🔥Excited to share our latest work: Agent KB .It achieves new open-source SOTA on the GAIA benchmark!. We enable agents to learn from each other's experiences across tasks through hierarchical experience sharing. Paper📜: Code🧑‍💻:

volunteer at Google booth #ICML25 as a research intern and then grab a drink!

Great work created by @YilunZhao_NLP @armancohan.

SciArena is such a cool platform: 🔬. Vote on base model outputs for real scientific literature tasks—long-form, citation-grounded answers based on actual papers. I’ve found it super useful. Go explore, vote, and give feedback! 🧠📚 .#AI4Science #AI2

🚀 Just discovered an amazing app for vibe coding! @_akhaliq .✨ Describe your app in plain English → Get complete HTML/CSS/JS code.🖼️ Upload UI designs → Auto-generate code.🌐 One-click deploy and Free to use with multiple LLMs! 👉 .

RT @ApollonVisual: @rohanpaul_ai those improvements are nothing to scoff at. really impressive "Claude-3.7 with Agent KB increased performa….

Thanks for sharing and promoting! @_akhaliq.

@wangchunshu @xinzhongderiyu3 @Chi_Wang_ @xingyaow_ @liujiaheng2 @espaiade @DanielStupid @richardxp888 @WUFang40615703 @GeZhang86038849 @MetaGPT_ @allhands_ai If you find this work interesting, would love your ⭐ on GitHub or 👍 on the HF paper!.

Thread 7/7.🙏 Huge thanks to all amazing collaborators: @wangchunshu @xinzhongderiyu3 @Chi_Wang_ @xingyaow_ @liujiaheng2 @espaiade @DanielStupid @richardxp888 @WUFang40615703 @GeZhang86038849 .@MetaGPT_ @allhands_ai.

Thread 6/7.🛠️ Framework-Agnostic: Agent KB works with different agent architectures (smolagents, OpenHands) and LLMs (GPT-4, Claude, o3-mini, etc.). It's designed as a modular infrastructure that any agent system can benefit from!

Thread 5/7.🔍 Key Innovation: Unlike existing memory systems that store raw execution logs, Agent KB maintains abstracted reasoning patterns that capture generalizable problem-solving strategies. This enables effective knowledge transfer even between dissimilar domains! 🎯

Thread 4/7.💻 SWE-bench Results (Code repair tasks):. Claude-3.7: 41.33% → 53.33% (+12.0pp).Significant improvements across all tested models. Our cross-domain knowledge transfer works beyond just Q&A - it enhances code understanding and debugging too!

Thread 3/7.📊 GAIA Results (Game-changing improvements!):. GPT-4.1: 55.15% → 73.94% (+18.79pp).Claude-3.7: 58.79% → 75.15% (+16.36pp).Level 3 (hardest): Claude-3.7 jumps from 38.46% to 57.69% (+19.23pp). This shows our approach excels, especially on complex reasoning tasks! 🚀

Thread 2/7.💡Our Solution: Agent KB introduces a Teacher-Student dual-phase retrieval with a novel Reason-Retrieve-Refine pipeline: .Student Agent: retrieves workflow-level patterns for overall strategy .Teacher Agent: retrieves step-level experiences for execution refinement

Thread 1/7.🤔 The Problem: Current agents can't learn from each other's experiences. They repeatedly rediscover similar problem-solving strategies when encountering new tasks, even when successful approaches from related domains could be adapted.

What we need is a healthier and more inclusive community infrastructure, not fear-based accountability through punishment.

Fear-driven reviewing can lead to anxiety and rushed, irresponsible decisions. In practice, it takes few minutes for an Area Chair to reassign a reviewer. If someone is unavailable, they should be encouraged to notify the AC early—not be coerced through punitive policies.

It's profoundly unethical for #NeurIPS to enforce reviewing obligations by threatening to desk-reject if a coauthor fails to submit reviews. Coauthors may feel pressured to urge their colleagues to rush reviews, resulting in low-quality reports.

**Biomedical Superintelligence** is the paradigm I have been actively advancing. It's a continuously evolving multi-agent system that autonomously integrates reasoning, hypothesis gen, experiment design, and feedback-driven learning across molecular, cellular, and clinical scales.

Truly embracing San Francisco culture.