Causal Inference with Video Features as Treatments
Kentaro Nakamura, Adam Breuer, Michael H. Crespin, Bryce J. Dietrich, Kosuke Imai
*Preprint PDF*

Summary: We develop the first statistical methodology for causal inference with video features as treatments. With our methodology, researchers can for the first time ask which visual features, appearing at which points in a video, influence audience responses.

Displacement Geometry Captures Platonic Shared Reality Across Models and Modalities
Chenming Shang, Yujin Tang, Jun Jie Ou Yang, Ruize Xu, Adam Breuer (co-PI), Nikhil Singh (co-PI)
Published in: NeurIPS’26 (forthcoming)
*Preprint PDF forthcoming*

Summary: The Platonic Representation Hypothesis (PRH) claims that independently trained models converge on a shared statistical model of reality, yet recent work finds only weak pointwise similarity between models. In this paper, we resolve this disagreement by showing that that what models share is not the location of samples in representation space, but the directions (displacement vectors) between them. Under a single orthogonal alignment limited to just rotation and reflection, these displacement vectors are substantially preserved across 44 independently trained vision and language encoders spanning modalities and asymmetric capability pairs. We trace this geometry to concept-level structure: within a model, parent concepts are orthogonal to their child variation vectors; across models, concept displacements are parallel. A major implication is that capabilities can be imported from one model to another using a single cached forward pass through the source.

FU-AUTHOR: Attribution-Based Compensation Is Incentive-Incompatible for LLMs
Matthew Keating, Rasmus Torp, Adam Breuer
Published in: NeurIPS’26 ATTRIB (forthcoming)
*Preprint PDF forthcoming*

Summary: We show that attribution-based compensation mechanisms proposed to pay authors for their work can perversely reward AI companies for paying them less. AI companies can retain value from authors' works while shifting attribution, and therefore compensation, toward copyright-free data of their choosing. In one illustrative attack, a model generates Star Wars-like text that TRAK attributes to Shakespeare. We show that AI companies can accomplish this without subterfuge, corruption, or alteration of authors' data. Our `attacks' on fair compensation require neither knowledge of the exact attribution mechanism nor queries to it. Nor do they require an AI company to break the proposed attribution rules. Instead, we show that proposed attribution mechanisms are incentive-incompatible, even under unrealistically idealistic assumptions.

Reducing information dependency does not cause training data privacy. Adversarially non-robust features do
Rasmus Torp, Shailen Smith, Adam Breuer
Published in: ICLR’26
*Download PDF*

Summary: Prevailing theory holds that state-of-the-art models leak sensitive private information, such as individuals’ faces or medical data, because models memorize training data and/or learn training data information dependencies. We begin by presenting a variety of surprising results that overturn this prevailing view. Specifically, we show that models can catastrophically leak private data even when training data is 98% censored (i.e. absent memorization), violating the assumptions and protections promised by recent information-theoretic privacy bounds. Instead, our OpenAI-funded research shows that privacy leakage is caused by so-called non-robust features, revealing a never-before-observed link between state-of-the-art privacy defenses and adversarial examples. We introduce a novel algorithm, Anti-Adversarial Training (AT-AT), which leverages this connection to obtain superior privacy at higher accuracy than the state of the art. Our results revise the prevailing understanding of training data exposure and reveal a new privacy–robustness tradeoff.

Using AI to Summarize US Presidential Campaign TV Advertisement Videos, 1952-2012
Adam Breuer, Bryce J. Dietrich, Michael H. Crespin, Matthew Butler, J.A. Pyrse, Kosuke Imai
Published in: Nature: Scientific Data (2025)
*Download PDF*

Summary: This paper introduces the largest and most comprehensive dataset of US presidential campaign television advertisements, available in digital format. To facilitate a variety of research, we design a large-scale parallelized, AI-based analysis pipeline that automates the laborious process of preparing, transcribing, and summarizing videos, and we conduct extensive human evaluations to show that these transcripts and summaries match the quality of manually generated alternatives. We illustrate the value of this data by including an application that tracks the genesis and evolution of current focal issue areas over seven decades of presidential elections. Our analysis pipeline and codebase also show how to use LLM-based tools to obtain high-quality summaries for other video datasets.

Toward Interpretable LDA Topic Models with Strong Guarantees in Logarithmic Parallel Time
Adam Breuer
Published in: ICML'25
*Download PDF *

Summary: This paper presents the first combinatorial algorithms that provably solve LDA topic models with near-optimal posterior probability in logarithmic parallel computation time—exponentially faster than any known LDA algorithm. Our approach resolves several longstanding open problems in topic modeling by providing formal interpretability guarantees and the first provable guarantees for downstream causal inference applications, which are the primary use case for topic models in the social sciences. Empirically, our algorithms consistently produce topic models of higher semantic quality than state-of-the-art LDA algorithms, neural topic models, and LLM-based topic models.

Identifying Peer Effects on Participation in Collective Action: Evidence from a set of natural experiments in cell phone movement & communication metadata
Adam Breuer
(Awarded NSF SES Grant)

Algorithmic discovery of efficient envy-free division protocols for games of complete and incomplete information
Adam Breuer