Automated Pipeline Detects Unexpected LLM Behavioral Shifts

An automated pipeline for contrastive evaluation has been created by researchers to assess how interventions affect large language models' behavior. This technique examines the free-form, multi-token outputs of a base model, M1, against those of an intervention model, M2, within aligned prompt contexts. It generates natural-language hypotheses that are both human-readable and statistically validated, outlining the differences between the models, along with recurring themes that highlight patterns in the validated hypotheses. The pipeline was tested in a synthetic environment by introducing known behavioral changes, which it accurately identified. It was subsequently utilized for three real-world interventions: reasoning distillation, knowledge editing, and unlearning. The findings revealed both expected and unexpected behavioral changes, effectively differentiating between significant and minor interventions, without generating false differences when effects were absent or misaligned with the prompts. The research is documented in arXiv:2605.05090v1.