Bin Wu ✈️ ACL 2026

Looking forward to presenting our works!

@TEKnologyy Thanks, Danny! Really happy to co-organize the SIGIR workshop with you!

@binwu_cs Awesome Bin!

Excited to attend #ACL2026, #ICML2026 & #SIGIR2026 this July! 🎉 Presenting 3 papers at ACL & ICML and organizing the AgentSearch Workshop at SIGIR. If you’re attending too, let’s connect! Looking forward to meeting you there.

@NegarEmpr Excited! Looking forward to meeting you in person, Negar!

@binwu_cs See you in ICML and SIGIR then! 😀

📚SIGIR 2026 AgentSearch Workshop: Indexing, Retrieval, and Ranking of AI Agents agent-search.github.io/agentsearch-si…

📚ICML 2026 From Interaction Trajectories to Prompt Rules: Credit Assignment for Multi-Agent Prompt Optimization openreview.net/pdf?id=pXF0puB…

📚 ACL 2026 Beyond Static Toolsets: Self-Evolving LLM Tool Agents via Continual Documentation Adaptation aclanthology.org/2026.findings-… Structured Dialogue Refinement: Building Retrieval-Augmented Question Answering on Goal-Oriented Dialogues aclanthology.org/2026.findings-…

I’m a PhD student at UCL @uclcs , affiliated with the UCL NLP Group @ucl_nlp and the UCL Centre for AI @ai_ucl . My research focuses on LLM agents, orchestration systems, agent evaluation & diagnosis, and agent/tool search. Find more on my personal website.

@emine_yilm @qzhang_cs @QSFByte Also excited about: “InnoEval: On Research Idea Evaluation as a Knowledge-Grounded, Multi-Perspective Reasoning Problem” arxiv.org/pdf/2602.14367 #NLP #AIResearch

Two papers accepted to ICML 2026 🎉 Grateful to Emine Yilmaz @emine_yilm , and Qiang Zhang @qzhang_cs for their guidance, and Shuofei Qiao @QSFByte for leading the InnoEval work. More on my work below 👇 #ICML2026 #MachineLearning #LLM #NLP #AIAgent

@emine_yilm @qzhang_cs @QSFByte 1st author Paper: “From Interaction Trajectories to Prompt Rules: Credit Assignment for Multi-Agent Prompt Optimization” → Trajectory-aware credit assignment for multi-agent systems → Enables localized, iterative prompt rule refinement #MultiAgentSystems #LLM #AIAget

Resources: 🌐 Homepage: bingo-w.github.io/AgentSearchBen… 💻 Code & data: github.com/Bingo-W/AgentS… 🤗 Benchmark: huggingface.co/AgentSearch

🚀 New paper: AgentSearchBench — a benchmark for AI agent search in the wild. As agent ecosystems grow, a fundamental question emerges: how do we find the right agent for a task? We build a benchmark with ~10k real-world agents and show: • Semantic similarity ≠ agent performance • Description-based ranking often underestimates capable agents • Execution-aware probing improves ranking 📄 Paper: arxiv.org/abs/2604.22436 Joint work with Arastun Mammadli @arastunmammadli (co-first author) and Xiaoyu Zhang, Emine Yilmaz @emine_yilm at the UCL AI Center @ai_ucl and UCL Computer Science @uclcs #AI #Agents #LLM #InformationRetrieval

Excited to share that the deadline for our AgentSearch Workshop @ SIGIR 2026 has been extended to 📅 𝗠𝗮𝘆 𝟴, 𝟮𝟬𝟮𝟲 (𝗔𝗼𝗘) If you work on retrieval, AI agents, or tool use, we’d love to see your submissions. 🔗 agent-search.github.io/agentsearch-si…

Excited to share that 3 of my papers are accepted to #ACL2026 🎉 I’m fortunate to contribute as a main author (1st/2nd author) across these works: • 𝗕𝗲𝘆𝗼𝗻𝗱 𝗦𝘁𝗮𝘁𝗶𝗰 𝗧𝗼𝗼𝗹𝘀𝗲𝘁𝘀: 𝗦𝗲𝗹𝗳-𝗘𝘃𝗼𝗹𝘃𝗶𝗻𝗴 𝗟𝗟𝗠 𝗧𝗼𝗼𝗹 𝗔𝗴𝗲𝗻𝘁𝘀 𝘃𝗶𝗮 𝗖𝗼𝗻𝘁𝗶𝗻𝘂𝗮𝗹 𝗗𝗼𝗰𝘂𝗺𝗲𝗻𝘁𝗮𝘁𝗶𝗼𝗻 𝗔𝗱𝗮𝗽𝘁𝗮𝘁𝗶𝗼𝗻 (Findings) Follow-up to my ACL work last year (paper: aclanthology.org/2025.findings-…), we study how LLM agents adapt to evolving tool ecosystems via continual documentation adaptation. • 𝗦𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗲𝗱 𝗗𝗶𝗮𝗹𝗼𝗴𝘂𝗲 𝗥𝗲𝗳𝗶𝗻𝗲𝗺𝗲𝗻𝘁: 𝗕𝘂𝗶𝗹𝗱𝗶𝗻𝗴 𝗥𝗔𝗚 𝗼𝗻 𝗚𝗼𝗮𝗹-𝗢𝗿𝗶𝗲𝗻𝘁𝗲𝗱 𝗗𝗶𝗮𝗹𝗼𝗴𝘂𝗲𝘀 (Findings) My internship work at Bloomberg — we adapt goal-oriented dialogues for retrieval-augmented QA. • 𝗠𝗶𝘁𝗶𝗴𝗮𝘁𝗶𝗻𝗴 𝗖𝗼𝗻𝘁𝗲𝘅𝘁 𝗜𝗻𝘁𝗲𝗿𝗳𝗲𝗿𝗲𝗻𝗰𝗲 𝗳𝗼𝗿 𝗥𝗲𝗹𝗶𝗮𝗯𝗹𝗲 𝗮𝗻𝗱 𝗘𝗳𝗳𝗶𝗰𝗶𝗲𝗻𝘁 𝗦𝗲𝗮𝗿𝗰𝗵 𝗔𝗴𝗲𝗻𝘁𝘀 (Main) We investigate context interference in multi-turn search agents and propose a refinement-based solution. Also sharing: 𝗣𝗥𝗘𝗙: 𝗥𝗲𝗳𝗲𝗿𝗲𝗻𝗰𝗲-𝗙𝗿𝗲𝗲 𝗘𝘃𝗮𝗹𝘂𝗮𝘁𝗶𝗼𝗻 𝗼𝗳 𝗣𝗲𝗿𝘀𝗼𝗻𝗮𝗹𝗶𝘀𝗲𝗱 𝗧𝗲𝘅𝘁 𝗚𝗲𝗻𝗲𝗿𝗮𝘁𝗶𝗼𝗻 𝗶𝗻 𝗟𝗟𝗠𝘀 (not accepted this time) A framework for evaluating personalised LLM outputs without requiring references Grateful to my supervisors, collaborators, and teammates, especially my supervisor Emine @emine_yilm and Edgar @edgarmeij , my internship team lead @myahya , and my collaborators Boyang @BeyondHsueh 🙏 #ACL2026 #NLP #LLM #AIAgents #ToolUse #RAG #MachineLearning #AIResearch

@_sambhavkothari That’s an insightful observation. It makes me think CLI and code-MCP function more like planners, while token-MCP acts more like a task assigner. An orchestration layer could be the ideal way to bring out the best of both worlds.

@jmin__cho @NVIDIAAI @NVIDIAAIDev @JHUCompSci @HopkinsDSAI Congratulations!

Thanks @NVIDIAAI & @NVIDIAAIDev for supporting our research on multimodal, embodied AI through the NVIDIA Academic Grant Program! #NVIDIAGrant @JHUCompSci @HopkinsDSAI

@lateinteraction @CAISconf Really excited to see a new conference dedicated to agentic systems — and delighted to be serving as a PC member this year!

The inaugural ACM CAIS 2026 conference @CAISconf on AI systems is soliciting additional reviewers. If you are interested in reviewing, please fill out this short form. Reviewers should expect a load of 4-5 papers, with reviews due April 3, 2026. forms.gle/d3Dx3zRXPzLyod…

@HuaxiuYaoML Excited to see the combination of meta learning and agnetic system. Thanks for sharing, Huaxiu.

MetaClaw [v0.3.2] now supports more personal agents: OpenClaw, CoPaw, IronClaw, PicoClaw, ZeroClaw, and more. Let your agent learn and evolve from every conversation. On 346 real-world tasks over 30 simulated days: Kimi-K2.5 accuracy 21.4% → 40.6% (+90%), task completion 8.25×. Meta-learn your🦞 from every conversation in the wild. Detailed technical report dropping in 3 days. Stay tuned 🔥 Link to repo in 🧵👇 @richardxp888, @JimChenjw, @Xinyu2ML, @lillianwei423, @StephenQS0710, @HaoqinT, @JiaqiLiu835914, @zhengop, @cihangxie

@TengX6 Amazing work! Thanks for your sharing, Teng.

🚀 New work: Meta-Reinforcement Learning with Self-Reflection LLM agents shouldn't just solve problems. They should learn from their own attempts. Most current RL methods optimize single independent trajectories. Each attempt starts from scratch, with no mechanism to improve across attempts. But intelligent systems should get better after trying once. This raises a fundamental question: How do we train models to learn from their own attempts? We believe Meta-Reinforcement Learning may be a key paradigm for training future LLM agents, enabling models to adapt and improve across attempts and environments. In this work we introduce MR-Search, a training paradigm built around: 🧠 In-Context Meta-Reinforcement Learning 🪞 Self-Reflection 🔁 Learning to learn at test time 📄 Paper: arxiv.org/abs/2603.11327 💻 Code: github.com/tengxiao1/MR-S…