Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

Chain of Thought Still Thinks Fast: APriCoT Helps with Thinking Slow

Created by
  • Haebom

Author

Kyle Moore, Jesse Roberts, Thao Pham, Douglas Fisher

Outline

This paper investigates the impact of language model bias on answer choice preferences in the Massive Multi-Task Language Understanding (MMLU) task. The results show that language model bias predicts model preferences and reflects human test-taking strategies, even when using CoT inference. To address this issue, the authors introduce counterfactual prompting and indiscriminately primed CoT (APriCoT). While counterfactual prompting alone using CoT is insufficient to mitigate bias, APriCoT effectively reduces the influence of underlying probability and improves overall accuracy. CoT tends to reinforce fast-thinking model bias under certain prompting methods, suggesting that slow-thinking is necessary for bias mitigation. APriCoT represents a step toward developing a more robust and fair "slow-thinking" language model.

Takeaways, Limitations

•
Takeaways:
◦
We reveal that language model bias significantly influences answer selection in tasks such as MMLU.
◦
CoT alone does not fully address model bias, suggesting the need for a "slow thinking" process.
◦
We demonstrate that APriCoT effectively reduces bias and improves accuracy compared to existing methods.
•
Limitations:
◦
Further research is needed to determine whether the effects of APriCoT can be generalized to all types of bias or all language models.
◦
Further analysis of the computational cost and efficiency of APriCoT is needed.
◦
There is a lack of clear criteria for defining and measuring “slow thinking.”
👍
Made with Slashpage