LARC: Latent Reasoning in Continuous Space for Unified Multimodal Models

Byungwoo Jeon¹, Yoonwoo Jeong², Hyunseok Lee¹, Minsu Cho^2,3†, Jinwoo Shin^1,3†

¹KAIST²POSTECH³RLWRLD

^†Equal advising

NeurIPS 2026

TL;DR

Unified multimodal models reason in discrete tokens but draw images in a continuous space, so spending more on reasoning does not buy better images. LARC (Latent Reasoning in Continuous space) moves the reasoning into that continuous space: the model interleaves short text spans with latent steps built from its own hidden states. Curriculum fine-tuning turns token-level reasoning into latent reasoning, and self-evolving reinforcement learning then rewards each trace by how much it helps predict the target image, with no ground-truth chains or external critic. LARC improves text-to-image generation, image editing and visual reasoning over its base model and, to our knowledge, is the first reasoning-augmented unified model with clear test-time scaling.

Instruction-guided editing with test-time scaling (the example from Figure 4 of the paper). With more latent reasoning, the edit follows more of the instruction.

Method

Reasoning-augmented unified models think in discrete tokens while their image module works in a continuous space, so longer reasoning chains mostly rephrase the prompt. LARC lets the backbone reason over its own hidden states instead. After a <latent> token, the last hidden state is fed back as the next input embedding, so the trajectory alternates short text spans with continuous latent steps, and the image module reads the whole trajectory through cross-attention.

Two-stage training of LARC. Stage 1, supervised fine-tuning: the understanding expert learns to feed its last hidden states back as inputs, producing latent reasoning steps, while the generation expert produces the image. Stage 2, self-evolving reinforcement learning: the policy model rolls out reasoning trajectories for a context such as the prompt An Avocado Chair; a frozen reward model, the Stage 1 model, compares the reasoning-conditioned and reasoning-free distributions of the target image to compute an information-gain reward that drives the policy update. — **Stage 1 (curriculum SFT)** gradually introduces latent steps, in which the model's last hidden state becomes its next input embedding. **Stage 2 (self-evolving RL)** optimizes the policy with GRPO. The reward is the information gain a reasoning trace gives the frozen Stage 1 model (π_ref) about the target image, estimated from flow-matching losses under shared noise, plus a diversity term that prevents latent collapse.

Results

LARC is built on BAGEL and trained on 1M interleaved chain-of-thought samples from FLUX-Reason, WEAVE and ScaleEdit. It scores above both its base model BAGEL and the token-reasoning Uni-CoT on every metric below, across text-to-image generation, image editing and multimodal understanding.

The radar charts require JavaScript.

Test-time scaling

Under budget forcing, adding latent reasoning steps at inference improves image generation, image editing and visual reasoning, and LARC pulls ahead of best-of-N sampling as the budget grows. Prior reasoning-augmented unified models do not show this behavior.

Three line plots of score against the number of latent steps or sampled images, from 1 to 10, for image generation, image editing (CLIP score) and visual reasoning (matching score). In all three, latent reasoning scaling (red) rises faster than best-of-N sampling (blue).

Scaling cost

Best-of-N reruns the full generator for every extra image, so its cost grows linearly with N. Latent steps only extend the understanding expert's forward pass: going from 1 to 10 latent steps adds 1.2 seconds per sample.

The line chart requires JavaScript.

Seconds per sample on one NVIDIA A100 with batch size 1 (paper Table 11). Hover the chart to read both methods at a budget.

Qualitative test-time scaling

BAGEL with best-of-N sampling (left column) versus LARC (right column). Rows use N = L = 1, 4 and 10.

Text-to-image comparison for a prompt describing a tiger in a jungle that ends with looking front. BAGEL's three images show the tiger side-on; LARC's tiger turns toward the viewer as the budget grows, marked correct. — **Text-to-image.** The prompt ends with “looking front.” As the budget grows, LARC turns the tiger toward the viewer, while BAGEL keeps it side-on.

Image editing comparison on a black-and-white portrait with the instruction to add sunglasses, a mild smile and a cup of coffee. BAGEL adds a cup but no sunglasses; LARC adds the sunglasses from the second budget on, marked correct. — **Image editing.** “Make him wear sunglasses with mild smile and holding a cup of coffee.” BAGEL never adds the sunglasses.