CS + AI at Dartmouth. Previously NYU.
I study how misalignment emerges and spreads between LLM agents — and whether we can catch it in the activations.
Most of what we're deploying now is multi-agent: models debating, negotiating, and coordinating with each other. I'm interested in the failures that only appear in that setting — contagion, collusion, emergent convention — and in whether they leave a mechanistic signature you can probe for.
- Moral agentic alignment via game theory — with Tennant (Google DeepMind).
- Topology in multi-agent systems — a systematic study of how network structure governs the propagation of misalignment.
- Misalignment-Contagion — Can a misaligned minority shift the private beliefs of an aligned majority through debate? Minority influence, belief internalization, network topology.
- mach-games — A common-pool resource game that makes deception instrumentally attractive, then taxonomizes what the agents actually do.
- emergent-conventions — Do agents form path-dependent symbol↔referent conventions with no pretraining anchors?
- moral-personas-ipd — Moral personas in iterated prisoner's dilemma. Trajectories on HuggingFace.
- Emergent-Pathologies — Do behavioral pathologies emerge from BFI-conditioned personas in repeated social dilemmas?
- log-prob-adversarial-debate — A log-probability-guided attack on multi-agent debate that lands on frontier models.
- Atlas — Compositional safety evaluation: does agent safety degrade when safety dimensions are combined rather than tested in isolation?
- Compass — The scoring side of that question: does a safety score stay meaningful once trajectories combine several safety dimensions at once?
- cda-mech — Probing residual streams for collusive intent in continuous double auctions.
- fv_composition — Does composing function vectors compose their behaviors? Eight-stage causal pipeline with CIE head ranking.
- PolicyLLM — Neuro-symbolic extraction, validation, and runtime enforcement of organizational policy on LLM output.
- simplecoder-agent — ReAct-style CLI coding agent with deterministic editing and semantic search.
- VidExplainAgent — Making STEM video accessible to blind and low-vision learners through automated visual description and interactive Q&A.
- Autonomous-Steering — Transformer-based steering-angle prediction on DAVE-2, following MGTR and Motion Transformer.
Most repos are runnable end-to-end — configs, seeds, and analysis scripts included. If something doesn't reproduce, open an issue; I'd rather know.