arXiv：cs.CL· Omri Uzan, Ron Polonsky, Douwe Kiela, Christopher Potts·· 3 小时前AI 评分40
DocOpt：用强化学习优化文档以提升黑盒检索性能
Document Optimization for Black-Box Retrieval via Reinforcement Learning
AI 导读
DocOpt 通过 GRPO 以检索排名提升为奖励，训练 LLM 或 VLM 生成优化文档改写，仅需黑盒访问检索排名，适用于单向量、多向量和词法检索器。
正文
Abstract:Generative large language models (LLMs) are increasingly used as inference-time components in retrieval pipelines, for tasks such as query rewriting and document reranking. However, these online approaches place costly autoregressive computation directly on the latency-critical retrieval path. We explore an alternative axis: using LLMs to improve documents instead, rewriting them into better representations and shifting computation offline. Yet producing a useful document rewrite is not straightforward: retrieval is inherently discriminative, so an effective rewrite must make a document more similar to relevant queries than competing candidates under the retriever's notion of similarity. We therefore formulate document transformation as an optimization problem, directly training an LLM or VLM to produce rewrites that improve retrieval. Our approach, DocOpt, uses GRPO with retriever ranking improvements as rewards, requires only black-box access to retrieval ranks, and applies across single-vector, multi-vector, and lexical retrievers. We evaluate zero-shot LLM rewriting and DocOpt on code and visual retrieval tasks, finding that document rewriting can improve retrieval and that optimizing rewrites yields further gains. For example, OpenAI text-embedding-3-small achieves 58.35 nDCG@5 on average with direct retrieval; zero-shot rewriting improves this to 60.83 with GPT-5.4-mini, 63.75 with Claude Haiku 4.5, and 64.23 with Qwen3. DocOpt further improves performance to 67.94, surpassing the 6.5X more expensive text-embedding-3-large retriever at 66.15.
| Subjects: | Computation and Language (cs.CL); Information Retrieval (cs.IR) |
| Cite as: | arXiv:2604.05087 [cs.CL] |
| (or arXiv:2604.05087v4 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2604.05087 arXiv-issued DOI via DataCite |
Submission history
From: Omri Uzan [view email]
[v1]
Mon, 6 Apr 2026 18:41:40 UTC (5,156 KB)
[v2]
Sat, 20 Jun 2026 18:39:23 UTC (1,071 KB)
[v3]
Wed, 5 Aug 2026 03:31:30 UTC (1,070 KB)
[v4]
Wed, 7 Oct 2026 00:32:10 UTC (1,076 KB)
来源：arXiv：cs.CL · arxiv.org