Zhengliang Shi

Shen Gao, Zhengliang Shi, Minghang Zhu et al. · pku

Augmenting large language models (LLMs) with external tools has emerged as a promising approach to extending the capability of LLMs. Although some works employ open-source LLMs for the tool learning task, most of them are trained in a controlled environment in which LLMs only learn to execute the human-provided tools. However, selecting proper tools from the large toolset is also a crucial ability for the tool learning model to be applied in real-world applications. Existing methods usually directly employ self-instruction methods to train the model, which ignores differences in tool complexity. In this paper, we propose the Confucius, a novel tool learning framework to train LLM to use complicated tools in real-world scenarios, which contains two main phases: (1) We first propose a multi-stage learning method to teach the LLM to use various tools from an easy-to-difficult curriculum; (2) thenceforth, we propose the Iterative Self-instruct from Introspective Feedback (ISIF) to dynamically construct the dataset to improve the ability to use the complicated tool. Extensive experiments conducted on both controlled and real-world settings demonstrate the superiority of our tool learning framework in the real-world application scenarios compared to both tuning-free (e.g. ChatGPT, Claude) and tuning-based baselines (e.g. GPT4Tools).

Weiwei Sun, Zhengliang Shi, Shen Gao et al. · pku

Pre-trained language models (LMs) store knowledge in their parameters and can generate informative responses when used in conversational systems. However, LMs suffer from the problem of "hallucination:" they may generate plausible-looking statements that are irrelevant or factually incorrect. To address this problem, we propose a contrastive learning scheme, named MixCL. A novel mixed contrastive objective is proposed to explicitly optimize the implicit knowledge elicitation process of LMs, and thus reduce their hallucination in conversations. We also examine negative sampling strategies of retrieved hard negatives and model-generated negatives. We conduct experiments on Wizard-of-Wikipedia, a public, open-domain knowledge-grounded dialogue benchmark, and assess the effectiveness of MixCL. MixCL effectively reduces the hallucination of LMs in conversations and achieves the highest performance among LM-based dialogue agents in terms of relevancy and factuality. We show that MixCL achieves comparable performance to state-of-the-art KB-based approaches while enjoying notable advantages in terms of efficiency and scalability.

Zhengliang Shi, Weiwei Sun, Shuo Zhang et al.

Evaluating open-domain dialogue systems is challenging for reasons such as the one-to-many problem, i.e., many appropriate responses other than just the golden response. As of now, automatic evaluation methods need better consistency with humans, while reliable human evaluation can be time- and cost-intensive. To this end, we propose the Reference-Assisted Dialogue Evaluation (RADE) approach under the multi-task learning framework, which leverages the pre-created utterance as reference other than the gold response to relief the one-to-many problem. Specifically, RADE explicitly compares reference and the candidate response to predict their overall scores. Moreover, an auxiliary response generation task enhances prediction via a shared encoder. To support RADE, we extend three datasets with additional rated responses other than just a golden response by human annotation. Experiments on our three datasets and two existing benchmarks demonstrate the effectiveness of our method, where Pearson, Spearman, and Kendall correlations with human evaluation outperform state-of-the-art baselines.

Zhengliang Shi, Shen Gao, Xiuyi Chen et al. · baidu

Tool learning empowers large language models (LLMs) as agents to use external tools and extend their utility. Existing methods employ one single LLM-based agent to iteratively select and execute tools, thereafter incorporating execution results into the next action prediction. Despite their progress, these methods suffer from performance degradation when addressing practical tasks due to: (1) the pre-defined pipeline with restricted flexibility to calibrate incorrect actions, and (2) the struggle to adapt a general LLM-based agent to perform a variety of specialized actions. To mitigate these problems, we propose ConAgents, a Cooperative and interactive Agents framework, which coordinates three specialized agents for tool selection, tool execution, and action calibration separately. ConAgents introduces two communication protocols to enable the flexible cooperation of agents. To effectively generalize the ConAgents into open-source models, we also propose specialized action distillation, enhancing their ability to perform specialized actions in our framework. Our extensive experiments on three datasets show that the LLMs, when equipped with the ConAgents, outperform baselines with substantial improvement (i.e., up to 14% higher success rate).

Chengrui Huang, Zhengliang Shi, Yuntao Wen et al.

Tool learning methods have enhanced the ability of large language models (LLMs) to interact with real-world applications. Many existing works fine-tune LLMs or design prompts to enable LLMs to select appropriate tools and correctly invoke them to meet user requirements. However, it is observed in previous works that the performance of tool learning varies from tasks, datasets, training settings, and algorithms. Without understanding the impact of these factors, it can lead to inconsistent results, inefficient model deployment, and suboptimal tool utilization, ultimately hindering the practical integration and scalability of LLMs in real-world scenarios. Therefore, in this paper, we explore the impact of both internal and external factors on the performance of tool learning frameworks. Through extensive experiments on two benchmark datasets, we find several insightful conclusions for future work, including the observation that LLMs can benefit significantly from increased trial and exploration. We believe our empirical study provides a new perspective for future tool learning research.

Yiqun Chen, Jinyuan Feng, Wei Yang et al.

The inference overhead induced by redundant reasoning undermines the interactive experience and severely bottlenecks the deployment of Large Reasoning Models. Existing reinforcement learning (RL)-based solutions tackle this problem by coupling a length penalty with outcome-based rewards. This simplistic reward weighting struggles to reconcile brevity with accuracy, as enforcing brevity may compromise critical reasoning logic. In this work, we address this limitation by proposing a multi-agent RL framework that selectively penalizes redundant chunks, while preserving essential reasoning logic. Our framework, Self-Compression via MARL (SCMA), instantiates redundancy detection and evaluation through two specialized agents: \textbf{a Segmentation Agent} for decomposing the reasoning process into logical chunks, and \textbf{a Scoring Agent} for quantifying the significance of each chunk. The Segmentation and Scoring agents collaboratively define an importance-weighted length penalty during training, incentivizing \textbf{a Reasoning Agent} to prioritize essential logic without introducing inference overhead during deployment. Empirical evaluations across model scales demonstrate that SCMA reduces response length by 11.1\% to 39.0\% while boosting accuracy by 4.33\% to 10.02\%. Furthermore, ablation studies and qualitative analysis validate that the synergistic optimization within the MARL framework fosters emergent behaviors, yielding more powerful LRMs compared to vanilla RL paradigms.

Jiulong Wu, Zhengliang Shi, Shuaiqiang Wang et al. · baidu

Large Visual Language Models (LVLMs) have demonstrated impressive capabilities across multiple tasks. However, their trustworthiness is often challenged by hallucinations, which can be attributed to the modality misalignment and the inherent hallucinations of their underlying Large Language Models (LLMs) backbone. Existing preference alignment methods focus on aligning model responses with human preferences while neglecting image-text modality alignment, resulting in over-reliance on LLMs and hallucinations. In this paper, we propose Entity-centric Multimodal Preference Optimization (EMPO), which achieves enhanced modality alignment compared to existing human preference alignment methods. Besides, to overcome the scarcity of high-quality multimodal preference data, we utilize open-source instruction datasets to automatically construct high-quality preference data across three aspects: image, instruction, and response. Experiments on two human preference datasets and five multimodal hallucination benchmarks demonstrate the effectiveness of EMPO, e.g., reducing hallucination rates by 85.9\% on Object-HalBench and 49.8\% on MM-HalBench.

Dongsheng Zhu, Weixian Shi, Zhengliang Shi et al. · baidu

Although current Large Language Models (LLMs) exhibit impressive capabilities, performing complex real-world tasks still requires tool learning. Mainstream methods, such as CoT/ReAct, rely on step-by-step tool invocation to interact with external environments, but they are limited in perceptual scope and lack adequate task-planning capability. To address these limitations, other studies introduce the first Search-based Decision Tree (DFSDT), which still suffers from the high computational cost. In this paper, we introduce a novel parallel tool invocation paradigm, DTA-Llama (Divide-Then-Aggregate Llama). First, we transform traditional tree-based tool search paths into Directed Acyclic Graph (DAG) structure, generating a high-quality parallel tool invocation dataset. The DTA-Llama is then trained on the dataset to learn to iteratively divide the current task into several parallel tool invocation sub-tasks and aggregate the invocation results to decide the next actions. Furthermore, we introduce an efficient inference framework inspired by the Process/Threads mechanism when applying the DTA-Llama to practical tasks. Experimental results show that our approach substantially enhances task performance while reducing token consumption and inference time. Llama2-7B, using our method, is comparable to the official parallel function calling method of GPT-3.5. The relevant code, dataset, and model weights are available at https://corn0205.github.io/

Zhengliang Shi, Yuhan Wang, Lingyong Yan et al. · baidu

Tool learning aims to augment large language models (LLMs) with diverse tools, enabling them to act as agents for solving practical tasks. Due to the limited context length of tool-using LLMs, adopting information retrieval (IR) models to select useful tools from large toolsets is a critical initial step. However, the performance of IR models in tool retrieval tasks remains underexplored and unclear. Most tool-use benchmarks simplify this step by manually pre-annotating a small set of relevant tools for each task, which is far from the real-world scenarios. In this paper, we propose ToolRet, a heterogeneous tool retrieval benchmark comprising 7.6k diverse retrieval tasks, and a corpus of 43k tools, collected from existing datasets. We benchmark six types of models on ToolRet. Surprisingly, even the models with strong performance in conventional IR benchmarks, exhibit poor performance on ToolRet. This low retrieval quality degrades the task pass rate of tool-use LLMs. As a further step, we contribute a large-scale training dataset with over 200k instances, which substantially optimizes the tool retrieval ability of IR models.

Zhengliang Shi, Yiqun Chen, Haitao Li et al.

Large language models (LLMs) have rapidly evolved from text generators into powerful problem solvers. Yet, many open tasks demand critical thinking, multi-source, and verifiable outputs, which are beyond single-shot prompting or standard retrieval-augmented generation. Recently, numerous studies have explored Deep Research (DR), which aims to combine the reasoning capabilities of LLMs with external tools, such as search engines, thereby empowering LLMs to act as research agents capable of completing complex, open-ended tasks. This survey presents a comprehensive and systematic overview of deep research systems, including a clear roadmap, foundational components, practical implementation techniques, important challenges, and future directions. Specifically, our main contributions are as follows: (i) we formalize a three-stage roadmap and distinguish deep research from related paradigms; (ii) we introduce four key components: query planning, information acquisition, memory management, and answer generation, each paired with fine-grained sub-taxonomies; (iii) we summarize optimization techniques, including prompting, supervised fine-tuning, and agentic reinforcement learning; and (iv) we consolidate evaluation criteria and open challenges, aiming to guide and facilitate future development. As the field of deep research continues to evolve rapidly, we are committed to continuously updating this survey to reflect the latest progress in this area.

Ziyou Hu, Zhengliang Shi, Minghang Zhu et al.

Reward models (RMs) have become essential for aligning large language models (LLMs), serving as scalable proxies for human evaluation in both training and inference. However, existing RMs struggle on knowledge-intensive and long-form tasks, where evaluating correctness requires grounding beyond the model's internal knowledge. This limitation hinders them from reliably discriminating subtle quality differences, especially when external evidence is necessary. To address this, we introduce OpenRM, a tool-augmented long-form reward model that systematically judges open-ended responses by invoking external tools to gather relevant evidence. We train OpenRM with Group Relative Policy Optimization (GRPO) on over 27K synthesized pairwise examples generated through a controllable data synthesis framework. The training objective jointly supervises intermediate tool usage and final outcome accuracy, incentivizing our reward model to learn effective evidence-based judgment strategies. Extensive experiments on three newly-collected datasets and two widely-used benchmarks demonstrate that OpenRM substantially outperforms existing reward modeling approaches. As a further step, we integrate OpenRM into both inference-time response selection and training-time data selection. This yields consistent gains in downstream LLM alignment tasks, highlighting the potential of tool-augmented reward models for scaling reliable long-form evaluation.

Xinyu Guo, Zhengliang Shi, Minglai Yang et al.

This paper introduces a framework for relation extraction (RE) that enhances both accuracy and explainability. The framework has two key components: (i) a reasoning mechanism that formulates relation extraction as a series of text-processing steps inspired by cognitive science, and (ii) an optimization process driven by reinforcement learning (RL) with a novel reward function designed to improve both task accuracy and explanation quality. We call our approach CogRE. Our framework addresses the lack of supervision for language-based explanations in traditional RE by promoting outputs that include important relation keywords. These keywords are drawn from a high-quality dictionary that is automatically constructed using an LLM. We evaluate our approach for the task of one-shot RE using two LLMs and two RE datasets. Our experiments show that CogRE improves explanation quality by addressing two common failure patterns in one-shot RE: poor attention focus and limited one-shot learning capability. For example, our cognitive-structured reasoning with Qwen2.5-15B-Instruct on One-shot NYT29 achieves 24.65% F1, surpassing prior reasoning-based designs. Optimizing this approach with RL using our reward further improves performance by +23.46% (absolute). Finally, human evaluation shows that our best model generates relational keywords closely aligned with gold labels, increasing human explanation quality ratings by 54% (relative).

Zhengliang Shi, Ruotian Ma, Jen-tse Huang et al. · pku, tencent-ai

Large language models (LLMs) are increasingly entrusted with high-stakes decisions that affect human welfare. However, the principles and values that guide these models when distributing scarce societal resources remain largely unexamined. To address this, we introduce the Social Welfare Function (SWF) Benchmark, a dynamic simulation environment where an LLM acts as a sovereign allocator, distributing tasks to a heterogeneous community of recipients. The benchmark is designed to create a persistent trade-off between maximizing collective efficiency (measured by Return on Investment) and ensuring distributive fairness (measured by the Gini coefficient). We evaluate 20 state-of-the-art LLMs and present the first leaderboard for social welfare allocation. Our findings reveal three key insights: (i) A model's general conversational ability, as measured by popular leaderboards, is a poor predictor of its allocation skill. (ii) Most LLMs exhibit a strong default utilitarian orientation, prioritizing group productivity at the expense of severe inequality. (iii) Allocation strategies are highly vulnerable, easily perturbed by output-length constraints and social-influence framing. These results highlight the risks of deploying current LLMs as societal decision-makers and underscore the need for specialized benchmarks and targeted alignment for AI governance.

Yue Wang, Ruotian Ma, Xingyu Chen et al.

The rise of Large Language Models (LLMs) is reshaping multimodel models, with speech synthesis being a prominent application. However, existing approaches often underutilize the linguistic intelligence of these models, typically failing to leverage their powerful instruction-following capabilities. This limitation hinders the model's ability to follow text instructions for controllable Text-to-Speech~(TTS). To address this, we propose a new paradigm inspired by ``operationalism'' that decouples instruction understanding from speech generation. We introduce BatonVoice, a framework where an LLM acts as a ``conductor'', understanding user instructions and generating a textual ``plan'' -- explicit vocal features (e.g., pitch, energy). A separate TTS model, the ``orchestra'', then generates the speech from these features. To realize this component, we develop BatonTTS, a TTS model trained specifically for this task. Our experiments demonstrate that BatonVoice achieves strong performance in controllable and emotional speech synthesis, outperforming strong open- and closed-source baselines. Notably, our approach enables remarkable zero-shot cross-lingual generalization, accurately applying feature control abilities to languages unseen during post-training. This demonstrates that objectifying speech into textual vocal features can more effectively unlock the linguistic intelligence of LLMs.

Xinbei Ma, Ruotian Ma, Xingyu Chen et al. · pku, tencent-ai

LLM-based multi-agent systems demonstrate great potential for tackling complex problems, but how competition shapes their behavior remains underexplored. This paper investigates the over-competition in multi-agent debate, where agents under extreme pressure exhibit unreliable, harmful behaviors that undermine both collaboration and task performance. To study this phenomenon, we propose HATE, the Hunger Game Debate, a novel experimental framework that simulates debates under a zero-sum competition arena. Our experiments, conducted across a range of LLMs and tasks, reveal that competitive pressure significantly stimulates over-competition behaviors and degrades task performance, causing discussions to derail. We further explore the impact of environmental feedback by adding variants of judges, indicating that objective, task-focused feedback effectively mitigates the over-competition behaviors. We also probe the post-hoc kindness of LLMs and form a leaderboard to characterize top LLMs, providing insights for understanding and governing the emergent social dynamics of AI community.

Minghang Zhu, Zhengliang Shi, Zhiwei Xu et al.

The advancement of large language models (LLMs) has enabled the construction of multi-agent systems to solve complex tasks by dividing responsibilities among specialized agents, such as a planning agent for subgoal generation and a grounding agent for executing tool-use actions. Most existing methods typically fine-tune these agents independently, leading to capability gaps among them with poor coordination. To address this, we propose MOAT, a Multi-Agent Joint Alignment Tuning framework that improves agents collaboration through iterative alignment. MOAT alternates between two key stages: (1) Planning Agent Alignment, which optimizes the planning agent to generate subgoal sequences that better guide the grounding agent; and (2) Grounding Agent Improving, which fine-tunes the grounding agent using diverse subgoal-action pairs generated by the agent itself to enhance its generalization capablity. Theoretical analysis proves that MOAT ensures a non-decreasing and progressively convergent training process. Experiments across six benchmarks demonstrate that MOAT outperforms state-of-the-art baselines, achieving average improvements of 3.1% on held-in tasks and 4.4% on held-out tasks.

Minghang Zhu, Shen Gao, Zhengliang Shi et al.

A common training approach for language models involves using a large-scale language model to expand a human-provided dataset, which is subsequently used for model training.This method significantly reduces training costs by eliminating the need for extensive human data annotation. However, it still faces challenges such as high carbon emissions during data augmentation and the risk of data leakage when we use closed-source LLMs. To address these issues, we propose a self-evolution method for language models. First, we introduce the Multi-level Principle Generation, which enables a large-scale model to summarize task-completion principles based on a small amount of task data. Then, we propose the Principle-based Instance Generation, in which a smaller-scale language model uses these task principles to generate a large amount of data. This data is then used for model training. Experimental results show that our proposed method significantly improves model performance compared to directly using a smaller-scale language model to generate data. Additionally, since we only use the large-scale language model to generate the task-completion principles, the carbon emissions associated with training the model are greatly reduced.

Zhengliang Shi, Weiwei Sun, Shen Gao et al.

Multi-Hop Question Answering (MHQA) tasks present a significant challenge for large language models (LLMs) due to the intensive knowledge required. Current solutions, like Retrieval-Augmented Generation, typically retrieve potential documents from an external corpus to read an answer. However, the performance of this retrieve-then-read paradigm is constrained by the retriever and the inevitable noise in the retrieved documents. To mitigate these challenges, we introduce a novel generate-then-ground (GenGround) framework, synergizing the parametric knowledge of LLMs and external documents to solve a multi-hop question. GenGround empowers LLMs to alternate two phases until the final answer is derived: (1) formulate a simpler, single-hop question and directly generate the answer; (2) ground the question-answer pair in retrieved documents, amending any wrong predictions in the answer. We also propose an instructional grounding distillation method to generalize our method into smaller models. Extensive experiments conducted on four datasets illustrate the superiority of our method.