Rationale contrastive decoding reduces CoT query shortcuts
Original title: Mitigating Query Shortcuts in Chain-of-Thought with Context-Aware Decoding
The training-free method improved answer accuracy on five reasoning benchmarks and made answers depend more on the rationale.
The paper tests whether context-aware decoding can stop large language models from ignoring chain-of-thought rationales and using query-specific shortcuts. Rationale Contrastive Decoding contrasts next-token distributions with and without the rationale, amplifying tokens whose likelihood rises because of the rationale. The authors interpret the rule as KL-regularized reward maximization. Intervention tests with ground-truth and random rationales found greater behavioral dependence on rationale content.
- RCD is an application of existing context-aware decoding, not a new trained model.
- Accuracy gains are reported over standard chain-of-thought on five benchmarks.
- Ground-truth and random rationale interventions were used to measure dependence.
- The paper frames the method as simple and training-free.
From the original article
1 min read at arxiv.orgWhile Chain-of-Thought (CoT) prompting improves the reasoning performance of large language models (LLMs), recent studies have shown that final answers do not always depend on the generated rationales. In particular, LLMs may rely on query-specific shortcuts rather than the information provided by their reasoning steps.
By Shin'ya Yamaguchi · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.