Not All Prompts Are Equal: Exploration-Guided Prompt Scaffolding for Multimodal Reinforcement Post-Training
The paper argues that prompt choice can waste RL training budget, then proposes scoring and rewriting prompts to make rollouts more useful.
Its Exploration Potential Score uses on-policy rollout statistics to estimate which prompts are likely to teach the current multimodal model. Low-utility prompts are not dropped; a teacher model rewrites them while preserving the task intent. The method is tested with GRPO on Geo3K and MMK12, with reported gains both in-domain and on MathVision and MMMU-Pro. HF Daily Papers' note
Its Exploration Potential Score uses on-policy rollout statistics to estimate which prompts are likely to teach the current multimodal model. Low-utility prompts are not dropped; a teacher model rewrites them while preserving the task intent. The method is tested with GRPO on Geo3K and MMK12, with reported gains both in-domain and on MathVision and MMMU-Pro. HF Daily Papers' note
score 4