What it does

The repository is a living bibliography for the survey paper Large Language Models for Information Retrieval: A Survey. It catalogs papers where LLMs touch some part of the search pipeline—query rewriting, document retrieval, re-ranking, result reading, and increasingly, end-to-end search agents that plan their own information-seeking strategies.

The interesting bit

The taxonomy itself reveals how the field has shifted. Early work (2022–2023) focused on using LLMs to generate synthetic training data or rewrite queries. By 2024, researchers were building “search agents” with self-planning modules and benchmarking them against static pipelines. The version history in the README is arguably more informative than most paper abstracts—it shows the taxonomy being rebuilt in public.

Key highlights

• Query Rewriter: Prompting, fine-tuning, and distillation methods for query expansion and conversational search.
• Retriever: LLMs used to generate training data (InPars, Promptagator) or directly as retriever architectures (differentiable search indices, instruction-following encoders).
• Re-ranker: Supervised, unsupervised, and data-augmentation approaches; a newer “Reasoning-intensive Rerankers” category added in 2025.
• Reader: Passive and active readers, plus reference compression—essentially, how to stop LLMs from drowning in retrieved context.
• Search Agent: The newest section, covering static and dynamic agents with benchmarking and self-planning capabilities.
• Maintained by RUC-NLPIR; actively updated (four major versions since 2023, most recently September 2025).

Caveats

• The paper list is extensive but not filtered for quality—presence here does not equal endorsement.
• No code, no benchmarks, no reproduction scripts: this is purely a reading list with links.
• Some sections are thin (e.g., “Fine-tuning Methods” under Query Rewriter has exactly one paper).

Verdict

Worth bookmarking if you are entering LLM+IR research and need a structured literature map. Skip it if you want runnable baselines or critical synthesis—the curation is broad, not deep.