Rewriting training prompts lifts multimodal scores by up to 11.5%

single source· 1 articles · confidence: medium · first seen 2026-09-13 20:00 UTC

What this means for you

Nothing to act on yet: this is an abstract with no code, no weights and no evaluation date, and the gains are the authors' own. If you run online RL post-training on multimodal data, the reusable part is the scoring step — flagging saturated and near-impossible prompts from statistics you already collect, before spending rollouts on them.

A preprint posted to arXiv on 13 September argues that training prompts differ in how much a model can learn from them, and that the unhelpful ones should be rewritten rather than dropped. It scores each prompt from rollout statistics during reinforcement post-training — the stage where a model improves by generating answers and having them graded. A teacher model then rewrites the low-scoring prompts while keeping the task. Combined with GRPO on Geo3K and MMK12, the authors report up to 9.7% relative in-domain improvement, 11.5% on MathVision and 11.1% on MMMU-Pro.

Key facts

  • ·The paper is arXiv 2609.15051, posted 13 September 2026. source
  • ·Prompts are scored with an "Exploration Potential Score", derived from KL-regularised policy improvement theory and computed from on-policy rollout statistics with no additional overhead, according to the authors. source
  • ·Low-utility prompts are rewritten by a teacher model instead of being discarded, preserving the original task intent. source
  • ·The method is integrated with GRPO and evaluated on Geo3K and MMK12. source
  • ·Reported gains are up to 9.7% relative in-domain, 11.5% on MathVision and 11.1% on MMMU-Pro. source
  • ·The abstract gives no evaluation dates, no harness details and no indication that code or weights have been released. source

What the sources say

Sources

The original reporting. Follow these — they did the work.

← the wire