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.

Original post →

More from coding & agent

coding & agent channel →