Samuel Pfrommer
Robotic Foundation Models, Skild AI
Skild AI
Samuel Pfrommer builds robotic foundation models at Skild AI and was previously a PhD student at Berkeley EECS, advised by Somayeh Sojoudi. He is co-first author of Ranking Manipulation for Conversational Search Engines, an EMNLP 2024 oral paper on how prompt injections change which sources AI search engines rank first.
Berkeley and the Study of Robust Learning
Pfrommer's personal site lists his research interests as robust machine learning, safe reinforcement learning, and diffusion-based control. At Berkeley he worked with Somayeh Sojoudi on certified robustness, publishing "Asymmetric Certified Robustness via Feature-Convex Neural Networks" at NeurIPS 2023 and "Transport of Algebraic Structure to Latent Embeddings" as an ICML 2024 spotlight.
A Paper on Ranking in Conversational Search
With Yatong Bai, Tanmay Gautam, and Sojoudi, Pfrommer wrote Ranking Manipulation for Conversational Search Engines, which he and Bai share first authorship of and which was an oral presentation at EMNLP 2024. The authors built a dataset of real consumer product websites, showed that different LLMs weigh product name, document content, and context position differently, and then presented a tree-of-attacks technique that reliably promotes low-ranked products. The attacks transferred to perplexity.ai.
The paper's closing argument is about incentives: given the financial reward for ranking higher, the authors called the problem important for future robustness work. Chapter 7 of the GEO/AEO History covers it with the two other manipulation papers of 2024, including the Preference Manipulation Attacks work of Edoardo Debenedetti and Florian Tramèr.
From Search Robustness to Robotics
Pfrommer's earlier work includes ContactNets, a method for learning robot dynamics with discontinuous contact, presented at CoRL 2020. He is now at Skild AI, where he builds robotic foundation models.