Steering Foundation Model Generation for Training: Thesis Proposal скачать в хорошем качестве

Steering Foundation Model Generation for Training: Thesis Proposal

Thesis proposal by Yash Savani (advised by Prof. Zico Kolter, CMU). Foundation models trained via self-supervision have internalized implicit knowledge far beyond their training objectives — diffusion models understand 3D structure, LLMs develop reasoning from next-token prediction. This thesis develops principled tools to extract this implicit knowledge for downstream tasks, organized around three guiding principles: steer the generation dynamics, respect the problem structure, and optimize for training not just quality. These principles are demonstrated across four works: DDRep (NeurIPS 2024) samples differentiable 3D representations via diffusion dynamics, Stepwise-Flow-GRPO improves RL for flow models through temporal credit assignment, PODS exploits compute asymmetry in RLVR for 1.7× faster training, and ADS steers LLM decoding to poison distillation while preserving teacher performance.