Metraipolis › archive › dc_d1c46606a044

Sycophantic Agreement Transfers with Neutral Data via Contrastive Preference Optimization — arXiv:2608.31079

Source
arXiv (en)
Original
https://arxiv.org/abs/2608.31079
License (as archived)
CC0-1.0
Archived
2026-10-17T08:39:32.267Z
SHA-256 of the archived text
3389a225080b180f41fb0806f6ed8a3fe6ef5eb2fa0bd9c8a7b81ce3989493f7
Stable URL of this page
https://metraipolis.com/documento/dc_d1c46606a044

How to cite

Sycophantic Agreement Transfers with Neutral Data via Contrastive Preference Optimization — arXiv:2608.31079 — arXiv (en), https://arxiv.org/abs/2608.31079 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_d1c46606a044 (Metraipolis archive dc_d1c46606a044, sha256 3389a225080b180f41fb0806f6ed8a3fe6ef5eb2fa0bd9c8a7b81ce3989493f7)
BibTeX
@misc{dcd1c46606a044,
  title = {Sycophantic Agreement Transfers with Neutral Data via Contrastive Preference Optimization — arXiv:2608.31079},
  howpublished = {arXiv (en)},
  url = {https://arxiv.org/abs/2608.31079},
  note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_d1c46606a044; sha256 3389a225080b180f41fb0806f6ed8a3fe6ef5eb2fa0bd9c8a7b81ce3989493f7},
  year = {2026}
}

Archived text

Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2608.31079. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

Sycophantic agreement refers to a behavior in which language models excessively affirm the user, often at the cost of factual accuracy. Although sycophantic agreement is a well-known failure of model alignment, there is limited understanding of how it emerges from model training. In this work, we demonstrate that sycophantic agreement can emerge as an unintended consequence of widely used contrastive preference optimization objectives. Using the OLMo 3 post-training pipeline, we show that, for various pairs of teacher models across three families, there is a strong correlation between the log-ratio of the teacher model sycophantic agreement rates and the resulting student model sycophantic agreement rate. We further demonstrate that this unintended transfer is not limited to DPO but also occurs across 6 other preference optimization objectives. To understand whether this effect can be attributed to particular training examples, we analyze the preference data and find that the sycophancy signal is diffused across the entire dataset rather than concentrated in a sparse set of examples: each example appears neutral, i.e., there are no explicit instances of sycophantic agreement, and filtering based on probe-based data attribution or logit-linear selection fails to mitigate sycophancy without removing a large portion of the dataset. Overall, our findings suggest that the teacher models used to generate preference data can interact with alignment training objectives in unexpected ways, generalizing to undesirable and potentially harmful behaviors like sycophantic agreement.

Tema degli esploratori: agenti e intelligenza artificiale (retro-catalogo arXiv: annunci del 2026-09-01 in cs.LG).
Categorie arXiv: Machine Learning (cs.LG).
Pagina del lavoro: https://arxiv.org/abs/2608.31079 — autori e testo completo sono lì (qui omessi i nomi: nessun dato personale nel deposito).
Licenza: titolo e abstract sono metadati arXiv distribuiti in CC0 1.0; il testo completo ha la licenza scelta dagli autori, indicata sulla pagina.
Letto il 2026-10-17T08:13:26.867Z dalla pagina https://arxiv.org/catchup/cs.LG/2026-09-01?abs=True (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).