KAIST's LARC: Continuous Latent Reasoning Adds Test-Time Scaling to Unified Multimodal Models
kaist-ai · hf · 2026-11-02
LARC from KAIST/POSTECH addresses the modality mismatch in unified multimodal models: reasoning happens in discrete token space while image generation is continuous. It reasons over continuous hidden states via two training stages — curriculum SFT converting token-level reasoning into latent reasoning, then information-gain RL that self-evolves without ground truth. LARC achieves SOTA on GenEval with a 7pp gain over BAGEL and shows clear test-time scaling, beating token-based generation under both parallel and sequential strategies.
More from Multimodal
- Midjourney wallpaper recipe: sref + profile mix for eerie film-blur portraits — michaelrabone · 2026-11-02
- Cavallaero Teaser: A Painterly Rider Races Through a Golden Forest — CommunicationTop4244 · 2026-11-02
- Anima's snapref Offers a New Reference-Based Fix for Character Consistency — Intrepid_Swim_8669 · 2026-11-02
- Pencil-textured folding fan: surreal product concept render goes viral — Delahuntagram · 2026-11-02
- Free Midjourney SREF gallery adds 1,200+ handpicked style codes overnight — tisch_eins · 2026-11-02
- Adobe's EditAgent beats prior image-editing agents but needs ~10x the compute — emmanuelvivier · 2026-11-02