Contrastive decoding improves reasoning reliability in large language models
Mitigating Query Shortcuts in Chain-of-Thought with Context-Aware Decoding
Computation and LanguageArtificial IntelligenceMachine Learning
Summary
Sometimes, large language models give the right answers without actually using the detailed reasoning steps they explain. This means they might just take shortcuts based on the question instead of truly thinking through the problem. The authors studied a new way called rationale contrastive decoding, which helps the model rely more on its reasoning steps when deciding on answers. Their tests show this method improves accuracy and makes answers depend more on the reasoning provided. This approach works without needing to train the model further.
What this means in practice
- •For llm developers: Increase the reliability of reasoning in language models by integrating contrastive decoding to reduce shortcut answers.
- •For ai tool builders: Enhance AI assistant accuracy by applying rationale contrastive decoding to better link explanations with final answers.
Authors
Shin'ya Yamaguchi
Abstract
While 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. In this paper, we investigate whether context-aware decoding (CAD), an existing contrastive decoding method, can mitigate such shortcut behavior when applied to CoT reasoning. We study this application as Rationale Contrastive Decoding (RCD), which contrasts the next-token distributions conditioned on the query with and without its rationale, amplifying tokens whose likelihood increases due to the rationale. We provide a theoretical interpretation of this decoding rule via a KL-regularized reward-maximization objective. Experiments across five reasoning benchmarks demonstrate that RCD improves answer accuracy over standard CoT. Furthermore, intervention analyses using ground-truth and random rationales reveal that RCD increases the behavioral dependence of final answers on the rationale content. Our findings suggest that contrastive decoding offers a simple, training-free approach to mitigating query shortcuts.