Jun 2026
Why Struggle with Continuous Latents? Interpretable Discrete Latent Reasoning via Rendered Compression
Inspired by render-based compression, this work renders textual chains of thought into images, extract visual features, and construct a discrete latent vocabulary via clustering-based fine-tuning, and concludes that discrete latent tokens provide a controllable and interpretable basis for efficient latent reasoning.
Shuochen Chang, Qingyang Liu, Shaobo Wang et al.
· arXiv.org · 0 citations