Generative Retrieval (GR) treats search and recommendation as autoregressive generation of discrete semantic identifiers (SIDs). The design of SIDs largely determines GR quality, yet its discrete and interdependent components—SID construction, collision handling, decoding, and training—cannot be optimized end‑to‑end with gradients and remain largely hand‑crafted. This paper introduces a framework that leverages large language model (LLM)‑driven evolutionary program search to improve these components, the first known application of this approach to GR.
We employ a locally hosted 31B open‑weight model to mutate executable GR components, avoiding any proprietary LLM API. In parallel, a 1M‑parameter proxy decoder provides fast feedback without repeatedly training the 1B‑parameter target generator. On the NQ320K benchmark, we evaluate 21 SID programs; the proxy’s Recall@100 correlates strongly with the target’s Recall@100 (Spearman $\rho = 0.90$), indicating reliable performance prediction.
Guided by this proxy, evolution improves the seed GR on NQ320K, MSMARCO300K, and a multilingual e‑commerce dataset under a fixed training recipe. On NQ320K, jointly evolving SID construction and collision handling raises Recall@1 from 55.8% to 59.3% and Recall@100 from 85.8% to 90.6%. A component study reveals that proxy reliability depends on the search space: improvements to the semantic index transfer consistently, whereas changes to the generator’s training objective or the full pipeline do not always transfer reliably.
Consequently, provided the proxy’s reliability is established for the chosen search space, LLM‑driven evolutionary search becomes a practical tool for GR design.
Review: This work demonstrates the feasibility of using locally run large models for program‑level evolution, offering a systematic path to optimize generative retrieval systems.