Studying how LLMs reason, decide, and take risks
I run behavioral experiments on language models using game theory and causal inference to figure out how prompts shape the way AI makes decisions. My work connects economics, alignment, and empirical AI research.
LLM behavioral alignment through the lens of economics
Deterministic embodiment scoring environment on Qwen-2.5 14B. 283 prompt conditions, 31,638 logged responses. Reasoning mode predicts cooperation rate at r = -0.54: top 10% embodied defect 3.7%, bottom 10% defect 81.3%. Score runs deterministically per response, usable as a runtime monitor.
15,000 GPT-4o-mini observations per treatment in parallel to a 750-subject human dataset (Charness-Jackson Stag-Hare). Both populations cooperate 61.1% at baseline. A responsibility-framing intervention drives humans down 17 points and the model up 15 points: 32 points apart, opposite directions, identical surface starting behavior.
When LLMs play behavioral games, their output changes based on the persona they take on. Mapping which demographic dimensions shift AI strategic choices.
Quasi-experimental analysis of zero bail policy reform using RDD and DiD to measure real-world outcomes of criminal justice interventions.
Tools and systems I've built
Photo-to-listing tool for resellers. Snap a photo, AI identifies the item, pulls real sold prices from eBay, and generates optimized listings. Auto-posts to eBay and Facebook Marketplace.
Annotation software for YOLO training using SAM. Bootstraps object detection with CLIP transformer embeddings.
Agent-based simulation of free market dynamics. Models competitive equilibrium, price discovery, and emergent behavior.
150M police stops at 280 req/s. Python + Rust. Saved $200k/yr in API costs for criminal justice research.
Web interface for managing behavioral experiments. Annotation, survey, and data collection pipeline for MTurk and students.
View →Conversational game built on LLMs. Interactive dialogue exploring narrative generation and player-AI interaction.
View →Autonomous equipment rental on distributed ledgers. Raspberry Pi + IOTA. Won IOT2Tangle Hackathon.
View →Multi-agent bot swarm with C&C server. Recorded human mouse movements for realism. Docker isolation, custom scripting DSL.
Multiplayer educational capture-the-flag teaching web crawling. Matter.js WebGL, Flask backend.
View →