Batch-of-Thought: Cross-Instance Learning for Enhanced LLM Reasoning
Researchers have introduced Batch-of-Thought (BoT), a novel, training-free method designed to enhance the reasoning capabilities of Large Language Models (LLMs). Unlike current systems that process queries in isolation, BoT enables cross-instance learning by jointly processing related queries. This approach allows the system to identify high-quality reasoning templates, detect errors through consistency checks, and amortize computational costs. The method is instantiated within a multi-agent reflection architecture known as BoT-R, where a Reflector agent performs joint evaluations to unlock mutual information gains. Experimental results across three model families and six benchmarks demonstrate that BoT-R consistently improves accuracy and confidence calibration. Furthermore, the method significantly reduces inference costs by up to 61%. The study provides both theoretical and experimental analysis explaining the benefits of batch-aware reasoning in LLM systems. The research team has made their code publicly available to support further development and verification. This advancement addresses the inefficiency of discarding valuable cross-instance signals in traditional LLM processing.
Wire timeline
Batch-of-Thought: Cross-Instance Learning for Enhanced LLM Reasoning
Researchers have introduced Batch-of-Thought (BoT), a novel, training-free method designed to enhance the reasoning capabilities of Large Language Models (LLMs). Unlike current systems that process queries in isolation, BoT enables cross-instance learning by jointly processing related queries. This approach allows the system to identify high-quality reasoning templates, detect errors through consistency checks, and amortize computational costs. The method is instantiated within a multi-agent reflection architecture known as BoT-R, where a Reflector agent performs joint evaluations to unlock mutual information gains. Experimental results across three model families and six benchmarks demonstrate that BoT-R consistently improves accuracy and confidence calibration. Furthermore, the method significantly reduces inference costs by up to 61%. The study provides both theoretical and experimental analysis explaining the benefits of batch-aware reasoning in LLM systems. The research team has made their code publicly available to support further development and verification. This advancement addresses the inefficiency of discarding valuable cross-instance signals in traditional LLM processing.
cs.AI updates on arXiv.org