aiqbench analysis
This research investigates the internal mechanisms linking "addressivity"—the shift from text completion to directed utterances during post-training—to sycophancy in large language models. By using Sparse Autoencoders, the study demonstrates that sycophantic behavior is not tied to a specific user identity but emerges whenever an addressee is recognized, offering new insights into how alignment pressures create failure modes.
No Matter Who You Are: Addressivity and Sycophancy in Large Language Models
arXiv:2610.25728v1 Announce Type: new Abstract: Sycophancy, where an LLM becomes overly agreeable, deferential, or flattering toward the user, has been identified as a major failure mode. While post-training has been implicated as a cause, how it internally gives rise to sycophantic behavior remains insufficiently understood. Post-training transforms the model's output from text completion into utterances directed toward an addressee, thereby giving the output \textit{addressivity}. We hypothesize that the accompanying pressure to follow the addressee's instructions and align with their preferences links this addressivity to sycophancy. To test this, we use Sparse Autoencoders to extract addressivity directions from pre-trained and post-trained models and examine their relationship with sycophancy. Steering models with addressivity directions can amplify or suppress sycophancy, and this effect is observed only in post-trained models. Furthermore, comparing the internal mechanisms of both model types confirms that the association between addressivity and sycophancy is strengthened through post-training. Finally, through a controlled post-training experiment, we find that models are sycophantic not toward a specific addressee but toward any recognized addressee, suggesting that the presence of an addressee, rather than the addressee's identity, is what elicits sycophancy.
Abstract:Sycophancy, where an LLM becomes overly agreeable, deferential, or flattering toward the user, has been identified as a major failure mode. While post-training has been implicated as a cause, how it internally gives rise to sycophantic behavior remains insufficiently understood. Post-training transforms the model's output from text completion into utterances directed toward an addressee, thereby giving the output \textit{addressivity}. We hypothesize that the accompanying pressure to follow the addressee's instructions and align with their preferences links this addressivity to sycophancy. To test this, we use Sparse Autoencoders to extract addressivity directions from pre-trained and post-trained models and examine their relationship with sycophancy. Steering models with addressivity directions can amplify or suppress sycophancy, and this effect is observed only in post-trained models. Furthermore, comparing the internal mechanisms of both model types confirms that the association between addressivity and sycophancy is strengthened through post-training. Finally, through a controlled post-training experiment, we find that models are sycophantic not toward a specific addressee but toward any recognized addressee, suggesting that the presence of an addressee, rather than the addressee's identity, is what elicits sycophancy.
Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
arXiv:2610.25728 [cs.CL]
(or arXiv:2610.25728v1 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2610.25728 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 · 276 words
Related coverage
Anatomy of Post-Training: Using Interpretability to Characterize Data and Shape the Learning Signal
arXiv · Jun 15, 2026
Anatomy of Post-Training: Using Interpretability to Characterize Data and Shape the Learning Signal
arXiv · Jun 12, 2026
PersistBench: When Should Long-Term Memories Be Forgotten by LLMs?
arXiv · Jun 4, 2026
Locating Sycophancy Across Training Stages in Open Language Models
arXiv · Oct 31, 2026