Search agents treat eight paraphrases as eight sources; EchoSpec's two-stage training cuts eight-echo deference to 29–31%
Synopsis
Reinforcement-learned search agents treat repeated paraphrases of one claim as independent corroboration: on multi-hop Wikipedia questions with primary evidence intact, eight paraphrased echoes of a false claim overturn 58–73% of previously correct answers and drive abstention on deleted-evidence questions below 15%. EchoSpec specifies reliability across five corpus transformations, warm-starts the policy on bridge-entity verification trajectories, then trains with evidence-gated group reinforcement learning; on Qwen2.5-7B it cuts eight-echo deference to 29–31% while raising eight-echo evidence reach to 26–27%. The result holds across Qwen2.5-3B, Qwen2.5-7B and Llama-3.1-8B backbones and transfers zero-shot to unseen BGE and BM25 retrievers.
Interpretation
When a search agent meets several paraphrases of one claim, it treats them as independent sources and abandons an answer it had right. On multi-hop Wikipedia questions with the primary evidence still in the corpus, eight paraphrased echoes of a false claim overturn 58–73% of answers correct on the clean corpus, and exact-match reinforcement learning pushes abstention on deleted-evidence questions from 47% to below 15%. The comparison spans five base models from 3B to 32B and seven released reinforcement-learned search agents, and Search-R1 defers more than the 7B base model it was trained from at every echo count.
The failure sits in evidence acquisition rather than in judging evidence once it is in hand. Once the primary passage enters the context, deference at eight echoes is 25–34%, comparable to 27–39% at a single echo, whereas deference reaches 70–85% at eight echoes when the primary evidence stays out of view. Echoes fill the top retrieval ranks and end the search after one or two queries, cutting the share of trajectories that reach the primary evidence from 53–72% at one echo to 20–31% at eight, and all three tested agents show the pattern.
EchoSpec writes what to trust and when to answer into the search policy itself rather than filtering passages already fetched. On Qwen2.5-7B, eight-echo deference falls to 29–31%, 30–44 points below the 58.6–72.9% of seven released reinforcement-learned agents, while eight-echo evidence reach rises from 7.1% to 26–27% and standard QA precision rises to 47–55%. Eight-echo deference converges to 28.9–30.6% across three backbones, frozen checkpoints transfer zero-shot to BGE at 29.4% and BM25 at 25.1%, and larger budgets or forced multi-turn search remain at 55.5–61.8%.
Training runs in two stages: first learning to query bridge entities past echoes, then learning when to answer and when to abstain under an evidence-gated reward. Without the first-stage warm start, eight-echo evidence reach falls to 9.4% and deference returns to 69.4%; without the second-stage reinforcement learning, deference is 47.2%. Stage one fine-tunes LoRA adapters on multi-hop trajectories synthesized with an article-chain hint, and stage two runs group reinforcement learning on a reward that discounts unverified echo matches and prices wrong answers, with an outcome-independent acquisition bonus and scale-preserving advantages that keep the abstention threshold.
Perspective
The demonstrated setting is the 21M-passage wiki-18 Wikipedia index, e5-base-v2 dense retrieval and BM25 lexical retrieval, at most four searches per question, and multi-hop QA test sets built by two generator families, Qwen3-32B and Phi-4; the authors note that live web search adds dynamic rankings, search-engine-optimization manipulation and compromised sources, and that world construction requires multi-hop items with known evidence.
The authors identify advantage normalization in standard group reinforcement learning as the step that erases the abstention threshold, but the proposition characterizes the Bayes-optimal stopping payoff for a fixed risk penalty rather than guaranteeing calibrated posteriors after finite-sample training. The cross-generator comparison shows standalone eight-echo deference higher on Test Set B (37.7–38.8%) than on Test Set A (19.0–23.6%) while evidence reach transfers closely; nearly the whole gap appears when the primary evidence is not retrieved, which the authors read as fallback abstention on unretrieved items still benefiting from familiarity with training-generator phrasing.
