research
The Retriever Trains on the Mistakes It Will Actually See
By the third search step, roughly half or more of visited states held a wrong passage; on-policy training taught the encoder to recover.
Summary
By the third search step, roughly half or more of visited states held a wrong passage; on-policy training taught the encoder to recover.
Multi-hop retrieval models are commonly trained only on gold evidence histories, then deployed on chains polluted by their own mistakes. This study found that by the third step, about half or more of visited states contained a wrong passage and that those errors reduced the chance of finding missing gold evidence. The proposed encoder retrieves its own training chains each epoch, labels every visited state with an earliest missing gold passage and mixes those states with teacher-forced examples. It completed evidence chains more often than the compared systems and transferred to harder constructed variants of HotpotQA and 2WikiMultiHopQA. The evidence concerns benchmark retrieval, not open-web factual reliability.
Why it matters
By the third search step, roughly half or more of visited states held a wrong passage; on-policy training taught the encoder to recover.
Limits and context
- Multi-hop retrieval models are commonly trained only on gold evidence histories, then deployed on chains polluted by their own mistakes.
- The evidence concerns benchmark retrieval, not open-web factual reliability.
Key claims
By the third search step, roughly half or more of visited states held a wrong passage; on-policy training taught the encoder to recover.
Qualification: Multi-hop retrieval models are commonly trained only on gold evidence histories, then deployed on chains polluted by their own mistakes.
Evidence: source-2026-11-01-016
Sources
- arXiv preprint 2610.40409, original v1arXiv · primary research
Corrections
No corrections have been recorded for this story.