Overview
Application of sequential decision making to automated experimentation and related domains, where actions consume material irreversibly, instruments drift, and outcomes resolve after subsequent decisions must already be made. Connected to work in the AI Research organization at Lila Sciences.
Themes
- Embodied scientific reasoning — extending text-based scientific reasoning benchmarks to settings requiring the full experimental loop: hypothesis formation, execution, and interpretation under partial observability
- Decisions with unresolved outcomes — selecting the next experiment while prior experiments remain pending, and learning from incomplete outcome records
- Synthetic data for underrepresented situations — generating training data from simulators for rare configurations, with explicit measurement of transfer to real data
Relevant prior work
Robust Autonomy Emerges from Self-Play · ICML 2025
Cusumano-Towner et al., incl. Killian. Self-play in a massively parallel simulator with randomization over agent physical and behavioral characteristics, yielding robust driving policies. Relevant to what simulated experience can substitute for.
Efficient Agentic Reasoning Through Self-Regulated Simulative Planning · 2026
Deng et al., incl. Killian. SR²AM decomposes agent decision making into simulative reasoning via world model prediction, self-regulation over when to plan, and reactive execution, matching larger models with 25.8–95.3% fewer reasoning tokens.
Methods for this area draw substantially on the fundamental RL and foundation model post-training directions.