Representation Forcing for Bottleneck-Free Unified Multimodal Models

Hugging Face Daily Papers Papers

Summary

Introduces Representation Forcing (RF), a technique that enables unified multimodal models to perform both perception and generation end-to-end without external VAE latent spaces, matching state-of-the-art VAE-based models in image generation while improving understanding.

Unified multimodal models (UMMs) aim to handle perception and generation in a single model. Yet existing UMMs still rely on a frozen, separately pretrained VAE for image generation, imposing a structural bottleneck. Naively removing it introduces a quality gap, as the model must learn both high-level structure and low-level details from raw pixels. In this paper, we propose Representation Forcing (RF), a technique that closes this gap by making representation prediction a native capability of the model. Concretely, RF forces the decoder to autoregressively predict visual representations as intermediate tokens before pixels; these tokens then stay in context to guide pixel diffusion within the same backbone. By turning representations from perception outputs into generation targets, RF eliminates the need for any external generative latent space. We find that RF benefits both understanding and generation. On image generation, our pixel-space model with RF matches state-of-the-art VAE-based unified models. On image understanding, pixel-space RF generally outperforms its VAE-based variant. Together, these results offer an effective step toward end-to-end, bottleneck-free UMMs.
Original Article
View Cached Full Text

Cached at: 06/01/26, 03:18 AM

Paper page - Representation Forcing for Bottleneck-Free Unified Multimodal Models

Source: https://huggingface.co/papers/2605.31604 Published on May 29

#3 Paper of the day Authors:

,

,

,

,

,

,

,

,

,

,

,

Abstract

Representation Forcing enables unified multimodal models to perform both perception and generation tasks end-to-end without relying on external latent spaces, matching state-of-the-art performance in image generation while improving understanding capabilities.

Unified multimodal models(UMMs) aim to handle perception and generation in a single model. Yet existing UMMs still rely on a frozen, separately pretrained VAE for image generation, imposing a structural bottleneck. Naively removing it introduces a quality gap, as the model must learn both high-level structure and low-level details from raw pixels. In this paper, we proposeRepresentation Forcing(RF), a technique that closes this gap by making representation prediction a native capability of the model. Concretely, RF forces the decoder to autoregressively predictvisual representationsas intermediate tokens before pixels; these tokens then stay in context to guidepixel diffusionwithin the same backbone. By turning representations from perception outputs into generation targets, RF eliminates the need for any external generative latent space. We find that RF benefits both understanding and generation. On image generation, our pixel-space model with RF matches state-of-the-art VAE-based unified models. On image understanding, pixel-space RF generally outperforms its VAE-based variant. Together, these results offer an effective step toward end-to-end, bottleneck-free UMMs.

View arXiv pageView PDFProject pageAdd to collection

Get this paper in your agent:

hf papers read 2605\.31604

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2605.31604 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2605.31604 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2605.31604 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

How to setup a local coding agent on macOS

Hacker News Top

A detailed tutorial on setting up a local coding agent on macOS using Gemma 4 with MTP draft model and llama.cpp, achieving ~24% speed improvement through speculative decoding.

@GitHub_Daily: To dive deep into model research, you can't just stay at the application layer—you need to understand how the underlying system is trained and optimized. I stumbled upon LLMSys-PaperList, a carefully curated collection of papers related to large model systems. It is continuously updated from 2022 to the latest top conference papers in 2026, and organized by categories such as training, inference, multimodality...

X AI KOLs Timeline

A carefully curated collection of papers related to large model systems, covering training, inference, multimodality, and more. It is continuously updated and includes technical reports, frameworks, and courses, making it a valuable reference for researchers and developers.

MLUBench: A Benchmark for Lifelong Unlearning Evaluation in MLLMs

arXiv cs.AI

MLUBench is a large-scale benchmark for lifelong unlearning in multimodal large language models (MLLMs), featuring 127 entities across 9 classes. The paper identifies that existing unlearning methods suffer from cumulative degradation and proposes LUMoE to mitigate this, showing significant improvements.