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

French

Les modèles linguistiques peuvent modifier une réponse engagée pour accueillir l’opinion d’un utilisateur ou le pushback sans nouvelles preuves. Cependant, lorsque ce comportement se développe au cours de la formation et ce qu’un modèle libéré hérite des étapes antérieures reste incertain. Nous suivons 741 points de contrôle de 30 lignes de modèle ouvert, de la pré-entraînement aux modèles publiés. Nous plantons des jugements initiaux sur les scénarios moraux et mesurons jusqu’où un modèle suit l’opinion exprimée par un utilisateur ou abandonne son jugement sous pression. Dans la plupart des lignes, le comportement est présent avant la post-entraînement, et sa force moyenne change dans chacune des directions, à différentes étapes entre les lignes. Un modèle plus clair apparaît

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

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