Zhipeng Chen

Tianyi Tang, Junyi Li, Zhipeng Chen et al. · pku

To facilitate research on text generation, this paper presents a comprehensive and unified library, TextBox 2.0, focusing on the use of pre-trained language models (PLMs). To be comprehensive, our library covers $13$ common text generation tasks and their corresponding $83$ datasets and further incorporates $45$ PLMs covering general, translation, Chinese, dialogue, controllable, distilled, prompting, and lightweight PLMs. We also implement $4$ efficient training strategies and provide $4$ generation objectives for pre-training new PLMs from scratch. To be unified, we design the interfaces to support the entire research pipeline (from data loading to training and evaluation), ensuring that each step can be fulfilled in a unified way. Despite the rich functionality, it is easy to use our library, either through the friendly Python API or command line. To validate the effectiveness of our library, we conduct extensive experiments and exemplify four types of research scenarios. The project is released at the link: https://github.com/RUCAIBox/TextBox.

Junyi Li, Tianyi Tang, Zheng Gong et al. · pku

Nowadays, pretrained language models (PLMs) have dominated the majority of NLP tasks. While, little research has been conducted on systematically evaluating the language abilities of PLMs. In this paper, we present a large-scale empirical study on general language ability evaluation of PLMs (ElitePLM). In our study, we design four evaluation dimensions, i.e. memory, comprehension, reasoning, and composition, to measure ten widely-used PLMs within five categories. Our empirical results demonstrate that: (1) PLMs with varying training objectives and strategies are good at different ability tests; (2) fine-tuning PLMs in downstream tasks is usually sensitive to the data size and distribution; (3) PLMs have excellent transferability between similar tasks. Moreover, the prediction results of PLMs in our experiments are released as an open resource for more deep and detailed analysis on the language abilities of PLMs. This paper can guide the future work to select, apply, and design PLMs for specific tasks. We have made all the details of experiments publicly available at https://github.com/RUCAIBox/ElitePLM.

Wayne Xin Zhao, Kun Zhou, Junyi Li et al.

Language is essentially a complex, intricate system of human expressions governed by grammatical rules. It poses a significant challenge to develop capable AI algorithms for comprehending and grasping a language. As a major approach, language modeling has been widely studied for language understanding and generation in the past two decades, evolving from statistical language models to neural language models. Recently, pre-trained language models (PLMs) have been proposed by pre-training Transformer models over large-scale corpora, showing strong capabilities in solving various NLP tasks. Since researchers have found that model scaling can lead to performance improvement, they further study the scaling effect by increasing the model size to an even larger size. Interestingly, when the parameter scale exceeds a certain level, these enlarged language models not only achieve a significant performance improvement but also show some special abilities that are not present in small-scale language models. To discriminate the difference in parameter scale, the research community has coined the term large language models (LLM) for the PLMs of significant size. Recently, the research on LLMs has been largely advanced by both academia and industry, and a remarkable progress is the launch of ChatGPT, which has attracted widespread attention from society. The technical evolution of LLMs has been making an important impact on the entire AI community, which would revolutionize the way how we develop and use AI algorithms. In this survey, we review the recent advances of LLMs by introducing the background, key findings, and mainstream techniques. In particular, we focus on four major aspects of LLMs, namely pre-training, adaptation tuning, utilization, and capacity evaluation. Besides, we also summarize the available resources for developing LLMs and discuss the remaining issues for future directions.

Wayne Xin Zhao, Kun Zhou, Junyi Li et al.

Language is essentially a complex, intricate system of human expressions governed by grammatical rules. It poses a significant challenge to develop capable AI algorithms for comprehending and grasping a language. As a major approach, language modeling has been widely studied for language understanding and generation in the past two decades, evolving from statistical language models to neural language models. Recently, pre-trained language models (PLMs) have been proposed by pre-training Transformer models over large-scale corpora, showing strong capabilities in solving various NLP tasks. Since researchers have found that model scaling can lead to performance improvement, they further study the scaling effect by increasing the model size to an even larger size. Interestingly, when the parameter scale exceeds a certain level, these enlarged language models not only achieve a significant performance improvement but also show some special abilities that are not present in small-scale language models. To discriminate the difference in parameter scale, the research community has coined the term large language models (LLM) for the PLMs of significant size. Recently, the research on LLMs has been largely advanced by both academia and industry, and a remarkable progress is the launch of ChatGPT, which has attracted widespread attention from society. The technical evolution of LLMs has been making an important impact on the entire AI community, which would revolutionize the way how we develop and use AI algorithms. In this survey, we review the recent advances of LLMs by introducing the background, key findings, and mainstream techniques. In particular, we focus on four major aspects of LLMs, namely pre-training, adaptation tuning, utilization, and capacity evaluation. Besides, we also summarize the available resources for developing LLMs and discuss the remaining issues for future directions.

Jie Chen, Zhipeng Chen, Jiapeng Wang et al.

Continual pre-training (CPT) has been an important approach for adapting language models to specific domains or tasks. To make the CPT approach more traceable, this paper presents a technical report for continually pre-training Llama-3 (8B), which significantly enhances the Chinese language ability and scientific reasoning ability of the backbone model. To enhance the new abilities while retaining the original abilities, we design specific data mixture and curriculum strategies by utilizing existing datasets and synthesizing high-quality datasets. Specifically, we synthesize multidisciplinary scientific question and answer (QA) pairs based on related web pages, and subsequently incorporate these synthetic data to improve the scientific reasoning ability of Llama-3. We refer to the model after CPT as Llama-3-SynE (Synthetic data Enhanced Llama-3). We also present the tuning experiments with a relatively small model -- TinyLlama, and employ the derived findings to train the backbone model. Extensive experiments on a number of evaluation benchmarks show that our approach can largely improve the performance of the backbone models, including both the general abilities (+8.81 on C-Eval and +6.31 on CMMLU) and the scientific reasoning abilities (+12.00 on MATH and +4.13 on SciEval), without hurting the original capacities. Our model, data, and codes are available at https://github.com/RUC-GSAI/Llama-3-SynE.

Zhipeng Chen, Tao Qian, Wayne Xin Zhao et al.

Recently, scaling reinforcement learning with verifiable rewards (RLVR) for large language models (LLMs) has emerged as an effective training paradigm for significantly improving model capabilities, which requires guiding the model to perform extensive exploration and learning, leading to substantial computational overhead and becoming a key challenge. To reduce the number of training steps, Prior work performs linear extrapolation of model parameters. However, the dynamics of model parameter updates during RLVR training remain insufficiently understood. To further investigate the evolution of LLMs during RLVR training, we conduct empirical experiments and find that the rank-1 subspace of the model does not evolve linearly, and its dominance over the original parameters is further amplified during LoRA training. Based on the above insights, we propose the \textbf{N}onlinear \textbf{Ext}rapolation of low-rank trajectories (\textbf{NExt}), a novel framework that models and extrapolates low-rank parameter trajectories in a nonlinear manner. Concretely, we first train the model using LoRA and extract the rank-1 subspace of parameter differences at multiple training steps, which is then used for the subsequent nonlinear extrapolation. Afterward, we utilized the extracted rank-1 subspace to train a predictor, which can model the trajectory of parameter updates during RLVR, and then perform the predict-extend process to extrapolate model parameters, achieving the acceleration of RLVR. To further study and understand NExt, we conduct comprehensive experiments that demonstrate the effectiveness and robustness of the method. Our method reduces computational overhead by approximately 37.5\% while remaining compatible with a wide range of RLVR algorithms and tasks. We release our code in https://github.com/RUCAIBox/NExt.

Kun Zhou, Yutao Zhu, Zhipeng Chen et al.

Large language models~(LLMs) have greatly advanced the frontiers of artificial intelligence, attaining remarkable improvement in model capacity. To assess the model performance, a typical approach is to construct evaluation benchmarks for measuring the ability level of LLMs in different aspects. Despite that a number of high-quality benchmarks have been released, the concerns about the appropriate use of these benchmarks and the fair comparison of different models are increasingly growing. Considering these concerns, in this paper, we discuss the potential risk and impact of inappropriately using evaluation benchmarks and misleadingly interpreting the evaluation results. Specially, we focus on a special issue that would lead to inappropriate evaluation, \ie \emph{benchmark leakage}, referring that the data related to evaluation sets is occasionally used for model training. This phenomenon now becomes more common since pre-training data is often prepared ahead of model test. We conduct extensive experiments to study the effect of benchmark leverage, and find that it can dramatically boost the evaluation results, which would finally lead to an unreliable assessment of model performance. To improve the use of existing evaluation benchmarks, we finally present several guidelines for both LLM developers and benchmark maintainers. We hope this work can draw attention to appropriate training and evaluation of LLMs.

Wayne Xin Zhao, Kun Zhou, Beichen Zhang et al.

Although pre-trained language models~(PLMs) have recently advanced the research progress in mathematical reasoning, they are not specially designed as a capable multi-task solver, suffering from high cost for multi-task deployment (\eg a model copy for a task) and inferior performance on complex mathematical problems in practical applications. To address these issues, in this paper, we propose \textbf{JiuZhang~2.0}, a unified Chinese PLM specially for multi-task mathematical problem solving. Our idea is to maintain a moderate-sized model and employ the \emph{cross-task knowledge sharing} to improve the model capacity in a multi-task setting. Specially, we construct a Mixture-of-Experts~(MoE) architecture for modeling mathematical text, so as to capture the common mathematical knowledge across tasks. For optimizing the MoE architecture, we design \emph{multi-task continual pre-training} and \emph{multi-task fine-tuning} strategies for multi-task adaptation. These training strategies can effectively decompose the knowledge from the task data and establish the cross-task sharing via expert networks. In order to further improve the general capacity of solving different complex tasks, we leverage large language models~(LLMs) as complementary models to iteratively refine the generated solution by our PLM, via in-context learning. Extensive experiments have demonstrated the effectiveness of our model.

Zhipeng Chen, Xinheng Wang, Lun Xie et al.

Researchers have shown a growing interest in Audio-driven Talking Head Generation. The primary challenge in talking head generation is achieving audio-visual coherence between the lips and the audio, known as lip synchronization. This paper proposes a generic method, LPIPS-AttnWav2Lip, for reconstructing face images of any speaker based on audio. We used the U-Net architecture based on residual CBAM to better encode and fuse audio and visual modal information. Additionally, the semantic alignment module extends the receptive field of the generator network to obtain the spatial and channel information of the visual features efficiently; and match statistical information of visual features with audio latent vector to achieve the adjustment and injection of the audio content information to the visual information. To achieve exact lip synchronization and to generate realistic high-quality images, our approach adopts LPIPS Loss, which simulates human judgment of image quality and reduces instability possibility during the training process. The proposed method achieves outstanding performance in terms of lip synchronization accuracy and visual quality as demonstrated by subjective and objective evaluation results. The code for the paper is available at the following link: https://github.com/FelixChan9527/LPIPS-AttnWav2Lip

Chengshuo Shen, Wei Zheng, Yonghua Ding et al.

Disruption prediction has made rapid progress in recent years, especially in machine learning (ML)-based methods. Understanding why a predictor makes a certain prediction can be as crucial as the prediction's accuracy for future tokamak disruption predictors. The purpose of most disruption predictors is accuracy or cross-machine capability. However, if a disruption prediction model can be interpreted, it can tell why certain samples are classified as disruption precursors. This allows us to tell the types of incoming disruption and gives us insight into the mechanism of disruption. This paper designs a disruption predictor called Interpretable Disruption Predictor based On Physics-guided feature extraction (IDP-PGFE) on J-TEXT. The prediction performance of the model is effectively improved by extracting physics-guided features. A high-performance model is required to ensure the validity of the interpretation results. The interpretability study of IDP-PGFE provides an understanding of J-TEXT disruption and is generally consistent with existing comprehension of disruption. IDP-PGFE has been applied to the disruption due to continuously increasing density towards density limit experiments on J-TEXT. The time evolution of the PGFE features contribution demonstrates that the application of ECRH triggers radiation-caused disruption, which lowers the density at disruption. While the application of RMP indeed raises the density limit in J-TEXT. The interpretability study guides intuition on the physical mechanisms of density limit disruption that RMPs affect not only the MHD instabilities but also the radiation profile, which delays density limit disruption.

Zhipeng Chen, Zhongrui Zhang, Chao Zhang et al.

The advancement of Large Language Model (LLM)-powered agents has enabled automated task processing through reasoning and tool invocation capabilities. However, existing frameworks often operate under the idealized assumption that tool executions are invariably successful, relying solely on textual descriptions that fail to distinguish precise performance boundaries and cannot adapt to iterative tool updates. This gap introduces uncertainty in planning and execution, particularly in domains like visual content generation (AIGC), where nuanced tool performance significantly impacts outcomes. To address this, we propose PerfGuard, a performance-aware agent framework for visual content generation that systematically models tool performance boundaries and integrates them into task planning and scheduling. Our framework introduces three core mechanisms: (1) Performance-Aware Selection Modeling (PASM), which replaces generic tool descriptions with a multi-dimensional scoring system based on fine-grained performance evaluations; (2) Adaptive Preference Update (APU), which dynamically optimizes tool selection by comparing theoretical rankings with actual execution rankings; and (3) Capability-Aligned Planning Optimization (CAPO), which guides the planner to generate subtasks aligned with performance-aware strategies. Experimental comparisons against state-of-the-art methods demonstrate PerfGuard's advantages in tool selection accuracy, execution reliability, and alignment with user intent, validating its robustness and practical utility for complex AIGC tasks. The project code is available at https://github.com/FelixChan9527/PerfGuard.

Xinkun Ai, Wei Zheng, Ming Zhang et al.

The full understanding of plasma disruption in tokamaks is currently lacking, and data-driven methods are extensively used for disruption prediction. However, most existing data-driven disruption predictors employ supervised learning techniques, which require labeled training data. The manual labeling of disruption precursors is a tedious and challenging task, as some precursors are difficult to accurately identify, limiting the potential of machine learning models. To address this issue, commonly used labeling methods assume that the precursor onset occurs at a fixed time before the disruption, which may not be consistent for different types of disruptions or even the same type of disruption, due to the different speeds at which plasma instabilities escalate. This leads to mislabeled samples and suboptimal performance of the supervised learning predictor. In this paper, we present a disruption prediction method based on anomaly detection that overcomes the drawbacks of unbalanced positive and negative data samples and inaccurately labeled disruption precursor samples. We demonstrate the effectiveness and reliability of anomaly detection predictors based on different algorithms on J-TEXT and EAST to evaluate the reliability of the precursor onset time inferred by the anomaly detection predictor. The precursor onset times inferred by these predictors reveal that the labeling methods have room for improvement as the onset times of different shots are not necessarily the same. Finally, we optimize precursor labeling using the onset times inferred by the anomaly detection predictor and test the optimized labels on supervised learning disruption predictors. The results on J-TEXT and EAST show that the models trained on the optimized labels outperform those trained on fixed onset time labels.

Zhipeng Chen, Yingqian Min, Beichen Zhang et al.

In this report, we present the third technical report on the development of slow-thinking models as part of the STILL project. As the technical pathway becomes clearer, scaling RL training has become a central technique for implementing such reasoning models. We systematically experiment with and document the effects of various factors influencing RL training, conducting experiments on both base models and fine-tuned models. Specifically, we demonstrate that our RL training approach consistently improves the Qwen2.5-32B base models, enhancing both response length and test accuracy. Furthermore, we show that even when a model like DeepSeek-R1-Distill-Qwen-1.5B has already achieved a high performance level, it can be further refined through RL training, reaching an accuracy of 39.33% on AIME 2024. Beyond RL training, we also explore the use of tool manipulation, finding that it significantly boosts the reasoning performance of large reasoning models. This approach achieves a remarkable accuracy of 86.67% with greedy search on AIME 2024, underscoring its effectiveness in enhancing model capabilities. We release our resources at the STILL project website: https://github.com/RUCAIBox/Slow_Thinking_with_LLMs.

Wei Zheng, Fengming Xue, Ming Zhang et al.

Predicting disruptions across different tokamaks is a great obstacle to overcome. Future tokamaks can hardly tolerate disruptions at high performance discharge. Few disruption discharges at high performance can hardly compose an abundant training set, which makes it difficult for current data-driven methods to obtain an acceptable result. A machine learning method capable of transferring a disruption prediction model trained on one tokamak to another is required to solve the problem. The key is a disruption prediction model containing a feature extractor that is able to extract common disruption precursor traces in tokamak diagnostic data, and a transferable disruption classifier. Based on the concerns above, the paper first presents a deep fusion feature extractor designed specifically for extracting disruption precursor features from common diagnostics on tokamaks according to currently known precursors of disruption, providing a promising foundation for transferable models. The fusion feature extractor is proved by comparing with manual feature extraction on J-TEXT. Based on the feature extractor trained on J-TEXT, the disruption prediction model was transferred to EAST data with mere 20 discharges from EAST experiment. The performance is comparable with a model trained with 1896 discharges from EAST. From the comparison among other model training scenarios, transfer learning showed its potential in predicting disruptions across different tokamaks.

Kun Zhou, Beichen Zhang, Jiapeng Wang et al.

Mathematical reasoning is an important capability of large language models~(LLMs) for real-world applications. To enhance this capability, existing work either collects large-scale math-related texts for pre-training, or relies on stronger LLMs (\eg GPT-4) to synthesize massive math problems. Both types of work generally lead to large costs in training or synthesis. To reduce the cost, based on open-source available texts, we propose an efficient way that trains a small LLM for math problem synthesis, to efficiently generate sufficient high-quality pre-training data. To achieve it, we create a dataset using GPT-4 to distill its data synthesis capability into the small LLM. Concretely, we craft a set of prompts based on human education stages to guide GPT-4, to synthesize problems covering diverse math knowledge and difficulty levels. Besides, we adopt the gradient-based influence estimation method to select the most valuable math-related texts. The both are fed into GPT-4 for creating the knowledge distillation dataset to train the small LLM. We leverage it to synthesize 6 million math problems for pre-training our JiuZhang3.0 model, which only needs to invoke GPT-4 API 9.3k times and pre-train on 4.6B data. Experimental results have shown that JiuZhang3.0 achieves state-of-the-art performance on several mathematical reasoning datasets, under both natural language reasoning and tool manipulation settings. Our code and data will be publicly released in \url{https://github.com/RUCAIBox/JiuZhang3.0}.

Chengshuo Shen, Jianchao Li, Yonghua Ding et al.

Measurement of locked mode (LM) is important for the physical research of Magnetohydrodynamic (MHD) instabilities and plasma disruption. The n = 0 pick-up need to be extracted and subtracted to calculate the amplitude and phase of the LM. A new method to extract this pick-up has been developed by predicting the n = 0 pick-up brn=0 by the LM detectors based on Neural Networks (NNs) in J-TEXT. An approach called Power Multiple Time Scale (PMTS) has been developed with outstanding regressing effect in multiple frequency ranges. Three models have been progressed based on PMTS NNs. PMTS could fit the brn=0 on the LM detectors with little errors both in time domain and frequency domain. The n>0 pick-up brn>0 generated by resonant magnetic perturbations (RMPs) can be obtained after subtracting the extracted brn=0. This new method uses only one LM instead of 4 LM detectors to extract brn=0. Therefore, the distribution of the LM detectors can also be optimized based on this new method.

Zhipeng Chen, Kun Zhou, Wayne Xin Zhao et al.

Reinforcement learning (RL) has been widely used in training large language models (LLMs) for preventing unexpected outputs, eg reducing harmfulness and errors. However, existing RL methods mostly adopt the instance-level reward, which is unable to provide fine-grained supervision for complex reasoning tasks, and can not focus on the few key tokens that lead to the incorrectness. To address it, we propose a new RL method named RLMEC that incorporates a generative model as the reward model, which is trained by the erroneous solution rewriting task under the minimum editing constraint, and can produce token-level rewards for RL training. Based on the generative reward model, we design the token-level RL objective for training and an imitation-based regularization for stabilizing RL process. And the both objectives focus on the learning of the key tokens for the erroneous solution, reducing the effect of other unimportant tokens. The experiment results on mathematical tasks and question-answering tasks have demonstrated the effectiveness of our approach. Our code and data are available at https://github.com/RUCAIBox/RLMEC.

Fei Bai, Zhipeng Chen, Chuan Hao et al.

Recently, reinforcement learning~(RL) has become an important approach for improving the capabilities of large language models~(LLMs). In particular, reinforcement learning from verifiable rewards~(RLVR) has emerged as a promising paradigm for reasoning tasks. However, existing RL-based training still remains only a rough approximation to human learning. Human learners leverage both external and internal experience to guide exploration and gradually internalize useful trajectories into stable knowledge. Motivated by this gap, we ask: how can LLMs better utilize and internalize experience during RLVR training? To answer this question, we propose \textbf{D}ual \textbf{G}uidance \textbf{O}ptimization~(\textbf{DGO}), a unified framework that leverages \emph{external} and \emph{internal experience} to improve training effectiveness. Specifically, DGO first constructs an experience bank from previously explored trajectories. The policy then performs exploration under the joint guidance of the experience bank and the model's internal knowledge. The resulting trajectories are further used to refine the experience bank and optimize model parameters, forming a closed loop of experience utilization and internalization. Experiments show that DGO consistently outperforms baseline methods, suggesting that better utilization and internalization of experience lead to more effective reasoning.

Huatong Song, Jinhao Jiang, Wenqing Tian et al.

Large Language Models (LLMs) are powerful but prone to hallucinations due to static knowledge. Retrieval-Augmented Generation (RAG) helps by injecting external information, but current methods often are costly, generalize poorly, or ignore the internal knowledge of the model. In this paper, we introduce R1-Searcher++, a novel framework designed to train LLMs to adaptively leverage both internal and external knowledge sources. R1-Searcher++ employs a two-stage training strategy: an initial SFT Cold-start phase for preliminary format learning, followed by RL for Dynamic Knowledge Acquisition. The RL stage uses outcome-supervision to encourage exploration, incorporates a reward mechanism for internal knowledge utilization, and integrates a memorization mechanism to continuously assimilate retrieved information, thereby enriching the model's internal knowledge. By leveraging internal knowledge and external search engine, the model continuously improves its capabilities, enabling efficient retrieval-augmented reasoning. Our experiments demonstrate that R1-Searcher++ outperforms previous RAG and reasoning methods and achieves efficient retrieval. The code is available at https://github.com/RUCAIBox/R1-Searcher-plus.

Haoxiang Sun, Yingqian Min, Zhipeng Chen et al.

In recent years, the rapid development of large reasoning models has resulted in the saturation of existing benchmarks for evaluating mathematical reasoning, highlighting the urgent need for more challenging and rigorous evaluation frameworks. To address this gap, we introduce OlymMATH, a novel Olympiad-level mathematical benchmark, designed to rigorously test the complex reasoning capabilities of LLMs. OlymMATH features 200 meticulously curated problems, each manually verified and available in parallel English and Chinese versions. The problems are systematically organized into two distinct difficulty tiers: (1) AIME-level problems (easy) that establish a baseline for mathematical reasoning assessment, and (2) significantly more challenging problems (hard) designed to push the boundaries of current state-of-the-art models. In our benchmark, these problems span four core mathematical fields, each including a verifiable numerical solution to enable objective, rule-based evaluation. Empirical results underscore the significant challenge presented by OlymMATH, with state-of-the-art models including DeepSeek-R1, OpenAI's o3-mini and Gemini 2.5 Pro Exp demonstrating notably limited accuracy on the hard subset. Furthermore, the benchmark facilitates comprehensive bilingual assessment of mathematical reasoning abilities-a critical dimension that remains largely unaddressed in mainstream mathematical reasoning benchmarks. We release the benchmark, evaluation code, detailed results and a data visualization tool at https://github.com/RUCAIBox/OlymMATH.

Fei Bai, Yingqian Min, Beichen Zhang et al.

In this paper, we investigate code-integrated reasoning, where models generate code when necessary and integrate feedback by executing it through a code interpreter. To acquire this capability, models must learn when and how to use external code tools effectively, which is supported by tool-augmented reinforcement learning (RL) through interactive learning. Despite its benefits, tool-augmented RL can still suffer from potential instability in the learning dynamics. In light of this challenge, we present a systematic approach to improving the training effectiveness and stability of tool-augmented RL for code-integrated reasoning. Specifically, we develop enhanced training strategies that balance exploration and stability, progressively building tool-use capabilities while improving reasoning performance. Through extensive experiments on five mainstream mathematical reasoning benchmarks, our model demonstrates significant performance improvements over multiple competitive baselines. Furthermore, we conduct an in-depth analysis of the mechanism and effect of code-integrated reasoning, revealing several key insights, such as the extension of model's capability boundaries and the simultaneous improvement of reasoning efficiency through code integration. All data and code for reproducing this work are available at: https://github.com/RUCAIBox/CIR.

Zican Dong, Peiyu Liu, Junyi Li et al.

Recently, large language models (LLMs) have shown remarkable reasoning abilities by producing long reasoning traces. However, as the sequence length grows, the key-value (KV) cache expands linearly, incurring significant memory and computation costs. Existing KV cache eviction methods mitigate this issue by discarding less important KV pairs, but often fail to capture complex KV dependencies, resulting in performance degradation. To better balance efficiency and performance, we introduce ForesightKV, a training-based KV cache eviction framework that learns to predict which KV pairs to evict during long-text generations. We first design the Golden Eviction algorithm, which identifies the optimal eviction KV pairs at each step using future attention scores. These traces and the scores at each step are then distilled via supervised training with a Pairwise Ranking Loss. Furthermore, we formulate cache eviction as a Markov Decision Process and apply the GRPO algorithm to mitigate the significant language modeling loss increase on low-entropy tokens. Experiments on AIME2024 and AIME2025 benchmarks of three reasoning models demonstrate that ForesightKV consistently outperforms prior methods under only half the cache budget, while benefiting synergistically from both supervised and reinforcement learning approaches.

Shiyi Xu, Yiwen Hu, Yingqian Min et al.

With the significant progress of large reasoning models in complex coding and reasoning tasks, existing benchmarks, like LiveCodeBench and CodeElo, are insufficient to evaluate the coding capabilities of large language models (LLMs) in real competition environments. Moreover, current evaluation metrics such as Pass@K fail to capture the reflective abilities of reasoning models. To address these challenges, we propose \textbf{ICPC-Eval}, a top-level competitive coding benchmark designed to probing the frontiers of LLM reasoning. ICPC-Eval includes 118 carefully curated problems from 11 recent ICPC contests held in various regions of the world, offering three key contributions: 1) A challenging realistic ICPC competition scenario, featuring a problem type and difficulty distribution consistent with actual contests. 2) A robust test case generation method and a corresponding local evaluation toolkit, enabling efficient and accurate local evaluation. 3) An effective test-time scaling evaluation metric, Refine@K, which allows iterative repair of solutions based on execution feedback. The results underscore the significant challenge in evaluating complex reasoning abilities: top-tier reasoning models like DeepSeek-R1 often rely on multi-turn code feedback to fully unlock their in-context reasoning potential when compared to non-reasoning counterparts. Furthermore, despite recent advancements in code generation, these models still lag behind top-performing human teams. We release the benchmark at: https://github.com/RUCAIBox/Slow_Thinking_with_LLMs

Huatong Song, Jinhao Jiang, Yingqian Min et al.

Existing Large Reasoning Models (LRMs) have shown the potential of reinforcement learning (RL) to enhance the complex reasoning capabilities of Large Language Models~(LLMs). While they achieve remarkable performance on challenging tasks such as mathematics and coding, they often rely on their internal knowledge to solve problems, which can be inadequate for time-sensitive or knowledge-intensive questions, leading to inaccuracies and hallucinations. To address this, we propose \textbf{R1-Searcher}, a novel two-stage outcome-based RL approach designed to enhance the search capabilities of LLMs. This method allows LLMs to autonomously invoke external search systems to access additional knowledge during the reasoning process. Our framework relies exclusively on RL, without requiring process rewards or distillation for a cold start. % effectively generalizing to out-of-domain datasets and supporting both Base and Instruct models. Our experiments demonstrate that our method significantly outperforms previous strong RAG methods, even when compared to the closed-source GPT-4o-mini.

Dong Guo, Faming Wu, Feida Zhu et al. · pku

We present Seed1.5-VL, a vision-language foundation model designed to advance general-purpose multimodal understanding and reasoning. Seed1.5-VL is composed with a 532M-parameter vision encoder and a Mixture-of-Experts (MoE) LLM of 20B active parameters. Despite its relatively compact architecture, it delivers strong performance across a wide spectrum of public VLM benchmarks and internal evaluation suites, achieving the state-of-the-art performance on 38 out of 60 public benchmarks. Moreover, in agent-centric tasks such as GUI control and gameplay, Seed1.5-VL outperforms leading multimodal systems, including OpenAI CUA and Claude 3.7. Beyond visual and video understanding, it also demonstrates strong reasoning abilities, making it particularly effective for multimodal reasoning challenges such as visual puzzles. We believe these capabilities will empower broader applications across diverse tasks. In this report, we mainly provide a comprehensive review of our experiences in building Seed1.5-VL across model design, data construction, and training at various stages, hoping that this report can inspire further research. Seed1.5-VL is now accessible at https://www.volcengine.com/ (Volcano Engine Model ID: doubao-1-5-thinking-vision-pro-250428)

Yingqian Min, Zhipeng Chen, Jinhao Jiang et al.

Recently, slow-thinking reasoning systems, such as o1, have demonstrated remarkable capabilities in solving complex reasoning tasks. These systems typically engage in an extended thinking process before responding to a query, allowing them to generate more thorough, accurate, and well-reasoned solutions. These systems are primarily developed and maintained by industry, with their core techniques not publicly disclosed. In response, an increasing number of studies from the research community aim to explore the technical foundations underlying these powerful reasoning systems. Building on these prior efforts, this paper presents a reproduction report on implementing o1-like reasoning systems. We introduce an ``imitate, explore, and self-improve'' framework, denoted as \textbf{STILL-2}, as our primary technical approach to train the reasoning model. In the initial phase, we use distilled long-form thought data to fine-tune the reasoning model, enabling it to invoke a slow-thinking mode. The model is then encouraged to explore challenging problems by generating multiple rollouts, which can result in increasingly more high-quality trajectories that lead to correct answers. Furthermore, the model undergoes self-improvement by iteratively refining its training dataset. To verify the effectiveness of this approach, we conduct extensive experiments on three challenging benchmarks. The experimental results demonstrate that our approach achieves competitive performance compared to industry-level reasoning systems on these benchmarks.

Yutao Zhu, Kun Zhou, Kelong Mao et al.

Large language models (LLMs) have become the foundation of many applications, leveraging their extensive capabilities in processing and understanding natural language. While many open-source LLMs have been released with technical reports, the lack of training details hinders further research and development. This paper presents the development of YuLan, a series of open-source LLMs with $12$ billion parameters. The base model of YuLan is pre-trained on approximately $1.7$T tokens derived from a diverse corpus, including massive English, Chinese, and multilingual texts. We design a three-stage pre-training method to enhance YuLan's overall capabilities. Subsequent phases of training incorporate instruction-tuning and human alignment, employing a substantial volume of high-quality synthesized data. To facilitate the learning of complex and long-tail knowledge, we devise a curriculum-learning framework throughout across these stages, which helps LLMs learn knowledge in an easy-to-hard manner. YuLan's training is finished on Jan, 2024 and has achieved performance on par with state-of-the-art LLMs across various English and Chinese benchmarks. This paper outlines a comprehensive technical roadmap for developing LLMs from scratch. Our model and codes are available at https://github.com/RUC-GSAI/YuLan-Chat.

Zhipeng Chen, Kun Zhou, Wayne Xin Zhao et al.

Large language models (LLMs) are still struggling in aligning with human preference in complex tasks and scenarios. They are prone to overfit into the unexpected patterns or superficial styles in the training data. We conduct an empirical study that only selects the top-10\% most updated parameters in LLMs for alignment training, and see improvements in the convergence process and final performance. It indicates the existence of redundant neurons in LLMs for alignment training. To reduce its influence, we propose a low-redundant alignment method named \textbf{ALLO}, focusing on optimizing the most related neurons with the most useful supervised signals. Concretely, we first identify the neurons that are related to the human preference data by a gradient-based strategy, then identify the alignment-related key tokens by reward models for computing loss. Besides, we also decompose the alignment process into the forgetting and learning stages, where we first forget the tokens with unaligned knowledge and then learn aligned knowledge, by updating different ratios of neurons, respectively. Experimental results on 10 datasets have shown the effectiveness of ALLO. Our code and data are available at \url{https://github.com/RUCAIBox/ALLO}.

Zhipeng Chen, Kun Zhou, Beichen Zhang et al.

Although large language models (LLMs) have achieved excellent performance in a variety of evaluation benchmarks, they still struggle in complex reasoning tasks which require specific knowledge and multi-hop reasoning. To improve the reasoning abilities, we propose ChatCoT, a tool-augmented chain-of-thought reasoning framework for chat-based LLMs (e.g., ChatGPT). In ChatCoT, we model the chain-of-thought (CoT) reasoning as multi-turn conversations, to utilize tools in a more natural way through chatting. At each turn, LLMs can either interact with tools or perform the reasoning. Our approach can effectively leverage the multi-turn conversation ability of chat-based LLMs, and integrate the thought chain following and tools manipulation in a unified way. Specially, we initialize the early turns of the conversation by the knowledge about tools, tasks, and reasoning format, and propose an iterative tool-augmented reasoning step to perform step-by-step tool-augmented reasoning. The experiment results on two complex reasoning datasets (MATH and HotpotQA) have shown the effectiveness of ChatCoT on complex reasoning tasks, achieving a 7.9% relative improvement over the state-of-the-art baseline. Our code and data are available at: \url{https://github.com/RUCAIBOX/ChatCoT}.

Junyi Li, Tianyi Tang, Gaole He et al.

In this paper, we release an open-source library, called TextBox, to provide a unified, modularized, and extensible text generation framework. TextBox aims to support a broad set of text generation tasks and models. In our library, we implement 21 text generation models on 9 benchmark datasets, covering the categories of VAE, GAN, and pretrained language models. Meanwhile, our library maintains sufficient modularity and extensibility by properly decomposing the model architecture, inference, and learning process into highly reusable modules, which allows users to easily incorporate new models into our framework. The above features make TextBox specially suitable for researchers and practitioners to quickly reproduce baseline models and develop new models. TextBox is implemented based on PyTorch, and released under Apache License 2.0 at https://github.com/RUCAIBox/TextBox.

Yiming Cui, Ting Liu, Ziqing Yang et al.

Owing to the continuous efforts by the Chinese NLP community, more and more Chinese machine reading comprehension datasets become available. To add diversity in this area, in this paper, we propose a new task called Sentence Cloze-style Machine Reading Comprehension (SC-MRC). The proposed task aims to fill the right candidate sentence into the passage that has several blanks. We built a Chinese dataset called CMRC 2019 to evaluate the difficulty of the SC-MRC task. Moreover, to add more difficulties, we also made fake candidates that are similar to the correct ones, which requires the machine to judge their correctness in the context. The proposed dataset contains over 100K blanks (questions) within over 10K passages, which was originated from Chinese narrative stories. To evaluate the dataset, we implement several baseline systems based on the pre-trained models, and the results show that the state-of-the-art model still underperforms human performance by a large margin. We release the dataset and baseline system to further facilitate our community. Resources available through https://github.com/ymcui/cmrc2019

Ziqing Yang, Yiming Cui, Zhipeng Chen et al.

In this paper, we introduce TextBrewer, an open-source knowledge distillation toolkit designed for natural language processing. It works with different neural network models and supports various kinds of supervised learning tasks, such as text classification, reading comprehension, sequence labeling. TextBrewer provides a simple and uniform workflow that enables quick setting up of distillation experiments with highly flexible configurations. It offers a set of predefined distillation methods and can be extended with custom code. As a case study, we use TextBrewer to distill BERT on several typical NLP tasks. With simple configurations, we achieve results that are comparable with or even higher than the public distilled BERT models with similar numbers of parameters. Our toolkit is available through: http://textbrewer.hfl-rc.com

Yiming Cui, Ting Liu, Wanxiang Che et al.

Machine Reading Comprehension (MRC) has become enormously popular recently and has attracted a lot of attention. However, the existing reading comprehension datasets are mostly in English. In this paper, we introduce a Span-Extraction dataset for Chinese machine reading comprehension to add language diversities in this area. The dataset is composed by near 20,000 real questions annotated on Wikipedia paragraphs by human experts. We also annotated a challenge set which contains the questions that need comprehensive understanding and multi-sentence inference throughout the context. We present several baseline systems as well as anonymous submissions for demonstrating the difficulties in this dataset. With the release of the dataset, we hosted the Second Evaluation Workshop on Chinese Machine Reading Comprehension (CMRC 2018). We hope the release of the dataset could further accelerate the Chinese machine reading comprehension research. Resources are available: https://github.com/ymcui/cmrc2018

Zhipeng Chen, Xiaobo Qin, Youbin Wu et al.

Reinforcement learning with verifiable rewards (RLVR), which typically adopts Pass@1 as the reward, has faced the issues in balancing exploration and exploitation, causing policies to prefer conservative actions, converging to a local optimum. Identifying an appropriate reward metric is therefore crucial. Regarding the prior work, although Pass@k has been used in evaluation, its connection to LLM exploration ability in RLVR remains largely overlooked. To investigate this, we first use Pass@k as the reward to train the policy model (i.e., $\textbf{Pass@k Training}$), and observe the improvement on its exploration ability. Next, we derive an analytical solution for the advantage of Pass@k Training, leading to an efficient and effective process. Building on this, our analysis reveals that exploration and exploitation are not inherently conflicting objectives, while they can mutually enhance each other. Moreover, Pass@k Training with analytical derivation essentially involves directly designing the advantage function. Inspired by this, we preliminarily explore the advantage design for RLVR, showing promising results and highlighting a potential future direction.

Jinhao Jiang, Zhipeng Chen, Yingqian Min et al.

Recently, test-time scaling has garnered significant attention from the research community, largely due to the substantial advancements of the o1 model released by OpenAI. By allocating more computational resources during the inference phase, large language models~(LLMs) can extensively explore the solution space by generating more thought tokens or diverse solutions, thereby producing more accurate responses. However, developing an o1-like reasoning approach is challenging, and researchers have been making various attempts to advance this open area of research. In this paper, we present a preliminary exploration into enhancing the reasoning abilities of LLMs through reward-guided tree search algorithms. This framework is implemented by integrating the policy model, reward model, and search algorithm. It is primarily constructed around a tree search algorithm, where the policy model navigates a dynamically expanding tree guided by a specially trained reward model. The implemented framework is denoted as \textbf{STILL-1}. We thoroughly explore various design considerations necessary for implementing this framework and provide a detailed report of the technical aspects. To assess the effectiveness of our approach, we focus on mathematical reasoning tasks and conduct extensive evaluations on four challenging datasets, significantly enhancing the reasoning abilities of LLMs.

Jia Deng, Jie Chen, Zhipeng Chen et al.

Recently, reinforcement learning with verifiable rewards (RLVR) has been widely used for enhancing the reasoning abilities of large language models (LLMs). A core challenge in RLVR involves managing the exchange between entropy and performance of policies. Despite the importance of this exchange, a fine-grained understanding of when and how this exchange operates most effectively remains limited. To bridge this gap, we conduct a systematic empirical analysis of the entropy-performance exchange mechanism of RLVR across different levels of granularity. Specifically, we first divide the training process into two distinct stages based on entropy dynamics, i.e., rising stage and plateau stage, and then systematically investigate how this mechanism varies across stage-level, instance-level, and token-level granularitiess. Our analysis reveals that, in the rising stage, entropy reduction in negative samples facilitates the learning of effective reasoning patterns, which in turn drives rapid performance gains. Moreover, in the plateau stage, learning efficiency strongly correlates with high-entropy tokens present in low-perplexity samples and those located at the end of sequences. Motivated by these findings, we propose two methods that dynamically adjust the reward signal using perplexity and positional information to focus RL updates on tokens that exhibit high learning potential, achieving improvements compared to the baseline methods on various LLMs.

Jia Deng, Jie Chen, Zhipeng Chen et al.

Reinforcement learning with verifiable rewards (RLVR) has emerged as a powerful paradigm for enhancing the reasoning capabilities of large language models (LLMs). Unlike traditional RL approaches, RLVR leverages rule-based feedback to guide LLMs in generating and refining complex reasoning chains -- a process critically dependent on effective exploration strategies. While prior work has demonstrated RLVR's empirical success, the fundamental mechanisms governing LLMs' exploration behaviors remain underexplored. This technical report presents a systematic investigation of exploration capacities in RLVR, covering four main aspects: (1) exploration space shaping, where we develop quantitative metrics to characterize LLMs' capability boundaries; (2) entropy-performance exchange, analyzed across training stages, individual instances, and token-level patterns; and (3) RL performance optimization, examining methods to effectively translate exploration gains into measurable improvements. By unifying previously identified insights with new empirical evidence, this work aims to provide a foundational framework for advancing RLVR systems.

Wenpeng Xing, Zhipeng Chen, Changting Lin et al.

Invoking external tools enables Large Language Models (LLMs) to perform complex, real-world tasks, yet selecting the correct tool from large, hierarchically-structured libraries remains a significant challenge. The limited context windows of LLMs and noise from irrelevant options often lead to low selection accuracy and high computational costs. To address this, we propose the Hierarchical Gaussian Mixture Framework (HGMF), a probabilistic pruning method for scalable tool invocation. HGMF first maps the user query and all tool descriptions into a unified semantic space. The framework then operates in two stages: it clusters servers using a Gaussian Mixture Model (GMM) and filters them based on the query's likelihood. Subsequently, it applies the same GMM-based clustering and filtering to the tools associated with the selected servers. This hierarchical process produces a compact, high-relevance candidate set, simplifying the final selection task for the LLM. Experiments on a public dataset show that HGMF significantly improves tool selection accuracy while reducing inference latency, confirming the framework's scalability and effectiveness for large-scale tool libraries.

Zhipeng Chen, Lan Yang, Yonggang Qi et al.

Despite the rapid advancements in text-to-image (T2I) synthesis, enabling precise visual control remains a significant challenge. Existing works attempted to incorporate multi-facet controls (text and sketch), aiming to enhance the creative control over generated images. However, our pilot study reveals that the expressive power of humans far surpasses the capabilities of current methods. Users desire a more versatile approach that can accommodate their diverse creative intents, ranging from controlling individual subjects to manipulating the entire scene composition. We present VersaGen, a generative AI agent that enables versatile visual control in T2I synthesis. VersaGen admits four types of visual controls: i) single visual subject; ii) multiple visual subjects; iii) scene background; iv) any combination of the three above or merely no control at all. We train an adaptor upon a frozen T2I model to accommodate the visual information into the text-dominated diffusion process. We introduce three optimization strategies during the inference phase of VersaGen to improve generation results and enhance user experience. Comprehensive experiments on COCO and Sketchy validate the effectiveness and flexibility of VersaGen, as evidenced by both qualitative and quantitative results.

Yerbolat Khassanov, Zhipeng Chen, Tianfeng Chen et al.

This paper addresses challenges in integrating new languages into a pre-trained multilingual automatic speech recognition (mASR) system, particularly in scenarios where training data for existing languages is limited or unavailable. The proposed method employs a dual-pipeline with low-rank adaptation (LoRA). It maintains two data flow pipelines-one for existing languages and another for new languages. The primary pipeline follows the standard flow through the pre-trained parameters of mASR, while the secondary pipeline additionally utilizes language-specific parameters represented by LoRA and a separate output decoder module. Importantly, the proposed approach minimizes the performance degradation of existing languages and enables a language-agnostic operation mode, facilitated by a decoder selection strategy. We validate the effectiveness of the proposed method by extending the pre-trained Whisper model to 19 new languages from the FLEURS dataset

Yiming Cui, Wei-Nan Zhang, Wanxiang Che et al.

Recurrent Neural Networks (RNN) are known as powerful models for handling sequential data, and especially widely utilized in various natural language processing tasks. In this paper, we propose Contextual Recurrent Units (CRU) for enhancing local contextual representations in neural networks. The proposed CRU injects convolutional neural networks (CNN) into the recurrent units to enhance the ability to model the local context and reducing word ambiguities even in bi-directional RNNs. We tested our CRU model on sentence-level and document-level modeling NLP tasks: sentiment classification and reading comprehension. Experimental results show that the proposed CRU model could give significant improvements over traditional CNN or RNN models, including bidirectional conditions, as well as various state-of-the-art systems on both tasks, showing its promising future of extensibility to other NLP tasks as well.

Zhipeng Chen, Yiming Cui, Wentao Ma et al.

Machine Reading Comprehension (MRC) with multiple-choice questions requires the machine to read given passage and select the correct answer among several candidates. In this paper, we propose a novel approach called Convolutional Spatial Attention (CSA) model which can better handle the MRC with multiple-choice questions. The proposed model could fully extract the mutual information among the passage, question, and the candidates, to form the enriched representations. Furthermore, to merge various attention results, we propose to use convolutional operation to dynamically summarize the attention values within the different size of regions. Experimental results show that the proposed model could give substantial improvements over various state-of-the-art systems on both RACE and SemEval-2018 Task11 datasets.

Zhipeng Chen, Yiming Cui, Wentao Ma et al.

This paper describes the system which got the state-of-the-art results at SemEval-2018 Task 11: Machine Comprehension using Commonsense Knowledge. In this paper, we present a neural network called Hybrid Multi-Aspects (HMA) model, which mimic the human's intuitions on dealing with the multiple-choice reading comprehension. In this model, we aim to produce the predictions in multiple aspects by calculating attention among the text, question and choices, and combine these results for final predictions. Experimental results show that our HMA model could give substantial improvements over the baseline system and got the first place on the final test set leaderboard with the accuracy of 84.13%.

Zhipeng Chen, Benedetta Tondi, Xiaolong Li et al.

We address the problem of data-driven image manipulation detection in the presence of an attacker with limited knowledge about the detector. Specifically, we assume that the attacker knows the architecture of the detector, the training data and the class of features V the detector can rely on. In order to get an advantage in his race of arms with the attacker, the analyst designs the detector by relying on a subset of features chosen at random in V. Given its ignorance about the exact feature set, the adversary attacks a version of the detector based on the entire feature set. In this way, the effectiveness of the attack diminishes since there is no guarantee that attacking a detector working in the full feature space will result in a successful attack against the reduced-feature detector. We theoretically prove that, thanks to random feature selection, the security of the detector increases significantly at the expense of a negligible loss of performance in the absence of attacks. We also provide an experimental validation of the proposed procedure by focusing on the detection of two specific kinds of image manipulations, namely adaptive histogram equalization and median filtering. The experiments confirm the gain in security at the expense of a negligible loss of performance in the absence of attacks.

Yiming Cui, Ting Liu, Zhipeng Chen et al.

Machine Reading Comprehension (MRC) has become enormously popular recently and has attracted a lot of attention. However, existing reading comprehension datasets are mostly in English. To add diversity in reading comprehension datasets, in this paper we propose a new Chinese reading comprehension dataset for accelerating related research in the community. The proposed dataset contains two different types: cloze-style reading comprehension and user query reading comprehension, associated with large-scale training data as well as human-annotated validation and hidden test set. Along with this dataset, we also hosted the first Evaluation on Chinese Machine Reading Comprehension (CMRC-2017) and successfully attracted tens of participants, which suggest the potential impact of this dataset.

Yiming Cui, Zhipeng Chen, Si Wei et al.

Cloze-style queries are representative problems in reading comprehension. Over the past few months, we have seen much progress that utilizing neural network approach to solve Cloze-style questions. In this paper, we present a novel model called attention-over-attention reader for the Cloze-style reading comprehension task. Our model aims to place another attention mechanism over the document-level attention, and induces "attended attention" for final predictions. Unlike the previous works, our neural network model requires less pre-defined hyper-parameters and uses an elegant architecture for modeling. Experimental results show that the proposed attention-over-attention model significantly outperforms various state-of-the-art systems by a large margin in public datasets, such as CNN and Children's Book Test datasets.

Yiming Cui, Ting Liu, Zhipeng Chen et al.

Reading comprehension has embraced a booming in recent NLP research. Several institutes have released the Cloze-style reading comprehension data, and these have greatly accelerated the research of machine comprehension. In this work, we firstly present Chinese reading comprehension datasets, which consist of People Daily news dataset and Children's Fairy Tale (CFT) dataset. Also, we propose a consensus attention-based neural network architecture to tackle the Cloze-style reading comprehension problem, which aims to induce a consensus attention over every words in the query. Experimental results show that the proposed neural network significantly outperforms the state-of-the-art baselines in several public datasets. Furthermore, we setup a baseline for Chinese reading comprehension task, and hopefully this would speed up the process for future research.