<ul><li>Steering vectors are a lightweight method for controlling text properties by adding a learned bias to language model activations at inference time.</li><li>Evaluation of steering vectors in adaptive free-form summarization tasks is explored beyond multiple-choice settings.</li><li>Steering vectors effectively control topical focus, sentiment, toxicity, and readability in abstractive summaries, but high steering strengths can degrade text quality.</li><li>Combining steering and prompting offers the strongest control over text properties with a favorable efficacy-quality trade-off at moderate steering strengths.</li></ul>

Beyond Multiple Choice: Evaluating Steering Vectors for Adaptive Free-Form Summarization

Discover more