Selected Talks

• Tractable Language Model Programming: Themes and Prospects — Probabilistic programming techniques for LLMs
• Analytica: Soft Propositional Reasoning for Robust and Scalable LLM-Driven Analysis — Probabilistic models for LLM research agents
• Language Modeling by Language Models — Autonomous agents for scientific discovery
• Understanding the Logic of Generative AI through Logic — Semantic models for post-training
• Language Model Programming: Themes and Prospects — Programmatic models for LLM decomposition
• Pushing the Limits of Rule Reasoning in Transformers — Formal methods for LLM evaluation

News Archive

2026

• NeurIPS: Two papers accepted: Operadic Consistency: A Label-Free Signal for Compositional Reasoning Failures in LLMs and CoTs as Probabilistic Programs: A Programmatic View of Thinking Step-by-Step in Language Models.

• EMNLP: From Token Probabilities to Semantic Constraints: Towards Declarative Probabilistic Evaluation of Language Models was accepted.

• TPM @ UAI: CoTs as Tractable Probabilistic Programs was accepted. I also gave a keynote, Tractable Language Model Programming: Themes and Prospects (video).

• COLM: Artifact Linker, a benchmark and environment for LLM-driven automated scientific discovery, was accepted.

• We released a new preprint on modeling question decomposition with operads, along with a shorter technical report presented at the Combining Theory and Benchmarks workshop at ICML.

• We released Probabilistic Programs of Thought, joint work with the UCLA StarAI Lab.

• ICLR: Two papers accepted: AstaBench: Rigorous Benchmarking of AI Agents with a Holistic Scientific Research Suite and Analytica: Soft Propositional Reasoning for Robust and Scalable LLM-Driven Analysis.

2025

• I gave an invited keynote at the International Workshop on Symbolic-Neural Learning in Osaka (slides).

• NeurIPS: Language Modeling by Language Models, our work on research agents for autonomous machine learning, was accepted as a spotlight paper.

• EMNLP: TinyScientist: An Interactive, Extensible, and Controllable Framework for Building Research Agents was accepted.

• We released the AstaBench leaderboard and accompanying technical paper for evaluating LLM agents across scientific tasks.

• I taught an updated version of our Language Model Programming course at ESSLLI, including new lectures on probabilistic programming for prompting and loss-function decompilation.

• I presented Language Modeling by Language Models, a talk on automated scientific discovery, at the NAACL Workshop on AI for Scientific Discovery (preprint).

• ICML: Two papers accepted: Understanding the Logic of Direct Preference Alignment through Logic (brief overview) and ZebraLogic: On the Scaling Limits of LLMs for Logical Reasoning.

For earlier work and a complete list, see my Google Scholar profile.

2026

2025

2024

2023

2022

2021