Dynamics Model

Directly optimizing control policies.

Optimizing continuous action sequences.

Modifying reward to accelerate learning.

Reward only given upon task completion.

A high-priority instruction layer setting overarching behavior constraints for a chat model.

Rules and controls around generation (filters, validators, structured outputs) to reduce unsafe or invalid behavior.

RL without explicit dynamics model.

RL using learned or known environment models.