Gumbel-Softmax Watermark Variant Can Hide LLM's True CoT Cryptographically

sytelus · x · 2026-08-31

The author proposes that a variant of the Gumbel-Softmax watermarking scheme, originally developed by Scott Aaronson at OpenAI, can be adapted to embed secret messages within the Chain of Thought (CoT) of LLMs. This suggests it is 100% possible for LLMs to completely conceal their actual reasoning in a cryptographically secure manner.

Original post →

More from Research

Research channel →