Scaffolding Minds: Optimizing Latent Visual Target Representations for Multimodal Reasoning
This paper proposes Scaffolding Minds, a dedicated scaffolding encoder that provides an optimized target in latent space, and learns both the mean and variance of the RL sampler, and shows that these two improvements are complementary, together yielding substantial gains over strong baselines.
Haoqiang Kang, Yin-Ping Chen, Lu-Yang Liu et al.
· 2 citations