주재원 retweeted
I'm in team matching for @Google's Summer 2027 Research Intern position! I build reliable & safe multi-agent LLM systems and work on automated research agents. Open to related topics as well as new directions!
Potential hosts, let's connect! DMs/emails welcome; reposts very much appreciated 🙏
==========More about me==========
🌐 Homepage: froilanchoi.com
📄 CV: drive.google.com/file/d/19og…
✉️ email: [email protected]
===============================
#Google #GoogleResearch #ResearchInternship #LLMAgents #MultiAgentSystems #AIResearch
💡 Millions of shared skills. Why should agents keep starting from scratch?
Excited to share RASO! 🚀
Retrieve shared skill knowledge, adapt it to the target harness, and use it for rollout-free initialization and feedback-driven updates.
#AgenticAI #Skill #Harness
5/5 🛰️ → 📊 From satellite imagery to spreadsheets
A retrieved imagery skill handles empty groups before aggregation. Adapting this procedure helps a spreadsheet agent leave missing results blank and avoid shifting values into the wrong rows.
Useful knowledge can cross domains
4/5 📈 Performance
Across 4 benchmarks × 2 LLMs, RASO outperforms TextGrad, GEPA, SkillOpt, and WikiSkill, averaged over 3 seeds.
Gains over the strongest competing baseline reach +6.31 pp on SpreadsheetBench and +11.30 pp on WebShop.
3/5 🔧 Cross-Harness Adaptation
Relevant knowledge may come from skills written for entirely different tools and environments.
We extract the transferable procedure and re-express it using the target harness’s tools, objects, and constraints.
2/5 🔄 RASU: Retrieval-Augmented Skill Update
Observing a failure doesn’t always tell you how to fix it.
RASU uses failed trajectories to guide retrieval, bringing external knowledge into iterative skill refinement alongside execution feedback.
1/5 🚀 RASI: Retrieval-Augmented Skill Initialization
Why wait for expensive rollouts to build a useful skill?
RASI retrieves procedural knowledge using only task and harness descriptions, then adapts and combines it into an initial skill—with zero agent rollouts.
🎉 Thrilled to announce that AgentGrad is accepted to #NeurIPS2026!
🎯 Intervention-guided prompt optimization for multi-agent systems
📈 SOTA: 5 benchmarks × 2 LLMs
⚡ 2.5× faster | 💰 21.8% lower cost
📄 [arxiv.org/pdf/2609.08572]
#MultiAgentSystem #Harness
주재원 retweeted
We are presenting our work "Inversion-based Latent Bayesian Optimization" in Vancouver, NeurIPS'24!
Time: 12/11 (Wed) 11 am - 2 pm PST
Location: West Ballroom A-D
Poster ID: 6205
Paper: arxiv.org/pdf/2411.05330v1
Video: youtu.be/I7qxl0VsaIQ?si=dB5Y…
#NeurIPS2024 #BayesianOptimization
주재원 retweeted
We are sharing our NeurIPS 23 paper "CoBO."
Paper: arxiv.org/pdf/2310.20258.pdf
Video: youtube.com/watch?v=h0ZLEFax…
We are gonna presenting our paper at NeurIPS 23 in New Orleans on 12/12 at 10:45am, Poster number 1111.
See you soon!
#NeurIPS2023 #BO #GP #MachineLearning