Christopher Potts - Creating and detecting subliminal learning effects on demand - IPAM at UCLA

Recorded 01 September 2026. Christopher Potts of Stanford University presents "Creating and detecting subliminal learning effects on demand" at IPAM's Foundations of Interpretability Workshop.
Abstract: This talk will argue for a data-centric approach to interpretability. I will do this through the lens of subliminal learning, in which language models acquire behaviors from examples that seem unrelated to those behaviors (Cloud, Le, et al. 2025). I will first focus on using Data Policy Gradients (DPG; Thrush et al. 2026) to generate these effects on demand. DPG is a framework for synthesizing examples that, when used as training data, cause the target model to do well on a chosen differentiable metric, including surprising metrics like embedding a QR code pattern in the model weights. DPG creates a setting in which we can study the causal links between examples and behaviors, for subliminal effects as well as more obvious ones. I will then turn to detecting subliminal effects on demand, across a range of adversarial settings, using soft-prompt learning (Hu and Potts 2026). Finally, I will present a case study of sycophancy in OLMo. In Blank et al. 2026, we trace the relevant sycophantic behaviors specifically to the DPO phase of OLMo training, where they emerge as a subliminal effect stemming from the relationship between two models used to generate preference pairs. This analysis gives us fine-grained control over the sycophantic behavior itself. Taken together, these perspectives on subliminal learning provide a rich picture of the origin and nature of the effects, and they show that, by focusing on the relationship between data and model behaviors, we obtain real power to detect, characterize, and modulate the effects before they become embedded in model representations.

Cloud, Alex; Minh Le; James Chua; Jan Betley; Anna Sztyber-Betley; Jacob Hilton; Samuel Marks; and Owain Evans. 2025. Subliminal Learning: Language models transmit behavioral traits via hidden signals in data. https://arxiv.org/abs/2507.14805

Hu, Nathan and Christopher Potts. 2026. Verbalizing hidden fine-tuning effects with text optimization. To appear on arXiv!

Thrush, Tristan; Sung Min Park; Herman Brunborg; Luke Bailey; Marcel Roed; Neil Band; Christopher Potts; and Tatsunori Hashimoto. 2026. Synthetic data for any differentiable target. To appear in Proceedings of CoLM. https://arxiv.org/abs/2604.08423

Blank, Camila; Zhuofan Ying; Christopher Potts; Peter Hase; and Jing Huang. 2026. Sycophantic agreement transfers with neutral data via contrastive preference optimization. To appear on arXiv and at the CoLM Actionable Interpretability Workshop!

Learn more online at: https://www.ipam.ucla.edu/programs/workshops/foundations-of-interpretability/?tab=overview Receive SMS online on sms24.me

Watch on YouTube

Subscribe on YouTubeReaderBot

TubeReader video aggregator is a website that collects and organizes online videos from the YouTube source. Video aggregation is done for different purposes, and TubeReader take different approaches to achieve their purpose.

Our try to collect videos of high quality or interest for visitors to view; the collection may be made by editors or may be based on community votes.

Another method is to base the collection on those videos most viewed, either at the aggregator site or at various popular video hosting sites.

TubeReader site exists to allow users to collect their own sets of videos, for personal use as well as for browsing and viewing by others; TubeReader can develop online communities around video sharing.

Our site allow users to create a personalized video playlist, for personal use as well as for browsing and viewing by others.

@YouTubeReaderBot allows you to subscribe to Youtube channels.

By using @YouTubeReaderBot Bot you agree with YouTube Terms of Service.

Use the @YouTubeReaderBot telegram bot to be the first to be notified when new videos are released on your favorite channels.

Look for new videos or channels and share them with your friends.

You can start using our bot from this video, subscribe now to Christopher Potts - Creating and detecting subliminal learning effects on demand - IPAM at UCLA

What is YouTube?

YouTube is a free video sharing website that makes it easy to watch online videos. You can even create and upload your own videos to share with others. Originally created in 2005, YouTube is now one of the most popular sites on the Web, with visitors watching around 6 billion hours of video every month.