RecLM eliminates LLM recommender hallucinations with unified framework

Large language models (LLMs) powering recommender systems often hallucinate by suggesting items outside a platform's catalog—so-called out-of-domain (OOD) recommendations. A new paper from Hao Liao and nine colleagues proposes RecLM, a unified framework that eliminates this problem entirely. RecLM bridges the gap between retrieval and generation by instantiating three grounding paradigms under a single architecture: embedding-based retrieval (using vector similarity), constrained generation over rewritten item titles (forcing the LLM to output only valid catalog entries), and discrete item-tokenizer generation (treating items as tokens). All variants achieve OOD@10 = 0 on public benchmarks, meaning zero hallucinated recommendations appear in the top 10 results.

The constrained generation variants—RecLM-cgen and RecLM-token—also achieve overall state-of-the-art accuracy compared to strong ID-based and LLM-based baselines. By using the same backbone LLM and prompts, the authors provide a systematic comparison across the three paradigms, offering a practical framework for integrating LLMs into recommendation tasks without the risk of irrelevant suggestions. The work addresses a critical bottleneck for deploying LLM-powered recommenders in production, where trust and catalog fidelity are paramount. Source code is available on GitHub, making the approach reproducible for practitioners.