REP-models / README.md
0x-YuAN's picture
Super-squash branch 'main' using huggingface_hub
1937932
|
Raw
History Blame Contribute Delete
2.83 kB
metadata
license: apache-2.0
base_model: Qwen/Qwen2.5-7B-Instruct
pipeline_tag: text-generation
language:
  - en
tags:
  - reasoning
  - distillation
  - math
  - reasoning-trace-exposure
  - rep
datasets:
  - Chia-Mu-Lab/REP-datasets

REP-models — students distilled from exposed reasoning traces

Checkpoints released with "Hidden Thoughts Are Not Secret: Reasoning-Trace Exposure in LLMs" (EMNLP 2026, arXiv:2606.00642). Code: https://github.com/0x-yuan/REP.

Every student is Qwen/Qwen2.5-7B-Instruct fine-tuned on one of the corpora in Chia-Mu-Lab/REP-datasets. One sub-folder per final model.

Sub-folders

Sub-folder Supervision Training data (config in REP-datasets)
qwen25-7b-rep-q3_14b-clean REP-exposed trace of Qwen3-14B, answer-correct rows distill_q3_14b_clean
qwen25-7b-rep-q3_14b-original REP-exposed trace of Qwen3-14B, all structurally valid rows distill_q3_14b_original
qwen25-7b-rep-q3_32b-clean REP-exposed trace of Qwen3-32B, answer-correct rows distill_q3_32b_clean
qwen25-7b-rep-q3_32b-original REP-exposed trace of Qwen3-32B, all structurally valid rows distill_q3_32b_original
qwen25-7b-oracle-q3_14b Qwen3-14B internal trace (oracle upper bound) Qwen3-14B internal traces (not released)
qwen25-7b-oracle-q3_32b Qwen3-32B internal trace (oracle upper bound) Qwen3-32B internal traces (not released)

Results (paper Table 2 protocol: n=3, T=0.5; JEE = math subset, strict / partial)

Model MATH500 AIME24 AIME25 JEE-Math LiveCodeBench
Qwen2.5-7B-Instruct (base) 71.0 8.9 2.2 32.2 / 35.9 15.8
rep-q3_14b-clean 75.8 14.4 13.3 35.2 / 39.5 19.0
rep-q3_14b-original 72.4 12.2 13.3 33.5 / 38.9 18.3
rep-q3_32b-clean 72.8 14.4 17.8 36.4 / 41.1 15.8
rep-q3_32b-original 73.9 13.3 13.3 38.1 / 42.2 16.5
oracle-q3_14b 70.3 14.4 13.3 48.5 / 51.2 14.7
oracle-q3_32b 70.0 16.7 15.6 46.4 / 49.3 15.8

Traces exposed through prompting alone match or exceed the oracle internal trace as distillation supervision on math; see the paper for the full analysis.

Responsible use

Research artifact for studying reasoning-trace exposure and defenses. All training data derive from open-weight models. See ETHICS.md in the code release.

Citation

@inproceedings{lu2026hiddenthoughts,
  title     = {Hidden Thoughts Are Not Secret: Reasoning-Trace Exposure in LLMs},
  author    = {Lu, Yu-An and Tsai, Ci-Yang and Tsai, Yu-Lin and Popa, Raluca Ada and Yu, Chia-Mu},
  booktitle = {Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing (EMNLP)},
  year      = {2026}
}