Shared Lexical Task Representations Explain Behavioral Variability in LLMs (arxiv.org)

A recent study has highlighted significant insights into the unpredictable performance of large language models (LLMs), particularly concerning their sensitivity to different prompting styles. Researchers explored the contrasts between instruction-based prompts, which articulate tasks in natural language, and example-based prompts that utilize few-shot demonstrations. Their findings reveal that despite varying performance outcomes, there are common underlying mechanisms at play within the models. Notably, they identified "lexical task heads," which are specific attention heads responsible for task representation and are shared across different prompting methods. This research is pivotal for the AI/ML community as it clarifies why LLMs can exhibit such behavioral variability in response to different prompts. By demonstrating that the performance inconsistency can be explained by the activation levels of these shared task representations, the study opens avenues for improving model design and reliability. It suggests that understanding and enhancing the activation of these lexical task heads could mitigate some of the performance issues developers face, leading to more robust and predictable LLMs in practical applications.