Zhuofeng Li

*Co-first Author

1. Arxiv

   

   Beyond Semantic Similarity: Rethinking Retrieval for Agentic Search via Direct Corpus Interaction

   Zhuofeng Li , Haoxiang Zhang , Cong Wei , Pan Lu , Ping Nie , Yi Lu , Yuyang Bai , Shangbin Feng , Hangxiao Zhu , Ming Zhong , Yuyu Zhang , Jianwen Xie , Yejin Choi , James Zou , Jiawei Han , Wenhu Chen , Jimmy Lin , Dongfu Jiang , and Yu Zhang

   In arxiv preprint, May 2026

   

   Modern retrieval systems, whether lexical or semantic, expose a corpus through a fixed similarity interface that compresses access into a single top-k retrieval step before reasoning. This abstraction is efficient, but for agentic search, it becomes a bottleneck: exact lexical constraints, sparse clue conjunctions, local context checks, and multi-step hypothesis refinement are difficult to implement by calling a conventional off-the-shelf retriever, and evidence filtered out early cannot be recovered by stronger downstream reasoning. Agentic tasks further exacerbate this limitation because they require agents to orchestrate multiple steps, including discovering intermediate entities, combining weak clues, and revising the plan after observing partial evidence. To tackle the limitation, we study direct corpus interaction (DCI), where an agent searches the raw corpus directly with general-purpose terminal tools (e.g., grep, file reads, shell commands, lightweight scripts), without any embedding model, vector index, or retrieval API. This approach requires no offline indexing and adapts naturally to evolving local corpora. Across IR benchmarks and end-to-end agentic search tasks, this simple setup substantially outperforms strong sparse, dense, and reranking baselines on several BRIGHT and BEIR datasets, and attains strong accuracy on BrowseComp-Plus and multi-hop QA without relying on any conventional semantic retriever. Our results indicate that as language agents become stronger, retrieval quality depends not only on reasoning ability but also on the resolution of the interface through which the model interacts with the corpus, with which DCI opens a broader interface-design space for agentic search.

   @inproceedings{li2026dci,
     title = {Beyond Semantic Similarity: Rethinking Retrieval for Agentic Search via Direct Corpus Interaction},
     author = {Li, Zhuofeng and Zhang, Haoxiang and Wei, Cong and Lu, Pan and Nie, Ping and Lu, Yi and Bai, Yuyang and Feng, Shangbin and Zhu, Hangxiao and Zhong, Ming and Zhang, Yuyu and Xie, Jianwen and Choi, Yejin and Zou, James and Han, Jiawei and Chen, Wenhu and Lin, Jimmy and Jiang, Dongfu and Zhang, Yu},
     booktitle = {arxiv preprint},
     month = may,
     year = {2026},
     selected = true,
   }

2. EMNLP 2026

   

   OpenResearcher: A Fully Open Pipeline for Long-Horizon Deep Research Trajectory Synthesis

   Zhuofeng Li ^* , Dongfu Jiang ^* , Xueguang Ma , Haoxiang Zhang , Ping Nie , Yuyu Zhang , Kai Zou , Jianwen Xie , Yu Zhang , and Wenhu Chen

   In arxiv preprint, Mar 2026

   

   Training deep research agents requires long-horizon trajectories that interleave search, evidence aggregation, and multi-step reasoning. However, existing data collection pipelines typically rely on proprietary web APIs, making large-scale trajectory synthesis costly, unstable, and difficult to reproduce. We present OpenResearcher, a reproducible pipeline that decouples one-time corpus bootstrapping from multi-turn trajectory synthesis and executes the search-and-browse loop entirely offline using three explicit browser primitives—search, open, and find—over a 15M-document corpus. Using GPT-OSS-120B as the teacher model, we synthesize over 97K trajectories, including a substantial long-horizon tail with 100+ tool calls. After supervised fine-tuning a 30B-A3B backbone, we achieve 54.8% accuracy on BrowseComp-Plus, a +34.0 point improvement over the base model, while remaining competitive on BrowseComp, GAIA, and xbench-DeepSearch. The offline, fully instrumented environment also enables controlled analysis, revealing practical insights into deep research pipeline design. We release the pipeline, synthesized trajectories, model checkpoints, and the offline search environment.

   @inproceedings{li2026openresearcher,
     title = {OpenResearcher: A Fully Open Pipeline for Long-Horizon Deep Research Trajectory Synthesis},
     author = {Li, Zhuofeng and Jiang, Dongfu and Ma, Xueguang and Zhang, Haoxiang and Nie, Ping and Zhang, Yuyu and Zou, Kai and Xie, Jianwen and Zhang, Yu and Chen, Wenhu},
     booktitle = {arxiv preprint},
     month = mar,
     year = {2026},
     github = {TIGER-AI-Lab/OpenResearcher},
     huggingface = {https://huggingface.co/OpenResearcher},
     twitter = {https://x.com/zhuofengli96475/status/2036475211063648414},
     selected = true,
     num_co_first_author = {2},
   }

3. ACL 2026

   

   ReviewGrounder: Improving Review Substantiveness with Rubric-Guided, Tool-Integrated Agents

   Zhuofeng Li ^* , Yi Lu ^* , Dongfu Jiang , Haoxiang Zhang , Yuyang Bai , Chuan Li , Yu Wang , Shuiwang Ji , Jianwen Xie , and Yu Zhang

   In ACL 2026, Apr 2026

   

   The rapid rise in AI conference submissions has driven increasing exploration of large language models (LLMs) for peer review support. However, LLM-based reviewers often generate superficial, formulaic comments lacking substantive, evidence-grounded feedback. We attribute this to the underutilization of two key components of human reviewing: explicit rubrics and contextual grounding in existing work. To address this, we introduce REVIEWBENCH, a benchmark evaluating review text according to paper-specific rubrics derived from official guidelines, the paper’s content, and human-written reviews. We further propose REVIEWGROUNDER, a rubric-guided, tool-integrated multi-agent framework that decomposes reviewing into drafting and grounding stages, enriching shallow drafts via targeted evidence consolidation. Experiments on REVIEWBENCH show that REVIEWGROUNDER, using a Phi-4-14B-based drafter and a GPT-OSS-120B-based grounding stage, consistently outperforms baselines with substantially stronger/larger backbones (e.g., GPT-4.1 and DeepSeek-R1-670B) in both alignment with human judgments and rubric-based review quality across 8 dimensions.

   @inproceedings{li2026reviewgrounder,
     title = {ReviewGrounder: Improving Review Substantiveness with Rubric-Guided, Tool-Integrated Agents},
     author = {Li, Zhuofeng and Lu, Yi and Jiang, Dongfu and Zhang, Haoxiang and Bai, Yuyang and Li, Chuan and Wang, Yu and Ji, Shuiwang and Xie, Jianwen and Zhang, Yu},
     booktitle = {ACL 2026},
     month = apr,
     year = {2026},
     selected = true,
     num_co_first_author = {2},
   }

4. ICLR 2026 Oral

   

   In-The-Flow Agentic System Optimization for Effective Planning and Tool Use

   Zhuofeng Li ^* , Haoxiang Zhang ^* , Seungju Han , Sheng Liu , Jianwen Xie , Yu Zhang , Yejin Choi , James Zou , and Pan Lu

   In ICLR 2026 Oral & NeurIPS 2025 Workshop Best Paper Runner-up, Oct 2025

   

   Outcome-driven reinforcement learning has advanced reasoning in large language models (LLMs), but prevailing tool-augmented approaches train a single, monolithic policy that interleaves thoughts and tool calls under full context; this scales poorly with long horizons and diverse tools and generalizes weakly to new scenarios. Agentic systems offer a promising alternative by decomposing work across specialized modules, yet most remain training-free or rely on offline training decoupled from the live dynamics of multi-turn interaction. We introduce AgentFlow, a trainable, in-the-flow agentic framework that coordinates four modules (planner, executor, verifier, generator) through an evolving memory and directly optimizes its planner inside the multi-turn loop. To train on-policy in live environments, we propose Flow-based Group Refined Policy Optimization (Flow-GRPO), which tackles long-horizon, sparse-reward credit assignment by converting multi-turn optimization into a sequence of tractable single-turn policy updates. It broadcasts a single, verifiable trajectory-level outcome to every turn to align local planner decisions with global success and stabilizes learning with group-normalized advantages. Across ten benchmarks, AgentFlow with a 7B-scale backbone outperforms top-performing baselines with average accuracy gains of 14.9% on search, 14.0% on agentic, 14.5% on mathematical, and 4.1% on scientific tasks, even surpassing larger proprietary models like GPT-4o. Further analyses confirm the benefits of in-the-flow optimization, showing improved planning, enhanced tool-calling reliability, and positive scaling with model size and reasoning turns.

   @inproceedings{li2025agentflow,
     title = {In-The-Flow Agentic System Optimization for Effective Planning and Tool Use},
     author = {Li, Zhuofeng and Zhang, Haoxiang and Han, Seungju and Liu, Sheng and Xie, Jianwen and Zhang, Yu and Choi, Yejin and Zou, James and Lu, Pan},
     booktitle = {ICLR 2026 Oral & NeurIPS 2025 Workshop Best Paper Runner-up},
     month = oct,
     year = {2025},
     github = {lupantech/AgentFlow},
     huggingface = {https://huggingface.co/AgentFlow},
     twitter = {https://x.com/lupantech/status/1976016000345919803},
     selected = true,
     num_co_first_author = {2},
   }

5. ICLR 2026 SPOT Best Paper

   

   VerlTool: Towards Holistic Agentic Reinforcement Learning with Tool Use

   Dongfu Jiang ^* , Yi Lu ^* , Zhuofeng Li ^* , Zhiheng Lyu ^* , Ping Nie , Haozhe Wang , Alex Su , Hui Chen , Kai Zou , Chao Du , Tianyu Pang , and Wenhu Chen

   In ICLR 2026 SPOT Best Paper, Sep 2025

   

   Reinforcement Learning with Verifiable Rewards (RLVR) has demonstrated success in enhancing LLM reasoning capabilities, but remains limited to single-turn interactions without tool integration. While recent Agentic Reinforcement Learning with Tool use (ARLT) approaches have emerged to address multi-turn tool interactions, existing works develop task-specific codebases that suffer from fragmentation, synchronous execution bottlenecks, and limited extensibility across domains. These inefficiencies hinder broader community adoption and algorithmic innovation. We introduce VerlTool, a unified and modular framework that addresses these limitations through systematic design principles. VerlTool provides four key contributions: (1) upstream alignment with VeRL ensuring compatibility and simplified maintenance, (2) unified tool management via standardized APIs supporting diverse modalities including code execution, search, SQL databases, and vision processing, (3) asynchronous rollout execution achieving near 2× speedup by eliminating synchronization bottlenecks, and (4) comprehensive evaluation demonstrating competitive performance across 6 ARLT domains. Our framework formalizes ARLT as multi-turn trajectories with multi-modal observation tokens (text/image/video), extending beyond single-turn RLVR paradigms. We train and evaluate models on mathematical reasoning, knowledge QA, SQL generation, visual reasoning, web search, and software engineering tasks, achieving results comparable to specialized systems while providing unified training infrastructure.

   @inproceedings{jiang2025verltool,
     title = {VerlTool: Towards Holistic Agentic Reinforcement Learning with Tool Use},
     author = {Jiang, Dongfu and Lu, Yi and Li, Zhuofeng and Lyu, Zhiheng and Nie, Ping and Wang, Haozhe and Su, Alex and Chen, Hui and Zou, Kai and Du, Chao and Pang, Tianyu and Chen, Wenhu},
     booktitle = {ICLR 2026 SPOT Best Paper},
     month = sep,
     year = {2025},
     github = {TIGER-AI-Lab/verl-tool},
     huggingface = {https://huggingface.co/papers/2509.01055},
     twitter = {https://x.com/zhuofengli96475/status/1963216180438814901},
     selected = true,
     num_co_first_author = {4},
   }

6. CIKM 2025

   

   GReF: A Unified Generative Framework for Efficient Reranking via Ordered Multi-token Prediction

   Zhijie Lin ^* , Zhuofeng Li ^* , Chenglei Dai , Wentian Bao , Shuai Lin , Enyun Yu , Haoxiang Zhang , and Liang Zhao

   In Proceedings of the 34th ACM International Conference on Information and Knowledge Management, Nov 2025

   

   In a multi-stage recommendation system, reranking plays a crucial role in modeling intra-list correlations among items. A key challenge lies in exploring optimal sequences within the combinatorial space of permutations. Recent research follows a two-stage (generator-evaluator) paradigm, where a generator produces multiple feasible sequences, and an evaluator selects the best one. In practice, the generator is typically implemented as an autoregressive model. However, these two-stage methods face two main challenges. First, the separation of the generator and evaluator hinders end-to-end training. Second, autoregressive generators suffer from inference efficiency. In this work, we propose a Unified Generative Efficient Reranking Framework (GReF) to address the two primary challenges. Specifically, we introduce Gen-Reranker, an autoregressive generator featuring a bidirectional encoder and a dynamic autoregressive decoder to generate causal reranking sequences. Subsequently, we pre-train Gen-Reranker on the item exposure order for high-quality parameter initialization. To eliminate the need for the evaluator while integrating sequence-level evaluation during training for end-to-end optimization, we propose post-training the model through Rerank-DPO. Moreover, for efficient autoregressive inference, we introduce ordered multi-token prediction (OMTP), which trains Gen-Reranker to simultaneously generate multiple future items while preserving their order, ensuring practical deployment in real-time recommender systems. Extensive offline experiments demonstrate that GReF outperforms state-of-the-art reranking methods while achieving latency that is nearly comparable to non-autoregressive models. Additionally, GReF has also been deployed in a real-world video app Kuaishou with over 300 million daily active users, significantly improving online recommendation quality.

   @inproceedings{lin2025gref,
     title = {GReF: A Unified Generative Framework for Efficient Reranking via Ordered Multi-token Prediction},
     author = {Lin, Zhijie and Li, Zhuofeng and Dai, Chenglei and Bao, Wentian and Lin, Shuai and Yu, Enyun and Zhang, Haoxiang and Zhao, Liang},
     booktitle = {Proceedings of the 34th ACM International Conference on Information and Knowledge Management},
     month = nov,
     year = {2025},
     selected = true,
     num_co_first_author = {2},
   }

7. TMLR 2025

   

   Avoiding Structural Pitfalls: Self-Supervised Low-Rank Feature Tuning for Graph Test-Time Adaptation

   Haoxiang Zhang ^* , Zhuofeng Li ^* , Qiannan Zhang , Ziyi Kou , Juncheng Li , and Shichao Pei

   In Transactions on Machine Learning Research (TMLR), Oct 2025, Oct 2025

   

   Graph Neural Networks (GNNs) have shown impressive performance on graph-structured data, but they often struggle to generalize when facing distribution shifts at test time. Existing test-time adaptation methods for graphs primarily focus on node feature adjustments while overlooking structural information, leading to suboptimal adaptation. We propose a novel self-supervised low-rank feature tuning approach that jointly considers both node features and graph structure during test-time adaptation. Our method leverages self-supervised learning objectives and low-rank matrix factorization to efficiently adapt GNNs to new graph distributions without requiring labeled data, avoiding common structural pitfalls that plague existing approaches.

   @inproceedings{zhang2025avoiding,
     title = {Avoiding Structural Pitfalls: Self-Supervised Low-Rank Feature Tuning for Graph Test-Time Adaptation},
     author = {Zhang, Haoxiang and Li, Zhuofeng and Zhang, Qiannan and Kou, Ziyi and Li, Juncheng and Pei, Shichao},
     booktitle = {Transactions on Machine Learning Research (TMLR), Oct 2025},
     month = oct,
     year = {2025},
     selected = true,
     num_co_first_author = {2},
   }

8. ICLR 2026

   

   ImagenWorld: Stress-Testing Image Generation Models with Explainable Human Evaluation on Open-ended Real-World Tasks

   Samin Mahdizadeh Sani , Max Ku , Nima Jamali , Matina Mahdizadeh Sani , Paria Khoshtab , Wei-Chieh Sun , Parnian Fazel , Zhi Rui Tam , Thomas Chong , Edisy Kin Wai Chan , Donald Wai Tong Tsang , Chiao-Wei Hsu , Ting Wai Lam , Ho Yin Sam Ng , Chiafeng Chu , Chak-Wing Mak , Keming Wu , Hiu Tung Wong , Yik Chun Ho , Chi Ruan , Zhuofeng Li , I-Sheng Fang , Shih-Ying Yeh , Ho Kei Cheng , Ping Nie , and Wenhu Chen

   In ICLR 2026, Jan 2026

   

   We introduce ImagenWorld, a benchmark of 3.6K condition sets spanning six core tasks (generation and editing, with single or multiple references) and six topical domains (artworks, photorealistic images, information graphics, textual graphics, computer graphics, and screenshots). The benchmark is supported by 20K fine-grained human annotations and an explainable evaluation schema that tags localized object-level and segment-level errors, complementing automated VLM-based metrics. Our large-scale evaluation of 14 models yields several insights: (1) models typically struggle more in editing tasks than in generation tasks, especially in local edits. (2) models excel in artistic and photorealistic settings but struggle with symbolic and text-heavy domains such as screenshots and information graphics. (3) closed-source systems lead overall, while targeted data curation narrows the gap in text-heavy cases. (4) modern VLM-based metrics achieve Kendall accuracies up to 0.79, approximating human ranking, but fall short of fine-grained, explainable error attribution.

   @inproceedings{sani2026imagenworld,
     title = {ImagenWorld: Stress-Testing Image Generation Models with Explainable Human Evaluation on Open-ended Real-World Tasks},
     author = {Sani, Samin Mahdizadeh and Ku, Max and Jamali, Nima and Sani, Matina Mahdizadeh and Khoshtab, Paria and Sun, Wei-Chieh and Fazel, Parnian and Tam, Zhi Rui and Chong, Thomas and Chan, Edisy Kin Wai and Tsang, Donald Wai Tong and Hsu, Chiao-Wei and Lam, Ting Wai and Ng, Ho Yin Sam and Chu, Chiafeng and Mak, Chak-Wing and Wu, Keming and Wong, Hiu Tung and Ho, Yik Chun and Ruan, Chi and Li, Zhuofeng and Fang, I-Sheng and Yeh, Shih-Ying and Cheng, Ho Kei and Nie, Ping and Chen, Wenhu},
     booktitle = {ICLR 2026},
     month = jan,
     year = {2026},
     selected = true,
   }

9. Arxiv

   

   VideoEval-Pro: Robust and Realistic Long Video Understanding Evaluation

   Wentao Ma , Weiming Ren , Yiming Jia , Zhuofeng Li , Ping Nie , Ge Zhang , and Wenhu Chen

   In arxiv preprint, May 2025

   

   Large multimodal models (LMMs) have recently emerged as a powerful tool for long video understanding (LVU), but existing benchmarks have significant limitations. Most current benchmarks rely heavily on multiple-choice questions (MCQs) with inflated evaluation results due to guessing, and many questions have strong priors that allow models to answer without watching the entire video. We propose VideoEval-Pro, a new benchmark with open-ended short-answer questions that assess segment-level and full-video understanding through perception and reasoning tasks. By evaluating 21 video LMMs, we discovered significant performance drops (>25\%) on open-ended questions compared to MCQs, higher MCQ scores do not correlate with higher open-ended scores, and VideoEval-Pro benefits more from increased input frames compared to other benchmarks. Our goal is to provide a more realistic and reliable measure of long video understanding.

   @inproceedings{ma2025videoeval,
     title = {VideoEval-Pro: Robust and Realistic Long Video Understanding Evaluation},
     author = {Ma, Wentao and Ren, Weiming and Jia, Yiming and Li, Zhuofeng and Nie, Ping and Zhang, Ge and Chen, Wenhu},
     booktitle = {arxiv preprint},
     month = may,
     year = {2025},
     github = {TIGER-AI-Lab/VideoEval-Pro},
     huggingface = {https://huggingface.co/papers/2505.14640},
     twitter = {https://x.com/zhuofengli96475/status/1963216180438814901},
     selected = true,
   }

10. TMLR 2025

    

    StructEval: Benchmarking LLMs’ Capabilities to Generate Structural Outputs

    Jialin Yang , Dongfu Jiang , Lipeng He , Sherman Siu , Yuxuan Zhang , Disen Liao , Zhuofeng Li , Huaye Zeng , Yiming Jia , Haozhe Wang , and others

    In Transactions on Machine Learning Research (TMLR), Dec 2025, May 2025

    

    As Large Language Models (LLMs) become integral to software development workflows, their ability to generate structured outputs has become critically important. We introduce StructEval, a comprehensive benchmark for evaluating LLMs’ capabilities in producing both non-renderable (JSON, YAML, CSV) and renderable (HTML, React, SVG) structured formats. Unlike prior benchmarks, StructEval systematically evaluates structural fidelity across diverse formats through two paradigms: generation tasks and conversion tasks. Our benchmark encompasses 18 formats and 44 types of task, with novel metrics for format adherence and structural correctness. Results reveal significant performance gaps, even state-of-the-art models like o1-mini achieve only 75.58 average score.

    @inproceedings{yang2025structeval,
      title = {StructEval: Benchmarking LLMs' Capabilities to Generate Structural Outputs},
      author = {Yang, Jialin and Jiang, Dongfu and He, Lipeng and Siu, Sherman and Zhang, Yuxuan and Liao, Disen and Li, Zhuofeng and Zeng, Huaye and Jia, Yiming and Wang, Haozhe and others},
      booktitle = {Transactions on Machine Learning Research (TMLR), Dec 2025},
      month = may,
      year = {2025},
      github = {https://github.com/TIGER-AI-Lab/StructEval},
      huggingface = {https://huggingface.co/papers/2505.20139},
      selected = true,
    }

11. NeurIPS 2024

    

    Teg-db: A comprehensive dataset and benchmark of textual-edge graphs

    Zhuofeng Li , Zixing Gou , Xiangnan Zhang , Zhongyuan Liu , Sirui Li , Yuntong Hu , Chen Ling , Zheng Zhang , and Liang Zhao

    In Advances in Neural Information Processing Systems, Dec 2024

    

    Text-Attributed Graphs (TAGs) augment graph structures with natural language descriptions, but existing datasets primarily feature textual information only at nodes. We introduce TEG-DB, a comprehensive benchmark of textual-edge datasets with rich descriptions on both nodes and edges, spanning domains like citation and social networks. We conducted benchmark experiments to evaluate how current techniques including pre-trained language models and graph neural networks can utilize textual node and edge information. Our goal is to advance research in textual-edge graph analysis and provide deeper insights into complex real-world networks.

    @inproceedings{li2024teg,
      title = {Teg-db: A comprehensive dataset and benchmark of textual-edge graphs},
      author = {Li, Zhuofeng and Gou, Zixing and Zhang, Xiangnan and Liu, Zhongyuan and Li, Sirui and Hu, Yuntong and Ling, Chen and Zhang, Zheng and Zhao, Liang},
      booktitle = {Advances in Neural Information Processing Systems},
      volume = {37},
      pages = {60980--60998},
      month = dec,
      year = {2024},
      github = {Zhuofeng-Li/TEG-Benchmark},
      huggingface = {https://huggingface.co/papers/2406.10310},
      selected = true,
    }

12. CIKM 2024

    

    Learning from novel knowledge: Continual few-shot knowledge graph completion

    Zhuofeng Li , Haoxiang Zhang , Qiannan Zhang , Ziyi Kou , and Shichao Pei

    In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management, Oct 2024

    

    Knowledge graph completion has been increasingly recognized as a vital approach for uncovering missing knowledge and addressing the incompleteness issue in KGs. To enhance inference on rare relations and mitigate the impact of the long-tail distribution, the dominant strategy designs few-shot models following the meta-learning paradigm. However, these approaches typically operate under the assumption that KGs are available instantly, disregarding the newly emerging relations during KG enrichment. We propose a novel framework designed to equip the few-shot model with the ability to learn sequentially from novel relations through data-level rehearsal and model-level modulation to address catastrophic forgetting, alongside multi-view relation augmentation aimed at resolving the issue of insufficient novel relations.

    @inproceedings{li2024learning,
      title = {Learning from novel knowledge: Continual few-shot knowledge graph completion},
      author = {Li, Zhuofeng and Zhang, Haoxiang and Zhang, Qiannan and Kou, Ziyi and Pei, Shichao},
      booktitle = {Proceedings of the 33rd ACM International Conference on Information and Knowledge Management},
      pages = {1326--1335},
      month = oct,
      year = {2024},
      selected = true,
    }