ReasonIR: Training Retrievers for Reasoning Tasks

TL;DR

ReasonIR addresses the gap that existing retrievers struggle with reasoning-intensive retrieval by introducing ReasonIR-8B, a bi-encoder retriever trained with ReasonIR-Synthesizer-generated varied-length and hard-query data alongside public datasets. The approach achieves state-of-the-art BRIGHT performance, strengthens open-domain RAG results on MMLU and GPQA, and demonstrates effective test-time scaling through query rewriting and a simple LLM-based reranker. Key contributions include a multi-type synthetic data pipeline (VL and HQ), a multi-turn hard negative generation strategy, and a tie-breaking reranking method (ReasonIR-Rerank) that balances compute with accuracy. The work offers a general, extensible training recipe, supported by open-source code, data, and models, with promising implications for reasoning-aware information retrieval and RAG applications.

Abstract

We present ReasonIR-8B, the first retriever specifically trained for general reasoning tasks. Existing retrievers have shown limited gains on reasoning tasks, in part because existing training datasets focus on short factual queries tied to documents that straightforwardly answer them. We develop a synthetic data generation pipeline that, for each document, our pipeline creates a challenging and relevant query, along with a plausibly related but ultimately unhelpful hard negative. By training on a mixture of our synthetic data and existing public data, ReasonIR-8B achieves a new state-of-the-art of 29.9 nDCG@10 without reranker and 36.9 nDCG@10 with reranker on BRIGHT, a widely-used reasoning-intensive information retrieval (IR) benchmark. When applied to RAG tasks, ReasonIR-8B improves MMLU and GPQA performance by 6.4% and 22.6% respectively, relative to the closed-book baseline, outperforming other retrievers and search engines. In addition, ReasonIR-8B uses test-time compute more effectively: on BRIGHT, its performance consistently increases with longer and more information-rich rewritten queries; it continues to outperform other retrievers when combined with an LLM reranker. Our training recipe is general and can be easily extended to future LLMs; to this end, we open-source our code, data, and model.

Figures (21)

• Figure 1: (a) Performance against test-time compute on the reasoning-intensive information retrieval (IR) benchmark BRIGHT. ReasonIR-8B achieves new state-of-the-art scores, demonstrating that efficient bi-encoders can outperform significantly more expensive reranker baselines. We also introduce an LLM reranking method with a simple yet effective tie-breaking technique, combined with which ReasonIR-8B achieves a new SOTA RBRIGHT score of 36.9 nDCG@10 (§\ref{['sec:method-reranker']}). (b) Performance on Retrieval-augmented generation (RAG) benchmarks MMLU and GPQA. ReasonIR-8B outperforms other retriever and search engine baselines. The IR benchmark directly measures retrieval quality using annotated target documents; RAG benchmarks measure the performance of LM responses that incorporate retrieved information.
• Figure 2: Query length scaling with query rewriting on BRIGHT. ReasonIR-8B is our trained retriever that will be described in Section \ref{['sec:exp']}.
• Figure 3: (a) Qualitative examples of the three types of training data used in the ReasonIR training recipe and the synthetic generation pipeline, ReasonIR-Synthesizer, used to generate varied-length data ($\overline{\underline{\mathrlap{\textsc{VL}}\space}}$) and hard query data ($\overline{\underline{\mathrlap{\textsc{HQ}}\space}}$). (b) Query length distribution of the public, $\overline{\underline{\mathrlap{\textsc{VL}}\space}}$, and $\overline{\underline{\mathrlap{\textsc{HQ}}\space}}$ data. (c) Difficulty of the public, $\overline{\underline{\mathrlap{\textsc{VL}}\space}}$, and $\overline{\underline{\mathrlap{\textsc{HQ}}\space}}$ data, measured by the error rates of BM25 and GRIT-7B (i.e., how frequently they assign a higher similarity score to the hard negative document than the positive document).
• Figure 4: System prompt for reasoning-intensive document-to-query data generation ($\overline{\underline{\mathrlap{\textsc{HQ}}\space}}$)
• Figure 5: User prompt for reasoning-intensive document-to-query data generation ($\overline{\underline{\mathrlap{\textsc{HQ}}\space}}$)