Search papers, labs, and topics across Lattice.
The paper introduces GCoT-decoding, a novel decoding strategy that extends CoT-decoding to free-form question answering tasks by generating and evaluating reasoning paths without manual prompts. GCoT-decoding uses a two-stage branching method with Fibonacci sampling and error backtracking to generate candidate paths, which are then split into reasoning and answer spans for confidence scoring. Experiments on six datasets show that GCoT-decoding maintains performance on fixed QA and significantly improves performance on free QA.
Unleashing chain-of-thought reasoning on free-form question answering tasks is now possible without manual prompts, thanks to a new decoding strategy that automatically generates and evaluates reasoning paths.
Chain-of-Thought reasoning can enhance large language models, but it requires manually designed prompts to guide the model. Recently proposed CoT-decoding enables the model to generate CoT-style reasoning paths without prompts, but it is only applicable to problems with fixed answer sets. To address this limitation, we propose a general decoding strategy GCoT-decoding that extends applicability to a broader range of question-answering tasks. GCoT-decoding employs a two-stage branching method combining Fibonacci sampling and heuristic error backtracking to generate candidate decoding paths. It then splits each path into a reasoning span and an answer span to accurately compute path confidence, and finally aggregates semantically similar paths to identify a consensus answer, replacing traditional majority voting. We conduct extensive experiments on six datasets covering both fixed and free QA tasks. Our method not only maintains strong performance on fixed QA but also achieves significant improvements on free QA, demonstrating its generality.