Lindsay Smith

I am a 5th-year physics PhD candidate and NSF GRFP fellow at Princeton University, advised by William Bialek and David Schwab. My research focuses on studying the science of AI, drawing on my background in physics and complex systems.

Current ML/AI research areas: AI for scientific discovery, mechanistic interpretability, in-context learning, LLM multi-agent interactions, AI safety.

I am wrapping up my internship at Polymathic AI (Mar-Aug 2026), where I am working on AI for scientific discovery (see our NeurIPS paper!).

Browse current work CV Email me

PAPERS/

* indicates equal contribution.

[10]Discover-Many-Worlds: Reinforcement Learning for Scientific Discovery in Automatically Generated Physics

Lindsay M. Smith*, Matt L. Wiemann*, Peter Melchior, Salman Rahman, Andrew Gordon Wilson, Pavel Izmailov, Carolina Cuesta-Lázaro (2026)

Under review.

[9]DiscoverPhysics: Benchmarking LLMs for Out-of-the-Box Scientific Thinking

Matt L. Wiemann*, Lindsay M. Smith*, Peter Melchior, Siddharth Mishra-Sharma, Andrew Gordon Wilson, Pavel Izmailov, Carolina Cuesta-Lázaro (2026)

NeurIPS 2026 (Evaluations & Datasets track)

Leaderboard/project page

[8]Chain-of-Thought Injection as an Inference-Time Safety Intervention

LMS*, Ananya Malik*, Edward James Young, Puria Radmard, Cameron Tice, Hannes Whittingham (2026)

ICLR 2026 Workshop on Logical Reasoning of LLMs

[7]A Computational Basis of Natural Intelligence

Alireza Karami, Veronica Chelu, Po-Chen Kuo, LMS*, Mia Whitefield, Jonathan D. Cohen (2026)

Proceedings of the Analytical Connectionism Schools 2023--2024, PMLR

[6]Large language models and the entropy of English

Colin Scheibner*, LMS*, William Bialek (2025)

Under review

[5]ALICE: An Interpretable Neural Architecture for Generalization in Substitution Ciphers

Jeff Shen* and LMS* (2025)

Preprint

Project page + demo

[4]When can in-context learning generalize out of task distribution?

Chase Goddard, LMS, Vudtiwat Ngampruetikorn*, David J. Schwab* (2025)

ICML 2025

[3]Model Recycling: Model component reuse to promote in-context learning

LMS, Chase Goddard, Vudtiwat Ngampruetikorn*, David J. Schwab* (2024)

NeurIPS 2024 SciForDL Workshop

[2]Specialization-generalization transition in exemplar-based in-context learning

Chase Goddard, LMS, Vudtiwat Ngampruetikorn*, David J. Schwab* (2024)

NeurIPS 2024 SciForDL Workshop

[1]Learning continuous chaotic attractors with a reservoir computer

LMS, Jason Z. Kim, Zhixin Lu, Dani S. Bassett (2022)

Chaos journal · Editor's Pick

Scilight press summary

RECENT_PRESENTATIONS/

Kempner Institute Workshop: Learning Dynamics in Natural & Artificial Intelligence (2026)

Talk: "Learning long-ranged structure in English: Code length and entropy dynamics during LLM training"

Workshop abstracts

American Physical Society (APS) Global Summit 2026

Talk: "Universality of LLM mechanisms across scale and diversity"

Conference abstract

Meta 2025 PhD Forum

Poster: "ALICE: An Interpretable Neural Architecture for Generalization in Substitution Ciphers"

View poster

APS Global Summit 2025

Topical Group on Statistical and Nonlinear Physics (GSNP) Student Speaker Award Finalist

Talk: "Multi-Agent Debate: Analyzing Consensus in Networks of LLM Agents"

Conference abstract

BACKGROUND/

Previously, I was an undergraduate at the University of Pennsylvania where I majored in physics (with honors), minored in mathematics and French and Francophone studies, and graduated cum laude. My undergrad research advisor was Dani Bassett, and I researched human white matter brain networks, the human perception of the stars in the night sky, and abstraction in reservoir computers (a type of RNN).

Before undergrad, I worked on a project at WSU designing a flux spectrometer for the Deep Underground Neutrino Experiment (DUNE) at Fermilab, advised by Holger Meyer.

CONTACT/

Email: lindsay.smith@princeton.edu