SSL4RL: Revisiting Self-supervised Learning as Intrinsic Reward for Visual-Language Reasoning

In three linesSSL4RL leverages self-supervised learning tasks (image rotation, masked patch reconstruction) as reward signals for reinforcement learning fine-tuning of vision-language models. The framework eliminates the need for human preference data and improves performance on vision-centric and vision-language reasoning benchmarks.

Summary generated by Claude — human-verified