Zhong Zhang

Zhong Zhang, Giacomo Acciarini, Dario Izzo et al.

Low-thrust trajectory design relies heavily on repeated evaluations of fuel consumption and transfer feasibility, which require expensive optimal control solutions. In this work, we show these quantities can be accurately approximated by machine learning surrogates, enabling fast and scalable evaluation across a wide range of scenarios. By increasing both dataset size and model capacity, we observe that low-thrust trajectory optimization follows a scaling law, with performance improving linearly with the logarithm of training data and network parameters, and no evidence of saturation within the explored regime. Guided by this observation, we construct a large-scale dataset using the proposed homotopy-ray strategy tailored to mission design requirements. A key is the introduction of a self-similar transformation, which allows generalization across semi-major axes, inclinations, and central bodies avoiding retraining. As a result, the same neural approximator can be applied to diverse orbital environments and mission classes. The proposed models accurately predict optimal fuel consumption and minimum transfer time for single- and multi-revolution transfers. Their performance and generalization are demonstrated on a public dataset, a multi-asteroid flyby problem from the Global Trajectory Optimization Competition, and an asteroid rendezvous mission design. The models and datasets are released as open-source to support the space community.

Jun Yu, Hao Chang, Keda Lu et al.

Multi-modal aerial view object classification (MAVOC) in Automatic target recognition (ATR), although an important and challenging problem, has been under studied. This paper firstly finds that fine-grained data, class imbalance and various shooting conditions preclude the representational ability of general image classification. Moreover, the MAVOC dataset has scene aggregation characteristics. By exploiting these properties, we propose Scene Clustering Based Pseudo-labeling Strategy (SCP-Label), a simple yet effective method to employ in post-processing. The SCP-Label brings greater accuracy by assigning the same label to objects within the same scene while also mitigating bias and confusion with model ensembles. Its performance surpasses the official baseline by a large margin of +20.57% Accuracy on Track 1 (SAR), and +31.86% Accuracy on Track 2 (SAR+EO), demonstrating the potential of SCP-Label as post-processing. Finally, we win the championship both on Track1 and Track2 in the CVPR 2022 Perception Beyond the Visible Spectrum (PBVS) Workshop MAVOC Challenge. Our code is available at https://github.com/HowieChangchn/SCP-Label.

Fushuo Huo, Wenchao Xu, Zhong Zhang et al.

While Large Vision-Language Models (LVLMs) have rapidly advanced in recent years, the prevalent issue known as the `hallucination' problem has emerged as a significant bottleneck, hindering their real-world deployments. Existing methods mitigate this issue mainly from two perspectives: One approach leverages extra knowledge like robust instruction tuning LVLMs with curated datasets or employing auxiliary analysis networks, which inevitable incur additional costs. Another approach, known as contrastive decoding, induces hallucinations by manually disturbing the vision or instruction raw inputs and mitigates them by contrasting the outputs of the disturbed and original LVLMs. However, these approaches rely on empirical holistic input disturbances and double the inference cost. To avoid these issues, we propose a simple yet effective method named Self-Introspective Decoding (SID). Our empirical investigation reveals that pretrained LVLMs can introspectively assess the importance of vision tokens based on preceding vision and text (both instruction and generated) tokens. We develop the Context and Text-aware Token Selection (CT2S) strategy, which preserves only unimportant vision tokens after early layers of LVLMs to adaptively amplify text-informed hallucination during the auto-regressive decoding. This approach ensures that multimodal knowledge absorbed in the early layers induces multimodal contextual rather than aimless hallucinations. Subsequently, the original token logits subtract the amplified vision-and-text association hallucinations, guiding LVLMs decoding faithfully. Extensive experiments illustrate SID generates less-hallucination and higher-quality texts across various metrics, without extra knowledge and much additional computation burdens.

Feng Xie, Zhong Zhang, Liang Li et al.

Epidemic forecasting is the key to effective control of epidemic transmission and helps the world mitigate the crisis that threatens public health. To better understand the transmission and evolution of epidemics, we propose EpiGNN, a graph neural network-based model for epidemic forecasting. Specifically, we design a transmission risk encoding module to characterize local and global spatial effects of regions in epidemic processes and incorporate them into the model. Meanwhile, we develop a Region-Aware Graph Learner (RAGL) that takes transmission risk, geographical dependencies, and temporal information into account to better explore spatial-temporal dependencies and makes regions aware of related regions' epidemic situations. The RAGL can also combine with external resources, such as human mobility, to further improve prediction performance. Comprehensive experiments on five real-world epidemic-related datasets (including influenza and COVID-19) demonstrate the effectiveness of our proposed method and show that EpiGNN outperforms state-of-the-art baselines by 9.48% in RMSE.

Minghang Deng, Zhong Zhang, Junming Shao · tsinghua

Large-scale pre-trained models (PTMs) such as BERT and GPT have achieved great success in diverse fields. The typical paradigm is to pre-train a big deep learning model on large-scale data sets, and then fine-tune the model on small task-specific data sets for downstream tasks. Although PTMs have rapidly progressed with wide real-world applications, they also pose significant risks of potential attacks. Existing backdoor attacks or data poisoning methods often build up the assumption that the attacker invades the computers of victims or accesses the target data, which is challenging in real-world scenarios. In this paper, we propose a novel framework for an invisible attack on PTMs with enhanced MD5 collision. The key idea is to generate two equal-size models with the same MD5 checksum by leveraging the MD5 chosen-prefix collision. Afterwards, the two ``same" models will be deployed on public websites to induce victims to download the poisoned model. Unlike conventional attacks on deep learning models, this new attack is flexible, covert, and model-independent. Additionally, we propose a simple defensive strategy for recognizing the MD5 chosen-prefix collision and provide a theoretical justification for its feasibility. We extensively validate the effectiveness and stealthiness of our proposed attack and defensive method on different models and data sets.

Wentong Chen, Xin Cong, Zhong Zhang et al.

With the continuous advancement of Large Language Models (LLMs), intelligent agents are becoming increasingly vital. However, these agents often fail in environments governed by implicit rules--hidden constraints that cannot be observed directly and must be inferred through interaction. This causes agents to fall into repetitive trial-and-error loops, ultimately leading to task failure. To address this challenge, we propose Test-Time Exploration (TTExplore), a framework where a thinker component analyzes interaction history to infer these implicit rules and guide an actor. Effective exploration in this setting critically depends on the reasoning ability of the thinker. However, evaluating deep reasoning trajectories is inherently unstable and difficult, which poses a major obstacle to effective training. To overcome this issue, we introduce a novel and stable reinforcement learning pipeline. The core idea is to use accurate task-level scores as indirect rewards to bypass the difficulty of evaluating intermediate reasoning, and to retain only a single thinking node per trajectory to alleviate reward sparsity. Using this pipeline, we train a specialized 7B model, Exp-Thinker. Experiments on five text-based embodied tasks show that TTExplore equipped with Exp-Thinker improves baseline agent performance by an average of $14$-$19$ points, demonstrating the effectiveness of explicitly reasoning about implicit rules.

Feng Xie, Zhong Zhang, Xuechen Zhao et al.

The ongoing COVID-19 pandemic has caused immeasurable losses for people worldwide. To contain the spread of the virus and further alleviate the crisis, various health policies (e.g., stay-at-home orders) have been issued which spark heated discussions as users turn to share their attitudes on social media. In this paper, we consider a more realistic scenario on stance detection (i.e., cross-target and zero-shot settings) for the pandemic and propose an adversarial learning-based stance classifier to automatically identify the public's attitudes toward COVID-19-related health policies. Specifically, we adopt adversarial learning that allows the model to train on a large amount of labeled data and capture transferable knowledge from source topics, so as to enable generalize to the emerging health policies with sparse labeled data. To further enhance the model's deeper understanding, we incorporate policy descriptions as external knowledge into the model. Meanwhile, a GeoEncoder is designed which encourages the model to capture unobserved background factors specified by each region and then represent them as non-text information. We evaluate the performance of a broad range of baselines on the stance detection task for COVID-19-related health policies, and experimental results show that our proposed method achieves state-of-the-art performance in both cross-target and zero-shot settings.

Qinyu Luo, Yining Ye, Shihao Liang et al. · tencent-ai, tsinghua

Generative models have demonstrated considerable potential in software engineering, particularly in tasks such as code generation and debugging. However, their utilization in the domain of code documentation generation remains underexplored. To this end, we introduce RepoAgent, a large language model powered open-source framework aimed at proactively generating, maintaining, and updating code documentation. Through both qualitative and quantitative evaluations, we have validated the effectiveness of our approach, showing that RepoAgent excels in generating high-quality repository-level documentation. The code and results are publicly accessible at https://github.com/OpenBMB/RepoAgent.

Giacomo Acciarini, Dario Izzo, Zhong Zhang

Reachability analysis plays a central role in low-thrust spacecraft trajectory optimization by identifying which target states can be achieved under constraints on time, thrust, and propellant. Classical approaches construct reachable sets by solving many optimal control problems over grids of terminal states, requiring extensive forward simulations with fixed initial conditions. While effective, this approach is computationally expensive and becomes impractical for high-dimensional systems or strongly nonlinear dynamics, such as those encountered in cislunar environments or solar sail missions. This work introduces a dual formulation of the reachability problem. Instead of computing reachable sets directly, we determine, for fixed transfer time and boundary conditions, the maximum allowable initial mass (or, for solar sails, a scalar sail-strength parameter) that permits a successful transfer. A target is reachable if the spacecraft's initial mass does not exceed this threshold. This reformulation reduces reachability assessment to a scalar optimization problem for each target, producing a smooth scalar field that encodes equivalent feasibility information to classical reachable sets. We develop indirect maximum-initial-mass (MIM) formulations for both electric low-thrust and solar-sail dynamics and show how they can serve as efficient reachability oracles. Building on this formulation, we construct data-driven surrogate models to approximate the MIM-based reachability indicator. We investigate fully connected neural networks and demonstrate that residual networks provide the best trade-off between accuracy, training stability, and model complexity. The resulting surrogates enable rapid reachability evaluation while preserving the numerical advantages of the dual formulation, offering a practical tool for preliminary mission design and feasibility assessment.

Yilin Zou, Zhong Zhang, Maxime Robic et al.

Real-time trajectory optimization for nonlinear constrained autonomous systems is critical and typically performed by CPU-based sequential solvers. Specifically, reliance on global sparse linear algebra or the serial nature of dynamic programming algorithms restricts the utilization of massively parallel computing architectures like GPUs. To bridge this gap, we introduce a fully GPU-native trajectory optimization framework that combines sequential convex programming with a consensus-based alternating direction method of multipliers. By applying a temporal splitting strategy, our algorithm decouples the optimization horizon into independent, per-node subproblems that execute massively in parallel. The entire process runs fully on the GPU, eliminating costly memory transfers and large-scale sparse factorizations. This architecture naturally scales to multi-trajectory optimization. We validate the solver on a quadrotor agile flight task and a Mars powered descent problem using an on-board edge computing platform. Benchmarks reveal a sustained 4x throughput speedup and a 51% reduction in energy consumption over a heavily optimized 12-core CPU baseline. Crucially, the framework saturates the hardware, maintaining over 96% active GPU utilization to achieve planning rates exceeding 100 Hz. Furthermore, we demonstrate the solver's extensibility to robust Model Predictive Control by jointly optimizing dynamically coupled scenarios under stochastic disturbances, enabling scalable and safe autonomy.

Xuechen Zhao, Jiaying Zou, Zhong Zhang et al.

Zero-shot stance detection is challenging because it requires detecting the stance of previously unseen targets in the inference phase. The ability to learn transferable target-invariant features is critical for zero-shot stance detection. In this work, we propose a stance detection approach that can efficiently adapt to unseen targets, the core of which is to capture target-invariant syntactic expression patterns as transferable knowledge. Specifically, we first augment the data by masking the topic words of sentences, and then feed the augmented data to an unsupervised contrastive learning module to capture transferable features. Then, to fit a specific target, we encode the raw texts as target-specific features. Finally, we adopt an attention mechanism, which combines syntactic expression patterns with target-specific features to obtain enhanced features for predicting previously unseen targets. Experiments demonstrate that our model outperforms competitive baselines on four benchmark datasets.

Feng Xie, Zhong Zhang, Xuechen Zhao et al.

The accurate forecasting of infectious epidemic diseases is the key to effective control of the epidemic situation in a region. Most existing methods ignore potential dynamic dependencies between regions or the importance of temporal dependencies and inter-dependencies between regions for prediction. In this paper, we propose an Inter- and Intra-Series Embeddings Fusion Network (SEFNet) to improve epidemic prediction performance. SEFNet consists of two parallel modules, named Inter-Series Embedding Module and Intra-Series Embedding Module. In Inter-Series Embedding Module, a multi-scale unified convolution component called Region-Aware Convolution is proposed, which cooperates with self-attention to capture dynamic dependencies between time series obtained from multiple regions. The Intra-Series Embedding Module uses Long Short-Term Memory to capture temporal relationships within each time series. Subsequently, we learn the influence degree of two embeddings and fuse them with the parametric-matrix fusion method. To further improve the robustness, SEFNet also integrates a traditional autoregressive component in parallel with nonlinear neural networks. Experiments on four real-world epidemic-related datasets show SEFNet is effective and outperforms state-of-the-art baselines.

Yaxi Lu, Shenzhi Yang, Cheng Qian et al. · tsinghua

Agents powered by large language models have shown remarkable abilities in solving complex tasks. However, most agent systems remain reactive, limiting their effectiveness in scenarios requiring foresight and autonomous decision-making. In this paper, we tackle the challenge of developing proactive agents capable of anticipating and initiating tasks without explicit human instructions. We propose a novel data-driven approach for this problem. Firstly, we collect real-world human activities to generate proactive task predictions. These predictions are then labeled by human annotators as either accepted or rejected. The labeled data is used to train a reward model that simulates human judgment and serves as an automatic evaluator of the proactiveness of LLM agents. Building on this, we develop a comprehensive data generation pipeline to create a diverse dataset, ProactiveBench, containing 6,790 events. Finally, we demonstrate that fine-tuning models with the proposed ProactiveBench can significantly elicit the proactiveness of LLM agents. Experimental results show that our fine-tuned model achieves an F1-Score of 66.47% in proactively offering assistance, outperforming all open-source and close-source models. These results highlight the potential of our method in creating more proactive and effective agent systems, paving the way for future advancements in human-agent collaboration.

Shengda Fan, Xin Cong, Yuepeng Fu et al.

Recent advancements in large language models (LLMs) have driven a revolutionary paradigm shift in process automation from Robotic Process Automation to Agentic Process Automation by automating the workflow orchestration procedure based on LLMs. However, existing LLMs (even the advanced OpenAI GPT-4o) are confined to achieving satisfactory capability in workflow orchestration. To address this limitation, we present WorkflowLLM, a data-centric framework elaborately designed to enhance the capability of LLMs in workflow orchestration. It first constructs a large-scale fine-tuning dataset WorkflowBench with 106,763 samples, covering 1,503 APIs from 83 applications across 28 categories. Specifically, the construction process can be divided into three phases: (1) Data Collection: we collect real-world workflow data from Apple Shortcuts and RoutineHub, transcribing them into Python-style code. We further equip them with generated hierarchical thought via ChatGPT. (2) Query Expansion: we prompt ChatGPT to generate more task queries to enrich the diversity and complexity of workflows. (3) Workflow Generation: we leverage an annotator model trained on collected data to generate workflows for synthesized queries. Finally, we merge the synthetic samples that pass quality confirmation with the collected samples to obtain the WorkflowBench. Based on WorkflowBench, we fine-tune Llama-3.1-8B to obtain WorkflowLlama. Our experiments show that WorkflowLlama demonstrates a strong capacity to orchestrate complex workflows, while also achieving notable generalization performance on previously unseen APIs. Additionally, WorkflowBench exhibits robust zero-shot generalization capabilities on an out-of-distribution task planning dataset, T-Eval. Our data and code are available at https://github.com/OpenBMB/WorkflowLLM.

Yishan Li, Wentong Chen, Yukun Yan et al.

Generating deep research reports requires large-scale information acquisition and the synthesis of insight-driven analysis, posing a significant challenge for current language models. Most existing approaches follow a plan-then-write paradigm, whose performance heavily depends on the quality of the initial outline. However, constructing a comprehensive outline itself demands strong reasoning ability, causing current deep research systems to rely almost exclusively on closed-source or online large models. This reliance raises practical barriers to deployment and introduces safety and privacy concerns for user-authored data. In this work, we present AgentCPM-Report, a lightweight yet high-performing local solution composed of a framework that mirrors the human writing process and an 8B-parameter deep research agent. Our framework uses a Writing As Reasoning Policy (WARP), which enables models to dynamically revise outlines during report generation. Under this policy, the agent alternates between Evidence-Based Drafting and Reasoning-Driven Deepening, jointly supporting information acquisition, knowledge refinement, and iterative outline evolution. To effectively equip small models with this capability, we introduce a Multi-Stage Agentic Training strategy, consisting of cold-start, atomic skill RL, and holistic pipeline RL. Experiments on DeepResearch Bench, DeepConsult, and DeepResearch Gym demonstrate that AgentCPM-Report outperforms leading closed-source systems, with substantial gains in Insight.

Haotian Chen, Xin Cong, Shengda Fan et al.

While Large Language Model (LLM)-based agents have shown remarkable potential for solving complex tasks, existing systems remain heavily reliant on large-scale models, leaving the capabilities of edge-scale models largely underexplored. In this paper, we present the first systematic study on training agentic models at the 4B-parameter scale. We identify three primary bottlenecks hindering the performance of edge-scale models: catastrophic forgetting during Supervised Fine-Tuning (SFT), sensitivity to reward signal noise during Reinforcement Learning (RL), and reasoning degradation caused by redundant information in long-context scenarios. To address the issues, we propose AgentCPM-Explore, a compact 4B agent model with high knowledge density and strong exploration capability. We introduce a holistic training framework featuring parameter-space model fusion, reward signal denoising, and contextual information refinement. Through deep exploration, AgentCPM-Explore achieves state-of-the-art (SOTA) performance among 4B-class models, matches or surpasses 8B-class SOTA models on four benchmarks, and even outperforms larger-scale models such as Claude-4.5-Sonnet or DeepSeek-v3.2 in five benchmarks. Notably, AgentCPM-Explore achieves 97.09% accuracy on GAIA text-based tasks under pass@64. These results provide compelling evidence that the bottleneck for edge-scale models is not their inherent capability ceiling, but rather their inference stability. Based on our well-established training framework, AgentCPM-Explore effectively unlocks the significant, yet previously underestimated, potential of edge-scale models.

Yupeng Huo, Yaxi Lu, Zhong Zhang et al.

Equipping agents with memory is essential for solving real-world long-horizon problems. However, most existing agent memory mechanisms rely on static and hand-crafted workflows. This limits the performance and generalization ability of these memory designs, which highlights the need for a more flexible, learning-based memory framework. In this paper, we propose AtomMem, which reframes memory management as a dynamic decision-making problem. We deconstruct high-level memory processes into fundamental atomic CRUD (Create, Read, Update, Delete) operations, transforming the memory workflow into a learnable decision process. By combining supervised fine-tuning with reinforcement learning, AtomMem learns an autonomous, task-aligned policy to orchestrate memory behaviors tailored to specific task demands. Experimental results across 3 long-context benchmarks demonstrate that the trained AtomMem-8B consistently outperforms prior static-workflow memory methods. Further analysis of training dynamics shows that our learning-based formulation enables the agent to discover structured, task-aligned memory management strategies, highlighting a key advantage over predefined routines.

Runchu Tian, Yanghao Li, Yuepeng Fu et al. · tsinghua

Positional bias in large language models (LLMs) hinders their ability to effectively process long inputs. A prominent example is the "lost in the middle" phenomenon, where LLMs struggle to utilize relevant information situated in the middle of the input. While prior research primarily focuses on single pieces of relevant information, real-world applications often involve multiple relevant information pieces. To bridge this gap, we present LongPiBench, a benchmark designed to assess positional bias involving multiple pieces of relevant information. Thorough experiments are conducted with five commercial and six open-source models. These experiments reveal that while most current models are robust against the "lost in the middle" issue, there exist significant biases related to the spacing of relevant information pieces. These findings highlight the importance of evaluating and reducing positional biases to advance LLM's capabilities.

Bohan Lyu, Xin Cong, Heyang Yu et al. · tencent-ai, tsinghua

Large Language Models (LLMs) excel in traditional natural language processing tasks but struggle with problems that require complex domain-specific calculations or simulations. While equipping LLMs with external tools to build LLM-based agents can enhance their capabilities, existing approaches lack the flexibility to address diverse and ever-evolving user queries in open domains. Currently, there is also no existing dataset that evaluates LLMs on open-domain knowledge that requires tools to solve. To this end, we introduce OpenAct benchmark to evaluate the open-domain task-solving capability, which is built on human expert consultation and repositories in GitHub. It comprises 339 questions spanning 7 diverse domains that need to be solved with domain-specific methods. In our experiments, even state-of-the-art LLMs and LLM-based agents demonstrate unsatisfactory success rates, underscoring the need for a novel approach. Furthermore, we present OpenAgent, a novel LLM-based agent system that can tackle evolving queries in open domains through autonomously integrating specialized tools from GitHub. OpenAgent employs 1) a hierarchical framework where specialized agents handle specific tasks and can assign tasks to inferior agents, 2) a bi-level experience learning mechanism to learn from both humans' and its own experiences to tackle tool flaws. Experiments demonstrate its superior effectiveness and efficiency, which significantly outperforms baselines. Our data and code are open-source at https://github.com/OpenBMB/OpenAct.

Rui Zhao, Xihui Li, Yizheng Zhang et al.

Multi-agent deep Reinforcement Learning (RL) has made significant progress in developing intelligent game-playing agents in recent years. However, the efficient training of collective robots using multi-agent RL and the transfer of learned policies to real-world applications remain open research questions. In this work, we first develop a comprehensive robotic system, including simulation, distributed learning framework, and physical robot components. We then propose and evaluate reinforcement learning techniques designed for efficient training of cooperative and competitive policies on this platform. To address the challenges of multi-agent sim-to-real transfer, we introduce Out of Distribution State Initialization (OODSI) to mitigate the impact of the sim-to-real gap. In the experiments, OODSI improves the Sim2Real performance by 20%. We demonstrate the effectiveness of our approach through experiments with a multi-robot car competitive game and a cooperative task in real-world settings.

Cheng Qian, Bingxiang He, Zhong Zhuang et al. · tencent-ai, tsinghua

Current language model-driven agents often lack mechanisms for effective user participation, which is crucial given the vagueness commonly found in user instructions. Although adept at devising strategies and performing tasks, these agents struggle with seeking clarification and grasping precise user intentions. To bridge this gap, we introduce Intention-in-Interaction (IN3), a novel benchmark designed to inspect users' implicit intentions through explicit queries. Next, we propose the incorporation of model experts as the upstream in agent designs to enhance user-agent interaction. Employing IN3, we empirically train Mistral-Interact, a powerful model that proactively assesses task vagueness, inquires user intentions, and refines them into actionable goals before starting downstream agent task execution. Integrating it into the XAgent framework, we comprehensively evaluate the enhanced agent system regarding user instruction understanding and execution, revealing that our approach notably excels at identifying vague user tasks, recovering and summarizing critical missing information, setting precise and necessary agent execution goals, and minimizing redundant tool usage, thus boosting overall efficiency. All the data and codes are released.

Haoyu Wang, Guozheng Ma, Ziqiao Meng et al.

Self-alignment is an effective way to reduce the cost of human annotation while ensuring promising model capability. However, most current methods complete the data collection and training steps in a single round, which may overlook the continuously improving ability of self-aligned models. This gives rise to a key query: What if we do multi-time bootstrapping self-alignment? Does this strategy enhance model performance or lead to rapid degradation? In this paper, our pioneering exploration delves into the impact of bootstrapping self-alignment on large language models. Our findings reveal that bootstrapping self-alignment markedly surpasses the single-round approach, by guaranteeing data diversity from in-context learning. To further exploit the capabilities of bootstrapping, we investigate and adjust the training order of data, which yields improved performance of the model. Drawing on these findings, we propose Step-On-Feet Tuning (SOFT) which leverages model's continuously enhanced few-shot ability to boost zero or one-shot performance. Based on easy-to-hard training recipe, we propose SOFT+ which further boost self-alignment's performance. Our experiments demonstrate the efficiency of SOFT (SOFT+) across various classification and generation tasks, highlighting the potential of bootstrapping self-alignment on continually enhancing model alignment performance.

Zhong Zhang, Yaxi Lu, Yikun Fu et al. · tsinghua

The recent progress of large language model agents has opened new possibilities for automating tasks through graphical user interfaces (GUIs), especially in mobile environments where intelligent interaction can greatly enhance usability. However, practical deployment of such agents remains constrained by several key challenges. Existing training data is often noisy and lack semantic diversity, which hinders the learning of precise grounding and planning. Models trained purely by imitation tend to overfit to seen interface patterns and fail to generalize in unfamiliar scenarios. Moreover, most prior work focuses on English interfaces while overlooks the growing diversity of non-English applications such as those in the Chinese mobile ecosystem. In this work, we present AgentCPM-GUI, an 8B-parameter GUI agent built for robust and efficient on-device GUI interaction. Our training pipeline includes grounding-aware pre-training to enhance perception, supervised fine-tuning on high-quality Chinese and English trajectories to imitate human-like actions, and reinforcement fine-tuning with GRPO to improve reasoning capability. We also introduce a compact action space that reduces output length and supports low-latency execution on mobile devices. AgentCPM-GUI achieves state-of-the-art performance on five public benchmarks and a new Chinese GUI benchmark called CAGUI, reaching $96.9\%$ Type-Match and $91.3\%$ Exact-Match. To facilitate reproducibility and further research, we publicly release all code, model checkpoint, and evaluation data.

Yu Xia, Jingru Fan, Weize Chen et al. · tsinghua

Existing LLM-based agents have achieved strong performance on held-in tasks, but their generalizability to unseen tasks remains poor. Hence, some recent work focus on fine-tuning the policy model with more diverse tasks to improve the generalizability. In this work, we find that finetuning a reward model to guide the policy model is more robust than directly finetuning the policy model. Based on this finding, we propose AgentRM, a generalizable reward model, to guide the policy model for effective test-time search. We comprehensively investigate three approaches to construct the reward model, including explicit reward modeling, implicit reward modeling and LLM-as-a-judge. We then use AgentRM to guide the answer generation with Best-of-N sampling and step-level beam search. On four types of nine agent tasks, AgentRM enhances the base policy model by $8.8$ points on average, surpassing the top general agent by $4.0$. Moreover, it demonstrates weak-to-strong generalization, yielding greater improvement of $12.6$ on LLaMA-3-70B policy model. As for the specializability, AgentRM can also boost a finetuned policy model and outperform the top specialized agent by $11.4$ on three held-in tasks. Further analysis verifies its effectiveness in test-time scaling. Codes will be released to facilitate the research in this area.

Yaxi Lu, Haolun Li, Xin Cong et al. · tsinghua

This study investigates the structured generation capabilities of large language models (LLMs), focusing on producing valid JSON outputs against a given schema. Despite the widespread use of JSON in integrating language models with programs, there is a lack of comprehensive analysis and benchmarking of these capabilities. We explore various aspects of JSON generation, such as structure understanding, escaping, and natural language description, to determine how to assess and enable LLMs to generate valid responses. Building upon this, we propose SchemaBench features around 40K different JSON schemas to obtain and assess models' abilities in generating valid JSON. We find that the latest LLMs are still struggling to generate a valid JSON string. Moreover, we demonstrate that incorporating reinforcement learning with a Fine-grained Schema Validator can further enhance models' understanding of JSON schema, leading to improved performance. Our models demonstrate significant improvement in both generating JSON outputs and downstream tasks.

Mengyang Li, Zhong Zhang

Aligning large language models with human preferences is crucial for their safe deployment. While Direct Preference Optimization (DPO) offers an efficient alternative to reinforcement learning from human feedback, traditional DPO methods are limited by their reliance on single preference pairs. Recent work like Curriculum-DPO integrates multiple pairs using a one-dimensional difficulty curriculum based on pairwise distinguishability (PD), but overlooks the complexity of the input prompt itself. To address this, we propose 2D-Curri-DPO, a novel framework employing a two-dimensional curriculum that jointly models Prompt Complexity (PC) and Pairwise Distinguishability. This framework introduces dual difficulty metrics to quantify prompt semantic complexity and response preference clarity, defines a curriculum strategy space encompassing multiple selectable strategies for task adaptation, and incorporates a KL-divergence-based adaptive mechanism for dynamic reference model updates to enhance training stability. Comprehensive experiments demonstrate that 2D-Curri-DPO significantly outperforms standard DPO and prior curriculum methods across multiple benchmarks, including MT-Bench, Vicuna Bench, and WizardLM. Our approach achieves state-of-the-art performance on challenging test sets like UltraFeedback. Ablation studies confirm the benefits of the 2D structure and adaptive mechanisms, while analysis provides guidance for strategy selection. These findings demonstrate that effective alignment requires modeling both prompt complexity and pairwise distinguishability, establishing adaptive, multi-dimensional curriculum learning as a powerful and interpretable new paradigm for preference-based language model optimization.

Zhicheng Chen, Shibo Feng, Xi Xiao et al.

Discrete Token Modeling (DTM), which employs vector quantization techniques, has demonstrated remarkable success in modeling non-natural language modalities, particularly in time series generation. While our prior work SDformer established the first DTM-based framework to achieve state-of-the-art performance in this domain, two critical limitations persist in existing DTM approaches: 1) their inability to capture multi-scale temporal patterns inherent to complex time series data, and 2) the absence of theoretical foundations to guide model optimization. To address these challenges, we proposes a novel multi-scale DTM-based time series generation method, called Multi-Scale Discrete Transformer (MSDformer). MSDformer employs a multi-scale time series tokenizer to learn discrete token representations at multiple scales, which jointly characterize the complex nature of time series data. Subsequently, MSDformer applies a multi-scale autoregressive token modeling technique to capture the multi-scale patterns of time series within the discrete latent space. Theoretically, we validate the effectiveness of the DTM method and the rationality of MSDformer through the rate-distortion theorem. Comprehensive experiments demonstrate that MSDformer significantly outperforms state-of-the-art methods. Both theoretical analysis and experimental results demonstrate that incorporating multi-scale information and modeling multi-scale patterns can substantially enhance the quality of generated time series in DTM-based approaches. The code will be released upon acceptance.

Haotian Chen, Zijun Song, Boye Niu et al. · tsinghua

Tool learning, which enables large language models (LLMs) to utilize external tools effectively, has garnered increasing attention for its potential to revolutionize productivity across industries. Despite rapid development in tool learning, key challenges and opportunities remain understudied, limiting deeper insights and future advancements. In this paper, we investigate the tool learning ability of 41 prevalent LLMs by reproducing 33 benchmarks and enabling one-click evaluation for seven of them, forming a Tool Learning Platform named ToLeaP. We also collect 21 out of 33 potential training datasets to facilitate future exploration. After analyzing over 3,000 bad cases of 41 LLMs based on ToLeaP, we identify four main critical challenges: (1) benchmark limitations induce both the neglect and lack of (2) autonomous learning, (3) generalization, and (4) long-horizon task-solving capabilities of LLMs. To aid future advancements, we take a step further toward exploring potential directions, namely (1) real-world benchmark construction, (2) compatibility-aware autonomous learning, (3) rationale learning by thinking, and (4) identifying and recalling key clues. The preliminary experiments demonstrate their effectiveness, highlighting the need for further research and exploration.

Yanxi Zhang, Xin Cong, Zhong Zhang et al. · pku, tsinghua

Genuine human-like causal reasoning is fundamental for strong artificial intelligence. Humans typically identify whether an event is part of the causal chain first, and then influenced by modulatory factors such as morality, normality, and intention to make the final judgment. These two stages naturally map to the fields of 1) actual causality that provides formalisms for causal chain membership and 2) causal judgment from cognitive science that studies psychological modulators that influence causal selection. However, these two domains have largely been studied in isolation, leaving a gap for a systematic method based on LLMs. Therefore, we introduce HCR-Reasoner, a framework that systematically integrates the theory of actual causality and causal judgment into LLMs for human-like causal reasoning. It simulates humans by using actual causality formalisms to filter for structurally necessary candidate causes and causal judgment factors to determine the psychologically selected cause. For fine-grained evaluation, we introduce HCR-Bench, a challenging benchmark with 1,093 annotated instances with detailed reasoning steps. Results show HCR-Reasoner consistently and significantly improves LLMs' causal alignment with humans, and that explicitly integrating theory-guided reasoning into LLMs is highly effective for achieving faithful human-like causal reasoning.

Hongyuan Liu, Qinli Yang, Wen Li et al.

Vision-Language Models (VLMs) such as CLIP learn a shared embedding space for images and text, yet their representations remain geometrically separated, a phenomenon known as the modality gap. This gap limits tasks requiring cross-modal interchangeability, such as captioning and joint clustering. Existing post-processing approaches can partially improve cross-modal compatibility; however, we show through geometric analysis that they primarily reduce the global centroid offset while leaving the underlying distributional mismatch intact. We decompose the modality gap into a Centroid Gap and a Distribution Gap, and demonstrate that the Distribution Gap is the true predictor of cross-modal task quality ($R^2 = 0.986$), whereas the commonly used Raw Gap is misleading ($R^2 = 0.691$). Motivated by this observation, we propose TPC-CMA (Three-Phase Curriculum for Cross-Modal Alignment), a fine-tuning framework that explicitly reduces both components. The proposed CMA jointly mitigates centroid offsets and reshapes the distributional structure, while a three-phase curriculum with gradient-aware scheduling progressively introduces alignment during training to enable stable optimization. Experiments demonstrate that our method significantly improves cross-modal alignment. With $α_{\text{target}}{=}0.05$, the modality gap is reduced by 66.6\% with only 4.84\% accuracy drop. Under stronger alignment ($α_{\text{target}}{=}0.5$), the gap is reduced by 82.3\%, clustering ARI improves from 0.318 to 0.516, and captioning CIDEr increases by 57.1\% over the original model. Our code and pre-trained models will be made publicly available upon acceptance.

Xiaohan Yi, Guikun Xu, Xi Xiao et al.

We present CrystalDiT, a diffusion transformer for crystal structure generation that achieves state-of-the-art performance by challenging the trend of architectural complexity. Instead of intricate, multi-stream designs, CrystalDiT employs a unified transformer that imposes a powerful inductive bias: treating lattice and atomic properties as a single, interdependent system. Combined with a periodic table-based atomic representation and a balanced training strategy, our approach achieves 9.62% SUN (Stable, Unique, Novel) rate on MP-20, substantially outperforming recent methods including FlowMM (4.38%) and MatterGen (3.42%). Notably, CrystalDiT generates 63.28% unique and novel structures while maintaining comparable stability rates, demonstrating that architectural simplicity can be more effective than complexity for materials discovery. Our results suggest that in data-limited scientific domains, carefully designed simple architectures outperform sophisticated alternatives that are prone to overfitting.

Zhong Zhang, Francesco Topputo

In trajectory design, fuel consumption and trajectory reachability are two key performance indicators for low-thrust missions. This paper proposes general-purpose pretrained neural networks to predict these metrics. The contributions of this paper are as follows: Firstly, based on the confirmation of the Scaling Law applicable to low-thrust trajectory approximation, the largest dataset is constructed using the proposed homotopy ray method, which aligns with mission-design-oriented data requirements. Secondly, the data are transformed into a self-similar space, enabling the neural network to adapt to arbitrary semi-major axes, inclinations, and central bodies. This extends the applicability beyond existing studies and can generalize across diverse mission scenarios without retraining. Thirdly, to the best of our knowledge, this work presents the current most general and accurate low-thrust trajectory approximator, with implementations available in C++, Python, and MATLAB. The resulting neural network achieves a relative error of 0.78% in predicting velocity increments and 0.63% in minimum transfer time estimation. The models have also been validated on a third-party dataset, multi-flyby mission design problem, and mission analysis scenario, demonstrating their generalization capability, predictive accuracy, and computational efficiency.

Song Wang, Zhong Zhang, Huan Yan et al.

H&E-to-IHC stain translation techniques offer a promising solution for precise cancer diagnosis, especially in low-resource regions where there is a shortage of health professionals and limited access to expensive equipment. Considering the pixel-level misalignment of H&E-IHC image pairs, current research explores the pathological consistency between patches from the same positions of the image pair. However, most of them overemphasize the correspondence between domains or patches, overlooking the side information provided by the non-corresponding objects. In this paper, we propose a Mix-Domain Contrastive Learning (MDCL) method to leverage the supervision information in unpaired H&E-to-IHC stain translation. Specifically, the proposed MDCL method aggregates the inter-domain and intra-domain pathology information by estimating the correlation between the anchor patch and all the patches from the matching images, encouraging the network to learn additional contrastive knowledge from mixed domains. With the mix-domain pathology information aggregation, MDCL enhances the pathological consistency between the corresponding patches and the component discrepancy of the patches from the different positions of the generated IHC image. Extensive experiments on two H&E-to-IHC stain translation datasets, namely MIST and BCI, demonstrate that the proposed method achieves state-of-the-art performance across multiple metrics.

Zhong Zhang, Bang Liu, Junming Shao

Pre-trained language models (PLMs) are known to be overly parameterized and have significant redundancy, indicating a small degree of freedom of the PLMs. Motivated by the observation, in this paper, we study the problem of re-parameterizing and fine-tuning PLMs from a new perspective: Discovery of intrinsic task-specific subspace. Specifically, by exploiting the dynamics of the fine-tuning process for a given task, the parameter optimization trajectory is learned to uncover its intrinsic task-specific subspace. A key finding is that PLMs can be effectively fine-tuned in the subspace with a small number of free parameters. Beyond, we observe some outlier dimensions emerging during fine-tuning in the subspace. Disabling these dimensions degrades the model performance significantly. This suggests that these dimensions are crucial to induce task-specific knowledge to downstream tasks.

Xiaoyan Cao, Yao Yao, Lanqing Li et al.

Agriculture is the foundation of human civilization. However, the rapid increase of the global population poses a challenge on this cornerstone by demanding more food. Modern autonomous greenhouses, equipped with sensors and actuators, provide a promising solution to the problem by empowering precise control for high-efficient food production. However, the optimal control of autonomous greenhouses is challenging, requiring decision-making based on high-dimensional sensory data, and the scaling of production is limited by the scarcity of labor capable of handling this task. With the advances of artificial intelligence (AI), the internet of things (IoT), and cloud computing technologies, we are hopeful to provide a solution to automate and smarten greenhouse control to address the above challenges. In this paper, we propose a smart agriculture solution named iGrow, for autonomous greenhouse control (AGC): (1) for the first time, we formulate the AGC problem as a Markov decision process (MDP) optimization problem; (2) we design a neural network-based simulator incorporated with the incremental mechanism to simulate the complete planting process of an autonomous greenhouse, which provides a testbed for the optimization of control strategies; (3) we propose a closed-loop bi-level optimization algorithm, which can dynamically re-optimize the greenhouse control strategy with newly observed data during real-world production. We not only conduct simulation experiments but also deploy iGrow in real scenarios, and experimental results demonstrate the effectiveness and superiority of iGrow in autonomous greenhouse simulation and optimal control. Particularly, compelling results from the tomato pilot project in real autonomous greenhouses show that our solution significantly increases crop yield (+10.15\%) and net profit (+92.70\%) with statistical significance compared to planting experts.

Zhong Zhang, Tao Han, Jia Pan et al.

To achieve human-like dexterity for anthropomorphic robotic hands, it is essential to understand the biomechanics and control strategies of the human hand, in order to reduce the number of actuators being used without loosing hand flexibility. To this end, in this article, we propose a new interpretation about the working mechanism of the metacarpal (MCP) joint's extension and the underlying control strategies of the human hand, based on which we further propose a highly flexible finger design to achieve independent movements of interphalangeal (IP) joints and MCP joint. Besides, we consider the hyperextension of fingertip into our design which helps robotic finger present compliant and adaptive posture for touching and pinching. In addition, human thumb muscle functions are reconstructed in the proposed robotic hand design, by replacing 9 human muscle tendons with 3 cables in the proposed task-oriented design, realizing all 33 static and stable grasping postures. Videos are available at https://sites.google.com/view/szwd