Reproduce “Chain-of-Sanitized-Thoughts: Plugging PII Leakage in CoT of Large Reasoning Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Arghyadeep Das, Sai Sreenivas Chintha, Rishiraj Girmal, Kinjal Pandey, Sharvi Endait
The Chain-of-Sanitized-Thoughts (CoST) framework enables large reasoning models to proactively avoid Personally Identifiable Information (PII) leakage during their step-by-step reasoning process. The University of Massachusetts research demonstrates that both prompt engineering and supervised fine-tuning can substantially reduce PII exposure in reasoning traces with minimal impact on task utility, with efficacy depending on the model's inherent capabilities.
Key insights
- The effectiveness of privacy interventions is capability-dependent, with prompt engineering being more beneficial for higher-capability models and supervised fine-tuning for less capable ones.
- Privacy-first reasoning can be induced with minimal trade-off against utility, as interventions significantly reduce PII exposure while maintaining high task performance.
- Even with interventions, models remain most vulnerable to adversarially framed questions designed to coerce PII leakage, indicating a persistent challenge in robust privacy defense.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.