Skip to content

Dataset · kotekjedi

Recovered hidden reasoning of 27 closed LLMs

kotekjedi/stolen-thoughts

Dataset for research on extracting hidden reasoning chains from proprietary APIs on complex tasks.

What's inside

Repository for a research paper (arXiv:2608.09867) containing 10K to 100K English records in text and tabular formats. Contains proprietary API responses and recovered hidden reasoning chains for 27 models across Humanity's Last Exam, AIME 2025, and Codeforces tasks. Also includes a separate fortress slice with answers from a malicious knowledge testing experiment.

What you can do with it

  • Analysis and comparison of hidden reasoning traces from closed models
  • Researching vulnerabilities and safety of reasoning LLMs via the fortress slice
  • Distilling competition and algorithmic problem-solving skills into open models
Tasks
Text generation
Languages
English
Size
10K–100K rows
Downloads this month
13.7K
Likes
8
Access
on request on Hugging Face

How to load it

from datasets import load_dataset

ds = load_dataset("kotekjedi/stolen-thoughts")