EchoCoT: Extracting Hidden Chain-of-Thought from Large Reasoning Models
Researchers have developed EchoCoT, a method that can extract hidden chain-of-thought (CoT) traces from large reasoning models. These CoTs are valuable assets, but extracting them directly from black-box models has been largely unexplored. EchoCoT uses API interactions to iteratively extract hidden CoTs with high accuracy, even for very long CoTs. The method also generalizes to unseen datasets and can be used on both open-source and proprietary LRMs.
Researchers have developed EchoCoT, a method that can extract hidden chain-of-thought (CoT) traces from large reasoning models. These CoTs are valuable assets, but extracting them directly from black-box models has been largely unexplored. EchoCoT uses API interactions to iteratively extract hidden CoTs with high accuracy, even for very long CoTs. The method also generalizes to unseen datasets and can be used on both open-source and proprietary LRMs.
---
Why it matters: This matters because it highlights the potential security risk of extracting valuable information from large reasoning models without their owners' knowledge or consent. Engineers working on these models need to consider how to protect this sensitive information.
Source: https://arxiv.org/abs/2608.20055
This article was originally published at: https://arxiv.org/abs/2608.20055