LLM evolutionary search improves generative retrieval semantic indexes
Original title: Evolving Semantic Indexes for Generative Retrieval
On NQ320K, jointly evolving SID construction and collision handling lifts Recall@1 from 55.8% to 59.3%.
A paper proposes LLM-driven evolutionary program search to improve hand-designed semantic identifier components in generative retrieval. A locally served open-weight 31B model mutates executable components, and a 1M-parameter proxy decoder scores candidates instead of retraining a 1B generator. Across 21 SID programs on NQ320K, proxy and target Recall@100 correlate at Spearman ρ=0.90. Guided evolution improves a seed system on NQ320K, MSMARCO300K, and a multilingual e-commerce dataset under a fixed training recipe. Proxy gains transfer reliably for the semantic index, not for the training objective or the full pipeline.
- No proprietary LLM API is used; mutation runs on a local 31B model.
- NQ320K Recall@100 rises from 85.8% to 90.6% in the joint evolution result.
- Proxy reliability is reported as strong for the semantic index only.
- Evaluation also covers MSMARCO300K and a multilingual e-commerce dataset.
From the original article
1 min read at arxiv.orgGenerative retrieval (GR) casts search and recommendation as autoregressive generation of discrete semantic identifiers (SIDs). SID design largely determines GR quality, but its discrete, interdependent components (SID construction, collision handling, decoding, and training) cannot be optimized end-to-end with gradients and remain largely hand-designed.
By Jinming Nian · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.