Automatically Finding and Validating Unexpected Side-Effects of Interventions on Language Models
It provides a statistically grounded, interpretable tool for auditing model interventions, addressing the need for post-hoc behavioral validation in LLM development.
The paper presents an automated pipeline to detect and validate unexpected behavioral changes in language models after interventions, recovering known changes in synthetic tests and surfacing both intended and unintended shifts in real-world interventions like reasoning distillation, knowledge editing, and unlearning.
We present an automated, contrastive evaluation pipeline for auditing the behavioral impact of interventions on large language models. Given a base model $M_1$ and an intervention model $M_2$, our method compares their free-form, multi-token generations across aligned prompt contexts and produces human-readable, statistically validated natural-language hypotheses describing how the models differ, along with recurring themes that summarize patterns across validated hypotheses. We evaluate the approach in synthetic setting by injecting known behavioral changes and showing that the pipeline reliably recovers them. We then apply it to three real-world interventions, reasoning distillation, knowledge editing and unlearning, demonstrating that the method surfaces both intended and unexpected behavioral shifts, distinguishes large from subtle interventions, and does not hallucinate differences when effects are absent or misaligned with the prompt bank. Overall, the pipeline provides a statistically grounded and interpretable tool for post-hoc auditing of intervention-induced changes in model behavior.