Probing the Misaligned Thinking Process of Language Models

This paper introduces a framework for detecting misaligned behaviors in large language models by identifying 18 fine-grained cognitive processes termed "misalignment indicators." Utilizing linear probes to analyze internal activations, the authors achieve a 0.935 AUROC in distinguishing misalignment across five behaviors while maintaining low false positive rates on benign inputs. This work is significant for practitioners as it provides a systematic approach to monitor and mitigate risks associated with deploying language models in sensitive applications.