RECON: Reasoning with Condensation for Efficient Retrieval-Augmented Generation

2025-10-12

AI Summary

RAG systems often suffer from high context costs and degraded performance in reinforcement learning (RL)-driven multi-step reasoning due to lengthy, noisy retrieved documents. To address this, we propose RECON, the first framework to integrate a learnable, explicit summarization module into the RL-based RAG inference loop. RECON employs a two-stage training strategy—relevance-aware pretraining followed by multi-dimensional factual knowledge distillation—to ensure summaries are both accurate and concise. Fully end-to-end integrated within the Search-R1 pipeline, RECON reduces average context length by 35% on multi-hop QA tasks, while significantly lowering inference latency and training time. On benchmark evaluations, it improves exact match (EM) scores by 14.5% and 3.0% for 3B- and 7B-parameter models, respectively. These results demonstrate that learned contextual compression enables synergistic optimization of both efficiency and effectiveness in RAG systems.

Technology Category

• Natural Language Processing: Summarization
• Data Mining & Knowledge Management: Conversational Systems for Recommendation & Retrieval
• Search and Optimization: Learning to Search

Application Category

• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Retrieval-augmented generation (RAG) systems trained using reinforcement learning (RL) with reasoning are hampered by inefficient context management, where long, noisy retrieved documents increase costs and degrade performance. We introduce RECON (REasoning with CONdensation), a framework that integrates an explicit summarization module to compress evidence within the reasoning loop. Our summarizer is trained via a two-stage process: relevance pretraining on QA datasets, followed by multi-aspect distillation from proprietary LLMs to ensure factuality and clarity. Integrated into the Search-R1 pipeline, RECON reduces total context length by 35%, leading to improved training speed and inference latency, while simultaneously improving RAG performance on downstream QA benchmarks. Notably, it boosts the average EM score of the 3B model by 14.5% and the 7B model by 3.0%, showing particular strength in multi-hop QA. RECON demonstrates that learned context compression is essential for building practical, scalable, and performant RAG systems. Our code implementation is made available at https://github.com/allfornancy/RECON.

Problem

• Compressing retrieved documents to reduce context length
• Improving training speed and inference latency in RAG systems
• Enhancing RAG performance on question-answering benchmarks

Innovation

• Integrates explicit summarization module for evidence compression
• Uses two-stage training with relevance pretraining and distillation
• Reduces context length by 35% improving speed and performance