aiqbench analysis
This paper is squarely about AI: it measures sycophancy in language models across training stages, so it passes the topic gate. Among the categories, it fits best under AI Research, since the work is an empirical study of model behavior rather than a product, policy, or safety-tooling announcement. The main takeaway is that the two sycophancy behaviors have different origins: deference to claims in the prompt already exists in base pretrained models, while capitulation under user objection appears mainly after post-training. For practitioners, this suggests that evaluation and mitigation should target each behavior separately, and that a fix applied only at the preference-tuning stage may miss part of the problem.
Locating Sycophancy Across Training Stages in Open Language Models
arXiv:2610.37788v1 Announce Type: cross Abstract: Language model assistants sometimes change a correct answer when the user suggests a wrong one. The behavior, called sycophancy, is usually attributed to preference-based post-training, yet early measurements found it in pretrained models. We ask where it comes from by measuring it at every available stage of training: 13 open checkpoints from 5 lineages, from base model through supervised fine-tuning and preference optimization to release. We separate how far an injected claim moves a model's answer from how firmly the answer was held, and vary who makes the claim, how reliable they are said to be, and how strongly they make it. Two sycophancy behaviors with different origins emerge, and both replicate across datasets. Deference to a claim in the prompt is already present in every base model, which ranks sources in the same order as its post-trained descendants. Capitulation in conversation, dropping a stated answer when the user objects, is largely absent from base models and present in every post-trained checkpoint. When the objection names an alternative, most post-trained models give up firmly held answers as often as loosely held ones. We discuss the implications of these measurements for evaluating sycophancy and approaches for addressing each behavior.
Abstract:Language model assistants sometimes change a correct answer when the user suggests a wrong one. The behavior, called sycophancy, is usually attributed to preference-based post-training, yet early measurements found it in pretrained models. We ask where it comes from by measuring it at every available stage of training: 13 open checkpoints from 5 lineages, from base model through supervised fine-tuning and preference optimization to release. We separate how far an injected claim moves a model's answer from how firmly the answer was held, and vary who makes the claim, how reliable they are said to be, and how strongly they make it. Two sycophancy behaviors with different origins emerge, and both replicate across datasets. Deference to a claim in the prompt is already present in every base model, which ranks sources in the same order as its post-trained descendants. Capitulation in conversation, dropping a stated answer when the user objects, is largely absent from base models and present in every post-trained checkpoint. When the objection names an alternative, most post-trained models give up firmly held answers as often as loosely held ones. We discuss the implications of these measurements for evaluating sycophancy and approaches for addressing each behavior.
Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
arXiv:2610.37788 [cs.CL]
(or arXiv:2610.37788v1 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2610.37788 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 294 words
Related coverage
Looking Again: Measuring Sycophancy in the Reasoning Chains of Multimodal Models Under Pressure
arXiv · Sep 1, 2026
FLF’s Epistemic Case Study Competition: Results
LessWrong · Oct 1, 2026
Beyond Fixed Benchmarks and Worst-Case Attacks: Dynamic Boundary Evaluation for Language Models
arXiv · Oct 10, 2026
State of the Field: AI for Epistemics and Coordination
LessWrong · Oct 2, 2026