Echoes fool RL search agents: 8 paraphrased false claims overturn 58-73% of correct answers
_reachsumit · x · 2026-11-18
The EchoSpec paper exposes a systematic weakness in RL-trained search agents: paraphrased copies of a false claim are treated as independent sources, while the single primary source gets displaced. Across five base models (3B-32B) and seven released RL agents on multi-hop Wikipedia questions, eight paraphrased echoes overturn 58-73% of previously correct answers, and exact-match RL drives deleted-evidence abstention below 15%. The failure localizes to evidence acquisition — once primary evidence enters context, agents overrule eight echoes as reliably as one — and bigger budgets, forced search turns, or passive filtering don't recover displaced evidence. EchoSpec trains in two stages (verification warm-start, then group RL with an evidence-gated λ-reward) across Qwen2.5-3B/7B and Llama-3.1-8B backbones on nine suites.
More from coding & agent
- 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
- DetouR follows hyperlinks for multi-hop QA, adding 4.0 EM on 2Wiki with zero extra calls — _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
- Building a Claude skill for anime-style emotional framing techniques — andrew_n_carr · 2026-11-18