Search papers, labs, and topics across Lattice.
The paper introduces ThinkGR, a novel generative retrieval framework that integrates chain-of-thought (CoT) reasoning with document ID generation. ThinkGR employs a hybrid decoding strategy, dynamically switching between unconstrained thought generation and constrained docid decoding, and is trained using a two-phase approach involving supervised fine-tuning and retrieval-grounded reinforcement learning. Experiments on multi-hop retrieval benchmarks demonstrate state-of-the-art performance, with an average improvement of 6.86%.
Generative retrieval gets a reasoning upgrade: ThinkGR interleaves chain-of-thought with docid generation, boosting multi-hop retrieval performance by 6.86%.
While generative retrieval (GR) demonstrates competitive performance on standard retrieval benchmarks, existing approaches directly map queries to document identifiers (docids) without intermediate deliberation, limiting their effectiveness for complex queries that require multi-step reasoning. As a preliminary study on integrating chain-of-thought (CoT) into generative retrieval, we introduce ThinkGR, a unified framework that interleaves CoT with docid generation, enabling iterative thinking and retrieval within a single generative process. To bridge the gap between free-form thought generation and structured retrieval targets, we design (1) a hybrid decoding strategy that dynamically switches between unconstrained thought generation and constrained docid decoding, and (2) a two-phase training approach that first aligns thought-retrieval patterns through supervised fine-tuning, then optimizes thought quality via retrieval-grounded reinforcement learning. Experiments on four multi-hop retrieval benchmarks demonstrate that ThinkGR achieves state-of-the-art performance with an average improvement of +6.86\%. Our work opens new avenues for enhancing generative retrieval with explicit deliberation capabilities, with promising implications for retrieval tasks requiring complex reasoning.
