Qiu, Chenjie, Zeng, Zixuan, Shi, Shaobin, Chen, Bingfeng, Luo, Dongsheng, Xu, Boyan · Oct 29, 2026
Retrieving from Your Own Mistakes: On-Policy Training of Embedding Models for Multi-Hop Retrieval-Augmented Generation
Markdown · JSON · Original source
In their paper, Chenjie Qiu and coauthors argue that a retriever for multi-hop should train on the retrieval states it reaches itself, including states containing mistakes. Chain embedding retrievers find evidence step by step, updating the question representation with passages already retrieved. The authors say training only on gold histories leaves a mismatch: at test , a retriever conditions on its own passages, and by the third step about half or more of its states contain a wrong passage. Their on-policy state training approach labels each visited state with the earliest missing gold passage, masks other missing passages, and trains on these states alongside teacher-forced ones. The authors report that this retrieves complete evidence more often than teacher-forced training and compared systems, transfers without further training to harder long-chain variants of HotpotQA and 2WikiMultiHopQA, and recovers from its own mistakes more often on the same states. The idea matters because errors early in a multi-step search can make later evidence harder to find; the paper tests whether practicing on those error-shaped states helps the retriever recover. The abstract does not give the size of the gains or enough detail to assess their practical significance. How large are the reported gains, and how do they vary across the evaluated datasets and.
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
cs.AI · Article
Source details
Original source date: Oct 29, 2026 (2026-10-29).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-13T07:03:07.331Z. First observed: 2026-10-30T06:42:53.592Z. Retrieved: 2026-10-31T09:31:07.007Z.
Referenced by
Recorded citations and public file attachments. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
The supplied evidence is the paper’s abstract, which does not state the size of the reported gains or provide detailed results by dataset or comparison system. The abstract describes harder long-chain variants of HotpotQA and 2WikiMultiHopQA that the authors construct; the supplied evidence does not describe those variants in detail.
Open question
How large are the reported gains, and how do they vary across the evaluated datasets and systems?
Source-stated
The authors say a retriever trained only on gold histories can face a mismatch at test time, when it conditions on its own retrieved passages, which may be wrong.
At test time the retriever conditions on its own retrievals, which soon contain mistakes
Source-stated
The authors report that by the third step, about half or more of the states visited contain a wrong passage, and that these passages reduce the chance of retrieving missing gold evidence.
by the third step, about half or more of the states it visits hold a wrong passage, and these passages lower its chance of retrieving a missing gold passage
Source-stated
The proposed on-policy state training trains the encoder using states reached by its current retrievals, including states built from its own mistakes.
We propose on-policy state training, which trains the encoder on the states it reaches itself.
Source-stated
The authors report that the resulting retriever retrieves complete evidence more often than teacher-forced training and the systems they compare.
The resulting retriever retrieves complete evidence more often than teacher-forced training and the systems we compare
Source-stated
The authors report transfer without further training to harder long-chain variants of HotpotQA and 2WikiMultiHopQA that they construct.
it transfers without further training to harder long-chain variants of HotpotQA and 2WikiMultiHopQA that we construct
Escape Meridian interpretation
The paper’s approach targets the problem that a retriever’s own earlier mistakes can make later missing evidence harder to retrieve.
these passages lower its chance of retrieving a missing gold passage well beyond what the difficulty of the states explains
Source and contributor context
Qiu, Chenjie, Zeng, Zixuan, Shi, Shaobin, Chen, Bingfeng, Luo, Dongsheng, Xu, Boyan · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Retrieval-augmented generation
A setup in which a model first looks up relevant documents and then answers using them.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
TIME
TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Editorial provenance
Version 1. Publication ID: d08256f4-99a6-41f9-817f-b6ef0a597c9c.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-13T07:03:07.331Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 8bdfd9e8beeb32c2897f4e18ecd0e7e6a0af4b3557f428dc243ebfae7f8bb900. Source text hash: c4b521a32436de807e84e36c0822b4ffd39efd14d51f2f3e33f7aef0faa335b4.
Revision and source-check history
Version 1 · 2026-11-13T07:03:07.331Z
Retrieving from Your Own Mistakes: On-Policy Training of Embedding Models for Multi-Hop Retrieval-Augmented Generation
In their paper, Chenjie Qiu and coauthors argue that a retriever for multi-hop should train on the retrieval states it reaches itself, including states containing mistakes. Chain embedding retrievers find evidence step by step, updating the question representation with passages already retrieved. The authors say training only on gold histories leaves a mismatch: at test , a retriever conditions on its own passages, and by the third step about half or more of its states contain a wrong passage. Their on-policy state training approach labels each visited state with the earliest missing gold passage, masks other missing passages, and trains on these states alongside teacher-forced ones. The authors report that this retrieves complete evidence more often than teacher-forced training and compared systems, transfers without further training to harder long-chain variants of HotpotQA and 2WikiMultiHopQA, and recovers from its own mistakes more often on the same states. The idea matters because errors early in a multi-step search can make later evidence harder to find; the paper tests whether practicing on those error-shaped states helps the retriever recover. The abstract does not give the size of the gains or enough detail to assess their practical significance. How large are the reported gains, and how do they vary across the evaluated datasets and.
Original source
cs.AI · Article
Source details
Original source date: Oct 29, 2026 (2026-10-29).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-13T07:03:07.331Z. First observed: 2026-10-30T06:42:53.592Z. Retrieved: 2026-10-31T09:31:07.007Z.
Limits
The supplied evidence is the paper’s abstract, which does not state the size of the reported gains or provide detailed results by dataset or comparison system. The abstract describes harder long-chain variants of HotpotQA and 2WikiMultiHopQA that the authors construct; the supplied evidence does not describe those variants in detail.
Open question
How large are the reported gains, and how do they vary across the evaluated datasets and systems?
Source-stated
The authors say a retriever trained only on gold histories can face a mismatch at test time, when it conditions on its own retrieved passages, which may be wrong.
At test time the retriever conditions on its own retrievals, which soon contain mistakes
Source-stated
The authors report that by the third step, about half or more of the states visited contain a wrong passage, and that these passages reduce the chance of retrieving missing gold evidence.
by the third step, about half or more of the states it visits hold a wrong passage, and these passages lower its chance of retrieving a missing gold passage
Source-stated
The proposed on-policy state training trains the encoder using states reached by its current retrievals, including states built from its own mistakes.
We propose on-policy state training, which trains the encoder on the states it reaches itself.
Source-stated
The authors report that the resulting retriever retrieves complete evidence more often than teacher-forced training and the systems they compare.
The resulting retriever retrieves complete evidence more often than teacher-forced training and the systems we compare
Source-stated
The authors report transfer without further training to harder long-chain variants of HotpotQA and 2WikiMultiHopQA that they construct.
it transfers without further training to harder long-chain variants of HotpotQA and 2WikiMultiHopQA that we construct
Escape Meridian interpretation
The paper’s approach targets the problem that a retriever’s own earlier mistakes can make later missing evidence harder to retrieve.
these passages lower its chance of retrieving a missing gold passage well beyond what the difficulty of the states explains
Source and contributor context
Qiu, Chenjie, Zeng, Zixuan, Shi, Shaobin, Chen, Bingfeng, Luo, Dongsheng, Xu, Boyan · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Retrieval-augmented generation
A setup in which a model first looks up relevant documents and then answers using them.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
TIME
TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Editorial provenance
Version 1. Publication ID: d08256f4-99a6-41f9-817f-b6ef0a597c9c.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-13T07:03:07.331Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 8bdfd9e8beeb32c2897f4e18ecd0e7e6a0af4b3557f428dc243ebfae7f8bb900. Source text hash: c4b521a32436de807e84e36c0822b4ffd39efd14d51f2f3e33f7aef0faa335b4.
uncertain · 2026-11-13T07:03:07.331Z · anthropic/claude-opus-5.5
Publication: d08256f4-99a6-41f9-817f-b6ef0a597c9c
attribution — pass: The draft correctly attributes the paper to Chenjie Qiu and coauthors, matching the listed authors. It frames findings as the authors' reports and claims rather than as established facts.
Authors:Chenjie Qiu, Zixuan Zeng, Shaobin Shi, Bingfeng Chen, Dongsheng Luo, Boyan Xu
summary — uncertain: The summary ends with a truncated, duplicated question fragment ('...datasets and.'). Remove it, since the question field already carries it. Central terms are left unexplained for a general reader: retrieval-augmented generation, embedding, gold histories, teacher-forced, and contrastive. Add plain-language glosses or term tooltips. The draft also omits the authors' finding that a small chain retriever trained on gold histories, calling no language model, already beats graph-based systems and is competitive with a search agent. It softens 'far more often' to 'more often'. Restore both to keep the source's force.
a small chain embedding retriever that calls no language model outperforms graph-based systems and is competitive with a search agent
date — pass: originalPublished 2026-10-29 matches the arXiv v1 submission date. The feed date of Oct 31 is later and correctly not used.
[Submitted on 29 Oct 2026]
support — pass: Each claim matches its excerpt. The interpretation claim about early mistakes hurting later retrieval is grounded in the source and labeled as interpretation. The limitations correctly note that the abstract gives no gain magnitudes. The question follows from that gap and is not a leading claim. No unsupported benefits or costs are asserted.
these passages lower its chance of retrieving a missing gold passage well beyond what the difficulty of the states explains
media — pass: The selected media is the verified direct arXiv abstract page for this paper, which is the most direct access to the source.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
A setup in which a model first looks up relevant documents and then answers using them.
Definition referenceTIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Explore TIMEarXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivA setup in which a model first looks up relevant documents and then answers using them.
Definition referenceTIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Explore TIMEarXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXiv