IC-941CLIP reward model quality scales with model size, with a sharp phase transition between ViT-H/14 and ViT-BigG/14 for the humanoid kneeling task

Juan Rocamonde, Victoriano Montesinos, Elvis Nava, Ethan Perez, David Lindner

SourceVision-Language Models are Zero-Shot Reward Models for Reinforcement Learning

The authors evaluate four CLIP model sizes (RN50, ViT-L/14, ViT-H/14, ViT-BigG/14) as reward models on the humanoid kneeling task. The EPIC distance to human labels improves smoothly and approximately log-linearly with model size. However, the human-evaluated success rate of the trained agent shows a sharp phase transition: RN50, ViT-L/14, and ViT-H/14 all achieve 0% success, while ViT-BigG/14 achieves 100%. Goal-baseline regularization improves the reward model across all sizes but is more impactful for smaller models. The authors note that predicting the exact point where the RL agent can learn the task is difficult, a common pattern in evaluating large foundation models.

Evidence
correlational
Key metric
success rate: RN50 0%, ViT-L/14 0%, ViT-H/14 0%, ViT-BigG/14 100% (over 2 seeds); EPIC distance close to log-linear in log10 of number of parameters
Caveat
Evaluated on a single task (kneeling) and a single environment (humanoid); the phase transition point may differ for other tasks. Only 2 seeds used for the success rate comparison.
Model
CLIP / CLIP-ViT (LC)
Concepts
Scale-dependent behaviour
Datasets
MuJoCo Humanoid-v4 [eval]
Methods
EPIC distance [eval]
Related findings
IC-939, IC-940
Extraction
automatic-extraction