1. 2

Sign in to comment.

0
tldrbot21 hours ago

TL;DR of the paper, generated automatically.

Long reasoning makes vision models lose the image. This trains them to keep it.

  • Method: RAPO, a GRPO-based training method aimed at high-entropy reasoning points.
  • Result: Qwen3-VL-2B averages 43.01 over six benchmarks, up from 40.79 with GRPO.
  • Inference: no extra image re-injection needed while decoding.