Yuntao Li

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

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

ByteDance Seed, Jiaze Chen, Tiantian Fan et al. · bytedance

We introduce Seed1.5-Thinking, capable of reasoning through thinking before responding, resulting in improved performance on a wide range of benchmarks. Seed1.5-Thinking achieves 86.7 on AIME 2024, 55.0 on Codeforces and 77.3 on GPQA, demonstrating excellent reasoning abilities in STEM and coding. Beyond reasoning tasks, the method demonstrates notable generalization across diverse domains. For instance, it surpasses DeepSeek R1 by 8% in win rate on non-reasoning tasks, indicating its broader applicability. Compared to other state-of-the-art reasoning models, Seed1.5-Thinking is a Mixture-of-Experts (MoE) model with a relatively small size, featuring 20B activated and 200B total parameters. As part of our effort to assess generalized reasoning, we develop two internal benchmarks, BeyondAIME and Codeforces, both of which will be publicly released to support future research. Model trial link: https://www.volcengine.com/experience/ark.

Zichun Xu, Zhaomin Wang, Yuntao Li et al.

For peg-in-hole tasks, humans rely on binocular visual perception to locate the peg above the hole surface and then proceed with insertion. This paper draws insights from this behavior to enable agents to learn efficient assembly strategies through visual reinforcement learning. Hence, we propose a Separate Primitive Policy (S2P) to learn how to derive location and insertion actions simultaneously. S2P is compatible with model-free reinforcement learning algorithms. Ten insertion tasks featuring different polygons are developed as benchmarks for evaluations. Simulation experiments show that S2P can boost the sample efficiency and success rate even with force constraints. Real-world experiments are also performed to verify the feasibility of S2P. Ablations are finally given to discuss the generalizability of S2P and some factors that affect its performance.

Yuntao Li, Zhenpeng Su, Yutian Li et al.

Translating natural language queries into SQLs in a seq2seq manner has attracted much attention recently. However, compared with abstract-syntactic-tree-based SQL generation, seq2seq semantic parsers face much more challenges, including poor quality on schematical information prediction and poor semantic coherence between natural language queries and SQLs. This paper analyses the above difficulties and proposes a seq2seq-oriented decoding strategy called SR, which includes a new intermediate representation SSQL and a reranking method with score re-estimator to solve the above obstacles respectively. Experimental results demonstrate the effectiveness of our proposed techniques and T5-SR-3b achieves new state-of-the-art results on the Spider dataset.

Sirui Wang, Di Liang, Jian Song et al.

Transformer-based pre-trained language models such as BERT have achieved remarkable results in Semantic Sentence Matching. However, existing models still suffer from insufficient ability to capture subtle differences. Minor noise like word addition, deletion, and modification of sentences may cause flipped predictions. To alleviate this problem, we propose a novel Dual Attention Enhanced BERT (DABERT) to enhance the ability of BERT to capture fine-grained differences in sentence pairs. DABERT comprises (1) Dual Attention module, which measures soft word matches by introducing a new dual channel alignment mechanism to model affinity and difference attention. (2) Adaptive Fusion module, this module uses attention to learn the aggregation of difference and affinity features, and generates a vector describing the matching details of sentence pairs. We conduct extensive experiments on well-studied semantic matching and robustness test datasets, and the experimental results show the effectiveness of our proposed method.

Jian Song, Di Liang, Rumei Li et al.

Transformer-based pre-trained models like BERT have achieved great progress on Semantic Sentence Matching. Meanwhile, dependency prior knowledge has also shown general benefits in multiple NLP tasks. However, how to efficiently integrate dependency prior structure into pre-trained models to better model complex semantic matching relations is still unsettled. In this paper, we propose the \textbf{D}ependency-Enhanced \textbf{A}daptive \textbf{F}usion \textbf{A}ttention (\textbf{DAFA}), which explicitly introduces dependency structure into pre-trained models and adaptively fuses it with semantic information. Specifically, \textbf{\emph{(i)}} DAFA first proposes a structure-sensitive paradigm to construct a dependency matrix for calibrating attention weights. It adopts an adaptive fusion module to integrate the obtained dependency information and the original semantic signals. Moreover, DAFA reconstructs the attention calculation flow and provides better interpretability. By applying it on BERT, our method achieves state-of-the-art or competitive performance on 10 public datasets, demonstrating the benefits of adaptively fusing dependency structure in semantic matching task.

Zirui Liang, Yuntao Li, Tianjin Huang et al.

Graph neural networks (GNNs) have shown promise in addressing graph-related problems, including node classification. However, conventional GNNs assume an even distribution of data across classes, which is often not the case in real-world scenarios, where certain classes are severely underrepresented. This leads to suboptimal performance of standard GNNs on imbalanced graphs. In this paper, we introduce a unique approach that tackles imbalanced classification on graphs by considering graph heterophily. We investigate the intricate relationship between class imbalance and graph heterophily, revealing that minority classes not only exhibit a scarcity of samples but also manifest lower levels of homophily, facilitating the propagation of erroneous information among neighboring nodes. Drawing upon this insight, we propose an efficient method, called Fast Im-GBK, which integrates an imbalance classification strategy with heterophily-aware GNNs to effectively address the class imbalance problem while significantly reducing training time. Our experiments on real-world graphs demonstrate our model's superiority in classification performance and efficiency for node classification tasks compared to existing baselines.

Sirui Wang, Kaiwen Wei, Hongzhi Zhang et al.

Demonstration learning aims to guide the prompt prediction via providing answered demonstrations in the few shot settings. Despite achieving promising results, existing work only concatenates the answered examples as demonstrations to the prompt template (including the raw context) without any additional operation, neglecting the prompt-demonstration dependencies. Besides, prior research found that randomly replacing the labels of demonstrations marginally hurts performance, illustrating that the model could not properly learn the knowledge brought by the demonstrations. Inspired by the human learning process, in this paper, we introduce Imitation DEMOnstration Learning (Imitation-Demo) to strengthen demonstration learning via explicitly imitating human review behaviour, which includes: (1) contrastive learning mechanism to concentrate on the similar demonstrations. (2) demonstration-label re-prediction method to consolidate known knowledge. Experiment results show that our proposed method achieves state-of-the-art performance on 11 out of 14 classification corpora. Further studies also prove that Imitation-Demo strengthen the association between prompt and demonstrations, which could provide the basis for exploring how demonstration learning works.

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

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

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)

Chenwei Lou, Zewei Sun, Xinnian Liang et al.

Large Language Models (LLMs) have demonstrated remarkable capabilities but often face challenges with tasks requiring sophisticated reasoning. While Chain-of-Thought (CoT) prompting significantly enhances reasoning, it indiscriminately generates lengthy reasoning steps for all queries, leading to substantial computational costs and inefficiency, especially for simpler inputs. To address this critical issue, we introduce AdaCoT (Adaptive Chain-of-Thought), a novel framework enabling LLMs to adaptively decide when to invoke CoT. AdaCoT framed adaptive reasoning as a Pareto optimization problem that seeks to balance model performance with the costs associated with CoT invocation (both frequency and computational overhead). We propose a reinforcement learning (RL) based method, specifically utilizing Proximal Policy Optimization (PPO), to dynamically control the CoT triggering decision boundary by adjusting penalty coefficients, thereby allowing the model to determine CoT necessity based on implicit query complexity. A key technical contribution is Selective Loss Masking (SLM), designed to counteract decision boundary collapse during multi-stage RL training, ensuring robust and stable adaptive triggering. Experimental results demonstrate that AdaCoT successfully navigates the Pareto frontier, achieving substantial reductions in CoT usage for queries not requiring elaborate reasoning. For instance, on our production traffic testset, AdaCoT reduced CoT triggering rates to as low as 3.18\% and decreased average response tokens by 69.06%, while maintaining high performance on complex tasks.

Lei Zhuang, Jingdong Zhao, Yuntao Li et al.

Sampling-based motion planning (SBMP) algorithms are renowned for their robust global search capabilities. However, the inherent randomness in their sampling mechanisms often result in inconsistent path quality and limited search efficiency. In response to these challenges, this work proposes a novel deep learning-based motion planning framework, named Transformer-Enhanced Motion Planner (TEMP), which synergizes an Environmental Information Semantic Encoder (EISE) with a Motion Planning Transformer (MPT). EISE converts environmental data into semantic environmental information (SEI), providing MPT with an enriched environmental comprehension. MPT leverages an attention mechanism to dynamically recalibrate its focus on SEI, task objectives, and historical planning data, refining the sampling node generation. To demonstrate the capabilities of TEMP, we train our model using a dataset comprised of planning results produced by the RRT*. EISE and MPT are collaboratively trained, enabling EISE to autonomously learn and extract patterns from environmental data, thereby forming semantic representations that MPT could more effectively interpret and utilize for motion planning. Subsequently, we conducted a systematic evaluation of TEMP's efficacy across diverse task dimensions, which demonstrates that TEMP achieves exceptional performance metrics and a heightened degree of generalizability compared to state-of-the-art SBMPs.

Zhenpeng Su, Leiyu Pan, Minxuan Lv et al.

Reinforcement learning (RL) has become a powerful paradigm for optimizing large language models (LLMs) to handle complex reasoning tasks. A core challenge in this process lies in managing policy entropy, which reflects the balance between exploration and exploitation during training. Existing methods, such as proximal policy optimization (PPO) and its variants, discard valuable gradient signals from low-probability tokens due to the clipping mechanism. We systematically analyze the entropy dynamics and reveal that these clipped tokens play a critical yet overlooked role in regulating entropy evolution. We propose \textbf{C}oordinating \textbf{E}ntropy via \textbf{G}radient-\textbf{P}reserving \textbf{P}olicy \textbf{O}ptimization (CE-GPPO), a novel algorithm that reintroduces gradients from clipped tokens in native PPO in a gentle and bounded manner. By controlling the magnitude of gradients from tokens outside the clipping interval, CE-GPPO is able to achieve an exploration-exploitation trade-off. We provide theoretical justification and empirical evidence showing that CE-GPPO effectively mitigates entropy instability. Extensive experiments on mathematical reasoning benchmarks show that CE-GPPO consistently outperforms strong baselines across different model scales.

Xiaoyu Liu, Di Liang, Chang Dai et al.

Reward Models (RMs) are key components for evaluating and guiding language model outputs. However, traditional scalar RMs often struggle with incorporating contextual and background information during inference, leading to incomplete evaluations. Generative RMs (GRMs) attempt to address these limitations by generating intermediate reasoning steps. Yet, their uncontrolled black-box nature and inefficiency due to sequential decoding hinder their industrial deployment. Industrial scenarios, such as search and recommendation systems, often involve single-domain tasks requiring evaluation along specific dimensions. In such contexts, diagnosing "bad cases" necessitates structured feedback to identify and optimize dimension-specific issues. In this paper, we propose the Structural Reward Model (SRM), a modular and interpretable framework integrating side-branch models as auxiliary feature generators. By introducing fine-grained dimensions, SRMs enable interpretable and efficient evaluation, facilitating targeted diagnostics and optimization. This structured approach ensures adaptability and scalability for industrial applications. Through comprehensive experiments, we demonstrate that SRMs outperform scalar RMs and GRMs in robustness and alignment with human preferences. The modular design further supports efficient optimization for practical scenarios, allowing SRM to provide a practical reward modeling solution for industry.

Zhenpeng Su, Leiyu Pan, Minxuan Lv et al.

Large language model post-training relies on reinforcement learning to improve model capability and alignment quality. However, the off-policy training paradigm introduces distribution shift, which often pushes the policy beyond the trust region, leading to training instabilities manifested as fluctuations in policy entropy and unstable gradients. Although PPO-Clip mitigates this issue through importance clipping, it still overlooks the global distributional shift of actions. To address these challenges, we propose using the entropy ratio between the current and previous policies as a new global metric that effectively quantifies the relative change in policy exploration throughout updates. Building on this metric, we introduce an \textbf{Entropy Ratio Clipping} (ERC) mechanism that imposes bidirectional constraints on the entropy ratio. This stabilizes policy updates at the global distribution level and compensates for the inability of PPO-clip to regulate probability shifts of un-sampled actions. We integrate ERC into both DAPO and GPPO reinforcement learning algorithms. Experiments across multiple benchmarks show that ERC consistently improves performance.

Zichun Xu, Yuntao Li, Zhaomin Wang et al.

Depth information is robust to scene appearance variations and inherently carries 3D spatial details. In this paper, a visual backbone based on the vision transformer is proposed to fuse RGB and depth modalities for enhancing generalization. Different modalities are first processed by separate CNN stems, and the combined convolutional features are delivered to the scalable vision transformer to obtain visual representations. Moreover, a contrastive unsupervised learning scheme is designed with masked and unmasked tokens to accelerate the sample efficiency during the reinforcement learning process. Simulation results demonstrate that our visual backbone can focus more on task-related regions and exhibit better generalization in unseen scenarios. For sim2real transfer, a flexible curriculum learning schedule is developed to deploy domain randomization over training processes. Finally, the feasibility of our model is validated to perform real-world manipulation tasks via zero-shot transfer.

Yuntao Li, Hanchu Zhang, Yutian Li et al.

Conversational text-to-SQL aims at converting multi-turn natural language queries into their corresponding SQL (Structured Query Language) representations. One of the most intractable problems of conversational text-to-SQL is modelling the semantics of multi-turn queries and gathering the proper information required for the current query. This paper shows that explicitly modelling the semantic changes by adding each turn and the summarization of the whole context can bring better performance on converting conversational queries into SQLs. In particular, we propose two conversational modelling tasks in both turn grain and conversation grain. These two tasks simply work as auxiliary training tasks to help with multi-turn conversational semantic parsing. We conducted empirical studies and achieved new state-of-the-art results on the large-scale open-domain conversational text-to-SQL dataset. The results demonstrate that the proposed mechanism significantly improves the performance of multi-turn semantic parsing.

Yuntao Li, Can Xu, Huang Hu et al.

Retrieve-based dialogue response selection aims to find a proper response from a candidate set given a multi-turn context. Pre-trained language models (PLMs) based methods have yielded significant improvements on this task. The sequence representation plays a key role in the learning of matching degree between the dialogue context and the response. However, we observe that different context-response pairs sharing the same context always have a greater similarity in the sequence representations calculated by PLMs, which makes it hard to distinguish positive responses from negative ones. Motivated by this, we propose a novel \textbf{F}ine-\textbf{G}rained \textbf{C}ontrastive (FGC) learning method for the response selection task based on PLMs. This FGC learning strategy helps PLMs to generate more distinguishable matching representations of each dialogue at fine grains, and further make better predictions on choosing positive responses. Empirical studies on two benchmark datasets demonstrate that the proposed FGC learning method can generally and significantly improve the model performance of existing PLM-based matching models.

Meiqi Chen, Yuan Zhang, Xiaoyu Kou et al.

Graph Attention Network (GAT) focuses on modelling simple undirected and single relational graph data only. This limits its ability to deal with more general and complex multi-relational graphs that contain entities with directed links of different labels (e.g., knowledge graphs). Therefore, directly applying GAT on multi-relational graphs leads to sub-optimal solutions. To tackle this issue, we propose r-GAT, a relational graph attention network to learn multi-channel entity representations. Specifically, each channel corresponds to a latent semantic aspect of an entity. This enables us to aggregate neighborhood information for the current aspect using relation features. We further propose a query-aware attention mechanism for subsequent tasks to select useful aspects. Extensive experiments on link prediction and entity classification tasks show that our r-GAT can model multi-relational graphs effectively. Also, we show the interpretability of our approach by case study.

Yuntao Li, Bei Chen, Qian Liu et al.

In Natural Language Interfaces to Databases systems, the text-to-SQL technique allows users to query databases by using natural language questions. Though significant progress in this area has been made recently, most parsers may fall short when they are deployed in real systems. One main reason stems from the difficulty of fully understanding the users' natural language questions. In this paper, we include human in the loop and present a novel parser-independent interactive approach (PIIA) that interacts with users using multi-choice questions and can easily work with arbitrary parsers. Experiments were conducted on two cross-domain datasets, the WikiSQL and the more complex Spider, with five state-of-the-art parsers. These demonstrated that PIIA is capable of enhancing the text-to-SQL performance with limited interaction turns by using both simulation and human evaluation.

Xiaoyu Kou, Yankai Lin, Yuntao Li et al.

Knowledge graph embedding (KGE), aiming to embed entities and relations into low-dimensional vectors, has attracted wide attention recently. However, the existing research is mainly based on the black-box neural models, which makes it difficult to interpret the learned representation. In this paper, we introduce DisenE, an end-to-end framework to learn disentangled knowledge graph embeddings. Specially, we introduce an attention-based mechanism that enables the model to explicitly focus on relevant components of entity embeddings according to a given relation. Furthermore, we introduce two novel regularizers to encourage each component of the entity representation to independently reflect an isolated semantic aspect. Experimental results demonstrate that our proposed DisenE investigates a perspective to address the interpretability of KGE and is proved to be an effective way to improve the performance of link prediction tasks.

Michael Bernico, Yuntao Li, Dingchao Zhang

Transfer learning allows practitioners to recognize and apply knowledge learned in previous tasks (source task) to new tasks or new domains (target task), which share some commonality. The two important factors impacting the performance of transfer learning models are: (a) the size of the target dataset, and (b) the similarity in distribution between source and target domains. Thus far, there has been little investigation into just how important these factors are. In this paper, we investigate the impact of target dataset size and source/target domain similarity on model performance through a series of experiments. We find that more data is always beneficial, and model performance improves linearly with the log of data size, until we are out of data. As source/target domains differ, more data is required and fine tuning will render better performance than feature extraction. When source/target domains are similar and data size is small, fine tuning and feature extraction renders equivalent performance. Our hope is that by beginning this quantitative investigation on the effect of data volume and domain similarity in transfer learning we might inspire others to explore the significance of data in developing more accurate statistical models.