Newest papers

The Development of Sycophancy in Language Models: Tracing Vulnerabilities from Pretraining to Post-Training

Haonan Huang, Chengkun Cai · 2026 · arXiv (Cornell University)

DOI: 10.48550/arxiv.2610.26655 · cited by 0 works (OpenAlex) · open-access PDF

Russian

Модели языка могут изменить приверженный ответ, чтобы удовлетворить мнение пользователя или pushback без новых доказательств. Однако, когда это поведение развивается во время обучения и что выпущенная модель наследует от более ранних этапов остается неясным. Мы следим за 741 контрольным пунктом из 30 открытых модельных линий, от ранней предварительной подготовки до выпущенных моделей. Мы высаживаем первоначальные решения по нравственным сценариям и измеряем, насколько модель следует высказанному мнению пользователя или отказывается от своего решения под давлением. В большинстве линий поведение присутствует до послеобучения, а его средняя сила изменяется в любом направлении, на разных этапах между линиями. Более ясная модель возникает на уровне индивидуальных вопросов. Помимо ответов без давления,

Machine translation (Meta M2M100); the English original follows.

Abstract

Language models can change a committed answer to accommodate a user's opinion or pushback without new evidence. However, when this behaviour develops during training and what a released model inherits from earlier stages remain unclear. We follow 741 checkpoints of 30 open model lineages, from early pretraining to released models. We plant initial verdicts on moral scenarios and measure how far a model follows a user's stated opinion or abandons its verdict under pressure. In most lineages the behaviour is present before post-training, and its average strength changes in either direction, at different stages across lineages. A clearer pattern emerges at the level of individual questions. Beyond responses without pressure, pretraining behaviour predicts little of which questions a released model yields on. That pattern forms during post-training: on routes with supervised fine-tuning, the fine-tuning checkpoint predicts part of it and the preference-optimization checkpoint most of it. A sibling trained from the same base along another route predicts far less: for Tülu-3 8B, its own preference-stage checkpoint removes 85% of the remaining prediction error, a sibling 19%. Internal states add predictive information without changing this ranking. Similar average scores conceal different vulnerabilities, and training history matters for auditing a released model.

Read in: English · Spanish · Chinese · Hindi · Arabic · Portuguese · French · German · Japanese · Russian · Korean · Indonesian · Vietnamese

Topics: Computer science, Measure (data warehouse), Psychology, Audit, Artificial intelligence, Cognitive psychology

Is this paper real? This page was built from OpenAlex and arXiv records, not generated. Software can verify any reference in one call: GET https://store.agentexchange.work/gw/papers/verify?title=...&year=... (0.003 USD), fetch this record as JSON with /gw/papers/paper?id=2610.26655, or list what it cites with /gw/papers/references?id=.... Pay per call in USDC (x402) or with a 19 USD card key for 1,000 calls. Catalogue · llms.txt

Sources: OpenAlex (CC0), arXiv, Crossref. Agent Exchange