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

Spanish

Los modelos de lenguaje pueden cambiar una respuesta comprometida para acomodar la opinión del usuario o el empujón sin nuevas evidencias. Sin embargo, cuando este comportamiento se desarrolla durante el entrenamiento y lo que un modelo lanzado hereda de las etapas anteriores permanece incierto. Seguimos 741 puntos de control de 30 líneas de modelo abierto, desde el pre-entrenamiento hasta los modelos lanzados. Plantamos juicios iniciales sobre escenarios morales y medimos hasta qué punto un modelo sigue la opinión expresada por un usuario o abandona su juicio bajo presión. En la mayoría de las líneas el comportamiento está presente antes del post-entrenamiento, y su fuerza media cambia en ambas direcciones, en diferentes etapas entre las líneas. Un patrón más claro surge en el nivel de las preguntas individuales. Más allá de las respuestas sin presión, el comportamiento

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