wagwan

This paper teaches LLMs to avoid showing personally identifiable information (PII) in chain of thought, without hurting answers much. The problem is that chain of thought, meaning the model's step by step notes, can spill names, health details, or account data even if the final reply is cleaned. Instead of removing private bits after the model thinks, the authors try to make it reason with placeholders so sensitive details never get written out. They created a dataset of 350 prompts with privacy safe reasoning traces, then tested 2 approaches, prompt engineering, meaning stricter instructions, and supervised fine tuning, meaning training on examples. They measured leakage by counting how much input PII shows up in the reasoning text and by using another LLM as a judge, and they found prompts help strong models most while fine tuning helps weaker ones. That matters because chain of thought often gets logged inside apps, so private reasoning reduces a big hidden privacy risk. ---- Paper Link – arxiv. org/abs/2601.05076 Paper Title: "Chain-of-Sanitized-Thoughts: Plugging PII Leakage in CoT of Large Reasoning Models"

Cow posting will continue indefinitely.

nikhil kamath has successfully built the most overstaffed podcast in history

USING Claude Opus 4.7 TO CENTER A DIV