CoLT teaches multimodal models to think in latent steps and cuts inference 10.1x
jiqizhixin · x · 2026-07-23
Researchers from NTU and collaborators introduce CoLT, a method that makes multimodal models reason through a chain of latent thoughts instead of verbose text.
What it does
- Uses hidden representations as intermediate reasoning steps.
- Adds a lightweight decoder to supervise each latent step for stable training.
- Avoids the need for costly visual annotations used by some prior latent-reasoning approaches.
Reported gains
- Beats prior latent reasoning methods such as CODI and SIM-CoT.
- Cuts inference time by 10.1×.
- Reduces text decoding by 22.6×.
The post links the paper, code, and a report, positioning CoLT as a more efficient way for VLMs to think without generating long textual chains.
More from Multimodal
- Magnific teases October 28 release: 'We're making it move' — cuenca · 2026-10-27
- Midjourney --sref demo: retro Americana Batman art series — gizakdag · 2026-10-27
- AI video is a new medium, not cinema or animation — and it's perfect for depicting dreams — Kyrannio · 2026-10-27
- Dad turns his toddler's 31-year-old quote into a song via SUNO kids' choir — NathanpmYoung · 2026-10-27
- Band's local-AI music video hits 200k views: 900 generations kept 64, 3-second clips beat face drift — Hwang0500 · 2026-10-27
- Adobe Firefly video editor walkthrough: one-sentence Quick Cut turns 32s into 20s — LudovicCreator · 2026-10-26