Evaluating Steering Techniques using Human Similarity Judgments
This work addresses the need for cognitive alignment in LLM steering evaluations, which is incremental as it builds on existing steering methods by introducing a human-centered benchmark.
The study tackled the problem of evaluating LLM steering techniques by assessing their alignment with human cognition using a triadic similarity judgment task, finding that prompt-based methods outperformed others in steering accuracy and model-to-human alignment, with LLMs showing a bias towards 'kind' similarity and struggling with 'size' alignment.
Current evaluations of Large Language Model (LLM) steering techniques focus on task-specific performance, overlooking how well steered representations align with human cognition. Using a well-established triadic similarity judgment task, we assessed steered LLMs on their ability to flexibly judge similarity between concepts based on size or kind. We found that prompt-based steering methods outperformed other methods both in terms of steering accuracy and model-to-human alignment. We also found LLMs were biased towards 'kind' similarity and struggled with 'size' alignment. This evaluation approach, grounded in human cognition, adds further support to the efficacy of prompt-based steering and reveals privileged representational axes in LLMs prior to steering.