Latent Reasoning in Continuous Space for Unified Multimodal Models

Latent Reasoning in Continuous Space for Unified Multimodal Models. By Byungwoo Jeon and four others, from KAIST, POSTECH, and one other institution. Published on the thirtieth of October.

Unified multimodal models, which handle both understanding and generation, have done well across a wide range of tasks. Yet they still struggle to make effective use of language-based chain-of-thought reasoning when generating images. As a result, existing reasoning-augmented models of this kind often fail to inherit the test-time scaling and fine-grained controllability seen in large language models. The authors trace this to a modality mismatch. Reasoning happens in a discrete token space. Image generation happens in a continuous one. Chain-of-thought written out as words has to pass through that discrete bottleneck before it can inform an image, and on this account that is where the benefits seen in language models get lost.

LARC, a Latent Reasoning framework in Continuous space, moves the reasoning into the kind of space generation already works in. It reasons over continuous hidden states rather than tokens, easing the token-level bottleneck. The reasoning no longer has to be spelled out as text at each step. It stays in the same kind of representation the image is produced from.

Training comes in two stages. The first is curriculum supervised fine-tuning, which gradually converts token-level reasoning into latent reasoning, step by step rather than all at once. The second is information-gain reinforcement learning, in which the latent reasoning self-evolves without ground truth supervision. The curriculum moves the model off tokens, and the reinforcement stage carries the latent reasoning further without needing labelled answers.

Across diverse text-to-image generation benchmarks, LARC consistently improves generation quality, prompt following, and compositional fidelity. On GenEval it reaches state-of-the-art performance, a gain of seven percentage points over BAGEL. It also shows clear test-time scaling behavior, outperforming token-based generation under both parallel and sequential test-time scaling strategies. That last result speaks directly to the opening complaint, since test-time scaling was one of the properties reasoning-augmented models had failed to inherit.

Diagnosis and remedy line up closely. If the obstacle was reasoning in tokens for a task that lives in a continuous space, the remedy is to reason in that space, and the return of test-time scaling is the evidence offered that the remedy took.

Latent Reasoning in Continuous Space for Unified Multimodal Models. By Byungwoo Jeon and four others, from KAIST, POSTECH, and one other institution. Published on the thirtieth of October. This is the long version: the whole paper rather than the abstract.

A unified multimodal model is a language model with an image generator built in. Language models get better when you let them think longer. These models, the paper observes, do not. Give a reasoning-augmented unified model a bigger thinking budget before it draws, and the smooth gains seen on language benchmarks fail to appear. The authors call the cause a modality mismatch. The reasoning happens in discrete tokens, words from a fixed vocabulary, while the image is generated in a continuous space. Words struggle to carry the dense, spatially grounded information a picture needs, so reasoning collapses into rewording the prompt. Their claim is that reasoning itself has to move into continuous space.

Two ideas are needed first. The base model here is BAGEL, which has an understanding expert and a generation expert, and its image module is a flow-matching head: it builds the image in continuous space and can score how well it reconstructs a target image, through a loss.

The second idea is Coconut, earlier work on language models. Instead of turning the model's last hidden state into a word, you feed that hidden state straight back in as the next input. The thought never passes through the vocabulary. LARC carries that trick into image generation. The obstacle is training it with reinforcement learning, which the authors name in two parts. Coconut-style reasoning is deterministic, while policy-gradient methods need varied rollouts to compare. And a latent thought has no evaluator: nobody can read a hidden state and say whether it was a good thought.

The trajectory alternates between two modes. In language mode the model samples ordinary text tokens, and that sampling is where the randomness for reinforcement learning comes from. When it emits a special latent token, it switches to latent mode: for a fixed number of steps, the last hidden state is fed back as the next input. This repeats until an end token. The image module then reads the whole sequence of hidden states through cross-attention and draws.

Stage one is supervised fine-tuning on one million interleaved chain-of-thought samples, with a latent marker inserted at every reasoning step boundary. Turning on all the latent steps at once is unstable, so a curriculum does it gradually. First the model trains on plain text reasoning. Then only the first latent step is switched on, then the first two, and so on, each new one building on a stable prefix, with the learning rate lowered at each transition. The latent steps get no target of their own.

Stage two is reinforcement learning with GRPO, and its reward is the key idea. Freeze a copy of the stage one model. Ask it how likely the target image is given the prompt alone, and how likely given the prompt plus the reasoning trace. The reward is the difference: how much the thought helped. Its expectation is a conditional mutual information, which cannot go below zero. Since the image module gives no direct likelihood, the authors use the gap between two flow-matching losses, computed with the same noise sample so that a shared constant cancels. A second reward term pushes the latent steps apart from each other, so they don't collapse into one repeated vector. Only the understanding expert is updated. The generation expert stays fixed, because it is the judge.

On GenEval, LARC scores point eight eight overall, against point eight one for BAGEL: the seven point gain from the abstract. The largest jump is attribute binding, from point six three to point seven eight. On OneIG-Bench it gains ten point two points over BAGEL. On MIRA, a visual reasoning benchmark, it rises from seven point five to eleven point seven.

The comparison that tests the thesis is the ablation on reasoning space. Both arms start from the same base model, after supervised training only. Ordinary token-level chain of thought barely moves image generation, and on MIRA it drops below the untouched baseline, to seven point one. Latent reasoning reaches nine point seven. The training-stage ablation then shows that most of the headline gain arrives with reinforcement learning, which lifts GenEval from point eight three to point eight eight.

On cost, the appendix measures wall-clock time on one A100. Best-of-N on BAGEL grows almost linearly, to four hundred and sixty-nine seconds at ten samples. LARC with ten latent steps takes forty-eight seconds, only one point two seconds more than with one. Without the diversity reward, latent steps reach a cosine similarity of point nine seven, nearly identical, and every benchmark falls.

A short limitations section sits at the end of the appendix and names two limits. The framework assumes an image module that generates in continuous space, a flow-matching head; models that generate discrete image tokens are left for future work. And the latent steps add inference latency over the base model, though the paper's own timing puts that at about a second.

What the evidence does not show is also visible. Every result is a single number: greedy decoding, one image per prompt, no error bars or repeated runs. Several ablation gaps are two hundredths of a point. The test-time scaling claim rests on a figure whose underlying numbers are not tabulated in the text. And there is one base model, BAGEL.

The idea to carry away is where the thinking lives. If an image is made in continuous space, a thought written in words has to be squeezed through a vocabulary and then expanded again. LARC lets the model think in the same kind of space it draws in, and then trains those thoughts on a single question it can answer about itself: did this thought make the right image more likely? That self-scored signal, with no critic and no labelled reasoning, is what lets the thinking improve, and what lets more of it help.