Large language models improve text search by preserving reasoning quality
Reasoning Quality Matters: Combating Reasoning Collapse in LLM-based Embedding Learning
Information Retrieval
Summary
Finding the right text in huge databases can be tricky because language models used to create search clues sometimes forget how to reason well. The authors call this problem reasoning collapse, where the model either stops reasoning properly or generates irrelevant information. They propose a method called CoFree that gently guides the model through two steps to keep its reasoning sharp while improving its ability to represent text for search. This leads to more accurate search results in tests across many datasets and real-world applications.
What this means in practice
- •For search engine developers: Improve search engine accuracy by integrating reasoning-preserving embedding methods for better query and document matching.
- •For knowledge management teams: Enhance information retrieval systems in enterprises by applying embedding techniques that maintain reasoning to retrieve more relevant documents.
- •For customer support platform builders: Build customer support bots that retrieve precise answers by using embeddings guided by high-quality reasoning processes.$Commercial implications: This enables commercial chatbots and support tools to provide more relevant answers, improving user satisfaction and support efficiency.
Authors
Zihan Gong, Xiaohan Ye, Jiangchao Yao, Jinsong Lan, Xiaoyong Zhu, Xu Chen
Abstract
Large Language Models (LLMs) have recently shown strong potential for producing context-rich text embeddings for retrieval. Most existing methods either treat embedding learning as passive feature extraction or exploit LLM reasoning through instruction following for better embedding optimization. However, specialization toward embedding objectives can suppress useful reasoning generation or produce retrieval-irrelevant text. We refer to these two forms of degradation as reasoning collapse. To address this issue, we propose CoFree (Collapse-Free Reasoning Embedding), a two-stage framework that progressively integrates LLM reasoning into query and document embedding optimization while preserving reasoning quality. At the first stage, CoFree applies reference-guided supervised fine-tuning to restore the reasoning ability and retain representational strength of the foundation embedding model. At the second stage, we introduce dual rewards, an embedding-oriented reward and a reasoning-oriented reward, to guarantee fine-grained reasoning of the relevance toward the embedding goal in reinforcement learning. This endpoint-coupled optimization transforms embedding learning from static alignment into a high-quality reasoning-guided search process for retrieval. Extensive experiments demonstrate the effectiveness of CoFree, with CoFree-4B achieving an average absolute improvement of 2.8 nDCG@10 points over Qwen3-Embedding-4B across 22 datasets from MTEB and BRIGHT. Online experiments in a real-world retrieval system further show consistent gains. Code, RTED, and model checkpoints will be made publicly available.