Interpretability tools can transfer across models

1. 00:43
   arXiv: Jailbreaks, Prompt Injection, Poisoning & DefensesSelected

   Interpretability is not model-specific

   University of Houston researchers found that interpretability tools such as refusal directions, correctness directions, and sentiment probes obtained on one model can be transferred to other models via a linear mapping fitted using only unlabeled prompts. The experiments covered all 110 ordered pairs of 11 instruction-tuned models (1B to 72B parameters) across five families, as well as between base and instruction-tuned models. Mappings fitted with ordinary questions recovered 54% of the target model's own refusal steering effect and 16% to 77% of the correctness gain; after adding unlabeled prompts in the tool's region of effect, this rose to 102% and 88% to 98%, and refusal transfer required only 16 prompts per category, while control directions recovered at most 1%. The paper proves that transfer necessarily holds when two models correspond linearly in the region covered by the mapping data, and the resulting unlabeled coverage score tracks recovery rates across more than 16,000 runs and explains previous reports of cross-domain mapping failure; transfer requires white-box access to the target model's activations.

Current attention 40·Peak in the comparable period 43 (Nov 10 22:00)·Change over the comparable past 24 hours 0%