Research

Current Projects

Provocations for Human-AI Interaction: A Community Forum Case Study
with S. Verma, T. Pénigaud, A. Siu, and M. Revel
We are exploring user preferences and concerns based on results of the Meta Community Forum 2023 hosted by the Stanford Deliberative Democracy Lab.

The Power of Vague Conspiracy Theories: Why Expansive Voter Fraud Theories Spread, Are Difficult to Refute, and Create Cycles of Distrust
with Justin Grimmer, J. Mernyk, K. Rhee, and Brandon M. Stewart
We use quantitative text analysis of online voter fraud discourse to study why vague, expansive conspiracy theories spread and why they are so difficult to refute.

Work in progress on:

• Learning policies that decide how to steer language model generation at inference time.
• Evaluating how robust language models’ answers are when challenged by a devil’s advocate over multiple turns of debate.
• Revising text by searching over structured edit operations with tree search.

Published Work

Zachary Bamberger, Till Raphael Saenger, G. Morad, Ofra Amir, Brandon M. Stewart, and Amir Feder. 2026. STATe-of-Thoughts: Structured Action Templates for Tree-of-Thoughts. Conference on Language Modeling (COLM 2026); also presented at the COLM 2026 Workshop on Agent Behavior.
→ An interpretable inference-time compute method that searches over high-level reasoning patterns: a controller selects actions encoding reasoning choices, a generator produces steps conditioned on those choices, and an evaluator scores candidates to guide search.
→ Software release here

Till Raphael Saenger, Musashi Hinck, Justin Grimmer, and Brandon M. Stewart. 2024. AutoPersuade: A Framework for Evaluating and Explaining Persuasive Arguments. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 16325–16342, Miami, Florida, USA. Association for Computational Linguistics.
→ Associated Python package: suntopicmodel

Till Raphael Saenger, Ethan Kapstein, and Ronnie Sircar. 2024. Estimating the Collapse of Afghanistan’s Economy Using Nightlights Data. In PLOS ONE, 19(12): e0315337.

Dissertation

Till Raphael Saenger. 2026. From Discovery to Control: Causal Inference and Controllable Generation for Understanding Persuasion. Doctoral dissertation, Princeton University.

Previous Work

Till Raphael Saenger. Exploring the Cyclical Predictability of Sector-Specific Premia. 2020.

Till Raphael Saenger. On the Limitations of Cryptocurrencies. 2019.