Uday Tyagi

We're measuring whether a model's stated values survive contact with itself: hand an LLM its own constitution, let it rewrite it, repeat, and track semantic drift, value collapse, and self-reinforcing edits across rounds.

I built the experimental harness with two collaborators: multi-round self-editing pipelines across model families, embedding- and judge-based drift metrics, and behavioral evals that check whether an edited constitution actually changes what the model does downstream, or only what it says.

I trained a diffusion-based motion planner to replace ORCA for real-time multi-drone collision avoidance: a DiT generating velocity command sequences from ego and neighbor states, with DPM-Solver++ inference optimization and low-temperature reranking. It reaches ~80% mean scenario completion against ORCA's 59%.

Around it I built the deconfliction stack: a ROS 2 planner plugin, scenario-based evaluation pipelines with behavioral analytics, and a ~200k-trajectory training dataset spanning six collision scenarios.

I did the engineering and implementation work on the multi-agent RAG system for automated ESG carbon reporting: LangChain orchestration, RL for factual grounding.

I also designed the agent observability layer: audit-trail logging and behavioral guardrails that catch hallucinated citations before they land in a regulatory filing.

Evaluated hierarchical and parallel AI control protocols on the ControlArena benchmark, measuring how the safety-usefulness tradeoff moves as the capability gap between trusted and untrusted models widens.

Extended ControlArena with red-team/blue-team protocols and Elo-based scoring to find the point where oversight mechanisms fail under adversarial prompting.