Minghao Li

GLM-5 Team, Aohan Zeng, Xin Lv et al. · tsinghua

We present GLM-5, a next-generation foundation model designed to transition the paradigm of vibe coding to agentic engineering. Building upon the agentic, reasoning, and coding (ARC) capabilities of its predecessor, GLM-5 adopts DSA to significantly reduce training and inference costs while maintaining long-context fidelity. To advance model alignment and autonomy, we implement a new asynchronous reinforcement learning infrastructure that drastically improves post-training efficiency by decoupling generation from training. Furthermore, we propose novel asynchronous agent RL algorithms that further improve RL quality, enabling the model to learn from complex, long-horizon interactions more effectively. Through these innovations, GLM-5 achieves state-of-the-art performance on major open benchmarks. Most critically, GLM-5 demonstrates unprecedented capability in real-world coding tasks, surpassing previous baselines in handling end-to-end software engineering challenges. Code, models, and more information are available at https://github.com/zai-org/GLM-5.

Feifan Song, Bowen Yu, Minghao Li et al. · pku

Large language models (LLMs) often contain misleading content, emphasizing the need to align them with human values to ensure secure AI systems. Reinforcement learning from human feedback (RLHF) has been employed to achieve this alignment. However, it encompasses two main drawbacks: (1) RLHF exhibits complexity, instability, and sensitivity to hyperparameters in contrast to SFT. (2) Despite massive trial-and-error, multiple sampling is reduced to pair-wise contrast, thus lacking contrasts from a macro perspective. In this paper, we propose Preference Ranking Optimization (PRO) as an efficient SFT algorithm to directly fine-tune LLMs for human alignment. PRO extends the pair-wise contrast to accommodate preference rankings of any length. By iteratively contrasting candidates, PRO instructs the LLM to prioritize the best response while progressively ranking the rest responses. In this manner, PRO effectively transforms human alignment into aligning the probability ranking of n responses generated by LLM with the preference ranking of humans towards these responses. Experiments have shown that PRO outperforms baseline algorithms, achieving comparable results to ChatGPT and human responses through automatic-based, reward-based, GPT-4, and human evaluations.

Minghao Li, Yingxiu Zhao, Bowen Yu et al. · pku

Recent research has demonstrated that Large Language Models (LLMs) can enhance their capabilities by utilizing external tools. However, three pivotal questions remain unanswered: (1) How effective are current LLMs in utilizing tools? (2) How can we enhance LLMs' ability to utilize tools? (3) What obstacles need to be overcome to leverage tools? To address these questions, we introduce API-Bank, a groundbreaking benchmark, specifically designed for tool-augmented LLMs. For the first question, we develop a runnable evaluation system consisting of 73 API tools. We annotate 314 tool-use dialogues with 753 API calls to assess the existing LLMs' capabilities in planning, retrieving, and calling APIs. For the second question, we construct a comprehensive training set containing 1,888 tool-use dialogues from 2,138 APIs spanning 1,000 distinct domains. Using this dataset, we train Lynx, a tool-augmented LLM initialized from Alpaca. Experimental results demonstrate that GPT-3.5 exhibits improved tool utilization compared to GPT-3, while GPT-4 excels in planning. However, there is still significant potential for further improvement. Moreover, Lynx surpasses Alpaca's tool utilization performance by more than 26 pts and approaches the effectiveness of GPT-3.5. Through error analysis, we highlight the key challenges for future research in this field to answer the third question.

Qifei Jia, Xintong Yao, Minghao Li et al.

Traditional Image Aesthetic Assessment (IAA) methods mainly rely on regressing absolute Mean Opinion Scores (MOS). However, such a paradigm overlooks the inherently dynamic nature of human aesthetic perception, which relies on subconscious comparison against implicit visual references. Consequently, the lack of causal reasoning regarding aesthetic differences prevents models from learning generalizable aesthetic principles, thus limiting their generalization across diverse scenarios. In this work, we rethink the IAA task and propose Relative Edit-induced Difference Aesthetic learning (RED-Aes), a novel framework that leverages controllable image editing models to simulate the human aesthetic reasoning process. Instead of fitting absolute score distributions, RED-Aes explicitly learns the visual factors that drive aesthetic changes. To support this paradigm, we construct the RED-20k dataset, which comprises editing-based image pairs, quantitative aesthetic differences, and Chain-of-Thought (CoT) reasoning. Furthermore, we introduce a three-stage training strategy guided by a relative ranking consistency reward, optimizing the model solely via relative supervision. Extensive experiments demonstrate that RED-Aes achieves state-of-the-art performance on multiple public benchmarks, exhibiting superior generalization capabilities.

Minghao Li, Ran Ben Basat, Shay Vargaftik et al.

Deep neural networks (DNNs) are the de facto standard for essential use cases, such as image classification, computer vision, and natural language processing. As DNNs and datasets get larger, they require distributed training on increasingly larger clusters. A main bottleneck is the resulting communication overhead where workers exchange model updates (i.e., gradients) on a per-round basis. To address this bottleneck and accelerate training, a widely-deployed approach is compression. However, previous deployments often apply bi-directional compression schemes by simply using a uni-directional gradient compression scheme in each direction. This results in significant computational overheads at the parameter server and increased compression error, leading to longer training and lower accuracy. We introduce Tensor Homomorphic Compression (THC), a novel bi-directional compression framework that enables the direct aggregation of compressed values and thus eliminating the aforementioned computational overheads. Moreover, THC is compatible with in-network aggregation (INA), which allows for further acceleration. Our evaluation shows that training representative vision and language models with THC reaches target accuracy by 1.40x to 1.47x faster using INA and 1.28x to 1.33x faster using a software PS compared with state-of-the-art systems.

Liujie Zhang, Benzhe Ning, Rui Yang et al.

Reinforcement learning (RL) post-training has proven effective at unlocking reasoning, self-reflection, and tool-use capabilities in large language models. As models extend to omni-modal inputs and agentic multi-turn workflows, RL training systems face three interdependent challenges: heterogeneous data flows, operational robustness at scale, and the staleness -- throughput tradeoff. We present \textbf{Relax} (Reinforcement Engine Leveraging Agentic X-modality), an open-source RL training engine that addresses these challenges through three co-designed architectural layers. First, an \emph{omni-native architecture} builds multimodal support into the full stack -- from data preprocessing and modality-aware parallelism to inference generation -- rather than retrofitting it onto a text-centric pipeline. Second, each RL role runs as an independent, fault-isolated service that can be scaled, recovered, and upgraded without global coordination. Third, service-level decoupling enables asynchronous training via the TransferQueue data bus, where a single staleness parameter smoothly interpolates among on-policy, near-on-policy, and fully asynchronous execution. Relax achieves a 1.20$\times$ end-to-end speedup over veRL on Qwen3-4B on-policy training. Its fully async mode delivers a 1.76$\times$ speedup over colocate on Qwen3-4B and a 2.00$\times$ speedup on Qwen3-Omni-30B, while all modes converge to the same reward level. Relax supports R3 (Rollout Routing Replay)~\cite{ma2025r3} for MoE models with only 1.9\% overhead, compared to 32\% degradation in veRL under the same configuration. It further demonstrates stable omni-modal RL convergence on Qwen3-Omni across image, text, and audio, sustaining over 2{,}000 steps on video without degradation. Relax is available at https://github.com/rednote-ai/Relax.

Wenhui Tan, Minghao Li, Xiaoqian Ma et al.

Long chain-of-thought reasoning has made autoregressive decoding the dominant inference cost of modern large language models. Existing methods target either the input side (latent compression) or the output side (speculative decoding and multi-token prediction, MTP), but the two lines of work have been pursued independently. Moreover, output-side methods must incur an expensive verifier pass to validate the unreliable draft tokens predicted by MTP. To address these issues, we propose \textbf{Pair-In, Pair-Out (PIPO)}, which unifies both sides by viewing a latent compressor and an MTP head as mirror-image operations: the compressor folds two input tokens into one latent representation, while the MTP head unfolds one hidden state into one additional output token. To remove the verifier cost without sacrificing reliability, PIPO trains a lightweight confidence head that decides whether draft tokens should be accepted. We observe that On-Policy Distillation (OPD) naturally matches the rejection-sampling criterion of speculative decoding, so the confidence head can be trained alongside OPD with negligible extra cost. Experiments on AIME 2025, GPQA-Diamond, LiveCodeBench v6, and LongBench v2 with Qwen3.5-4B and 9B backbones show that PIPO improves pass@4 over regular decoding by up to $+7.15$ points, while delivering up to $2.64\times$ first-token-latency and $2.07\times$ per-token-latency speedups.

Shikun Feng, Minghao Li, Yinjun Jia et al.

The binding between proteins and ligands plays a crucial role in the realm of drug discovery. Previous deep learning approaches have shown promising results over traditional computationally intensive methods, but resulting in poor generalization due to limited supervised data. In this paper, we propose to learn protein-ligand binding representation in a self-supervised learning manner. Different from existing pre-training approaches which treat proteins and ligands individually, we emphasize to discern the intricate binding patterns from fine-grained interactions. Specifically, this self-supervised learning problem is formulated as a prediction of the conclusive binding complex structure given a pocket and ligand with a Transformer based interaction module, which naturally emulates the binding process. To ensure the representation of rich binding information, we introduce two pre-training tasks, i.e.~atomic pairwise distance map prediction and mask ligand reconstruction, which comprehensively model the fine-grained interactions from both structure and feature space. Extensive experiments have demonstrated the superiority of our method across various binding tasks, including protein-ligand affinity prediction, virtual screening and protein-ligand docking.

GLM-4. 5 Team, Aohan Zeng, Xin Lv et al.

We present GLM-4.5, an open-source Mixture-of-Experts (MoE) large language model with 355B total parameters and 32B activated parameters, featuring a hybrid reasoning method that supports both thinking and direct response modes. Through multi-stage training on 23T tokens and comprehensive post-training with expert model iteration and reinforcement learning, GLM-4.5 achieves strong performance across agentic, reasoning, and coding (ARC) tasks, scoring 70.1% on TAU-Bench, 91.0% on AIME 24, and 64.2% on SWE-bench Verified. With much fewer parameters than several competitors, GLM-4.5 ranks 3rd overall among all evaluated models and 2nd on agentic benchmarks. We release both GLM-4.5 (355B parameters) and a compact version, GLM-4.5-Air (106B parameters), to advance research in reasoning and agentic AI systems. Code, models, and more information are available at https://github.com/zai-org/GLM-4.5.

Dezhang Kong, Zhuxi Wu, Shiqi Liu et al.

LLM-based web agents have become increasingly popular for their utility in daily life and work. However, they exhibit critical vulnerabilities when processing malicious URLs: accepting a disguised malicious URL enables subsequent access to unsafe webpages, which can cause severe damage to service providers and users. Despite this risk, no benchmark currently targets this emerging threat. To address this gap, we propose MalURLBench, the first benchmark for evaluating LLMs' vulnerabilities to malicious URLs. MalURLBench contains 61,845 attack instances spanning 10 real-world scenarios and 7 categories of real malicious websites. Experiments with 12 popular LLMs reveal that existing models struggle to detect elaborately disguised malicious URLs. We further identify and analyze key factors that impact attack success rates and propose URLGuard, a lightweight defense module. We believe this work will provide a foundational resource for advancing the security of web agents. Our code is available at https://github.com/JiangYingEr/MalURLBench.

Minghao Li, Alicia Golden, Samuel Hsia et al.

The rapid scaling of large language model training requires distributing GPU resources across multiple data center buildings and regions. We refer to such paradigm as "scale-across" training. As infrastructure expands, the system design space becomes increasingly intricate, encompassing new model architectures, hardware heterogeneity, and evolving communication patterns. Drawing from Meta's production experience, we highlight the complexities of deploying training jobs across a few data centers housing hundreds of thousands of GPUs. To accelerate exploration of the large design space and to enable efficient training for frontier model development, we conduct in-depth characterization of three key design dimensions: parallelism placement, parallelism scheduling, and network layer technologies. We then propose ScaleAcross Explorer, an optimizer that considers the interplay of design dimensions and holistically optimizes scale-across training. Testbed experiments and simulations demonstrate up to 64.62% training speedups over production configuration and up to 37.59% training speedups over the state-of-the-art baseline across a wide range of design points.

Xiangyu Peng, Zangwei Zheng, Chenhui Shen et al.

Video generation models have achieved remarkable progress in the past year. The quality of AI video continues to improve, but at the cost of larger model size, increased data quantity, and greater demand for training compute. In this report, we present Open-Sora 2.0, a commercial-level video generation model trained for only $200k. With this model, we demonstrate that the cost of training a top-performing video generation model is highly controllable. We detail all techniques that contribute to this efficiency breakthrough, including data curation, model architecture, training strategy, and system optimization. According to human evaluation results and VBench scores, Open-Sora 2.0 is comparable to global leading video generation models including the open-source HunyuanVideo and the closed-source Runway Gen-3 Alpha. By making Open-Sora 2.0 fully open-source, we aim to democratize access to advanced video generation technology, fostering broader innovation and creativity in content creation. All resources are publicly available at: https://github.com/hpcaitech/Open-Sora.

Jiale Zhao, Guoxin Chen, Fanzhe Meng et al.

Achieving mastery in real world software engineering tasks is fundamentally bottlenecked by the scarcity of large scale, high quality training data. Scaling such data has been limited by the complexity of environment setup, unit test generation, and problem statement curation. In this paper, we propose ScaleSWE, an automated, sandboxed multi agent workflow designed to construct high quality SWE data at scale. The system coordinates three specialized agents for environment setup, test creation, and problem description synthesis to process 6 million pull requests across 5200 repositories, producing Scale SWE Data: 100k verified SWE instances, the largest such dataset to date. It substantially surpasses existing real world datasets in repository diversity and reflects realistic task complexity. We further demonstrate the dataset utility for training by distilling 71498 high quality trajectories and finetuning Qwen30BA3BInstruct to produce ScaleSWE Agent. Our agent achieves a 64 resolve rate on SWE Bench Verified a nearly three fold improvement over the base model. ScaleSWE provides a scalable, reproducible approach for data construction to advance LLM based software engineering. Scale SWE will be publicly available.

Weifan Jiang, Rana Shahout, Minghao Li et al.

Multi-Agent Debate (MAD) improves LLM-agent accuracy but suffers from rapid context growth, limiting scalability in larger multi-agent settings. Existing methods prune low-utility communications using prior signals, such as token-level log-likelihoods or LLM self-reported confidence. However, these signals become unreliable under hallucination, degrading the accuracy of MAD methods that rely on them. We propose SVR-MAD, a Bayesian-inspired MAD framework that treats pre-debate signals as priors and debate outcomes as posterior-style evidence for estimating agent correctness. SVR-MAD uses this evidence to incrementally construct the communication graph, prioritizing agents whose answers survive peer challenges. Experiments across multiple LLMs and benchmarks show that SVR-MAD reduces token cost by up to 61% while matching or improving accuracy relative to the most accurate competing MAD baseline.

Guoxin Chen, Fanzhe Meng, Jiale Zhao et al.

Current benchmarks for code agents primarily assess narrow, repository-specific fixes, overlooking critical real-world challenges such as cross-repository reasoning, domain-specialized problem solving, dependency-driven migration, and full-repository generation. To address this gap, we introduce BeyondSWE, a comprehensive benchmark that broadens existing evaluations along two axes - resolution scope and knowledge scope - using 500 real-world instances across four distinct settings. Experimental results reveal a significant capability gap: even frontier models plateau below 45% success, and no single model performs consistently across task types. To systematically investigate the role of external knowledge, we develop SearchSWE, a framework that integrates deep search with coding abilities. Our experiments show that search augmentation yields inconsistent gains and can in some cases degrade performance, highlighting the difficulty of emulating developer-like workflows that interleave search and reasoning during coding tasks. This work offers both a realistic, challenging evaluation benchmark and a flexible framework to advance research toward more capable code agents.

Minghao Li, Ying Zeng, Zhihao Cheng et al.

The advent of Deep Research agents has substantially reduced the time required for conducting extensive research tasks. However, these tasks inherently demand rigorous standards of factual accuracy and comprehensiveness, necessitating thorough evaluation before widespread adoption. In this paper, we propose ReportBench, a systematic benchmark designed to evaluate the content quality of research reports generated by large language models (LLMs). Our evaluation focuses on two critical dimensions: (1) the quality and relevance of cited literature, and (2) the faithfulness and veracity of the statements within the generated reports. ReportBench leverages high-quality published survey papers available on arXiv as gold-standard references, from which we apply reverse prompt engineering to derive domain-specific prompts and establish a comprehensive evaluation corpus. Furthermore, we develop an agent-based automated framework within ReportBench that systematically analyzes generated reports by extracting citations and statements, checking the faithfulness of cited content against original sources, and validating non-cited claims using web-based resources. Empirical evaluations demonstrate that commercial Deep Research agents such as those developed by OpenAI and Google consistently generate more comprehensive and reliable reports than standalone LLMs augmented with search or browsing tools. However, there remains substantial room for improvement in terms of the breadth and depth of research coverage, as well as factual consistency. The complete code and data will be released at the following link: https://github.com/ByteDance-BandAI/ReportBench

Zheng Li, Siyao Song, Jingyuan Ma et al.

Large language models (LLMs) show promise as teaching assistants, yet their teaching capability remains insufficiently evaluated. Existing benchmarks mainly focus on problem-solving or problem-level guidance, leaving knowledge-centered teaching underexplored. We propose a syllabus-grounded evaluation framework that measures LLM teaching capability via student performance improvement after multi-turn instruction. By restricting teacher agents to structured knowledge points and example problems, the framework avoids information leakage and enables reuse of existing benchmarks. We instantiate the framework on Gaokao data across multiple subjects. Experiments reveal substantial variation in teaching effectiveness across models and domains: some models perform well in mathematics, while teaching remains challenging in physics and chemistry. We also find that incorporating example problems does not necessarily improve teaching, as models often shift toward example-specific error correction. Overall, our results highlight teaching ability as a distinct and measurable dimension of LLM behavior.

Minghao Li, Ruihang Wang, Rui Tan et al.

Modern data centers (DCs) hosting artificial intelligence (AI)-dedicated devices operate at high power densities with rapidly varying workloads, making minute-level adaptation essential for safe and energy-efficient operation. However, manually designing piecewise deep reinforcement learning (DRL) agents cannot keep pace with frequent dynamics shifts and service-level agreement (SLA) changes of an evolving DC. This specification-to-policy lag causes a lack of timely, effective control policies, which may lead to service outages. To bridge the gap, we present DCoPilot, a hybrid framework for generative control policies in dynamic DC operation. DCoPilot synergizes two distinct generative paradigms, i.e., a large language model (LLM) that performs symbolic generation of structured reward forms, and a hypernetwork that conducts parametric generation of policy weights. DCoPilot operates through three coordinated phases: (i) simulation scale-up, which stress-tests reward candidates across diverse simulation-ready (SimReady) scenes; (ii) meta policy distillation, where a hypernetwork is trained to output policy weights conditioned on SLA and scene embeddings; and (iii) online adaptation, enabling zero-shot policy generation in response to updated specifications. Evaluated across five control task families spanning diverse DC components, DCoPilot achieves near-zero constraint violations and outperforms all baselines across specification variations. Ablation studies validate the effectiveness of LLM-based unified reward generation in enabling stable hypernetwork convergence.

Gaoyi Chen, Minghao Li, Weishi Shi et al.

Metric differential privacy (mDP) strengthens local differential privacy (LDP) by scaling noise to semantic distance, but many machine learning (ML) systems are consumed under joint observation, where model-agnostic, per-record guarantees can miss leakage from evidence aggregation. We introduce metric-normalized posterior leakage (mPL), an attacker-aligned, distance-calibrated measure of posterior-odds shift induced by releases, and show that for single or independent releases, uniformly bounding mPL is equivalent to mDP. Under joint observation, however, satisfying mDP may still leave mPL high because learned aggregators compound evidence across correlated items. To make control practical, we formalize probabilistically bounded mPL (PBmPL), which limits how often mPL may exceed a target budget, and we operationalize it via Adaptive mPL (AmPL), a trust-and-verify framework that perturbs, audits with a learned attacker, and adapts parameters (with optional Bayesian remapping) to balance privacy and utility. In a word-embedding case study, neural adversaries violate mPL under joint consumption despite per-record mDP perturbations, whereas AmPL substantially lowers the frequency of such violations with low utility loss, indicating PBmPL as a practical, certifiable protection for joint-consumption settings.

Zhaojing Zhou, Xunchao Li, Minghao Li et al.

The rapid scaling of Large Language Models (LLMs) elevates inference costs and compounds substantial deployment barriers. While quantization to 8 or 4 bits mitigates this, sub-3-bit methods face severe accuracy, scalability, and efficiency degradation. We propose Convolutional Code Quantization (CCQ), an inference-optimized quantization approach compressing LLMs to 2.0-2.75 bits with minimal accuracy loss. Departing from error-prone scalar quantization or slow vector quantization, CCQ integrates a hardware-aware bit-shift encoding and decoding solution with Convolutional Code, Hybrid Encoding, and Code Cluster, jointly overcoming accuracy-speed bottlenecks. We construct a lookup-free encoding space, enabling a linear mapping between the codebook and weight vectors, thereby optimizing inference performance. Meanwhile, by drawing on the concept of data mapping from vector quantization, we minimize the performance degradation of the model under extremely low-bit conditions. Experiments demonstrate that CCQ achieves outstanding performance on LLMs across various benchmarks. We compress DeepSeek-V3 (671B total parameters) to 184GB and ERNIE-4.5-300B-A47B to 89GB, enabling single-GPU deployment of ERNIE 4.5 and eliminating inter-card communication. The 2-bit ERNIE-4.5-300B-A47B model and inference engine have been open-sourced.

Huaixiao Tou, Ying Zeng, Yuemeng Li et al.

We present ShoppingComp, a challenging real-world benchmark for comprehensively evaluating LLM-powered shopping agents on three core capabilities: precise product retrieval, expert-level report generation, and safety critical decision making. Unlike prior e-commerce benchmarks, ShoppingComp introduces difficult product discovery queries with many constraints, while guaranteeing open-world products and enabling easy verification of agent outputs. The benchmark comprises 145 instances and 558 scenarios, curated by 35 experts to reflect authentic shopping needs. Results reveal stark limitations of current LLMs: even state-of-the-art models achieve low performance (e.g., 17.76\% for GPT-5.2, 15.82\% for Gemini-3-Pro).Error analysis reflects limitations in core agent competencies, including information grounding in open-world environments, reliable verification of multi-constraint requirements, consistent reasoning over noisy and conflicting evidence, and risk-aware decision making. By exposing these capability gaps, ShoppingComp characterizes the trust threshold that AI systems must cross before they can be proactively trusted for reliable real-world decision making. Our code and dataset are available at https://github.com/ByteDance-BandAI/ShoppingComp.

Yun Yue, Yongchao Liu, Suo Tong et al.

We develop a novel framework that adds the regularizers of the sparse group lasso to a family of adaptive optimizers in deep learning, such as Momentum, Adagrad, Adam, AMSGrad, AdaHessian, and create a new class of optimizers, which are named Group Momentum, Group Adagrad, Group Adam, Group AMSGrad and Group AdaHessian, etc., accordingly. We establish theoretically proven convergence guarantees in the stochastic convex settings, based on primal-dual methods. We evaluate the regularized effect of our new optimizers on three large-scale real-world ad click datasets with state-of-the-art deep learning models. The experimental results reveal that compared with the original optimizers with the post-processing procedure which uses the magnitude pruning method, the performance of the models can be significantly improved on the same sparsity level. Furthermore, in comparison to the cases without magnitude pruning, our methods can achieve extremely high sparsity with significantly better or highly competitive performance. The code is available at https://github.com/intelligent-machine-learning/tfplus/tree/main/tfplus.

Minghao Li, Yiheng Xu, Lei Cui et al.

Document layout analysis usually relies on computer vision models to understand documents while ignoring textual information that is vital to capture. Meanwhile, high quality labeled datasets with both visual and textual information are still insufficient. In this paper, we present \textbf{DocBank}, a benchmark dataset that contains 500K document pages with fine-grained token-level annotations for document layout analysis. DocBank is constructed using a simple yet effective way with weak supervision from the \LaTeX{} documents available on the arXiv.com. With DocBank, models from different modalities can be compared fairly and multi-modal approaches will be further investigated and boost the performance of document layout analysis. We build several strong baselines and manually split train/dev/test sets for evaluation. Experiment results show that models trained on DocBank accurately recognize the layout information for a variety of documents. The DocBank dataset is publicly available at \url{https://github.com/doc-analysis/DocBank}.

Minghao Li, Lei Cui, Shaohan Huang et al.

We present TableBank, a new image-based table detection and recognition dataset built with novel weak supervision from Word and Latex documents on the internet. Existing research for image-based table detection and recognition usually fine-tunes pre-trained models on out-of-domain data with a few thousand human-labeled examples, which is difficult to generalize on real-world applications. With TableBank that contains 417K high quality labeled tables, we build several strong baselines using state-of-the-art models with deep neural networks. We make TableBank publicly available and hope it will empower more deep learning approaches in the table detection and recognition task. The dataset and models are available at \url{https://github.com/doc-analysis/TableBank}.

Siqi Fan, Minghao Li, Xiaoqian Ma et al.

Large reasoning models achieve high accuracy through extended chain-of-thought but generate 5--8 more tokens than necessary, applying verbose reasoning uniformly regardless of problem difficulty. We propose Hint Tuning, a data-efficient approach that teaches models to calibrate reasoning depth. Our key insight: the corresponding instruct model serves as an ideal difficulty probe. By testing what the instruct model can solve with varying guidance, we automatically construct training data across three states: No-Hint (direct answer), Sparse-Hint (minimal prefix), and Full-Hint (complete reasoning). This converts the abstract challenge of difficulty labeling into a measurable consistency check between the instruct and reasoning models. With only 1K self-annotated samples, Hint Tuning achieves 24--66% token reduction (31.5% average) across mainstream reasoning models (Qwen3-Thinking, DeepSeek-R1-Distill) at multiple scales (4B--32B) while maintaining competitive accuracy on five benchmarks. Unlike methods requiring massive distillation datasets or expensive RL, we achieve superior efficiency through simple alignment with the instruct model's capabilities.

Shi Qiu, Shaoyang Guo, Zhuo-Yang Song et al.

Current benchmarks for evaluating the reasoning capabilities of Large Language Models (LLMs) face significant limitations: task oversimplification, data contamination, and flawed evaluation items. These deficiencies necessitate more rigorous assessment methods. To address these limitations, we introduce PHYBench, a benchmark of 500 original physics problems ranging from high school to Physics Olympiad difficulty. PHYBench addresses data contamination through original content and employs a systematic curation pipeline to eliminate flawed items. Evaluations show that PHYBench activates more tokens and provides stronger differentiation between reasoning models compared to other baselines like AIME 2024, OlympiadBench and GPQA. Even the best-performing model, Gemini 2.5 Pro, achieves only 36.9% accuracy compared to human experts' 61.9%. To further enhance evaluation precision, we introduce the Expression Edit Distance (EED) Score for mathematical expression assessment, which improves sample efficiency by 204% over binary scoring. Moreover, PHYBench effectively elicits multi-step and multi-condition reasoning, providing a platform for examining models' reasoning robustness, preferences, and deficiencies. The benchmark results and dataset are publicly available at https://www.phybench.cn/.

Shikun Feng, Yuyan Ni, Minghao Li et al.

Recently, a noticeable trend has emerged in developing pre-trained foundation models in the domains of CV and NLP. However, for molecular pre-training, there lacks a universal model capable of effectively applying to various categories of molecular tasks, since existing prevalent pre-training methods exhibit effectiveness for specific types of downstream tasks. Furthermore, the lack of profound understanding of existing pre-training methods, including 2D graph masking, 2D-3D contrastive learning, and 3D denoising, hampers the advancement of molecular foundation models. In this work, we provide a unified comprehension of existing pre-training methods through the lens of contrastive learning. Thus their distinctions lie in clustering different views of molecules, which is shown beneficial to specific downstream tasks. To achieve a complete and general-purpose molecular representation, we propose a novel pre-training framework, named UniCorn, that inherits the merits of the three methods, depicting molecular views in three different levels. SOTA performance across quantum, physicochemical, and biological tasks, along with comprehensive ablation study, validate the universality and effectiveness of UniCorn.

Wenpeng Xing, Minghao Li, Mohan Li et al.

Embodied AI systems, including robots and autonomous vehicles, are increasingly integrated into real-world applications, where they encounter a range of vulnerabilities stemming from both environmental and system-level factors. These vulnerabilities manifest through sensor spoofing, adversarial attacks, and failures in task and motion planning, posing significant challenges to robustness and safety. Despite the growing body of research, existing reviews rarely focus specifically on the unique safety and security challenges of embodied AI systems. Most prior work either addresses general AI vulnerabilities or focuses on isolated aspects, lacking a dedicated and unified framework tailored to embodied AI. This survey fills this critical gap by: (1) categorizing vulnerabilities specific to embodied AI into exogenous (e.g., physical attacks, cybersecurity threats) and endogenous (e.g., sensor failures, software flaws) origins; (2) systematically analyzing adversarial attack paradigms unique to embodied AI, with a focus on their impact on perception, decision-making, and embodied interaction; (3) investigating attack vectors targeting large vision-language models (LVLMs) and large language models (LLMs) within embodied systems, such as jailbreak attacks and instruction misinterpretation; (4) evaluating robustness challenges in algorithms for embodied perception, decision-making, and task planning; and (5) proposing targeted strategies to enhance the safety and reliability of embodied AI systems. By integrating these dimensions, we provide a comprehensive framework for understanding the interplay between vulnerabilities and safety in embodied AI.

Minghao Li, Dmitrii Avdiukhin, Rana Shahout et al.

Federated Learning (FL) trains deep models across edge devices without centralizing raw data, preserving user privacy. However, client heterogeneity slows down convergence and limits global model accuracy. Clustered FL (CFL) mitigates this by grouping clients with similar representations and training a separate model for each cluster. In practice, client data evolves over time, a phenomenon we refer to as data drift, which breaks cluster homogeneity and degrades performance. Data drift can take different forms depending on whether changes occur in the output values, the input features, or the relationship between them. We propose FIELDING, a CFL framework for handling diverse types of data drift with low overhead. FIELDING detects drift at individual clients and performs selective re-clustering to balance cluster quality and model performance, while remaining robust to malicious clients and varying levels of heterogeneity. Experiments show that FIELDING improves final model accuracy by 1.9-5.9% and achieves target accuracy 1.16x-2.23x faster than existing state-of-the-art CFL methods.

Minghao LI, Ruihang Wang, Rui Tan et al.

Data center (DC) infrastructure serves as the backbone to support the escalating demand for computing capacity. Traditional design methodologies that blend human expertise with specialized simulation tools scale poorly with the increasing system complexity. Recent studies adopt generative artificial intelligence to design plausible human-centric indoor layouts. However, they do not consider the underlying physics, making them unsuitable for the DC design that sets quantifiable operational objectives and strict physical constraints. To bridge the gap, we propose Phythesis, a novel framework that synergizes large language models (LLMs) and physics-guided evolutionary optimization to automate simulation-ready (SimReady) scene synthesis for energy-efficient DC design. Phythesis employs an iterative bi-level optimization architecture, where (i) the LLM-driven optimization level generates physically plausible three-dimensional layouts and self-criticizes them to refine the scene topology, and (ii) the physics-informed optimization level identifies the optimal asset parameters and selects the best asset combination. Experiments on three generation scales show that Phythesis achieves 57.3% generation success rate increase and 11.5% power usage effectiveness (PUE) improvement, compared with the vanilla LLM-based solution.

Siyao Song, Cong Ma, Zhihao Cheng et al.

Large language models (LLMs) have recently advanced in reasoning when optimized with reinforcement learning (RL) under verifiable rewards. Existing methods primarily rely on outcome-based supervision to strengthen internal LLM reasoning, often leading to inefficient exploration and sparse rewards. To mitigate this issue, we propose Expert-Assisted Policy Optimization (EAPO), a novel RL framework that enhances exploration by incorporating multi-turn interactions with external experts during training. Unlike prior methods, where policies reason in isolation, EAPO incentivizes the policy to adaptively determine when and how to consult experts, yielding richer reward signals and more reliable reasoning trajectories. External assistance ultimately internalizes expert knowledge into the policy model, amplifying the model's inherent reasoning capabilities. During evaluation, the policy model has been well-optimized to solve questions independently, producing improved reasoning paths and more accurate solutions. Experiments on mathematical reasoning benchmarks, including AIME 2024, AIME 2025, and AIMO 2025, show that EAPO consistently outperforms expert-assisted workflow, expert-distilled models, and RL baselines, with an average gain of 5 points over self-exploratory models.

Ruihang Wang, Minghao Li, Zhiwei Cao et al.

The explosion in artificial intelligence (AI) applications is pushing the development of AI-dedicated data centers (AIDCs), creating management challenges that traditional methods and standalone AI solutions struggle to address. While digital twins are beneficial for AI-based design validation and operational optimization, current AI methods for their creation face limitations. Specifically, physical AI (PhyAI) aims to capture the underlying physical laws, which demands extensive, case-specific customization, and generative AI (GenAI) can produce inaccurate or hallucinated results. We propose Fusion Intelligence, a novel framework synergizing GenAI's automation with PhyAI's domain grounding. In this dual-agent collaboration, GenAI interprets natural language prompts to generate tokenized AIDC digital twins. Subsequently, PhyAI optimizes these generated twins by enforcing physical constraints and assimilating real-time data. Case studies demonstrate the advantages of our framework in automating the creation and validation of AIDC digital twins. These twins deliver predictive analytics to support power usage effectiveness (PUE) optimization in the design stage. With operational data collected, the digital twin accuracy is further improved compared with pure physics-based models developed by human experts. Fusion Intelligence offers a promising pathway to accelerate digital transformation. It enables more reliable and efficient AI-driven digital transformation for a broad range of mission-critical infrastructures.

Zhiwei Cao, Minghao Li, Feng Lin et al.

Data centers (DCs) as mission-critical infrastructures are pivotal in powering the growth of artificial intelligence (AI) and the digital economy. The evolution from Internet DC to AI DC has introduced new challenges in operating and managing data centers for improved business resilience and reduced total cost of ownership. As a result, new paradigms, beyond the traditional approaches based on best practices, must be in order for future data centers. In this research, we propose and develop a novel Physical AI (PhyAI) framework for advancing DC operations and management. Our system leverages the emerging capabilities of state-of-the-art industrial products and our in-house research and development. Specifically, it presents three core modules, namely: 1) an industry-grade in-house simulation engine to simulate DC operations in a highly accurate manner, 2) an AI engine built upon NVIDIA PhysicsNemo for the training and evaluation of physics-informed machine learning (PIML) models, and 3) a digital twin platform built upon NVIDIA Omniverse for our proposed 5-tier digital twin framework. This system presents a scalable and adaptable solution to digitalize, optimize, and automate future data center operations and management, by enabling real-time digital twins for future data centers. To illustrate its effectiveness, we present a compelling case study on building a surrogate model for predicting the thermal and airflow profiles of a large-scale DC in a real-time manner. Our results demonstrate its superior performance over traditional time-consuming Computational Fluid Dynamics/Heat Transfer (CFD/HT) simulation, with a median absolute temperature prediction error of 0.18 °C. This emerging approach would open doors to several potential research directions for advancing Physical AI in future DC operations.

Jingning Xu, Benlai Tang, Mingjie Wang et al.

Recently, talking face generation has drawn ever-increasing attention from the research community in computer vision due to its arduous challenges and widespread application scenarios, e.g. movie animation and virtual anchor. Although persevering efforts have been undertaken to enhance the fidelity and lip-sync quality of generated talking face videos, there is still large room for further improvements of synthesis quality and efficiency. Actually, these attempts somewhat ignore the explorations of fine-granularity feature extraction/integration and the consistency between probability distributions of landmarks, thereby recurring the issues of local details blurring and degraded fidelity. To mitigate these dilemmas, in this paper, a novel CLIP-based Attention and Probability Map Guided Network (CPNet) is delicately designed for inferring high-fidelity talking face videos. Specifically, considering the demands of fine-grained feature recalibration, a clip-based attention condenser is exploited to transfer knowledge with rich semantic priors from the prevailing CLIP model. Moreover, to guarantee the consistency in probability space and suppress the landmark ambiguity, we creatively propose the density map of facial landmark as auxiliary supervisory signal to guide the landmark distribution learning of generated frame. Extensive experiments on the widely-used benchmark dataset demonstrate the superiority of our CPNet against state of the arts in terms of image and lip-sync quality. In addition, a cohort of studies are also conducted to ablate the impacts of the individual pivotal components.

Minghao Li, Tengchao Lv, Jingye Chen et al.

Text recognition is a long-standing research problem for document digitalization. Existing approaches are usually built based on CNN for image understanding and RNN for char-level text generation. In addition, another language model is usually needed to improve the overall accuracy as a post-processing step. In this paper, we propose an end-to-end text recognition approach with pre-trained image Transformer and text Transformer models, namely TrOCR, which leverages the Transformer architecture for both image understanding and wordpiece-level text generation. The TrOCR model is simple but effective, and can be pre-trained with large-scale synthetic data and fine-tuned with human-labeled datasets. Experiments show that the TrOCR model outperforms the current state-of-the-art models on the printed, handwritten and scene text recognition tasks. The TrOCR models and code are publicly available at \url{https://aka.ms/trocr}.

Minghao Li, Yingrui Jie, Yang Kong et al.

The problem of multi-robot navigation of connectivity maintenance is challenging in multi-robot applications. This work investigates how to navigate a multi-robot team in unknown environments while maintaining connectivity. We propose a reinforcement learning (RL) approach to develop a decentralized policy, which is shared among multiple robots. Given range sensor measurements and the positions of other robots, the policy aims to generate control commands for navigation and preserve the global connectivity of the robot team. We incorporate connectivity concerns into the RL framework as constraints and introduce behavior cloning to reduce the exploration complexity of policy optimization. The policy is optimized with all transition data collected by multiple robots in random simulated scenarios. We validate the effectiveness of the proposed approach by comparing different combinations of connectivity constraints and behavior cloning. We also show that our policy can generalize to unseen scenarios in both simulation and holonomic robots experiments.

Yuchuan Gou, Qiancheng Wu, Minghao Li et al.

In this paper, we propose a novel generative network (SegAttnGAN) that utilizes additional segmentation information for the text-to-image synthesis task. As the segmentation data introduced to the model provides useful guidance on the generator training, the proposed model can generate images with better realism quality and higher quantitative measures compared with the previous state-of-art methods. We achieved Inception Score of 4.84 on the CUB dataset and 3.52 on the Oxford-102 dataset. Besides, we tested the self-attention SegAttnGAN which uses generated segmentation data instead of masks from datasets for attention and achieved similar high-quality results, suggesting that our model can be adapted for the text-to-image synthesis task.

Yiheng Xu, Minghao Li, Lei Cui et al.

Pre-training techniques have been verified successfully in a variety of NLP tasks in recent years. Despite the widespread use of pre-training models for NLP applications, they almost exclusively focus on text-level manipulation, while neglecting layout and style information that is vital for document image understanding. In this paper, we propose the \textbf{LayoutLM} to jointly model interactions between text and layout information across scanned document images, which is beneficial for a great number of real-world document image understanding tasks such as information extraction from scanned documents. Furthermore, we also leverage image features to incorporate words' visual information into LayoutLM. To the best of our knowledge, this is the first time that text and layout are jointly learned in a single framework for document-level pre-training. It achieves new state-of-the-art results in several downstream tasks, including form understanding (from 70.72 to 79.27), receipt understanding (from 94.02 to 95.24) and document image classification (from 93.07 to 94.42). The code and pre-trained LayoutLM models are publicly available at \url{https://aka.ms/layoutlm}.