ITER: Interaction-aware retrieval improves deep-research agents
_reachsumit · x · 2026-08-31
Paper presents ITER, a dense retriever designed for deep-research agents that conditions search on prior interactions.
- Problem: Existing retrievers rely only on the current sub-query, ignoring history and visited docs, leading to redundant results.
- Solution: ITER encodes the main question, prior sub-queries, and visited docs using trajectory-relative learning signals.
- Results: Outperforms LRAT by 7.5% on InfoSeek-Eval and 13.5% on BrowseComp-Plus, with strong cross-agent robustness.
More from coding & agent
- Open-Source GitHub AI Code Reviewer: Self-Hosted Models, No SaaS Fees — Arindam_1729 · 2026-08-31
- Library recreates Claude Code, Codex, and Grok interfaces as shadcn components — RichardsonDx · 2026-08-31
- The Post AI Data Stack: Infrastructure Challenges — sh_reya · 2026-08-31
- OpenClaw 2.0 Released: Rebuilds Installation and Browser Experience — heyneighbor · 2026-08-31
- MCP Server Lets LLM Agents Play Civilization VI Natively — jia_seed · 2026-08-31
- Hermes Agent running GLM-5.2 via OpenRouter on a 2GB VPS — granawkins · 2026-08-31