single source · post-training
Reasoning-action consistency reward
The Reasoning-action consistency reward is a signal introduced during post-training that measures behavioral alignment between a generated reasoning trace and a corresponding predicted ego trajectory.1 It uses rule-based matching between inferred intended behavior and motion primitives to ensure the reasoning translates into coherent, physically consistent actions.2
Trade-off The consistency reward is crucial for anchoring reasoning to physically realizable behaviors, ensuring that improvements in interpretability do not come at the expense of control fidelity.3
How this definition was written
Summarised by glm-5.3-flash from 1 source; every sentence is checked against the quote it cites.
we introduce a CoC–action consistency reward that measures behavioral alignment between the generated reasoning trace and the corresponding predicted ego trajectory
— Alpamayo-R1: Bridging Reasoning and Action Prediction for Generalizable Autonomous Driving in the Long TailWe then parse the generated reasoning trace to infer the ego’s intended behavior and compare it against the meta-actions derived from the predicted trajectory using rule-based matching
— Alpamayo-R1: Bridging Reasoning and Action Prediction for Generalizable Autonomous Driving in the Long TailThe consistency reward is therefore crucial for anchoring reasoning to physically realizable behaviors, ensuring that improvements in interpretability do not come at the expense of control fidelity
— Alpamayo-R1: Bridging Reasoning and Action Prediction for Generalizable Autonomous Driving in the Long Tail
Also called consistency reward, CoC–action consistency reward.
Filed under post-training › reward modelling