Saved in:
| Main Authors: | , , , , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2026
|
| Subjects: | |
| Online Access: | https://arxiv.org/abs/2605.27186 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866918524746727424 |
|---|---|
| author | Zheng, Haoyu Zhu, Yun Yuan, Shu Chen, Shangming Wang, Qing Zhang, Wenqiao Xiao, Jun Zhuang, Yueting |
| author_facet | Zheng, Haoyu Zhu, Yun Yuan, Shu Chen, Shangming Wang, Qing Zhang, Wenqiao Xiao, Jun Zhuang, Yueting |
| contents | Large language models often solve tasks from a fully specified prompt but degrade when the same requirements unfold over multiple turns, known as the lost-in-conversation (LiC) gap. We trace part of this degradation to self-contamination: intermediate assistant replies enter later context and carry early deviations forward. Motivated by this mechanism, we propose MAIGO, an on-policy self-distillation method that reduces this contamination using history-cleaned references from the model's own policy. For middle turns, MAIGO removes prior assistant replies while preserving the user-visible sharded prefix; for answer turns, it distills from paired full-view references conditioned on the completed user-side dialogue. A reliability weight downweights middle-turn samples that disagree with the clean reference. MAIGO requires no verifier rewards, state labels, or inference-time scaffolding. Under the LiC paired-view protocol with deterministic verifiers, MAIGO improves Qwen2.5-7B-Instruct SHARDED accuracy from 52.8 to 66.1 and the SHARDED/FULL ratio from 66.5% to 84.1%, while keeping FULL accuracy within 2.3 points. These results show that self-contamination is a trainable component of the LiC gap. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2605_27186 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | MAIGO: Mitigating Lost-in-Conversation with History-Cleaned On-Policy Self-Distillation Zheng, Haoyu Zhu, Yun Yuan, Shu Chen, Shangming Wang, Qing Zhang, Wenqiao Xiao, Jun Zhuang, Yueting Computation and Language Large language models often solve tasks from a fully specified prompt but degrade when the same requirements unfold over multiple turns, known as the lost-in-conversation (LiC) gap. We trace part of this degradation to self-contamination: intermediate assistant replies enter later context and carry early deviations forward. Motivated by this mechanism, we propose MAIGO, an on-policy self-distillation method that reduces this contamination using history-cleaned references from the model's own policy. For middle turns, MAIGO removes prior assistant replies while preserving the user-visible sharded prefix; for answer turns, it distills from paired full-view references conditioned on the completed user-side dialogue. A reliability weight downweights middle-turn samples that disagree with the clean reference. MAIGO requires no verifier rewards, state labels, or inference-time scaffolding. Under the LiC paired-view protocol with deterministic verifiers, MAIGO improves Qwen2.5-7B-Instruct SHARDED accuracy from 52.8 to 66.1 and the SHARDED/FULL ratio from 66.5% to 84.1%, while keeping FULL accuracy within 2.3 points. These results show that self-contamination is a trainable component of the LiC gap. |
| title | MAIGO: Mitigating Lost-in-Conversation with History-Cleaned On-Policy Self-Distillation |
| topic | Computation and Language |
| url | https://arxiv.org/abs/2605.27186 |