Auditing Reasoning-Trace Memorization Claims after Unlearning with Head-Conditioned Canaries
Fuente:
arXiv
Enregistré dans:
| Auteurs principaux: | , , |
|---|---|
| Format: | Preprint |
| Publié: |
2026
|
| Sujets: | |
| Accès en ligne: | |
| Tags: |
Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
|
| _version_ | 1866909054976131072 |
|---|---|
| author | Li, Yanhang Fan, Zhichao Zhuang, Zexin |
| author_facet | Li, Yanhang Fan, Zhichao Zhuang, Zexin |
| contents | Evaluations of unlearning on reasoning models sometimes show a bypass pattern. The answer side looks unlearned, but the model's own thinking trace keeps emitting the forgotten content, and the gap is taken as evidence that the weights still remember. We audit this reading on DeepSeek-R1-Distill-Qwen-7B with LoRA-memorized fictional authors and NPO unlearning, conditioned on a six-token canary head. On one seed, swapping the thinking trace for a short non-canary prefill on the same weights drops the answer rate by as much as the bypass gap itself, whether the prefill mimics the training template or not. On a second seed the bypass gap shrinks rather than vanishing, and the prefill swap reverses direction and brings the answer rate to ceiling. A positive parser-split bypass gap thus does not by itself identify hidden weight-level memorization, and does not rule it out either. On a different distillate the same metric flips sign because the parser cannot find the closing tag. We recommend a decode-time template swap as a cheap sanity check alongside the canonical audit. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2605_18891 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | Auditing Reasoning-Trace Memorization Claims after Unlearning with Head-Conditioned Canaries Li, Yanhang Fan, Zhichao Zhuang, Zexin Machine Learning Artificial Intelligence Evaluations of unlearning on reasoning models sometimes show a bypass pattern. The answer side looks unlearned, but the model's own thinking trace keeps emitting the forgotten content, and the gap is taken as evidence that the weights still remember. We audit this reading on DeepSeek-R1-Distill-Qwen-7B with LoRA-memorized fictional authors and NPO unlearning, conditioned on a six-token canary head. On one seed, swapping the thinking trace for a short non-canary prefill on the same weights drops the answer rate by as much as the bypass gap itself, whether the prefill mimics the training template or not. On a second seed the bypass gap shrinks rather than vanishing, and the prefill swap reverses direction and brings the answer rate to ceiling. A positive parser-split bypass gap thus does not by itself identify hidden weight-level memorization, and does not rule it out either. On a different distillate the same metric flips sign because the parser cannot find the closing tag. We recommend a decode-time template swap as a cheap sanity check alongside the canonical audit. |
| title | Auditing Reasoning-Trace Memorization Claims after Unlearning with Head-Conditioned Canaries |
| topic | Machine Learning Artificial Intelligence |
| url | https://arxiv.org/abs/2605.18891 |