LARC: Latent Reasoning in Continuous Space for Unified Multimodal Models

Forum topic · 小凯 · 2026-11-03

Summary

Researchers from a 2026 arXiv paper (2610.42149) introduce LARC, a Latent Reasoning framework in Continuous space for Unified Multimodal Models (UMMs). The authors argue that reasoning-augmented UMMs underperform because of a fundamental modality mismatch: chain-of-thought reasoning operates in a discrete token space while image generation operates in a continuous space. LARC bridges this gap by reasoning over continuous hidden states instead of discrete tokens. Its training pipeline has two stages: (i) curriculum-based supervised fine-tuning (SFT) that gradually converts token-level reasoning into latent reasoning, and (ii) information-gain reinforcement learning (RL) that self-evolves latent reasoning without ground-truth supervision. Experiments across text-to-image generation benchmarks show consistent improvements in generation quality, prompt following, and compositional fidelity. LARC achieves state-of-the-art performance on GenEval, exceeding BAGEL by 7 percentage points, and exhibits clear test-time scaling behavior, outperforming token-based generation under both parallel and sequential test-time scaling strategies.

Paper Overview

Field: NLP / Multimodal Models Authors: Byungwoo Jeon, Yoonwoo Jeong, Hyunseok Lee, Minsu Cho, Jinwoo Shin Published: 2026-10-30 arXiv: 2610.42149

Summary

Unified Multimodal Models (UMMs) achieve remarkable success on diverse multimodal understanding and generation tasks, yet they still struggle to effectively leverage language-based chain-of-thought reasoning for image generation. As a result, existing reasoning-augmented UMMs often fail to inherit the test-time scaling and fine-grained controllability observed in large language models.

The authors attribute this failure to a fundamental modality mismatch: reasoning is performed in a discrete token space, whereas image generation operates in a continuous space.

Method: LARC

To bridge this gap, the paper introduces LARC (Latent Reasoning in Continuous space), which alleviates the token-level bottleneck by reasoning over continuous hidden states. The training framework consists of two stages:

1. Curriculum-based Supervised Fine-Tuning (SFT): gradually transforms token-level reasoning into latent reasoning. 2. Information-Gain Reinforcement Learning (RL): self-evolves latent reasoning without requiring ground-truth label supervision.

Results

Consistent improvements in generation quality, prompt following, and compositional fidelity across multiple text-to-image generation benchmarks.
State-of-the-art performance on GenEval, outperforming BAGEL by 7 percentage points.
Clear test-time scaling behavior: LARC outperforms token-based generation under both parallel and sequential test-time scaling strategies.

LARC: Latent Reasoning in Continuous Space for Unified Multimodal Models

Summary

Paper Overview

Summary

Method: LARC

Results

Links

Tags