Minimally Invasive Steering of Language Models
MISVO steers frozen language models toward a test-time reward while limiting distribution shift via local KL geometry.
Researchers propose Minimally Invasive Steering Vector Optimization (MISVO), which adds vectors to a frozen language model's final hidden states and penalizes interventions using the local KL geometry of the induced token distribution. An analytic Fisher gradient is computed through the frozen model head, and a frozen-reference surrogate keeps parameters unchanged. On preference and code-generation tasks with models of about 1B to 14B parameters, MISVO achieved the highest mean reward in six of seven settings while diversity and coherence stayed near Best-of-N.
- MISVO penalizes steering using the Fisher geometry of the token distribution.
- Model parameters stay frozen; only position-specific steering vectors are optimized.
- Highest mean reward in six of seven model-task settings.
- Diversity and coherence remain close to Best-of-N sampling.
Full article172 words · extracted from arxiv.org · click to collapse
Pre-logit steering adapts a frozen language model to a test-time reward by adding vectors to its final hidden states. Unregularized reward optimization can substantially alter the output distribution and degrade generation quality. We propose Minimally Invasive Steering Vector Optimization (MISVO), which penalizes interventions using the local KL geometry of the induced token distribution. The resulting Fisher quadratic measures distributional sensitivity and admits an analytic gradient computed through matrix--vector products with the frozen language-model head. We derive an exact decomposition of the sequence-level KL gradient into an analytic Fisher term and a suffix score-function term. For a fixed generation horizon, we show that the suffix term is second order in the steering magnitude and that three Fisher surrogates agree with the full KL gradient to first order. MISVO uses the frozen-reference surrogate to optimize position-specific interventions without updating model parameters. Across preference and code-generation tasks on models with approximately 1B--14B parameters, MISVO achieves the highest mean reward in six of seven model--task settings, with diversity and coherence scores close to those of Best-of-N.
Text extracted automatically; images, tables and formatting may be missing. Original: https://arxiv.org/abs/2609.30218