Mitigating Query Shortcuts in Chain-of-Thought with Context-Aware Decoding

Key points

• 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.1
• 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.1
• Experiments across five reasoning benchmarks demonstrate that RCD improves answer accuracy over standard CoT.1
• Our findings suggest that contrastive decoding offers a simple, training-free approach to mitigating query shortcuts.1

Sources (1)

• [1]Mitigating Query Shortcuts in Chain-of-Thought with Context-Aware Decoding
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 21, 10:05 AM

  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 this paper, we investigate whether context-aware decoding (CAD), an existing contrastive decoding method, can mitigate such shortcut behavior when applied to CoT reasoning.

Extractive summary: sentences quoted from the sources.