EMNLP 2026

Masking Stale Observations Helps Search Agents – Until It Doesn't: A Regime Map and Its Mechanism

Haoxiang Zhang^{*Equal contribution}, Qixin Xu^{*Equal contribution}, Yu Zhang^{✝Co-senior author}, Julian McAuley^{✝Co-senior author}, and 3 more authors Haoxiang Zhang^{*Equal contribution}, Qixin Xu^{*Equal contribution}, Zhuofeng Li, Lei Zhang, Pengcheng Jiang, Yu Zhang^{✝Co-senior author}, Julian McAuley^{✝Co-senior author}

Paper | Code | Abstract | BibTeX

GitHub ⭐

Long-horizon search agents accumulate large amounts of retrieved content across many tool calls, making context-budget efficiency increasingly important. We study stale-observation masking across agent backbones from 4B to 284B parameters and multiple retrievers. Its accuracy gain follows an asymmetric inverted-U: masking adds little under weak retrievers, helps most when strong retrieval is paired with a mid-capacity model, and can hurt saturated models by removing evidence they would otherwise use. The results frame context management as a regime-dependent intervention shaped jointly by retriever recall and a model's implicit filtering capacity.

@article{zhang2026masking, title={Masking Stale Observations Helps Search Agents -- Until It Doesn't: A Regime Map and Its Mechanism}, author={Zhang, Haoxiang and Xu, Qixin and Li, Zhuofeng and Zhang, Lei and Jiang, Pengcheng and Zhang, Yu and McAuley, Julian}, journal={arXiv preprint arXiv:2606.00408}, year={2026} }

NIPS 2026 Spotlight

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

Zhuofeng Li, Haoxiang Zhang, Dongfu Jiang^{✝Corresponding author}, Yu Zhang^{✝Corresponding author}, and 15 more authors 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^{✝Corresponding author}, Yu Zhang^{✝Corresponding author}

Paper | Code | Demo | Abstract | BibTeX

GitHub ⭐

Modern retrieval systems expose a corpus through a fixed similarity interface that compresses access into a single top-k retrieval step before reasoning, limiting agentic search capabilities requiring exact lexical matching, sparse clue combinations, and iterative hypothesis refinement. We propose Direct Corpus Interaction (DCI), enabling agents to search raw corpora using general-purpose tools like grep and shell commands without embedding models or vector indices. DCI substantially outperforms traditional sparse, dense, and reranking approaches on BRIGHT and BEIR benchmarks, while achieving strong results on BrowseComp-Plus and multi-hop question answering tasks.

@article{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}, journal={arXiv preprint arXiv:2605.05242}, year={2026} }

arXiv 2026

Dr-DCI: Scaling Direct Corpus Interaction via Dynamic Workspace Expansion

Yi Lu, Zhuofeng Li, Yu Zhang^{✝Corresponding author}, and 7 more authors Yi Lu, Zhuofeng Li, Ping Nie, Haoxiang Zhang, Yuyu Zhang, Kai Zou, Wenhu Chen, Jimmy Lin, Dongfu Jiang^{✝Corresponding author}, Yu Zhang^{✝Corresponding author}

Paper | Code | Abstract | BibTeX

GitHub ⭐

DR-DCI scales Direct Corpus Interaction by making retrieval an agent-callable action that dynamically expands a local workspace. Agents retrieve relevant documents as needed, then use shell-based DCI operations inside that workspace for precise filtering, comparison, and verification. This design preserves scalable corpus exploration while avoiding the latency and instability of running terminal commands over the full corpus. Across BrowseComp-Plus and large-scale Wiki-18 QA evaluations, DR-DCI improves accuracy and efficiency over raw DCI, retrieval-only methods, and trained search-agent baselines, while remaining effective as corpus size grows from 100K to tens of millions of documents.

@article{lu2026drdci, title={Dr-DCI: Scaling Direct Corpus Interaction via Dynamic Workspace Expansion}, author={Lu, Yi and Li, Zhuofeng and Nie, Ping and Zhang, Haoxiang and Zhang, Yuyu and Zou, Kai and Chen, Wenhu and Lin, Jimmy and Jiang, Dongfu and Zhang, Yu}, journal={arXiv preprint arXiv:2606.14885}, year={2026} }

COLM 2026

CocoaBench: Evaluating Unified Digital Agents in the Wild

Shibo Hao, Zhining Zhang, Zhiqi Liang, Haoxiang Zhang, and 27 more authors Zhiqi Liang, Tianyang Liu, Yuheng Zha, Qiyue Gao, Jixuan Chen, Zilong Wang, Zhoujun Cheng, Haoxiang Zhang, Junli Wang, Hexi Jin, Boyuan Zheng, Kun Zhou, Yu Wang, Feng Yao, Licheng Liu, Yijiang Li, Zhifei Li, Zhengtao Han, Pracha Promthaw, Tommaso Cerruti, Xiaohan Fu, Ziqiao Ma, Jingbo Shang, Lianhui Qin, Julian McAuley, Eric P. Xing, Zhengzhong Liu, Rupesh Kumar Srivastava, Zhiting Hu

Homepage | Paper | Code | Blog | Examples | Abstract | BibTeX

GitHub ⭐

Current LLM agents excel in isolated domains like software engineering and GUI automation, but evaluations rarely test integrated capabilities. CocoaBench introduces 153 human-designed, long-horizon tasks requiring flexible composition of vision, search, and coding skills. Tasks are specified only by an instruction and an automatic evaluation function over the final output, enabling reliable and scalable evaluation. Top systems achieve only 45.1% success, revealing significant gaps in reasoning, tool execution, and visual understanding.

@article{hao2026cocoabench, title={CocoaBench: Evaluating Unified Digital Agents in the Wild}, author={Hao, Shibo and Zhang, Zhining and Liang, Zhiqi and Liu, Tianyang and Zha, Yuheng and Gao, Qiyue and Chen, Jixuan and Wang, Zilong and Cheng, Zhoujun and Zhang, Haoxiang and Wang, Junli and Jin, Hexi and Zheng, Boyuan and Zhou, Kun and Wang, Yu and Yao, Feng and Liu, Licheng and Li, Yijiang and Li, Zhifei and Han, Zhengtao and Promthaw, Pracha and Cerruti, Tommaso and Fu, Xiaohan and Ma, Ziqiao and Shang, Jingbo and Qin, Lianhui and McAuley, Julian and Xing, Eric P. and Liu, Zhengzhong and Srivastava, Rupesh Kumar and Hu, Zhiting}, journal={arXiv preprint arXiv:2604.11201}, year={2026} }

ACL 2026

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

Zhuofeng Li^{*Equal contribution}, Yi Lu^{*Equal contribution}, Dongfu Jiang, Haoxiang Zhang, Yu Zhang^{✝Corresponding author}, and 5 more authors Dongfu Jiang, Haoxiang Zhang, Yuyang Bai, Chuan Li, Yu Wang, Shuiwang Ji, Jianwen Xie^{✝Corresponding author}, Yu Zhang^{✝Corresponding author}

Homepage | Paper | Code | Demo | Abstract | BibTeX

Peer reviews generated by language models tend to be superficial, lacking paper-specific evidence. We introduce REVIEWBENCH, a benchmark that evaluates review text according to paper-specific rubrics derived from official guidelines, the paper's content, and human-written reviews. We also propose REVIEWGROUNDER, a framework that decomposes reviewing into drafting and grounding stages, enriching shallow drafts via targeted evidence consolidation. REVIEWGROUNDER outperforms larger models including GPT-4.1 and DeepSeek-R1-670B across multiple quality 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={Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics}, year={2026} }

EMNLP 2026

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

Zhuofeng Li^{*Equal contribution}, Dongfu Jiang^{*Equal contribution}, Xueguang Ma, Haoxiang Zhang, and 6 more authors Dongfu Jiang^{*Equal contribution}, Xueguang Ma, Haoxiang Zhang, Ping Nie, Yuyu Zhang, Kai Zou, Jianwen Xie, Yu Zhang^{✝Corresponding author}, Wenhu Chen^{✝Corresponding author}

Paper | Code | Dataset | Model | Demo | Video | Eval Logs | Abstract | BibTeX

Hugging Face Daily Paper #2 • Adopted by NVIDIA's Nemotron family of models
GitHub ⭐

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. Supervised fine-tuning a 30B-A3B backbone on these trajectories achieves 54.8% accuracy on BrowseComp-Plus, a +34.0 point improvement over the base model, while remaining competitive on BrowseComp, GAIA, and xbench-DeepSearch. Because the environment is offline and fully instrumented, it also enables controlled analysis, where our study reveals practical insights into deep research pipeline design, including data filtering strategies, agent configuration choices, and how retrieval success relates to final answer accuracy.

@article{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}, journal={arXiv preprint arXiv:2603.20278}, year={2026} }

bioRxiv 2026

Eubiota: Modular Agentic AI for Autonomous Discovery in the Gut Microbiome

Pan Lu^{*Equal contribution}, Yifan Gao^{*Equal contribution}, William G. Peng, Haoxiang Zhang, and 11 more authors William G. Peng, Haoxiang Zhang, Kunlun Zhu, Elektra K. Robinson, Qixin Xu, Masakazu Kotaka, Harrison G. Zhang, Bingxuan Li, Anthony L. Shiver, Yejin Choi, Kerwyn Casey Huang, Justin L. Sonnenburg^{✝Corresponding author}, James Zou^{✝Corresponding author}

Website | Paper | App | Code | HuggingFace | Discord | Abstract | BibTeX

GitHub ⭐

Eubiota is an open-source modular agentic AI framework for autonomous discovery in the gut microbiome. It features specialized agents for planning, execution, verification, and synthesis that collaborate through shared memory and domain-specific tools. Using reinforcement learning to enhance multi-turn reasoning, Eubiota achieves 87.7% benchmark accuracy, outperforming GPT-4.1 by 10.4%. The system identified the uvr-ruv DNA repair pathway critical for bacterial fitness under inflammatory stress, designed a four-strain bacterial consortium that reduced colitis severity in mice, created an antibiotic combination that preserves beneficial gut bacteria, and discovered diet-associated metabolites that suppress NF-κB signaling.

@article{lu2026eubiota, title={Eubiota: Modular Agentic AI for Autonomous Discovery in the Gut Microbiome}, author={Lu, Pan and Gao, Yifan and Peng, William G. and Zhang, Haoxiang and Zhu, Kunlun and Robinson, Elektra K. and Xu, Qixin and Kotaka, Masakazu and Zhang, Harrison G. and Li, Bingxuan and Shiver, Anthony L. and Choi, Yejin and Huang, Kerwyn Casey and Sonnenburg, Justin L. and Zou, James}, journal={bioRxiv}, year={2026}, doi={10.64898/2026.02.27.708412}, url={https://www.biorxiv.org/content/10.64898/2026.02.27.708412v1} }

ICLR 2026 Oral & NIPS 2025 ER Workshop

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

Zhuofeng Li^{*Equal contribution}, Haoxiang Zhang^{*Equal contribution}, James Zou^{✝Corresponding author}, Pan Lu^{✝Corresponding author}, and 5 more authors Seungju Han, Sheng Liu, Jianwen Xie, Yu Zhang, Yejin Choi, James Zou^{✝Corresponding author}, Pan Lu^{✝Corresponding author}

Paper | Homepage | Demo | Code | Slack | Abstract | BibTeX

Hugging Face Daily Paper #2 • GitHub ⭐

This paper introduces AgentFlow, a trainable agentic framework that coordinates four specialized modules through evolving memory to optimize planning within multi-turn interactions. The authors propose Flow-GRPO, a training method that converts long-horizon optimization into tractable single-turn updates. Testing across ten benchmarks shows the 7B-scale model achieves substantial improvements, including 14.9% on search, 14.0% on agentic, 14.5% on mathematical, and 4.1% on scientific tasks, even surpassing GPT-4o on some evaluations.

@article{li2025flow, 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}, journal={arXiv preprint arXiv:2510.05592}, year={2025} }

TMLR

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

Haoxiang Zhang, Zhuofeng Li, Shichao Pei^{✝Corresponding author}, and 3 more authors Qiannan Zhang, Ziyi Kou, Juncheng Li, Shichao Pei^{✝Corresponding author}

Paper | Code | Abstract | BibTeX

Pre-trained graph neural networks (GNNs) have demonstrated significant success in leveraging large-scale graph data to learn transferable representations. However, their performance often degrades under distribution shifts, particularly in real-world scenarios where test labels are unavailable. To address this challenge, we propose Graph Optimization via Augmented Transformations (GOAT), a novel self-supervised test-time tuning paradigm that adapts pre-trained GNNs to distribution-shifted test data by focusing exclusively on node feature transformations. By avoiding complex and often suboptimal graph structure transformations, GOAT overcomes the limitations of existing data-centric methods. To further address the issue of transformation collapse, where feature transformations converge to trivial solutions—such as when the test-time learned data-centric transformation degenerates into a constant or identity mapping across different inputs, we introduce a parameter-efficient low-rank adapter that generates diverse transformations tailored to individual input graphs. This design not only enhances adaptation performance but also improves interpretability by avoiding modifications to the graph structure. Through extensive experiments on six real-world datasets with diverse distribution shifts, we demonstrate that GOAT achieves consistent performance improvements across different pre-trained GNN backbones, outperforming state-of-the-art test-time adaptation methods.

@article{zhang2025avoiding, title={Avoiding Structural Pitfalls: Self-Supervised Low-Rank Feature Tuning for Graph Test-Time Adaptation}, author={Haoxiang Zhang and Zhuofeng Li and Qiannan Zhang and Ziyi Kou and Juncheng Li and Shichao Pei}, journal={Transactions on Machine Learning Research}, issn={2835-8856}, year={2025}, url={https://openreview.net/forum?id=yiS6q42LLt}, note={} }

CIKM 2024

Learning from Novel Knowledge: Continual Few-shot Knowledge Graph Completion

Zhuofeng Li^{*Equal contribution}, Haoxiang Zhang^{*Equal contribution}, Shichao Pei^{✝Corresponding author}, and 2 more authors Qiannan Zhang, Ziyi Kou, Shichao Pei^{✝Corresponding author}

Paper | Code | Abstract | BibTeX

This work addresses the challenge of continual few-shot learning in knowledge graph completion, enabling models to learn from novel knowledge while preserving previously learned information.

@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}, year={2024} }

Dataset

Sci-Bench-AIME25: A Multi-Modal Chain-of-Thought Dataset for Advanced Tool-Intergrated Mathematical Reasoning

Haoxiang Zhang^{❧Corresponding author}, Siyuan Wang^{❧Corresponding author}, Xueji Fang, and 5 more authors Xueji Fang, Xinkai Zou, Tiange Lyu, Siyuan Cao, Jingyuan Huang, Jie Xing

Dataset | Abstract | BibTeX

A comprehensive multi-modal dataset designed for advancing mathematical reasoning capabilities through chain-of-thought methodologies and tool integration.

@dataset{zhang2025scibench_aime25, title = {{Sci-Bench-AIME25}: A Multi-Modal Chain-of-Thought Dataset for Advanced Tool-Intergrated Mathematical Reasoning}, author = {Zhang, Haoxiang and Wang, Siyuan and Fang, Xueji and Zou, Xinkai and Lyu, Tiange and Cao, Siyuan and Huang, Jingyuan and Xing, Jie}, year = {2025}, publisher = {Hugging Face}, doi = {10.5281/zenodo.17112481}, url = {https://huggingface.co/datasets/IPF/AIME25-CoT-CN}, note = {Also available on Zenodo} }

CIKM 2025

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

Zhijie Lin^{*Equal contribution}, Zhuofeng Li^{*Equal contribution}, Haoxiang Zhang, and 4 more authors Chenglei Dai, Wentian Bao, Enyun Yu, Haoxiang Zhang, Liang Zhao^{❧Corresponding author}

Paper | Abstract | BibTeX

A novel generative framework that enhances reranking efficiency through ordered multi-token prediction, providing improved performance in information retrieval tasks.

@article{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}, journal={arXiv preprint arXiv:2510.25220}, year={2025} }

Mathematics

SGNet: Efficient Snow Removal Deep Network with a Global Windowing Transformer

Shan Lie, Haoxiang Zhang, Bodong Cheng^{❧Corresponding author}

Paper | Abstract | BibTeX

An efficient deep learning approach for snow removal in images using a global windowing transformer architecture, achieving state-of-the-art performance in image restoration tasks.

@article{shan2024sgnet, title={SGNet: Efficient Snow Removal Deep Network with a Global Windowing Transformer}, author={Shan, Lie and Zhang, Haoxiang and Cheng, Bodong}, journal={Mathematics}, volume={12}, number={10}, pages={1424}, year={2024}, publisher={MDPI} }

×