research Self-distillation for LLM reasoning IBM Research · research intern · Jun–Aug 2026. Failure modes of self-distillation pipelines and where privileged-teacher signal lives in a rollout. RL for generative models MIT-IBM Watson AI Lab · May 2025–May 2026. RL fine-tuning of flow models for molecules and crystals, and the diversity analysis behind Tailor (ACL 2026). Sketch-to-3D retrieval for generative car design MIT DeCoDE Lab · Jan 2025. A 2D sketch → nearest 3D car mesh retrieval pipeline inside a multi-agent engineering-design framework (ASME IDETC 2025). Attestation for negotiating AI agents MIT Media Lab · Sep 2024–May 2025. zkSNARK proofs that an agent negotiation inside a GPU-backed TEE ran as specified, plus tool-using agents and an evaluation harness. Influence prediction in collaboration networks MIT PRIMES-USA · 2023–2024. Predicting future vital nodes in weighted collaboration networks by combining link prediction with influence-maximization heuristics (Physica A 2026, arXiv 2024). coursework Hash grid collisions in Instant-NGP 6.8300 Advances in Computer Vision · Spring 2026. Where Instant-NGP's hash collisions land, a two-choice hashing scheme that spreads them out, and a router that tests whether the MLP cares. TTARM — training-time adaptive rank modulation 6.7960 Deep Learning · Fall 2025. Tracking how the effective rank of transformer activations evolves during training, and scheduling regularizers to steer it. RL distillation of a diffusion motion planner 6.7920 Reinforcement Learning · Fall 2025. Distilling a text-to-motion diffusion planner into a lightweight autoregressive generator–selector trained with RL, and a careful post-mortem of why it's hard.