Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

RaDeR: Reasoning-aware Dense Retrieval Models

Debrup Das, Sam O' Nuallain, Razieh Rahimi

TL;DR

RaDeR presents a reasoning-aware dense retrieval framework that couples a first-stage dense retriever with a re-ranker, trained on synthetic data generated by a retrieval-augmented MCTS process for mathematical problems. By incorporating self-reflection and diverse query types, RaDeR yields substantial gains on reasoning-intensive benchmarks such as BRIGHT and MMTEB-RAR-b while preserving performance on traditional IR tasks, achieving data-efficient training relative to concurrent methods. The approach demonstrates the critical role of reasoning-aware retrieval for augmenting reasoning LLMs and offers a scalable path to improved relevance prediction in complex domains. These results suggest that integrating targeted reasoning dynamics into the retrieval stage can significantly enhance both retrieval quality and downstream QA performance.

Abstract

We propose RaDeR, a set of reasoning-based dense retrieval models trained with data derived from mathematical problem solving using large language models (LLMs). Our method leverages retrieval-augmented reasoning trajectories of an LLM and self-reflective relevance evaluation, enabling the creation of both diverse and hard-negative samples for reasoning-intensive relevance. RaDeR retrievers, trained for mathematical reasoning, effectively generalize to diverse reasoning tasks in the BRIGHT and RAR-b benchmarks, consistently outperforming strong baselines in overall performance. Notably, RaDeR achieves significantly higher performance than baselines on the Math and Coding splits. In addition, RaDeR presents the first dense retriever that outperforms BM25 when queries are Chain-of-Thought reasoning steps, underscoring the critical role of reasoning-based retrieval to augment reasoning language models. Furthermore, RaDeR achieves comparable or superior performance while using only 2.5% of the training data used by the concurrent work REASONIR, highlighting the quality of our synthesized training data.

RaDeR: Reasoning-aware Dense Retrieval Models

TL;DR

RaDeR presents a reasoning-aware dense retrieval framework that couples a first-stage dense retriever with a re-ranker, trained on synthetic data generated by a retrieval-augmented MCTS process for mathematical problems. By incorporating self-reflection and diverse query types, RaDeR yields substantial gains on reasoning-intensive benchmarks such as BRIGHT and MMTEB-RAR-b while preserving performance on traditional IR tasks, achieving data-efficient training relative to concurrent methods. The approach demonstrates the critical role of reasoning-aware retrieval for augmenting reasoning LLMs and offers a scalable path to improved relevance prediction in complex domains. These results suggest that integrating targeted reasoning dynamics into the retrieval stage can significantly enhance both retrieval quality and downstream QA performance.

Abstract

We propose RaDeR, a set of reasoning-based dense retrieval models trained with data derived from mathematical problem solving using large language models (LLMs). Our method leverages retrieval-augmented reasoning trajectories of an LLM and self-reflective relevance evaluation, enabling the creation of both diverse and hard-negative samples for reasoning-intensive relevance. RaDeR retrievers, trained for mathematical reasoning, effectively generalize to diverse reasoning tasks in the BRIGHT and RAR-b benchmarks, consistently outperforming strong baselines in overall performance. Notably, RaDeR achieves significantly higher performance than baselines on the Math and Coding splits. In addition, RaDeR presents the first dense retriever that outperforms BM25 when queries are Chain-of-Thought reasoning steps, underscoring the critical role of reasoning-based retrieval to augment reasoning language models. Furthermore, RaDeR achieves comparable or superior performance while using only 2.5% of the training data used by the concurrent work REASONIR, highlighting the quality of our synthesized training data.

Paper Structure

This paper contains 40 sections, 8 equations, 13 figures, 14 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. RaDeR: Reasoning-aware Retrievers
  4. Generating Retrieval Training Data
  5. Retrieval-Augmented Search-based Reasoning
  6. Synthesizing Training Data
  7. Experimental Settings
  8. Experimental Results
  9. Bright Retrieval Performance
  10. Retrieval Performance on RAR-b
  11. Retrieval on Traditional IR Benchmarks
  12. Bright Reranking Performance
  13. RAG Performance using RaDeR
  14. Conclusion
  15. Limitations
  16. ...and 25 more sections

Figures (13)

  • Figure 1: An example based on sample 'TheoremQA_jianyuxu/pigeonhole3' of Bright, where term matching retrievers face challenges in retrieving the relevant theorem w.r.t. both questions and CoT reasoning.
  • Figure 2: An overview of the RaDeR data generation pipeline. The OST action stands for one step thought generation, and CRS stands for complete remaining solution steps action.
  • Figure 3: Examples of different query types from our retrieval training dataset built for the given math question.
  • Figure 4: Example of RaDeR success case compared to Qwen2, from TheoremQA theorems of Bright.
  • Figure 5: Example of RaDeR Failure cases from TheoremQA theorems BRIGHT
  • ...and 8 more figures