Reconstruction Alignment Improves Unified Multimodal Models

Ji Xie; Trevor Darrell; Luke Zettlemoyer; XuDong Wang

← RecA project

Reconstruction Alignment Improves Unified Multimodal Models

Ji Xie¹, Trevor Darrell¹, Luke Zettlemoyer², XuDong Wang^1,3

¹ UC Berkeley · ² University of Washington · ³ Duke University

International Conference on Learning Representations (ICLR), 2026 · Vol. 2026, pp. 120095–120137

Abstract

Unified multimodal models (UMMs) unify visual understanding and generation within a single architecture. However, conventional training relies on image-text pairs (or sequences) whose captions are typically sparse and miss fine-grained visual details, even when they use hundreds of words to describe a simple image. We introduce Reconstruction Alignment (RECA), a resource-efficient post-training method that leverages visual understanding encoder embeddings as dense "text prompts", providing rich supervision without captions. Concretely, RECA conditions a UMM on its own visual understanding embeddings and optimizes it to reconstruct the input image with a self-supervised reconstruction loss, thereby realigning understanding and generation. Despite its simplicity, RECA is broadly applicable: across autoregressive, masked-autoregressive, and diffusion-based UMMs, it consistently improves generation and editing fidelity. With only 27 GPU hours, post-training with RECA substantially improves image generation performance on GenEval (0.73 → 0.90) and DPGBench (80.93 → 88.15), while also boosting editing benchmarks (ImgEdit 3.38 → 3.75, GEdit 6.94 → 7.27). Notably, RECA surpasses much larger open-source models and applies broadly across diverse UMM architectures, establishing it as an efficient and general post-training alignment strategy for UMMs.

Published abstract, arXiv:2509.07295v4.

Experimental scope. The 27 A100 GPU-hour budget refers to the BAGEL experiment. The highest Harmon generation scores (0.90 GenEval and 88.15 DPG-Bench) use a separate SFT → RecA recipe with GPT-4o-Image distillation data. See model-specific results and the paper’s experiments and Appendix E for the full settings.

Citation

@inproceedings{xie2026reconstruction,
  title={Reconstruction alignment improves unified multimodal models},
  author={Xie, Ji and Zettlemoyer, Luke and Wang, Xudong and others},
  booktitle={International Conference on Learning Representations},
  volume={2026},
  pages={120095--120137},
  year={2026}
}