MultiSearch: Parallel Queries and Explicit Merging Boost LLM Reasoning
A new framework called MultiSearch improves large language model reasoning by generating multiple queries per step and explicitly merging retrieved information. Developed by researchers and detailed in arXiv:2605.13534, MultiSearch uses reinforcement learning to expand information coverage and reduce noise, addressing limitations of single-query retrieval methods. The approach enhances signal-to-noise ratios and reasoning accuracy.
Key facts
- MultiSearch is an RL-based framework for LLM reasoning.
- It generates queries from multiple perspectives at each reasoning step.
- Retrieval is performed in parallel to expand information coverage.
- Explicit merging consolidates and refines retrieved information.
- It improves signal-to-noise ratio and reasoning accuracy.
- The paper is published on arXiv with ID 2605.13534.
- Existing methods often use a single query per step, limiting coverage.
- Single-query retrieval introduces high noise and unnecessary steps.
Entities
Institutions
- arXiv