Jie Tang

Xiao Liu, Hao Yu, Hanchen Zhang et al. · berkeley, microsoft-research

The potential of Large Language Model (LLM) as agents has been widely acknowledged recently. Thus, there is an urgent need to quantitatively \textit{evaluate LLMs as agents} on challenging tasks in interactive environments. We present AgentBench, a multi-dimensional benchmark that consists of 8 distinct environments to assess LLM-as-Agent's reasoning and decision-making abilities. Our extensive test over \num API-based and open-sourced (OSS) LLMs shows that, while top commercial LLMs present a strong ability of acting as agents in complex environments, there is a significant disparity in performance between them and many OSS competitors that are no larger than 70B. We identify the typical reasons of failures in environments and LLMs, showing that poor long-term reasoning, decision-making, and instruction following abilities are the main obstacles for developing usable LLM agents. Improving instruction following and training on high quality multi-round alignment data could improve agent performance. And different from existing assumptions, training on code present ambivalent impacts on different agent tasks. Datasets, environments, and an integrated evaluation package for AgentBench are released at https://github.com/THUDM/AgentBench.

Zhuoyi Yang, Jiayan Teng, Wendi Zheng et al. · tsinghua

We present CogVideoX, a large-scale text-to-video generation model based on diffusion transformer, which can generate 10-second continuous videos aligned with text prompt, with a frame rate of 16 fps and resolution of 768 * 1360 pixels. Previous video generation models often had limited movement and short durations, and is difficult to generate videos with coherent narratives based on text. We propose several designs to address these issues. First, we propose a 3D Variational Autoencoder (VAE) to compress videos along both spatial and temporal dimensions, to improve both compression rate and video fidelity. Second, to improve the text-video alignment, we propose an expert transformer with the expert adaptive LayerNorm to facilitate the deep fusion between the two modalities. Third, by employing a progressive training and multi-resolution frame pack technique, CogVideoX is adept at producing coherent, long-duration, different shape videos characterized by significant motions. In addition, we develop an effective text-video data processing pipeline that includes various data preprocessing strategies and a video captioning method, greatly contributing to the generation quality and semantic alignment. Results show that CogVideoX demonstrates state-of-the-art performance across both multiple machine metrics and human evaluations. The model weight of both 3D Causal VAE, Video caption model and CogVideoX are publicly available at https://github.com/THUDM/CogVideo.

Xiao Liu, Hanyu Lai, Hao Yu et al. · tsinghua

We present WebGLM, a web-enhanced question-answering system based on the General Language Model (GLM). Its goal is to augment a pre-trained large language model (LLM) with web search and retrieval capabilities while being efficient for real-world deployments. To achieve this, we develop WebGLM with strategies for the LLM-augmented retriever, bootstrapped generator, and human preference-aware scorer. Specifically, we identify and address the limitations of WebGPT (OpenAI), through which WebGLM is enabled with accuracy, efficiency, and cost-effectiveness advantages. In addition, we propose systematic criteria for evaluating web-enhanced QA systems. We conduct multi-dimensional human evaluation and quantitative ablation studies, which suggest the outperformance of the proposed WebGLM designs over existing systems. WebGLM with the 10-billion-parameter GLM (10B) is shown to perform better than the similar-sized WebGPT (13B) and even comparably to WebGPT (175B) in human evaluation. The code, demo, and data are at \url{https://github.com/THUDM/WebGLM}.

Xiao Liu, Haoyun Hong, Xinghao Wang et al. · tsinghua

Entity alignment, aiming to identify equivalent entities across different knowledge graphs (KGs), is a fundamental problem for constructing Web-scale KGs. Over the course of its development, the label supervision has been considered necessary for accurate alignments. Inspired by the recent progress of self-supervised learning, we explore the extent to which we can get rid of supervision for entity alignment. Commonly, the label information (positive entity pairs) is used to supervise the process of pulling the aligned entities in each positive pair closer. However, our theoretical analysis suggests that the learning of entity alignment can actually benefit more from pushing unlabeled negative pairs far away from each other than pulling labeled positive pairs close. By leveraging this discovery, we develop the self-supervised learning objective for entity alignment. We present SelfKG with efficient strategies to optimize this objective for aligning entities without label supervision. Extensive experiments on benchmark datasets demonstrate that SelfKG without supervision can match or achieve comparable results with state-of-the-art supervised baselines. The performance of SelfKG suggests that self-supervised learning offers great potential for entity alignment in KGs. The code and data are available at https://github.com/THUDM/SelfKG.

Jiale Cheng, Xiao Liu, Kehan Zheng et al. · tsinghua

Large language models (LLMs) have shown impressive success in various applications. However, these models are often not well aligned with human intents, which calls for additional treatments on them; that is, the alignment problem. To make LLMs better follow user instructions, existing alignment methods primarily focus on further training them. However, the extra training of LLMs is usually expensive in terms of GPU computing; even worse, some LLMs are not accessible for user-demanded training, such as GPTs. In this work, we take a different perspective -- Black-Box Prompt Optimization (BPO) -- to perform alignments. The idea is to optimize user prompts to suit LLMs' input understanding, so as to best realize users' intents without updating LLMs' parameters. BPO leverages human preferences to optimize prompts, thus making it superior to LLM (e.g., ChatGPT) as a prompt engineer. Moreover, BPO is model-agnostic, and the empirical results demonstrate that the BPO-aligned ChatGPT yields a 22% increase in the win rate against its original version and 10% for GPT-4. Notably, the BPO-aligned LLMs can outperform the same models aligned by PPO and DPO, and it also brings additional performance gains when combining BPO with PPO or DPO. Code and datasets are released at https://github.com/thu-coai/BPO.

Yushi Bai, Jiajie Zhang, Xin Lv et al. · tsinghua

Current long context large language models (LLMs) can process inputs up to 100,000 tokens, yet struggle to generate outputs exceeding even a modest length of 2,000 words. Through controlled experiments, we find that the model's effective generation length is inherently bounded by the sample it has seen during supervised fine-tuning (SFT). In other words, their output limitation is due to the scarcity of long-output examples in existing SFT datasets. To address this, we introduce AgentWrite, an agent-based pipeline that decomposes ultra-long generation tasks into subtasks, enabling off-the-shelf LLMs to generate coherent outputs exceeding 20,000 words. Leveraging AgentWrite, we construct LongWriter-6k, a dataset containing 6,000 SFT data with output lengths ranging from 2k to 32k words. By incorporating this dataset into model training, we successfully scale the output length of existing models to over 10,000 words while maintaining output quality. We also develop LongBench-Write, a comprehensive benchmark for evaluating ultra-long generation capabilities. Our 9B parameter model, further improved through DPO, achieves state-of-the-art performance on this benchmark, surpassing even much larger proprietary models. In general, our work demonstrates that existing long context LLM already possesses the potential for a larger output window--all you need is data with extended output during model alignment to unlock this capability. Our code & models are at: https://github.com/THUDM/LongWriter.

Zhen Yang, Tinglin Huang, Ming Ding et al. · tsinghua

In-Batch contrastive learning is a state-of-the-art self-supervised method that brings semantically-similar instances close while pushing dissimilar instances apart within a mini-batch. Its key to success is the negative sharing strategy, in which every instance serves as a negative for the others within the mini-batch. Recent studies aim to improve performance by sampling hard negatives \textit{within the current mini-batch}, whose quality is bounded by the mini-batch itself. In this work, we propose to improve contrastive learning by sampling mini-batches from the input data. We present BatchSampler\footnote{The code is available at \url{https://github.com/THUDM/BatchSampler}} to sample mini-batches of hard-to-distinguish (i.e., hard and true negatives to each other) instances. To make each mini-batch have fewer false negatives, we design the proximity graph of randomly-selected instances. To form the mini-batch, we leverage random walk with restart on the proximity graph to help sample hard-to-distinguish instances. BatchSampler is a simple and general technique that can be directly plugged into existing contrastive learning models in vision, language, and graphs. Extensive experiments on datasets of three modalities show that BatchSampler can consistently improve the performance of powerful contrastive models, as shown by significant improvements of SimCLR on ImageNet-100, SimCSE on STS (language), and GraphCL and MVGRL on graph datasets.

Wenyi Hong, Weihan Wang, Ming Ding et al.

Beginning with VisualGLM and CogVLM, we are continuously exploring VLMs in pursuit of enhanced vision-language fusion, efficient higher-resolution architecture, and broader modalities and applications. Here we propose the CogVLM2 family, a new generation of visual language models for image and video understanding including CogVLM2, CogVLM2-Video and GLM-4V. As an image understanding model, CogVLM2 inherits the visual expert architecture with improved training recipes in both pre-training and post-training stages, supporting input resolution up to $1344 \times 1344$ pixels. As a video understanding model, CogVLM2-Video integrates multi-frame input with timestamps and proposes automated temporal grounding data construction. Notably, CogVLM2 family has achieved state-of-the-art results on benchmarks like MMBench, MM-Vet, TextVQA, MVBench and VCGBench. All models are open-sourced in https://github.com/THUDM/CogVLM2 and https://github.com/THUDM/GLM-4, contributing to the advancement of the field.

Jay Zhangjie Wu, Xiuyu Li, Difei Gao et al. · berkeley

Humans watch more than a billion hours of video per day. Most of this video was edited manually, which is a tedious process. However, AI-enabled video-generation and video-editing is on the rise. Building on text-to-image models like Stable Diffusion and Imagen, generative AI has improved dramatically on video tasks. But it's hard to evaluate progress in these video tasks because there is no standard benchmark. So, we propose a new dataset for text-guided video editing (TGVE), and we run a competition at CVPR to evaluate models on our TGVE dataset. In this paper we present a retrospective on the competition and describe the winning method. The competition dataset is available at https://sites.google.com/view/loveucvpr23/track4.

Xiao Liu, Shiyu Zhao, Kai Su et al. · tsinghua

Knowledge graph (KG) embeddings have been a mainstream approach for reasoning over incomplete KGs. However, limited by their inherently shallow and static architectures, they can hardly deal with the rising focus on complex logical queries, which comprise logical operators, imputed edges, multiple source entities, and unknown intermediate entities. In this work, we present the Knowledge Graph Transformer (kgTransformer) with masked pre-training and fine-tuning strategies. We design a KG triple transformation method to enable Transformer to handle KGs, which is further strengthened by the Mixture-of-Experts (MoE) sparse activation. We then formulate the complex logical queries as masked prediction and introduce a two-stage masked pre-training strategy to improve transferability and generalizability. Extensive experiments on two benchmarks demonstrate that kgTransformer can consistently outperform both KG embedding-based baselines and advanced encoders on nine in-domain and out-of-domain reasoning tasks. Additionally, kgTransformer can reason with explainability via providing the full reasoning paths to interpret given answers.

Zhen Yang, Jinhao Chen, Zhengxiao Du et al.

Large language models (LLMs) have demonstrated significant capabilities in mathematical reasoning, particularly with text-based mathematical problems. However, current multi-modal large language models (MLLMs), especially those specialized in mathematics, tend to focus predominantly on solving geometric problems but ignore the diversity of visual information available in other areas of mathematics. Moreover, the geometric information for these specialized mathematical MLLMs is derived from several public datasets, which are typically limited in diversity and complexity. To address these limitations, we aim to construct a fine-tuning dataset named MathVL, and develop a series of specialized mathematical MLLMs termed MathGLM-Vision by conducting Supervised Fine-Tuning (SFT) on MathVL with various parameter-scale backbones. To extensively evaluate the effectiveness of MathGLM-Vision, we conduct experiments on several public benchmarks and our curated MathVL-test consisting of 2,000 problems. Experimental results demonstrate that MathGLM-Vision achieves significant improvements compared with some existing models, including backbone models and open-source mathematical MLLMs. These findings indicate the importance of diversity dataset in enhancing the mathematical reasoning abilities of MLLMs.

Huanjing Zhao, Beining Yang, Yukuo Cen et al. · tsinghua

The text-attributed graph (TAG) is one kind of important real-world graph-structured data with each node associated with raw texts. For TAGs, traditional few-shot node classification methods directly conduct training on the pre-processed node features and do not consider the raw texts. The performance is highly dependent on the choice of the feature pre-processing method. In this paper, we propose P2TAG, a framework designed for few-shot node classification on TAGs with graph pre-training and prompting. P2TAG first pre-trains the language model (LM) and graph neural network (GNN) on TAGs with self-supervised loss. To fully utilize the ability of language models, we adapt the masked language modeling objective for our framework. The pre-trained model is then used for the few-shot node classification with a mixed prompt method, which simultaneously considers both text and graph information. We conduct experiments on six real-world TAGs, including paper citation networks and product co-purchasing networks. Experimental results demonstrate that our proposed framework outperforms existing graph few-shot learning methods on these datasets with +18.98% ~ +35.98% improvements.

Wei Wang, Dan Zhang, Tao Feng et al.

Large Language Models (LLMs) are becoming increasingly powerful and capable of handling complex tasks, e.g., building single agents and multi-agent systems. Compared to single agents, multi-agent systems have higher requirements for the collaboration capabilities of language models. Many benchmarks are proposed to evaluate their collaborative abilities. However, these benchmarks lack fine-grained evaluations of LLM collaborative capabilities. Additionally, multi-agent collaborative and competitive scenarios are ignored in existing works. To address these two problems, we propose a benchmark, called BattleAgentBench, which defines seven sub-stages of three varying difficulty levels and conducts a fine-grained evaluation of language models in terms of single-agent scenario navigation capabilities, paired-agent task execution abilities, and multi-agent collaboration and competition capabilities. We conducted extensive evaluations on leading four closed-source and seven open-source models. Experimental results indicate that API-based models perform excellently on simple tasks but open-source small models struggle with simple tasks. Regarding difficult tasks that require collaborative and competitive abilities, although API-based models have demonstrated some collaborative capabilities, there is still enormous room for improvement.

Jinfeng Zhou, Zhuang Chen, Dazhen Wan et al.

In this paper, we present CharacterGLM, a series of models built upon ChatGLM, with model sizes ranging from 6B to 66B parameters. Our CharacterGLM is designed for generating Character-based Dialogues (CharacterDial), which aims to equip a conversational AI system with character customization for satisfying people's inherent social desires and emotional needs. On top of CharacterGLM, we can customize various AI characters or social agents by configuring their attributes (identities, interests, viewpoints, experiences, achievements, social relationships, etc.) and behaviors (linguistic features, emotional expressions, interaction patterns, etc.). Our model outperforms most mainstream close-source large langauge models, including the GPT series, especially in terms of consistency, human-likeness, and engagement according to manual evaluations. We will release our 6B version of CharacterGLM and a subset of training data to facilitate further research development in the direction of character-based dialogue generation.

Jiayi Gui, Yiming Liu, Jiale Cheng et al.

Large Language Models (LLMs) have demonstrated notable capabilities across various tasks, showcasing complex problem-solving abilities. Understanding and executing complex rules, along with multi-step planning, are fundamental to logical reasoning and critical for practical LLM agents and decision-making systems. However, evaluating LLMs as effective rule-based executors and planners remains underexplored. In this paper, we introduce LogicGame, a novel benchmark designed to evaluate the comprehensive rule understanding, execution, and planning capabilities of LLMs. Unlike traditional benchmarks, LogicGame provides diverse games that contain a series of rules with an initial state, requiring models to comprehend and apply predefined regulations to solve problems. We create simulated scenarios in which models execute or plan operations to achieve specific outcomes. These game scenarios are specifically designed to distinguish logical reasoning from mere knowledge by relying exclusively on predefined rules. This separation allows for a pure assessment of rule-based reasoning capabilities. The evaluation considers not only final outcomes but also intermediate steps, providing a comprehensive assessment of model performance. Moreover, these intermediate steps are deterministic and can be automatically verified. LogicGame defines game scenarios with varying difficulty levels, from simple rule applications to complex reasoning chains, in order to offer a precise evaluation of model performance on rule understanding and multi-step execution. Utilizing LogicGame, we test various LLMs and identify notable shortcomings in their rule-based logical reasoning abilities.

Shicheng Tan, Weng Lam Tam, Yuanchun Wang et al.

The large scale of pre-trained language models poses a challenge for their deployment on various devices, with a growing emphasis on methods to compress these models, particularly knowledge distillation. However, current knowledge distillation methods rely on the model's intermediate layer features and the golden labels (also called hard labels), which usually require aligned model architecture and enough labeled data respectively. Moreover, the parameters of vocabulary are usually neglected in existing methods. To address these problems, we propose a general language model distillation (GLMD) method that performs two-stage word prediction distillation and vocabulary compression, which is simple and surprisingly shows extremely strong performance. Specifically, GLMD supports more general application scenarios by eliminating the constraints of dimension and structure between models and the need for labeled datasets through the absence of intermediate layers and golden labels. Meanwhile, based on the long-tailed distribution of word frequencies in the data, GLMD designs a strategy of vocabulary compression through decreasing vocabulary size instead of dimensionality. Experimental results show that our method outperforms 25 state-of-the-art methods on the SuperGLUE benchmark, achieving an average score that surpasses the best method by 3%.

Hanyu Lai, Xiao Liu, Iat Long Iong et al. · tsinghua

Large language models (LLMs) have fueled many intelligent web agents, but most existing ones perform far from satisfying in real-world web navigation tasks due to three factors: (1) the complexity of HTML text data (2) versatility of actions on webpages, and (3) task difficulty due to the open-domain nature of the web. In light of these challenges, we develop the open AutoWebGLM based on ChatGLM3-6B. AutoWebGLM can serve as a powerful automated web navigation agent that outperform GPT-4. Inspired by human browsing patterns, we first design an HTML simplification algorithm to represent webpages with vital information preserved succinctly. We then employ a hybrid human-AI method to build web browsing data for curriculum training. Finally, we bootstrap the model by reinforcement learning and rejection sampling to further facilitate webpage comprehension, browser operations, and efficient task decomposition by itself. For comprehensive evaluation, we establish a bilingual benchmark -- AutoWebBench -- for real-world web navigation tasks. We evaluate AutoWebGLM across diverse web navigation benchmarks, demonstrating its potential to tackle challenging tasks in real environments. Related code, model, and data are released at \url{https://github.com/THUDM/AutoWebGLM}.

Yushi Bai, Xin Lv, Jiajie Zhang et al. · tsinghua

Extending large language models to effectively handle long contexts requires instruction fine-tuning on input sequences of similar length. To address this, we present LongAlign -- a recipe of the instruction data, training, and evaluation for long context alignment. First, we construct a long instruction-following dataset using Self-Instruct. To ensure the data diversity, it covers a broad range of tasks from various long context sources. Second, we adopt the packing and sorted batching strategies to speed up supervised fine-tuning on data with varied length distributions. Additionally, we develop a loss weighting method to balance the contribution to the loss across different sequences during packing training. Third, we introduce the LongBench-Chat benchmark for evaluating instruction-following capabilities on queries of 10k-100k in length. Experiments show that LongAlign outperforms existing recipes for LLMs in long context tasks by up to 30\%, while also maintaining their proficiency in handling short, generic tasks. The code, data, and long-aligned models are open-sourced at https://github.com/THUDM/LongAlign.

Jiazheng Xu, Yu Huang, Jiale Cheng et al. · tsinghua

Visual generative models have achieved remarkable progress in synthesizing photorealistic images and videos, yet aligning their outputs with human preferences across critical dimensions remains a persistent challenge. Though reinforcement learning from human feedback offers promise for preference alignment, existing reward models for visual generation face limitations, including black-box scoring without interpretability and potentially resultant unexpected biases. We present VisionReward, a general framework for learning human visual preferences in both image and video generation. Specifically, we employ a hierarchical visual assessment framework to capture fine-grained human preferences, and leverages linear weighting to enable interpretable preference learning. Furthermore, we propose a multi-dimensional consistent strategy when using VisionReward as a reward model during preference optimization for visual generation. Experiments show that VisionReward can significantly outperform existing image and video reward models on both machine metrics and human evaluation. Notably, VisionReward surpasses VideoScore by 17.2% in preference prediction accuracy, and text-to-video models with VisionReward achieve a 31.6% higher pairwise win rate compared to the same models using VideoScore. All code and datasets are provided at https://github.com/THUDM/VisionReward.

Wendi Zheng, Jiayan Teng, Zhuoyi Yang et al.

Recent advancements in text-to-image generative systems have been largely driven by diffusion models. However, single-stage text-to-image diffusion models still face challenges, in terms of computational efficiency and the refinement of image details. To tackle the issue, we propose CogView3, an innovative cascaded framework that enhances the performance of text-to-image diffusion. CogView3 is the first model implementing relay diffusion in the realm of text-to-image generation, executing the task by first creating low-resolution images and subsequently applying relay-based super-resolution. This methodology not only results in competitive text-to-image outputs but also greatly reduces both training and inference costs. Our experimental results demonstrate that CogView3 outperforms SDXL, the current state-of-the-art open-source text-to-image diffusion model, by 77.0\% in human evaluations, all while requiring only about 1/2 of the inference time. The distilled variant of CogView3 achieves comparable performance while only utilizing 1/10 of the inference time by SDXL.

Dan Zhang, Yangliao Geng, Wenwen Gong et al.

Self-supervised learning (SSL) has recently achieved great success in mining the user-item interactions for collaborative filtering. As a major paradigm, contrastive learning (CL) based SSL helps address data sparsity in Web platforms by contrasting the embeddings between raw and augmented data. However, existing CL-based methods mostly focus on contrasting in a batch-wise way, failing to exploit potential regularity in the feature dimension. This leads to redundant solutions during the representation learning of users and items. In this work, we investigate how to employ both batch-wise CL (BCL) and feature-wise CL (FCL) for recommendation. We theoretically analyze the relation between BCL and FCL, and find that combining BCL and FCL helps eliminate redundant solutions but never misses an optimal solution. We propose a dual contrastive learning recommendation framework -- RecDCL. In RecDCL, the FCL objective is designed to eliminate redundant solutions on user-item positive pairs and to optimize the uniform distributions within users and items using a polynomial kernel for driving the representations to be orthogonal; The BCL objective is utilized to generate contrastive embeddings on output vectors for enhancing the robustness of the representations. Extensive experiments on four widely-used benchmarks and one industry dataset demonstrate that RecDCL can consistently outperform the state-of-the-art GNNs-based and SSL-based models (with an improvement of up to 5.65\% in terms of Recall@20). The source code is publicly available (https://github.com/THUDM/RecDCL).

Yifan Xu, Xiao Liu, Xueqiao Sun et al. · tsinghua

Autonomous agents have become increasingly important for interacting with the real world. Android agents, in particular, have been recently a frequently-mentioned interaction method. However, existing studies for training and evaluating Android agents lack systematic research on both open-source and closed-source models. In this work, we propose AndroidLab as a systematic Android agent framework. It includes an operation environment with different modalities, action space, and a reproducible benchmark. It supports both large language models (LLMs) and multimodal models (LMMs) in the same action space. AndroidLab benchmark includes predefined Android virtual devices and 138 tasks across nine apps built on these devices. By using the AndroidLab environment, we develop an Android Instruction dataset and train six open-source LLMs and LMMs, lifting the average success rates from 4.59% to 21.50% for LLMs and from 1.93% to 13.28% for LMMs. AndroidLab is open-sourced and publicly available at https://github.com/THUDM/Android-Lab.

Lilong Xue, Dan Zhang, Yuxiao Dong et al.

Large Language Models (LLMs) have demonstrated exceptional abilities in comprehending and generating text, motivating numerous researchers to utilize them for Information Extraction (IE) purposes, including Relation Extraction (RE). Nonetheless, most existing methods are predominantly designed for Sentence-level Relation Extraction (SentRE) tasks, which typically encompass a restricted set of relations and triplet facts within a single sentence. Furthermore, certain approaches resort to treating relations as candidate choices integrated into prompt templates, leading to inefficient processing and suboptimal performance when tackling Document-Level Relation Extraction (DocRE) tasks, which entail handling multiple relations and triplet facts distributed across a given document, posing distinct challenges. To overcome these limitations, we introduce AutoRE, an end-to-end DocRE model that adopts a novel RE extraction paradigm named RHF (Relation-Head-Facts). Unlike existing approaches, AutoRE does not rely on the assumption of known relation options, making it more reflective of real-world scenarios. Additionally, we have developed an easily extensible RE framework using a Parameters Efficient Fine Tuning (PEFT) algorithm (QLoRA). Our experiments on the RE-DocRED dataset showcase AutoRE's best performance, achieving state-of-the-art results, surpassing TAG by 10.03\% and 9.03\% respectively on the dev and test set. The code is available at https://github.com/THUDM/AutoRE and the demonstration video is provided at https://www.youtube.com/watch?v=IhKRsZUAxKk.

Ji Qi, Ming Ding, Weihan Wang et al. · tsinghua

Vision-Language Models (VLMs) have demonstrated their broad effectiveness thanks to extensive training in aligning visual instructions to responses. However, such training of conclusive alignment leads models to ignore essential visual reasoning, further resulting in failures in meticulous visual problems and unfaithful responses. Drawing inspiration from human cognition in solving visual problems (e.g., marking, zoom in), this paper introduces Chain of Manipulations, a mechanism that enables VLMs to solve problems step-by-step with evidence. After training, models can solve various visual problems by eliciting intrinsic manipulations (e.g., grounding, zoom in) with results (e.g., boxes, image) actively without involving external tools, while also allowing users to trace error causes. We study the roadmap to implement this mechanism, including (1) a flexible design of manipulations upon extensive analysis, (2) an efficient automated data generation pipeline, (3) a compatible VLM architecture capable of multi-turn multi-image, and (4) a model training process for versatile capabilities. With the design, we also manually annotate 6K high-quality samples for the challenging graphical mathematical problems. Our trained model, \textbf{CogCoM}, equipped with this mechanism with 17B parameters achieves state-of-the-art performance across 9 benchmarks from 4 categories, demonstrating the effectiveness while preserving the interpretability. Our code, model weights, and collected data are publicly available at https://github.com/THUDM/CogCoM.

Dan Zhang, Ziniu Hu, Sining Zhoubian et al.

Large Language Models (LLMs) have shown promise in assisting scientific discovery. However, such applications are currently limited by LLMs' deficiencies in understanding intricate scientific concepts, deriving symbolic equations, and solving advanced numerical calculations. To bridge these gaps, we introduce SciInstruct, a suite of scientific instructions for training scientific language models capable of college-level scientific reasoning. Central to our approach is a novel self-reflective instruction annotation framework to address the data scarcity challenge in the science domain. This framework leverages existing LLMs to generate step-by-step reasoning for unlabelled scientific questions, followed by a process of self-reflective critic-and-revise. Applying this framework, we curated a diverse and high-quality dataset encompassing physics, chemistry, math, and formal proofs. We analyze the curated SciInstruct from multiple interesting perspectives (e.g., domain, scale, source, question type, answer length, etc.). To verify the effectiveness of SciInstruct, we fine-tuned different language models with SciInstruct, i.e., ChatGLM3 (6B and 32B), Llama3-8B-Instruct, and Mistral-7B: MetaMath, enhancing their scientific and mathematical reasoning capabilities, without sacrificing the language understanding capabilities of the base model. We release all codes and SciInstruct at https://github.com/THUDM/SciGLM.

Jinhao Chen, Zhen Yang, Jianxin Shi et al.

Multimodal large language models (MLLMs) have demonstrated remarkable capabilities in vision-language answering tasks. Despite their strengths, these models often encounter challenges in achieving complex reasoning tasks such as mathematical problem-solving. Previous works have focused on fine-tuning on specialized mathematical datasets. However, these datasets are typically distilled directly from teacher models, which capture only static reasoning patterns and leaving substantial gaps compared to student models. This reliance on fixed teacher-derived datasets not only restricts the model's ability to adapt to novel or more intricate questions that extend beyond the confines of the training data, but also lacks the iterative depth needed for robust generalization. To overcome these limitations, we propose \textbf{\method}, a \textbf{Math}ematical \textbf{S}elf-\textbf{E}volving framework for MLLMs. In contrast to traditional one-shot fine-tuning paradigms, \method iteratively refines the model through cycles of inference, reflection, and reward-based feedback. Specifically, we leverage iterative fine-tuning by incorporating correct reasoning paths derived from previous-stage inference and integrating reflections from a specialized Outcome Reward Model (ORM). To verify the effectiveness of \method, we evaluate it on a suite of challenging benchmarks, demonstrating significant performance gains over backbone models. Notably, our experimental results on MathVL-test surpass the leading open-source multimodal mathematical reasoning model QVQ. Our code and models are available at \texttt{https://zheny2751\allowbreak-dotcom.github.io/\allowbreak MathSE.github.io/}.

Zhuoyi Yang, Heyang Jiang, Wenyi Hong et al.

Diffusion models have shown remarkable performance in image generation in recent years. However, due to a quadratic increase in memory during generating ultra-high-resolution images (e.g. 4096*4096), the resolution of generated images is often limited to 1024*1024. In this work. we propose a unidirectional block attention mechanism that can adaptively adjust the memory overhead during the inference process and handle global dependencies. Building on this module, we adopt the DiT structure for upsampling and develop an infinite super-resolution model capable of upsampling images of various shapes and resolutions. Comprehensive experiments show that our model achieves SOTA performance in generating ultra-high-resolution images in both machine and human evaluation. Compared to commonly used UNet structures, our model can save more than 5x memory when generating 4096*4096 images. The project URL is https://github.com/THUDM/Inf-DiT.

Xiaokang Zhang, Zijun Yao, Jing Zhang et al.

Detecting non-factual content is a longstanding goal to increase the trustworthiness of large language models (LLMs) generations. Current factuality probes, trained using humanannotated labels, exhibit limited transferability to out-of-distribution content, while online selfconsistency checking imposes extensive computation burden due to the necessity of generating multiple outputs. This paper proposes PINOSE, which trains a probing model on offline self-consistency checking results, thereby circumventing the need for human-annotated data and achieving transferability across diverse data distributions. As the consistency check process is offline, PINOSE reduces the computational burden of generating multiple responses by online consistency verification. Additionally, it examines various aspects of internal states prior to response decoding, contributing to more effective detection of factual inaccuracies. Experiment results on both factuality detection and question answering benchmarks show that PINOSE achieves surpassing results than existing factuality detection methods. Our code and datasets are publicly available on this anonymized repository.

Bosi Wen, Pei Ke, Yufei Sun et al.

Since the adoption of large language models (LLMs) for text evaluation has become increasingly prevalent in the field of natural language processing (NLP), a series of existing works attempt to optimize the prompts for LLM evaluators to improve their alignment with human judgment. However, their efforts are limited to optimizing individual factors of evaluation prompts, such as evaluation criteria or output formats, neglecting the combinatorial impact of multiple factors, which leads to insufficient optimization of the evaluation pipeline. Nevertheless, identifying well-behaved prompting strategies for adjusting multiple factors requires extensive enumeration. To this end, we comprehensively integrate 8 key factors for evaluation prompts and propose a novel automatic prompting strategy optimization method called Heuristic Prompting Strategy Search (HPSS). Inspired by the genetic algorithm, HPSS conducts an iterative search to find well-behaved prompting strategies for LLM evaluators. A heuristic function is employed to guide the search process, enhancing the performance of our algorithm. Extensive experiments across four evaluation tasks demonstrate the effectiveness of HPSS, consistently outperforming both human-designed evaluation prompts and existing automatic prompt optimization methods. Our code is available at https://github.com/thu-coai/HPSS.

Sining Zhoubian, Dan Zhang, Jie Tang

With respect to improving the reasoning accuracy of LLMs, the representative reinforcement learning (RL) method GRPO faces failure due to insignificant reward variance, while verification methods based on process reward models (PRMs) suffer from difficulties with training data acquisition and verification effectiveness. To tackle these problems, this paper introduces ReST-RL, a unified LLM RL paradigm that significantly improves LLM's code reasoning ability by combining an improved GRPO algorithm with a meticulously designed test time decoding method assisted by a value model (VM). As the first stage of policy reinforcement, ReST-GRPO adopts an optimized ReST algorithm to filter and assemble high-value training data, increasing the reward variance of GRPO sampling, thus improving the effectiveness and efficiency of training. After the basic reasoning ability of LLM policy has been improved, we further propose a test time decoding optimization method called VM-MCTS. Through Monte-Carlo Tree Search (MCTS), we collect accurate value targets with no annotation required, on which VM training is based. When decoding, the VM is deployed by an adapted MCTS algorithm to provide precise process signals as well as verification scores, assisting the LLM policy to achieve high reasoning accuracy. We conduct extensive experiments on coding problems to verify the validity of the proposed RL paradigm. Upon comparison, our approach significantly outperforms other reinforcement training baselines (e.g., naive GRPO and ReST-DPO), as well as decoding and verification baselines (e.g., PRM-BoN and ORM-MCTS) on well-known coding benchmarks of various levels (e.g., APPS, BigCodeBench, and HumanEval), indicating its power to strengthen the reasoning ability of LLM policies. Codes for our project can be found at https://github.com/THUDM/ReST-RL.

Shu-Xun Yang, Cunxiang Wang, Yidong Wang et al.

Evaluating mathematical capabilities is critical for assessing the overall performance of large language models (LLMs). However, existing evaluation methods often focus solely on final answers, resulting in highly inaccurate and uninterpretable evaluation outcomes, as well as their failure to assess proof or open-ended problems. To address these issues, we propose a novel mathematical process evaluation agent based on Tree-of-Error, called StepMathAgent. This agent incorporates four internal core operations: logical step segmentation, step scoring, score aggregation and error tree generation, along with four external extension modules: difficulty calibration, simplicity evaluation, completeness validation and format assessment. Furthermore, we introduce StepMathBench, a benchmark comprising 1,000 step-divided process evaluation instances, derived from 200 high-quality math problems grouped by problem type, subject category and difficulty level. Experiments on StepMathBench show that our proposed StepMathAgent outperforms all state-of-the-art methods, demonstrating human-aligned evaluation preferences and broad applicability to various scenarios. Our data and code are available at https://github.com/SHU-XUN/StepMathAgent.

Jiale Cheng, Yusen Liu, Xinyu Zhang et al.

Large language models (LLMs) increasingly rely on long-context modeling for tasks such as document understanding, code analysis, and multi-step reasoning. However, scaling context windows to the million-token level brings prohibitive computational and memory costs, limiting the practicality of long-context LLMs. In this work, we take a different perspective-visual context scaling-to tackle this challenge. Instead of extending token-based sequences, we propose Glyph, a framework that renders long texts into images and processes them with vision-language models (VLMs). This approach substantially compresses textual input while preserving semantic information, and we further design an LLM-driven genetic search to identify optimal visual rendering configurations for balancing accuracy and compression. Through extensive experiments, we demonstrate that our method achieves 3-4x token compression while maintaining accuracy comparable to leading LLMs such as Qwen3-8B on various long-context benchmarks. This compression also leads to around 4x faster prefilling and decoding, and approximately 2x faster SFT training. Furthermore, under extreme compression, a 128K-context VLM could scale to handle 1M-token-level text tasks. In addition, the rendered text data benefits real-world multimodal tasks, such as document understanding. Our code and model are released at https://github.com/thu-coai/Glyph.

Bo Chen, Xingyi Cheng, Pan Li et al.

Protein language models have shown remarkable success in learning biological information from protein sequences. However, most existing models are limited by either autoencoding or autoregressive pre-training objectives, which makes them struggle to handle protein understanding and generation tasks concurrently. We propose a unified protein language model, xTrimoPGLM, to address these two types of tasks simultaneously through an innovative pre-training framework. Our key technical contribution is an exploration of the compatibility and the potential for joint optimization of the two types of objectives, which has led to a strategy for training xTrimoPGLM at an unprecedented scale of 100 billion parameters and 1 trillion training tokens. Our extensive experiments reveal that 1) xTrimoPGLM significantly outperforms other advanced baselines in 18 protein understanding benchmarks across four categories. The model also facilitates an atomic-resolution view of protein structures, leading to an advanced 3D structural prediction model that surpasses existing language model-based tools. 2) xTrimoPGLM not only can generate de novo protein sequences following the principles of natural ones, but also can perform programmable generation after supervised fine-tuning (SFT) on curated sequences. These results highlight the substantial capability and versatility of xTrimoPGLM in understanding and generating protein sequences, contributing to the evolving landscape of foundation models in protein science.

Shizhan Liu, Xinran Deng, Zhuoyi Yang et al.

Latent diffusion models pair VAEs with diffusion backbones, and the structure of VAE latents strongly influences the difficulty of diffusion training. However, existing video VAEs typically focus on reconstruction fidelity, overlooking latent structure. We present a statistical analysis of video VAE latent spaces and identify two spectral properties essential for diffusion training: a spatio-temporal frequency spectrum biased toward low frequencies, and a channel-wise eigenspectrum dominated by a few modes. To induce these properties, we propose two lightweight, backbone-agnostic regularizers: Local Correlation Regularization and Latent Masked Reconstruction. Experiments show that our Spectral-Structured VAE (SSVAE) achieves a $3\times$ speedup in text-to-video generation convergence and a 10\% gain in video reward, outperforming strong open-source VAEs. The code is available at https://github.com/zai-org/SSVAE.

Jiale Cheng, Yida Lu, Xiaotao Gu et al.

Although Large Language Models (LLMs) are becoming increasingly powerful, they still exhibit significant but subtle weaknesses, such as mistakes in instruction-following or coding tasks. As these unexpected errors could lead to severe consequences in practical deployments, it is crucial to investigate the limitations within LLMs systematically. Traditional benchmarking approaches cannot thoroughly pinpoint specific model deficiencies, while manual inspections are costly and not scalable. In this paper, we introduce a unified framework, AutoDetect, to automatically expose weaknesses in LLMs across various tasks. Inspired by the educational assessment process that measures students' learning outcomes, AutoDetect consists of three LLM-powered agents: Examiner, Questioner, and Assessor. The collaboration among these three agents is designed to realize comprehensive and in-depth weakness identification. Our framework demonstrates significant success in uncovering flaws, with an identification success rate exceeding 30% in prominent models such as ChatGPT and Claude. More importantly, these identified weaknesses can guide specific model improvements, proving more effective than untargeted data augmentation methods like Self-Instruct. Our approach has led to substantial enhancements in popular LLMs, including the Llama series and Mistral-7b, boosting their performance by over 10% across several benchmarks. Code and data are publicly available at https://github.com/thu-coai/AutoDetect.

Dan Zhang, Sining Zhoubian, Ziniu Hu et al.

Recent methodologies in LLM self-training mostly rely on LLM generating responses and filtering those with correct output answers as training data. This approach often yields a low-quality fine-tuning training set (e.g., incorrect plans or intermediate reasoning). In this paper, we develop a reinforced self-training approach, called ReST-MCTS*, based on integrating process reward guidance with tree search MCTS* for collecting higher-quality reasoning traces as well as per-step value to train policy and reward models. ReST-MCTS* circumvents the per-step manual annotation typically used to train process rewards by tree-search-based reinforcement learning: Given oracle final correct answers, ReST-MCTS* is able to infer the correct process rewards by estimating the probability this step can help lead to the correct answer. These inferred rewards serve dual purposes: they act as value targets for further refining the process reward model and also facilitate the selection of high-quality traces for policy model self-training. We first show that the tree-search policy in ReST-MCTS* achieves higher accuracy compared with prior LLM reasoning baselines such as Best-of-N and Tree-of-Thought, within the same search budget. We then show that by using traces searched by this tree-search policy as training data, we can continuously enhance the three language models for multiple iterations, and outperform other self-training algorithms such as ReST$^\text{EM}$ and Self-Rewarding LM. We release all code at https://github.com/THUDM/ReST-MCTS.

Yelin Chen, Fanjin Zhang, Jie Tang

Named Entity Recognition (NER) is a fundamental problem in natural language processing (NLP). However, the task of extracting longer entity spans (e.g., awards) from extended texts (e.g., homepages) is barely explored. Current NER methods predominantly fall into two categories: span-based methods and generation-based methods. Span-based methods require the enumeration of all possible token-pair spans, followed by classification on each span, resulting in substantial redundant computations and excessive GPU memory usage. In contrast, generation-based methods involve prompting or fine-tuning large language models (LLMs) to adapt to downstream NER tasks. However, these methods struggle with the accurate generation of longer spans and often incur significant time costs for effective fine-tuning. To address these challenges, this paper introduces a lightweight span-based NER method called SeNER, which incorporates a bidirectional arrow attention mechanism coupled with LogN-Scaling on the [CLS] token to embed long texts effectively, and comprises a novel bidirectional sliding-window plus-shaped attention (BiSPA) mechanism to reduce redundant candidate token-pair spans significantly and model interactions between token-pair spans simultaneously. Extensive experiments demonstrate that our method achieves state-of-the-art extraction accuracy on three long NER datasets and is capable of extracting entities from long texts in a GPU-memory-friendly manner. Code: https://github.com/THUDM/scholar-profiling/tree/main/sener

Qinkai Zheng, Xu Zou, Yuxiao Dong et al.

Adversarial attacks on graphs have posed a major threat to the robustness of graph machine learning (GML) models. Naturally, there is an ever-escalating arms race between attackers and defenders. However, the strategies behind both sides are often not fairly compared under the same and realistic conditions. To bridge this gap, we present the Graph Robustness Benchmark (GRB) with the goal of providing a scalable, unified, modular, and reproducible evaluation for the adversarial robustness of GML models. GRB standardizes the process of attacks and defenses by 1) developing scalable and diverse datasets, 2) modularizing the attack and defense implementations, and 3) unifying the evaluation protocol in refined scenarios. By leveraging the GRB pipeline, the end-users can focus on the development of robust GML models with automated data processing and experimental evaluations. To support open and reproducible research on graph adversarial learning, GRB also hosts public leaderboards across different scenarios. As a starting point, we conduct extensive experiments to benchmark baseline techniques. GRB is open-source and welcomes contributions from the community. Datasets, codes, leaderboards are available at https://cogdl.ai/grb/home.

Xiao Liu, Kaixuan Ji, Yicheng Fu et al.

Prompt tuning, which only tunes continuous prompts with a frozen language model, substantially reduces per-task storage and memory usage at training. However, in the context of NLU, prior work reveals that prompt tuning does not perform well for normal-sized pretrained models. We also find that existing methods of prompt tuning cannot handle hard sequence labeling tasks, indicating a lack of universality. We present a novel empirical finding that properly optimized prompt tuning can be universally effective across a wide range of model scales and NLU tasks. It matches the performance of finetuning while having only 0.1%-3% tuned parameters. Our method P-Tuning v2 is an implementation of Deep Prompt Tuning \cite{li2021prefix,qin2021learning} optimized and adapted for NLU. Given the universality and simplicity of P-Tuning v2, we believe it can serve as an alternative to finetuning and a strong baseline for future research.Our code and data are released at https://github.com/THUDM/P-tuning-v2.

Yijia Xiao, Jiezhong Qiu, Ziang Li et al.

Protein is linked to almost every life process. Therefore, analyzing the biological structure and property of protein sequences is critical to the exploration of life, as well as disease detection and drug discovery. Traditional protein analysis methods tend to be labor-intensive and time-consuming. The emergence of deep learning models makes modeling data patterns in large quantities of data possible. Interdisciplinary researchers have begun to leverage deep learning methods to model large biological datasets, e.g. using long short-term memory and convolutional neural network for protein sequence classification. After millions of years of evolution, evolutionary information is encoded in protein sequences. Inspired by the similarity between natural language and protein sequences, we use large-scale language models to model evolutionary-scale protein sequences, encoding protein biology information in representation. Significant improvements are observed in both token-level and sequence-level tasks, demonstrating that our large-scale model can accurately capture evolution information from pretraining on evolutionary-scale individual sequences. Our code and model are available at https://github.com/THUDM/ProteinLM.

Jiaao He, Jiezhong Qiu, Aohan Zeng et al.

Mixture-of-Expert (MoE) presents a strong potential in enlarging the size of language model to trillions of parameters. However, training trillion-scale MoE requires algorithm and system co-design for a well-tuned high performance distributed training system. Unfortunately, the only existing platform that meets the requirements strongly depends on Google's hardware (TPU) and software (Mesh Tensorflow) stack, and is not open and available to the public, especially GPU and PyTorch communities. In this paper, we present FastMoE, a distributed MoE training system based on PyTorch with common accelerators. The system provides a hierarchical interface for both flexible model design and easy adaption to different applications, such as Transformer-XL and Megatron-LM. Different from direct implementation of MoE models using PyTorch, the training speed is highly optimized in FastMoE by sophisticated high-performance acceleration skills. The system supports placing different experts on multiple GPUs across multiple nodes, enabling enlarging the number of experts linearly against the number of GPUs. The source of FastMoE is available at https://github.com/laekov/fastmoe under Apache-2 license.

Zhengyan Zhang, Xu Han, Hao Zhou et al.

Pre-trained Language Models (PLMs) have proven to be beneficial for various downstream NLP tasks. Recently, GPT-3, with 175 billion parameters and 570GB training data, drew a lot of attention due to the capacity of few-shot (even zero-shot) learning. However, applying GPT-3 to address Chinese NLP tasks is still challenging, as the training corpus of GPT-3 is primarily English, and the parameters are not publicly available. In this technical report, we release the Chinese Pre-trained Language Model (CPM) with generative pre-training on large-scale Chinese training data. To the best of our knowledge, CPM, with 2.6 billion parameters and 100GB Chinese training data, is the largest Chinese pre-trained language model, which could facilitate several downstream Chinese NLP tasks, such as conversation, essay generation, cloze test, and language understanding. Extensive experiments demonstrate that CPM achieves strong performance on many NLP tasks in the settings of few-shot (even zero-shot) learning. The code and parameters are available at https://github.com/TsinghuaAI/CPM-Generate.

Wenzheng Feng, Jie Zhang, Yuxiao Dong et al.

We study the problem of semi-supervised learning on graphs, for which graph neural networks (GNNs) have been extensively explored. However, most existing GNNs inherently suffer from the limitations of over-smoothing, non-robustness, and weak-generalization when labeled nodes are scarce. In this paper, we propose a simple yet effective framework -- GRAPH RANDOM NEURAL NETWORKS (GRAND) -- to address these issues. In GRAND, we first design a random propagation strategy to perform graph data augmentation. Then we leverage consistency regularization to optimize the prediction consistency of unlabeled nodes across different data augmentations. Extensive experiments on graph benchmark datasets suggest that GRAND significantly outperforms state-of-the-art GNN baselines on semi-supervised node classification. Finally, we show that GRAND mitigates the issues of over-smoothing and non-robustness, exhibiting better generalization behavior than existing GNNs. The source code of GRAND is publicly available at https://github.com/Grand20/grand.

Christopher Berner, Greg Brockman, Brooke Chan et al. · openai

On April 13th, 2019, OpenAI Five became the first AI system to defeat the world champions at an esports game. The game of Dota 2 presents novel challenges for AI systems such as long time horizons, imperfect information, and complex, continuous state-action spaces, all challenges which will become increasingly central to more capable AI systems. OpenAI Five leveraged existing reinforcement learning techniques, scaled to learn from batches of approximately 2 million frames every 2 seconds. We developed a distributed training system and tools for continual training which allowed us to train OpenAI Five for 10 months. By defeating the Dota 2 world champion (Team OG), OpenAI Five demonstrates that self-play reinforcement learning can achieve superhuman performance on a difficult task.

Zhengxiao Du, Chang Zhou, Ming Ding et al.

Inferring new facts from existing knowledge graphs (KG) with explainable reasoning processes is a significant problem and has received much attention recently. However, few studies have focused on relation types unseen in the original KG, given only one or a few instances for training. To bridge this gap, we propose CogKR for one-shot KG reasoning. The one-shot relational learning problem is tackled through two modules: the summary module summarizes the underlying relationship of the given instances, based on which the reasoning module infers the correct answers. Motivated by the dual process theory in cognitive science, in the reasoning module, a cognitive graph is built by iteratively coordinating retrieval (System 1, collecting relevant evidence intuitively) and reasoning (System 2, conducting relational reasoning over collected information). The structural information offered by the cognitive graph enables our model to aggregate pieces of evidence from multiple reasoning paths and explain the reasoning process graphically. Experiments show that CogKR substantially outperforms previous state-of-the-art models on one-shot KG reasoning benchmarks, with relative improvements of 24.3%-29.7% on MRR. The source code is available at https://github.com/THUDM/CogKR.

Fuli Feng, Xiangnan He, Jie Tang et al.

Recent efforts show that neural networks are vulnerable to small but intentional perturbations on input features in visual classification tasks. Due to the additional consideration of connections between examples (\eg articles with citation link tend to be in the same class), graph neural networks could be more sensitive to the perturbations, since the perturbations from connected examples exacerbate the impact on a target example. Adversarial Training (AT), a dynamic regularization technique, can resist the worst-case perturbations on input features and is a promising choice to improve model robustness and generalization. However, existing AT methods focus on standard classification, being less effective when training models on graph since it does not model the impact from connected examples. In this work, we explore adversarial training on graph, aiming to improve the robustness and generalization of models learned on graph. We propose Graph Adversarial Training (GraphAT), which takes the impact from connected examples into account when learning to construct and resist perturbations. We give a general formulation of GraphAT, which can be seen as a dynamic regularization scheme based on the graph structure. To demonstrate the utility of GraphAT, we employ it on a state-of-the-art graph neural network model --- Graph Convolutional Network (GCN). We conduct experiments on two citation graphs (Citeseer and Cora) and a knowledge graph (NELL), verifying the effectiveness of GraphAT which outperforms normal training on GCN by 4.51% in node classification accuracy. Codes are available via: https://github.com/fulifeng/GraphAT.

Yushi Bai, Shangqing Tu, Jiajie Zhang et al. · tsinghua

This paper introduces LongBench v2, a benchmark designed to assess the ability of LLMs to handle long-context problems requiring deep understanding and reasoning across real-world multitasks. LongBench v2 consists of 503 challenging multiple-choice questions, with contexts ranging from 8k to 2M words, across six major task categories: single-document QA, multi-document QA, long in-context learning, long-dialogue history understanding, code repository understanding, and long structured data understanding. To ensure the breadth and the practicality, we collect data from nearly 100 highly educated individuals with diverse professional backgrounds. We employ both automated and manual review processes to maintain high quality and difficulty, resulting in human experts achieving only 53.7% accuracy under a 15-minute time constraint. Our evaluation reveals that the best-performing model, when directly answers the questions, achieves only 50.1% accuracy. In contrast, the o1-preview model, which includes longer reasoning, achieves 57.7%, surpassing the human baseline by 4%. These results highlight the importance of enhanced reasoning ability and scaling inference-time compute to tackle the long-context challenges in LongBench v2. The project is available at https://longbench2.github.io.

Yu Gu, Yiheng Shu, Hao Yu et al. · microsoft-research, tsinghua

The applications of large language models (LLMs) have expanded well beyond the confines of text processing, signaling a new era where LLMs are envisioned as generalist agents capable of operating within complex environments. These environments are often highly expansive, making it impossible for the LLM to process them within its short-term memory. Motivated by recent research on extending the capabilities of LLMs with tools, we seek to investigate the intriguing potential of tools to augment LLMs in handling such complexity by introducing a novel class of tools, termed middleware, to aid in the proactive exploration within these massive environments. Such specialized tools can serve as a middleware layer shielding the LLM from environmental complexity. In two representative complex environments -- knowledge bases (KBs) and databases -- we demonstrate the significant potential of augmenting language agents with tools in complex environments. Notably, equipped with the middleware, GPT-4 achieves 2.8X the performance of the best baseline in tasks requiring access to database content and 2.2X in KB tasks. Our findings illuminate the path for advancing language agents in real-world applications.

Haoming Wang, Haoyang Zou, Huatong Song et al. · pku

The development of autonomous agents for graphical user interfaces (GUIs) presents major challenges in artificial intelligence. While recent advances in native agent models have shown promise by unifying perception, reasoning, action, and memory through end-to-end learning, open problems remain in data scalability, multi-turn reinforcement learning (RL), the limitations of GUI-only operation, and environment stability. In this technical report, we present UI-TARS-2, a native GUI-centered agent model that addresses these challenges through a systematic training methodology: a data flywheel for scalable data generation, a stabilized multi-turn RL framework, a hybrid GUI environment that integrates file systems and terminals, and a unified sandbox platform for large-scale rollouts. Empirical evaluation demonstrates that UI-TARS-2 achieves significant improvements over its predecessor UI-TARS-1.5. On GUI benchmarks, it reaches 88.2 on Online-Mind2Web, 47.5 on OSWorld, 50.6 on WindowsAgentArena, and 73.3 on AndroidWorld, outperforming strong baselines such as Claude and OpenAI agents. In game environments, it attains a mean normalized score of 59.8 across a 15-game suite-roughly 60% of human-level performance-and remains competitive with frontier proprietary models (e.g., OpenAI o3) on LMGame-Bench. Additionally, the model can generalize to long-horizon information-seeking tasks and software engineering benchmarks, highlighting its robustness across diverse agent tasks. Detailed analyses of training dynamics further provide insights into achieving stability and efficiency in large-scale agent RL. These results underscore UI-TARS-2's potential to advance the state of GUI agents and exhibit strong generalization to real-world interactive scenarios.

Fanjin Zhang, Shijie Shi, Yifan Zhu et al. · tsinghua

With the rapid proliferation of scientific literature, versatile academic knowledge services increasingly rely on comprehensive academic graph mining. Despite the availability of public academic graphs, benchmarks, and datasets, these resources often fall short in multi-aspect and fine-grained annotations, are constrained to specific task types and domains, or lack underlying real academic graphs. In this paper, we present OAG-Bench, a comprehensive, multi-aspect, and fine-grained human-curated benchmark based on the Open Academic Graph (OAG). OAG-Bench covers 10 tasks, 20 datasets, 70+ baselines, and 120+ experimental results to date. We propose new data annotation strategies for certain tasks and offer a suite of data pre-processing codes, algorithm implementations, and standardized evaluation protocols to facilitate academic graph mining. Extensive experiments reveal that even advanced algorithms like large language models (LLMs) encounter difficulties in addressing key challenges in certain tasks, such as paper source tracing and scholar profiling. We also introduce the Open Academic Graph Challenge (OAG-Challenge) to encourage community input and sharing. We envisage that OAG-Bench can serve as a common ground for the community to evaluate and compare algorithms in academic graph mining, thereby accelerating algorithm development and advancement in this field. OAG-Bench is accessible at https://www.aminer.cn/data/.