DetouR follows hyperlinks for multi-hop QA, adding 4.0 EM on 2Wiki with zero extra calls
_reachsumit · x · 2026-11-18
arXiv paper DetouR addresses a blind spot in multi-hop QA search agents: they only follow semantic routes (queries the model can name), while decisive evidence often sits one hyperlink away from already-retrieved pages. The plug-in reuses links arriving with retrieved pages, adding no search or LLM calls: Link Rerouting swaps near-duplicate queries for routes to linked pages via a lightweight learned scorer, and Chain-aware Context Selection runs personalized PageRank so evidence chains reach the reader intact. Across four agents it improves EM by 4.0 points on 2Wiki and 1.4 on HotpotQA on average. A counterintuitive finding: larger LLMs struggle with this non-semantic evidence, adding little over simple selectors or even falling behind.
More from coding & agent
- BizPeel adds 48 live competitor-research tools for Claude Code at under 2 cents per lookup — yihui_indie · 2026-11-18
- Ant Group's Repo-Premium shows BM25 matches dense retrieval in fair code-search tests — _reachsumit · 2026-11-18
- Scoble Tests Hy4 Agent in WorkBuddy on a 20-Step Real Office Job — Scobleizer · 2026-11-18
- Echoes fool RL search agents: 8 paraphrased false claims overturn 58-73% of correct answers — _reachsumit · 2026-11-18
- rust-anything rewrites legacy projects to Rust with AI, up to 20x faster — aigclink · 2026-11-18
- Jetriever ranks agent memories by utility not similarity, hitting gold in 84.8% of cases — _reachsumit · 2026-11-18