AgentIR retriever reads agent thinking tokens, lifts BrowseComp-Plus to 67% from 35%
CShorten30 · x · 2026-10-06
- The author will present AgentIR at COLM 2025 (Oct 7): a retriever that leverages an agent's explicit reasoning tokens to understand what the agent is "thinking".
- Unlike humans, agents expose their chain of thought directly; AgentIR puts those tokens to work for retrieval decisions.
- Results on BrowseComp-Plus: 35% (BM25) → 50% (Qwen3-Embedding) → 67% (AgentIR).
- The team reports early adoption from search providers (LightOnIO), vector DB vendors (Weaviate), and teams training search agents.
More from coding & agent
- Dev builds native Mac app to message all your AI agents at once, seeking testers — charlieholtz · 2026-10-06
- Persona Policies (PPol): evolutionary framework for diverse LLM user simulators — natashajaques · 2026-10-06
- Dev uses GitHub Copilot app to generate a rebuildable 60s hype video via PR — DanWahlin · 2026-10-06
- Oxylabs raises $130M at $3.6B valuation, first outside capital, betting on AI agents — Ronangmi · 2026-10-06
- PM open-sources GitDocket: manage AI coding epics, tasks and wikis in Git — bjorgen · 2026-10-06
- Amelia Wattenberger: AI is still in its terminal era, and GUI history holds lessons — Wattenberger · 2026-10-06