Thinking Forward and Backward: Multi-Objective Reinforcement Learning for Retrieval-Augmented Reasoning
This addresses the issue of degraded response quality in retrieval-augmented reasoning for question answering tasks, though it appears incremental as it builds on existing search-based RAG methods.
The paper tackled the problem of limited effectiveness of retrieval-augmented generation in complex, multi-step reasoning by proposing Bi-RAR, a framework that evaluates intermediate steps bidirectionally using a multi-objective reinforcement learning approach, resulting in superior performance on seven question answering benchmarks.
Retrieval-augmented generation (RAG) has proven to be effective in mitigating hallucinations in large language models, yet its effectiveness remains limited in complex, multi-step reasoning scenarios. Recent efforts have incorporated search-based interactions into RAG, enabling iterative reasoning with real-time retrieval. Most approaches rely on outcome-based supervision, offering no explicit guidance for intermediate steps. This often leads to reward hacking and degraded response quality. We propose Bi-RAR, a novel retrieval-augmented reasoning framework that evaluates each intermediate step jointly in both forward and backward directions. To assess the information completeness of each step, we introduce a bidirectional information distance grounded in Kolmogorov complexity, approximated via language model generation probabilities. This quantification measures both how far the current reasoning is from the answer and how well it addresses the question. To optimize reasoning under these bidirectional signals, we adopt a multi-objective reinforcement learning framework with a cascading reward structure that emphasizes early trajectory alignment. Empirical results on seven question answering benchmarks demonstrate that Bi-RAR surpasses previous methods and enables efficient interaction and reasoning with the search engine during training and inference.