Training Language Models to Explain Their Own Computations

The study titled 'Training Language Models to Explain Their Own Computations' investigates the ability of language models (LMs) to articulate their internal workings. By fine-tuning LMs with tens of thousands of example explanations, researchers found that these models can generate coherent natural language descriptions of their features and causal structures. Notably, LMs exhibit a unique advantage when explaining their own computations, outperforming other models, regardless of their capabilities. This finding is significant as it not only confirms that LMs can learn to explain their computations reliably but also suggests that such explanations can serve as a valuable tool for enhancing interpretability in artificial intelligence. The implications of this research extend to improving transparency in AI systems, making it easier for users to understand how these models arrive at their outputs.