When Collaboration Recovers Forgotten Knowledge: Robust Unlearning for LLM Teams

Hengsong Liu, Liangyu Teng, jingnan cai, Yuntian Shi, Jingyu Zhang, Yang Liu, Jing Liu, Liang SongOpenReview 2027 · Submission to ICLR 2027 · Withdrawn · OpenReview

Its authors withdrew this paper, so there is no full text to read or summarize.

Abstract

LLM unlearning supports privacy protection by removing unwanted knowledge without full retraining. LLM teams are widely adopted because collaboration improves answer quality. Despite advances in robustness to adversarial queries, repeated sampling, and model updates, unlearning under increasingly widespread team deployment remains underexplored. Our experiments show that intermediate messages can enable downstream recovery even when the final answer suppresses the target. We therefore propose Aggregation-Conditioned Unlearning (ACU) to directly supervise message recoverability. ACU trains a shared proposer– aggregator policy using native-output scores and fixed-reader reconstructions from individual messages, pooled proposals, and full trajectories. It prioritizes the most revealing audit, assigns credit through independent replacements, and constrains direct and collaborative utility. Experiments span three model families and two datasets, with recovery analyses across target models, readers, team sizes, and sampling protocols. On TOFU, ACU reduces full-trajectory recovery ROUGE from 41.8 to 19.3 relative to native-output training, and fact recovery by 10.9–15.6 percentage points under matched sampling budgets. These results demonstrate ef- fective collaborative unlearning while preserving retained capabilities.
BibTeX
@misc{liu2027when,
  author = {Hengsong Liu and Liangyu Teng and jingnan cai and Yuntian Shi and Jingyu Zhang and Yang Liu and Jing Liu and Liang Song},
  title = {{When Collaboration Recovers Forgotten Knowledge: Robust Unlearning for LLM Teams}},
  year = {2027},
  url = {https://openreview.net/forum?id=hl8kmkVpoo},
  howpublished = {OpenReview},
  note = {Submitted to ICLR 2027},
}

Loading related content...

Figure