Hi, this is Henry. If you're interested in similar problems, I'd love to chat
Simulating the Effects of Fields on Simulated Agents
By reinterpreting our DisCo-BN joint as a baseline energy surface, we can model influences external to agents as fields, and derive an update operator for agents, allowing populations to change dynamically in response to modeled stimuli
Distilled-Conditional Bayesian Networks
LLMs supply conditional structure but make poor coherent generators. We distill their conditionals, node by node, into an offline Bayesian network that generates arbitrarily large coherent synthetic populations after a one-time distillation cost.
Use the LLM to Build Your Simulator, Not to Be Your Simulator
LLM-based population simulation usually means sampling personas at inference time. We distill the LLM's response surface into a small log-linear model once, calibrate its absolute rates against observed toplines, and answer every later subgroup question offline.
The Mechanism Boundary in Automated Research
Automated research systems need more than better agent orchestration. Here, we argue that their reliability depends on where the research process can be made mechanically checkable, and what machinery can be built into the systems to take advantage.
The Architecture of Mechanized Research
SMAI separates agent reasoning from scientific verdicts by turning experiment designs into mechanically checkable commitments before any results are observed
SMAI: Scientific Method as Infrastructure
Auto-research agents can explore solution spaces faster than humans. But purely agent-driven approaches lose the structural rigor that pre-defined environments provide: methodology drift, post-hoc selection, inferential breakdowns. Structural verification provides the scaffolding that lets scaled systems be trustworthy.