On the Hardness of Devoted Chain-of-Thought Reasoning in Massive Language Fashions
Authors: Sree Harsha Tanneru, Dan Ley, Chirag Agarwal, Himabindu Lakkaraju
Summary: As Massive Language Fashions (LLMs) are more and more being employed in real-world functions in vital domains equivalent to healthcare, it is very important make sure that the Chain-of-Thought (CoT) reasoning generated by these fashions faithfully captures their underlying conduct. Whereas LLMs are recognized to generate CoT reasoning that’s interesting to people, prior research have proven that these explanations don’t precisely mirror the precise conduct of the underlying LLMs. On this work, we discover the promise of three broad approaches generally employed to steer the conduct of LLMs to boost the faithfulness of the CoT reasoning generated by LLMs: in-context studying, fine-tuning, and activation enhancing. Particularly, we introduce novel methods for in-context studying, fine-tuning, and activation enhancing geared toward bettering the faithfulness of the CoT reasoning. We then perform intensive empirical analyses with a number of benchmark datasets to discover the promise of those methods. Our analyses point out that these methods supply restricted success in bettering the faithfulness of the CoT reasoning, with solely slight efficiency enhancements in managed situations. Activation enhancing demonstrated minimal success, whereas fine-tuning and in-context studying achieved marginal enhancements that didn’t generalize throughout numerous reasoning and truthful question-answering benchmarks. In abstract, our work underscores the inherent problem in eliciting trustworthy CoT reasoning from LLMs, suggesting that the present array of approaches might not be enough to handle this complicated problem.
