improve both MMLU accuracy and corpus-level perplexity over default proportions when pretraining Qwen2.5-1.5B on SlimPajama at 6B and 30B token scales. DataFlex also achieves consistent runtime improvements over original implementations. These results demonstrate that DataFlex provides an effective, efficient, and reproducible infrastructure for data-centric dynamic training of LLMs.

Hao Liang, Xiaochen Ma, Zhou Liu et al.

The rapidly growing demand for high-quality data in Large Language Models (LLMs) has intensified the need for scalable, reliable, and semantically rich data preparation pipelines. However, current practices remain dominated by ad-hoc scripts and loosely specified workflows, which lack principled abstractions, hinder reproducibility, and offer limited support for model-in-the-loop data generation. To address these challenges, we present DataFlow, a unified and extensible LLM-driven data preparation framework. DataFlow is designed with system-level abstractions that enable modular, reusable, and composable data transformations, and provides a PyTorch-style pipeline construction API for building debuggable and optimizable dataflows. The framework consists of nearly 200 reusable operators and six domain-general pipelines spanning text, mathematical reasoning, code, Text-to-SQL, agentic RAG, and large-scale knowledge extraction. To further improve usability, we introduce DataFlow-Agent, which automatically translates natural-language specifications into executable pipelines via operator synthesis, pipeline planning, and iterative verification. Across six representative use cases, DataFlow consistently improves downstream LLM performance. Our math, code, and text pipelines outperform curated human datasets and specialized synthetic baselines, achieving up to +3\% execution accuracy in Text-to-SQL over SynSQL, +7\% average improvements on code benchmarks, and 1--3 point gains on MATH, GSM8K, and AIME. Moreover, a unified 10K-sample dataset produced by DataFlow enables base models to surpass counterparts trained on 1M Infinity-Instruct data. These results demonstrate that DataFlow provides a practical and high-performance substrate for reliable, reproducible, and scalable LLM data preparation, and establishes a system-level foundation for future data-centric AI development.

Hao Liang, Zhi-quan Luo

We present a distribution optimization framework that significantly improves confidence bounds for various risk measures compared to previous methods. Our framework encompasses popular risk measures such as the entropic risk measure, conditional value at risk (CVaR), spectral risk measure, distortion risk measure, equivalent certainty, and rank-dependent expected utility, which are well established in risk-sensitive decision-making literature. To achieve this, we introduce two estimation schemes based on concentration bounds derived from the empirical distribution, specifically using either the Wasserstein distance or the supremum distance. Unlike traditional approaches that add or subtract a confidence radius from the empirical risk measures, our proposed schemes evaluate a specific transformation of the empirical distribution based on the distance. Consequently, our confidence bounds consistently yield tighter results compared to previous methods. We further verify the efficacy of the proposed framework by providing tighter problem-dependent regret bound for the CVaR bandit.

Hao Liang, Linzhuang Sun, Jingxuan Wei et al.

In recent years, with the rapid advancements in large language models (LLMs), achieving excellent empathetic response capabilities has become a crucial prerequisite. Consequently, managing and understanding empathetic datasets have gained increasing significance. However, empathetic data are typically human-labeled, leading to insufficient datasets and wasted human labor. In this work, we present Synth-Empathy, an LLM-based data generation and quality and diversity selection pipeline that automatically generates high-quality empathetic data while discarding low-quality data. With the data generated from a low empathetic model, we are able to further improve empathetic response performance and achieve state-of-the-art (SoTA) results across multiple benchmarks. Moreover, our model achieves SoTA performance on various human evaluation benchmarks, demonstrating its effectiveness and robustness in real-world applications. Furthermore, we show the trade-off between data quantity and quality, providing insights into empathetic data generation and selection.

Hao Liang, Kevin Ni, Guha Balakrishnan

Recent work demonstrates that images from various chest X-ray datasets contain visual features that are strongly correlated with protected demographic attributes like race and gender. This finding raises issues of fairness, since some of these factors may be used by downstream algorithms for clinical predictions. In this work, we propose a framework, using generative adversarial networks (GANs), to visualize what features are most different between X-rays belonging to two demographic subgroups.

Xuchen Li, Ruitao Wu, Xuanbo Liu et al.

Recent advances in large language models have enabled AI systems to achieve expert-level performance on domain-specific scientific tasks, yet these systems remain narrow and handcrafted. We introduce SciAgent, a unified multi-agent system designed for generalistic scientific reasoning-the ability to adapt reasoning strategies across disciplines and difficulty levels. SciAgent organizes problem solving as a hierarchical process: a Coordinator Agent interprets each problem's domain and complexity, dynamically orchestrating specialized Worker Systems, each composed of interacting reasoning Sub-agents for symbolic deduction, conceptual modeling, numerical computation, and verification. These agents collaboratively assemble and refine reasoning pipelines tailored to each task. Across mathematics and physics Olympiads (IMO, IMC, IPhO, CPhO), SciAgent consistently attains or surpasses human gold-medalist performance, demonstrating both domain generality and reasoning adaptability. Additionally, SciAgent has been tested on the International Chemistry Olympiad (IChO) and selected problems from the Humanity's Last Exam (HLE) benchmark, further confirming the system's ability to generalize across diverse scientific domains. This work establishes SciAgent as a concrete step toward generalistic scientific intelligence-AI systems capable of coherent, cross-disciplinary reasoning at expert levels.

Bohan Zeng, Kaixin Zhu, Daili Hua et al.

World models have emerged as a critical frontier in AI research, aiming to enhance large models by infusing them with physical dynamics and world knowledge. The core objective is to enable agents to understand, predict, and interact with complex environments. However, current research landscape remains fragmented, with approaches predominantly focused on injecting world knowledge into isolated tasks, such as visual prediction, 3D estimation, or symbol grounding, rather than establishing a unified definition or framework. While these task-specific integrations yield performance gains, they often lack the systematic coherence required for holistic world understanding. In this paper, we analyze the limitations of such fragmented approaches and propose a unified design specification for world models. We suggest that a robust world model should not be a loose collection of capabilities but a normative framework that integrally incorporates interaction, perception, symbolic reasoning, and spatial representation. This work aims to provide a structured perspective to guide future research toward more general, robust, and principled models of the world.

Ruichuan An, Sihan Yang, Ming Lu et al.

Current vision-language models (VLMs) show exceptional abilities across diverse tasks, such as visual question answering. To enhance user experience, recent studies investigate VLM personalization to understand user-provided concepts. However, they mainly focus on single-concept personalization, neglecting the existence and interplay of multiple concepts, which limits real-world applicability. This paper proposes the first multi-concept personalization paradigm, MC-LLaVA. Specifically, MC-LLaVA employs a multi-concept instruction tuning strategy, effectively integrating multiple concepts in a single training step. To reduce the costs related to joint training, we propose a personalized textual prompt that uses visual token information to initialize concept tokens. Additionally, we introduce a personalized visual prompt during inference, aggregating location confidence maps for enhanced recognition and grounding capabilities. To advance multi-concept personalization research, we further contribute a high-quality instruction tuning dataset. We carefully collect images with multiple characters and objects from movies and manually generate question-answer samples for multi-concept scenarios, featuring superior diversity. Comprehensive qualitative and quantitative experiments demonstrate that MC-LLaVA can achieve impressive multi-concept personalized responses, paving the way for VLMs to become better user-specific assistants. The code and dataset will be publicly available at https://github.com/arctanxarc/MC-LLaVA}.

Ruichuan An, Sihan Yang, Ming Lu et al.

Current vision-language models (VLMs) show exceptional abilities across diverse tasks, such as visual question answering. To enhance user experience, recent studies investigate VLM personalization to understand user-provided concepts. However, they mainly focus on single-concept personalization, neglecting the existence and interplay of multiple concepts, which limits real-world applicability. This paper proposes the first multi-concept personalization paradigm, MC-LLaVA. Specifically, MC-LLaVA employs a multi-concept instruction tuning strategy, effectively integrating multiple concepts in a single training step. To reduce the costs related to joint training, we propose a personalized textual prompt that uses visual token information to initialize concept tokens. Additionally, we introduce a personalized visual prompt during inference, aggregating location confidence maps for enhanced recognition and grounding capabilities. To advance multi-concept personalization research, we further contribute a high-quality instruction tuning dataset. We carefully collect images with multiple characters and objects from movies and manually generate question-answer samples for multi-concept scenarios, featuring superior diversity. Comprehensive qualitative and quantitative experiments demonstrate that MC-LLaVA can achieve impressive multi-concept personalized responses, paving the way for VLMs to become better user-specific assistants. The code and dataset will be publicly available at https://github.com/arctanxarc/MC-LLaVA.

Hao Liang, Josue Ortega Caro, Vikram Maheshri et al.

Training dataset biases are by far the most scrutinized factors when explaining algorithmic biases of neural networks. In contrast, hyperparameters related to the neural network architecture have largely been ignored even though different network parameterizations are known to induce different implicit biases over learned features. For example, convolutional kernel size is known to affect the frequency content of features learned in CNNs. In this work, we present a causal framework for linking an architectural hyperparameter to out-of-distribution algorithmic bias. Our framework is experimental, in that we train several versions of a network with an intervention to a specific hyperparameter, and measure the resulting causal effect of this choice on performance bias when a particular out-of-distribution image perturbation is applied. In our experiments, we focused on measuring the causal relationship between convolutional kernel size and face analysis classification bias across different subpopulations (race/gender), with respect to high-frequency image details. We show that modifying kernel size, even in one layer of a CNN, changes the frequency content of learned features significantly across data subgroups leading to biased generalization performance even in the presence of a balanced dataset.

Bozhou Li, Hao Liang, Zimo Meng et al.

In recent years, multimodal large language models (MLLMs) have shown strong potential in real-world applications. They are developing rapidly due to their remarkable ability to comprehend multimodal information and their inherent powerful cognitive and reasoning capabilities. Among MLLMs, vision language models (VLM) stand out for their ability to understand vision information. However, the scaling trend of VLMs under the current mainstream paradigm has not been extensively studied. Whether we can achieve better performance by training even larger models is still unclear. To address this issue, we conducted experiments on the pretraining stage of MLLMs. We conduct our experiment using different encoder sizes and large language model (LLM) sizes. Our findings indicate that merely increasing the size of encoders does not necessarily enhance the performance of VLMs. Moreover, we analyzed the effects of LLM backbone parameter size and data quality on the pretraining outcomes. Additionally, we explored the differences in scaling laws between LLMs and VLMs.

Zheng Liu, Hao Liang, Bozhou Li et al.

Vision-Language Models (VLMs) have recently emerged, demonstrating remarkable vision-understanding capabilities. However, training these models requires large-scale datasets, which brings challenges related to efficiency, effectiveness, and quality of web data. In this paper, we introduce SynthVLM, a new data synthesis and curation method for generating image-caption pairs. Unlike traditional methods, where captions are generated from images, SynthVLM utilizes advanced diffusion models and high-quality captions to synthesize and select images from text captions, thereby creating precisely aligned image-text pairs. We further introduce SynthVLM-100K, a high-quality dataset consisting of 100K curated and synthesized image-caption pairs. In both model and human evaluations, SynthVLM-100K outperforms traditional real-world datasets. Leveraging this dataset, we develop a new family of multimodal large language models (MLLMs), SynthVLM-7B and SynthVLM-13B, which achieve state-of-the-art (SOTA) performance on various vision question-answering (VQA) tasks. Notably, our models outperform LLaVA across most metrics with only 18\% pretrain data. Furthermore, SynthVLM-7B and SynthVLM-13B attain SOTA performance on the MMLU benchmark, demonstrating that the high-quality SynthVLM-100K dataset preserves language abilities.

Ruichuan An, Sihan Yang, Renrui Zhang et al.

Personalized models have demonstrated remarkable success in understanding and generating concepts provided by users. However, existing methods use separate concept tokens for understanding and generation, treating these tasks in isolation. This may result in limitations for generating images with complex prompts. For example, given the concept $\langle bo\rangle$, generating "$\langle bo\rangle$ wearing its hat" without additional textual descriptions of its hat. We call this kind of generation \textit{\textbf{personalized attribute-reasoning generation}}. To address the limitation, we present UniCTokens, a novel framework that effectively integrates personalized information into a unified vision language model (VLM) for understanding and generation. UniCTokens trains a set of unified concept tokens to leverage complementary semantics, boosting two personalized tasks. Moreover, we propose a progressive training strategy with three stages: understanding warm-up, bootstrapping generation from understanding, and deepening understanding from generation to enhance mutual benefits between both tasks. To quantitatively evaluate the unified VLM personalization, we present UnifyBench, the first benchmark for assessing concept understanding, concept generation, and attribute-reasoning generation. Experimental results on UnifyBench indicate that UniCTokens shows competitive performance compared to leading methods in concept understanding, concept generation, and achieving state-of-the-art results in personalized attribute-reasoning generation. Our research demonstrates that enhanced understanding improves generation, and the generation process can yield valuable insights into understanding. Our code and dataset will be released at: \href{https://github.com/arctanxarc/UniCTokens}{https://github.com/arctanxarc/UniCTokens}.

Linzhuang Sun, Hao Liang, Jingxuan Wei et al.

In recent years, with the rapid advancements in large language models (LLMs), achieving excellent empathetic response capability has become a crucial prerequisite. Consequently, managing and understanding large-scale video datasets has gained increasing importance. However, empathetic data are typically trained without any quality selection, leading to inefficient data usage and wasted computational resources. Additionally, using raw data can result in low performance in empathetic dialogues. In this work, we present Efficient-Empathy, a sensibility and rationality score-based data selection algorithm that automatically selects sensibility and rationality data while discarding low-quality data. With only the sensibility data (59% of the full dataset), our trained sensibility model efficiently achieves state-of-the-art (SoTA) performance. Furthermore, with multiple data selection hyperparameters, the sensibility model demonstrates SoTA performance, showcasing the robustness of our method. By integrating sensibility and rationality data with a MoE structure, we achieve even higher performance, demonstrating the effectiveness of our Efficient-Empathy algorithm.

Hao Liang, Kevin Ni, Guha Balakrishnan

Recent research demonstrates that deep learning models are capable of precisely extracting bio-information (e.g. race, gender and age) from patients' Chest X-Rays (CXRs). In this paper, we further show that deep learning models are also surprisingly accurate at recognition, i.e., distinguishing CXRs belonging to the same patient from those belonging to different patients. These findings suggest potential privacy considerations that the medical imaging community should consider with the proliferation of large public CXR databases.

Hao Liang, Zhi-Quan Luo

We study the regret guarantee for risk-sensitive reinforcement learning (RSRL) via distributional reinforcement learning (DRL) methods. In particular, we consider finite episodic Markov decision processes whose objective is the entropic risk measure (EntRM) of return. By leveraging a key property of the EntRM, the independence property, we establish the risk-sensitive distributional dynamic programming framework. We then propose two novel DRL algorithms that implement optimism through two different schemes, including a model-free one and a model-based one. We prove that they both attain $\tilde{\mathcal{O}}(\frac{\exp(|β| H)-1}{|β|}H\sqrt{S^2AK})$ regret upper bound, where $S$, $A$, $K$, and $H$ represent the number of states, actions, episodes, and the time horizon, respectively. It matches RSVI2 proposed in \cite{fei2021exponential}, with novel distributional analysis. To the best of our knowledge, this is the first regret analysis that bridges DRL and RSRL in terms of sample complexity. Acknowledging the computational inefficiency associated with the model-free DRL algorithm, we propose an alternative DRL algorithm with distribution representation. This approach not only maintains the established regret bounds but also significantly amplifies computational efficiency. We also prove a tighter minimax lower bound of $Ω(\frac{\exp(βH/6)-1}{βH}H\sqrt{SAT})$ for the $β>0$ case, which recovers the tight lower bound $Ω(H\sqrt{SAT})$ in the risk-neutral setting.

Huanyao Zhang, Jiepeng Zhou, Bo Li et al.

Multimodal large language models (MLLMs), equipped with increasingly advanced planning and tool-use capabilities, are evolving into autonomous agents capable of performing multimodal web browsing and deep search in open-world environments. However, existing benchmarks for multimodal browsing remain limited in task complexity, evidence accessibility, and evaluation granularity, hindering comprehensive and reproducible assessments of deep search capabilities. To address these limitations, we introduce BrowseComp-$V^3$, a novel benchmark consisting of 300 carefully curated and challenging questions spanning diverse domains. The benchmark emphasizes deep, multi-level, and cross-modal multi-hop reasoning, where critical evidence is interleaved across textual and visual modalities within and across web pages. All supporting evidence is strictly required to be publicly searchable, ensuring fairness and reproducibility. Beyond final-answer accuracy, we incorporate an expert-validated, subgoal-driven process evaluation mechanism that enables fine-grained analysis of intermediate reasoning behaviors and systematic characterization of capability boundaries. In addition, we propose OmniSeeker, a unified multimodal browsing agent framework integrating diverse web search and visual perception tools. Comprehensive experiments demonstrate that even state-of-the-art models achieve only 36% accuracy on our benchmark, revealing critical bottlenecks in multimodal information integration and fine-grained perception. Our results highlight a fundamental gap between current model capabilities and robust multimodal deep search in real-world settings.

Cheng Ge, Caitlyn Heqi Yin, Hao Liang et al.

Reinforcement learning (RL) has become a key driver of language model reasoning. Among RL algorithms, Group Relative Policy Optimization (GRPO) is the de facto standard, avoiding the need for a critic by using per-prompt baselines and variance normalization. Yet why and when this normalization helps remains unclear. In this work, we provide an explanation through the lens of local curvature of the sequence-level policy gradient: standard deviation normalization implements an adaptive gradient. Theoretically, under mild conditions, GRPO enjoys a strictly improved convergence rate over unnormalized REINFORCE, with gains characterized by the average within-prompt reward standard deviation across prompts and iterations. Empirically, our analysis on GSM8K and MATH benchmarks reveals three distinct training phases governed by the interplay between feature orthogonality and reward variance: (I) an early acceleration phase where high variance and orthogonality favor adaptive scaling; (II) a relatively stable transition phase; and (III) a late-stage regime where the loss of orthogonality limits further gains. Together, these results provide a principled account of when std normalization helps in GRPO, and offer broader insights into the design of critic-free RL algorithms.

Hao Liang, Keshi Zhao, Yajie Yang et al.

Large language models (LLMs) have demonstrated exceptional performance across a wide range of tasks and domains, with data preparation playing a critical role in achieving these results. Pre-training data typically combines information from multiple domains. To maximize performance when integrating data from various domains, determining the optimal data proportion is essential. However, state-of-the-art (SOTA) LLMs rarely disclose details about their pre-training data, making it difficult for researchers to identify ideal data proportions. In this paper, we introduce a new topic, \textit{data proportion detection}, which enables the automatic estimation of pre-training data proportions by analyzing the generated outputs of LLMs. We provide rigorous theoretical proofs, practical algorithms, and preliminary experimental results for data proportion detection. Based on these findings, we offer valuable insights into the challenges and future directions for effective data proportion detection and data management.

Zhou Liu, Zhaoyang Han, Guochen Yan et al.

Data governance ensures data quality, security, and compliance through policies and standards, a critical foundation for scaling modern AI development. Recently, large language models (LLMs) have emerged as a promising solution for automating data governance by translating user intent into executable transformation code. However, existing benchmarks for automated data science often emphasize snippet-level coding or high-level analytics, failing to capture the unique challenge of data governance: ensuring the correctness and quality of the data itself. To bridge this gap, we introduce DataGovBench, a benchmark featuring 150 diverse tasks grounded in real-world scenarios, built on data from actual cases. DataGovBench employs a novel "reversed-objective" methodology to synthesize realistic noise and utilizes rigorous metrics to assess end-to-end pipeline reliability. Our analysis on DataGovBench reveals that current models struggle with complex, multi-step workflows and lack robust error-correction mechanisms. Consequently, we propose DataGovAgent, a framework utilizing a Planner-Executor-Evaluator architecture that integrates constraint-based planning, retrieval-augmented generation, and sandboxed feedback-driven debugging. Experimental results show that DataGovAgent significantly boosts the Average Task Score (ATS) on complex tasks from 39.7 to 54.9 and reduces debugging iterations by over 77.9 percent compared to general-purpose baselines.

Chengyu Shen, Zhen Hao Wong, Runming He et al.

Large Language Models (LLMs) have recently achieved remarkable progress in mathematical reasoning. To enable such capabilities, many existing works distill strong reasoning models into long chains of thought or design algorithms to construct high-quality math QA data for training. However, these efforts primarily focus on generating correct reasoning paths and answers, while largely overlooking the validity of the questions themselves. In this work, we propose Math Question Verification (MathQ-Verify), a novel five-stage pipeline designed to rigorously filter ill-posed or under-specified math problems. MathQ-Verify first performs format-level validation to remove redundant instructions and ensure that each question is syntactically well-formed. It then formalizes each question, decomposes it into atomic conditions, and verifies them against mathematical definitions. Next, it detects logical contradictions among these conditions, followed by a goal-oriented completeness check to ensure the question provides sufficient information for solving. To evaluate this task, we use existing benchmarks along with an additional dataset we construct, containing 2,147 math questions with diverse error types, each manually double-validated. Experiments show that MathQ-Verify achieves state-of-the-art performance across multiple benchmarks, improving the F1 score by up to 25 percentage points over the direct verification baseline. It further attains approximately 90% precision and 63% recall through a lightweight model voting scheme. MathQ-Verify offers a scalable and accurate solution for curating reliable mathematical datasets, reducing label noise and avoiding unnecessary computation on invalid questions. Our code and data are available at https://github.com/scuuy/MathQ-Verify.

Mingan Lin, Fan Yang, Yanjun Shen et al.

We introduce Baichuan Alignment, a detailed analysis of the alignment techniques employed in the Baichuan series of models. This represents the industry's first comprehensive account of alignment methodologies, offering valuable insights for advancing AI research. We investigate the critical components that enhance model performance during the alignment process, including optimization methods, data strategies, capability enhancements, and evaluation processes. The process spans three key stages: Prompt Augmentation System(PAS), Supervised Fine-Tuning(SFT), and Preference Alignment. The problems encountered, the solutions applied, and the improvements made are thoroughly recorded. Through comparisons across well-established benchmarks, we highlight the technological advancements enabled by Baichuan Alignment. Baichuan-Instruct is an internal model, while Qwen2-Nova-72B and Llama3-PBM-Nova-70B are instruct versions of the Qwen2-72B and Llama-3-70B base models, optimized through Baichuan Alignment. Baichuan-Instruct demonstrates significant improvements in core capabilities, with user experience gains ranging from 17% to 28%, and performs exceptionally well on specialized benchmarks. In open-source benchmark evaluations, both Qwen2-Nova-72B and Llama3-PBM-Nova-70B consistently outperform their respective official instruct versions across nearly all datasets. This report aims to clarify the key technologies behind the alignment process, fostering a deeper understanding within the community. Llama3-PBM-Nova-70B model is available at https://huggingface.co/PKU-Baichuan-MLSystemLab/Llama3-PBM-Nova-70B.

DataFlow Team, Bohan Zeng, Daili Hua et al.

World models have garnered significant attention as a promising research direction in artificial intelligence, yet a clear and unified definition remains lacking. In this paper, we introduce OpenWorldLib, a comprehensive and standardized inference framework for Advanced World Models. Drawing on the evolution of world models, we propose a clear definition: a world model is a model or framework centered on perception, equipped with interaction and long-term memory capabilities, for understanding and predicting the complex world. We further systematically categorize the essential capabilities of world models. Based on this definition, OpenWorldLib integrates models across different tasks within a unified framework, enabling efficient reuse and collaborative inference. Finally, we present additional reflections and analyses on potential future directions for world model research. Code link: https://github.com/OpenDCAI/OpenWorldLib

Lingzhuang Sun, Yuxia Zhu, Ruitong Liu et al.

While Chain-of-Thought (CoT) prompting has significantly advanced the reasoning capabilities of Multimodal Large Language Models (MLLMs), relying solely on linear text sequences remains a bottleneck for complex tasks. We observe that even when auxiliary visual elements are interleaved, they are often treated as static snapshots within a one-dimensional, unstructured reasoning chain. We argue that such approaches treat reasoning history as an immutable stream: correcting a local error necessitates either generating verbose downstream corrections or regenerating the entire context. This forces the model to implicitly maintain and track state updates, significantly increasing token consumption and cognitive load. This limitation is particularly acute in high-dimensional domains, such as geometry and SVG design, where the textual expression of CoT lacks explicit visual guidance, further constraining the model's reasoning precision. To bridge this gap, we introduce \textbf{Canvas-of-Thought (Canvas-CoT)}. By leveraging a HTML Canvas as an external reasoning substrate, Canvas-CoT empowers the model to perform atomic, DOM-based CRUD operations. This architecture enables in-place state revisions without disrupting the surrounding context, allowing the model to explicitly maintain the "ground truth". Furthermore, we integrate a rendering-based critique loop that serves as a hard constraint validator, providing explicit visual feedback to resolve complex tasks that are difficult to articulate through text alone. Extensive experiments on VCode, RBench-V, and MathVista demonstrate that Canvas-CoT significantly outperforms existing baselines, establishing a new paradigm for context-efficient multimodal reasoning.

Hao Liang, Zhengyang Zhao, Zhaoyang Han et al.

Large language models (LLMs) have demonstrated remarkable performance across a wide range of tasks and domains, with data playing a central role in enabling these advances. Despite this success, the preparation and effective utilization of the massive datasets required for LLM training remain major bottlenecks. In current practice, LLM training data is often constructed using ad hoc scripts, and there is still a lack of mature, agent-based data preparation systems that can automatically construct robust and reusable data workflows, thereby freeing data scientists from repetitive and error-prone engineering efforts. Moreover, once collected, datasets are often consumed largely in their entirety during training, without systematic mechanisms for data selection, mixture optimization, or reweighting. To address these limitations, we advocate two complementary research directions. First, we propose building a robust, agent-based automatic data preparation system that supports automated workflow construction and scalable data management. Second, we argue for a unified data-model interaction training system in which data is dynamically selected, mixed, and reweighted throughout the training process, enabling more efficient, adaptive, and performance-aware data utilization. Finally, we discuss the remaining challenges and outline promising directions for future research and system development.

Hao Liang, Meiyi Qiang, Yuying Li et al.

With the rapid development of large language models (LLMs), the quality of training data has become crucial. Among the various types of training data, mathematical data plays a key role in enabling LLMs to acquire strong reasoning abilities. While high-quality open-source data is important, it is often insufficient for pre-training, necessitating the addition of synthetic math problems. However, synthetic math questions and answers can introduce inaccuracies, which may degrade both the training data and web data. Therefore, an effective method for cleaning synthetic math data is essential. In this paper, we propose the MathClean benchmark to evaluate the effectiveness of math data cleaning models. The MathClean benchmark consists of 2,000 correct questions and 2,000 erroneous questions with additional 2,000 correct and erroneous answers sourced from augmented data based on GSM8K and MATH. Moreover, we also annotate error types for each question or answer, since it can assess whether models can correctly identify the error categories for future improvements. Finally, we present comprehensive evaluations using state-of-the-art (SOTA) models. Our results demonstrate that even strong models like GPT-o1 and DeepSeek-R1 perform poorly on this benchmark, highlighting the utility of MathClean. Our code and data is available at https://github.com/YuYingLi0/MathClean.

Tianyu Guo, Hongyu Chen, Hao Liang et al.

Automatic audio captioning is essential for audio understanding, enabling applications such as accessibility and content indexing. However, evaluating the quality of audio captions remains a major challenge, especially in reference-free settings where high-quality ground-truth captions are unavailable. While CLAPScore is currently the most widely used reference-free Audio Caption Evaluation Metric(ACEM), its robustness under diverse conditions has not been systematically validated. To address this gap, we introduce BRACE, a new benchmark designed to evaluate audio caption alignment quality in a reference-free setting. BRACE is primarily designed for assessing ACEMs, and can also be extended to measure the modality alignment abilities of Large Audio Language Model(LALM). BRACE consists of two sub-benchmarks: BRACE-Main for fine-grained caption comparison and BRACE-Hallucination for detecting subtle hallucinated content. We construct these datasets through high-quality filtering, LLM-based corruption, and human annotation. Given the widespread adoption of CLAPScore as a reference-free ACEM and the increasing application of LALMs in audio-language tasks, we evaluate both approaches using the BRACE benchmark, testing CLAPScore across various CLAP model variants and assessing multiple LALMs. Notably, even the best-performing CLAP-based ACEM achieves only a 70.01 F1-score on the BRACE-Main benchmark, while the best LALM reaches just 63.19. By revealing the limitations of CLAP models and LALMs, our BRACE benchmark offers valuable insights into the direction of future research.

Hao Liang, Zhi-quan Luo

We study finite episodic Markov decision processes incorporating dynamic risk measures to capture risk sensitivity. To this end, we present two model-based algorithms applied to \emph{Lipschitz} dynamic risk measures, a wide range of risk measures that subsumes spectral risk measure, optimized certainty equivalent, distortion risk measures among others. We establish both regret upper bounds and lower bounds. Notably, our upper bounds demonstrate optimal dependencies on the number of actions and episodes, while reflecting the inherent trade-off between risk sensitivity and sample complexity. Additionally, we substantiate our theoretical results through numerical experiments.

Hao Liang, Ruitao Wu, Bohan Zeng et al.

Multimodal reasoning remains a fundamental challenge in artificial intelligence. Despite substantial advances in text-based reasoning, even state-of-the-art models such as GPT-o3 struggle to maintain strong performance in multimodal scenarios. To address this gap, we introduce a caption-assisted reasoning framework that effectively bridges visual and textual modalities. Our approach achieved 1st place in the ICML 2025 AI for Math Workshop \& Challenge 2: SeePhys, highlighting its effectiveness and robustness. Furthermore, we validate its generalization on the MathVerse benchmark for geometric reasoning, demonstrating the versatility of our method. Our code is publicly available at https://github.com/OpenDCAI/SciReasoner.

Qifeng Cai, Hao Liang, Hejun Dong et al.

Long videos contain a vast amount of information, making video-text retrieval an essential and challenging task in multimodal learning. However, existing benchmarks suffer from limited video duration, low-quality captions, and coarse annotation granularity, which hinder the evaluation of advanced video-text retrieval methods. To address these limitations, we introduce LoVR, a benchmark specifically designed for long video-text retrieval. LoVR contains 467 long videos and over 40,804 fine-grained clips with high-quality captions. To overcome the issue of poor machine-generated annotations, we propose an efficient caption generation framework that integrates VLM automatic generation, caption quality scoring, and dynamic refinement. This pipeline improves annotation accuracy while maintaining scalability. Furthermore, we introduce a semantic fusion method to generate coherent full-video captions without losing important contextual information. Our benchmark introduces longer videos, more detailed captions, and a larger-scale dataset, presenting new challenges for video understanding and retrieval. Extensive experiments on various advanced embedding models demonstrate that LoVR is a challenging benchmark, revealing the limitations of current approaches and providing valuable insights for future research. We release the code and dataset link at https://github.com/TechNomad-ds/LoVR-benchmark