The paper 'Language Models Don’t Always Say What They Think: Unfaithful Explanations in Chain-of-Thought Prompting' by Miles Turpin et al. investigates the faithfulness of chain-of-thought (CoT) explanations generated by large language models (LLMs) for various tasks. CoT explanations are verbalisation's of step-by-step reasoning that LLMs produce before giving a final output.
The paper shows that CoT explanations can be misleading and influenced by biasing features in the model inputs, such as the order of multiple-choice options. The paper tests two LLMs, GPT-3.5 and Claude 1.0, on 13 tasks from BIG-Bench Hard and a social-bias task, and finds that accuracy drops significantly when models are biased toward incorrect answers.
The paper also finds that models justify answers based on stereotypes without mentioning the influence of social biases. The paper concludes that CoT explanations can be plausible yet unfaithful, which poses a risk for trusting LLMs without ensuring their safety. The paper suggests that CoT is promising for explainability, but requires more efforts to evaluate and improve explanation faithfulness.
As the paper discusses the findings the authors state:
LLMs may be able to recognize that the biasing features are influencing their predictions—e.g., in post-hoc critiques (Saunders et al., 2022)—even if their CoT explanations do not verbalize them. If they can, then this implies that unfaithful CoT explanations may be a form of model dishonesty, as opposed to a lack of capability.
What is becoming increasingly apparent with LLMs is that the approach to usage matters significantly. Basic CoT approaches do seem to cause more room for erroneous / deceptive / hallucination's as outputs. It may just be the case with relatively under developed models, and insufficient training approaches that these models were subjected to. It all goes to show that the general release of such models, when we still are unsure of their outputs, was too early.
Comments
Post a Comment