UniCorn: Towards Self-Improving Unified Multimodal Models through Self-Generated Supervision

TL;DR

UniCorn addresses Conduction Aphasia, the gap between multimodal understanding and generation in Unified Multimodal Models, by introducing a fully self-supervised post-training framework. It partitions a single model into Proposer, Solver, and Judge to generate diverse prompts, synthesize images, and internally judge quality, then distills this interaction through Cognitive Pattern Reconstruction into training signals. A Text→Image→Text cycle, UniCycle, provides a training-free measure of semantic coherence to verify genuine multimodal intelligence. Empirically, UniCorn achieves substantial gains across six T2I benchmarks, attaining SOTA on UniCycle and strong improvements on TIIF, WISE, OneIG, and others, while remaining robust under out-of-distribution conditions. The work demonstrates a scalable path toward fully self-contained unified multimodal intelligence by leveraging internal signals rather than external supervision or curated data.

Abstract

While Unified Multimodal Models (UMMs) have achieved remarkable success in cross-modal comprehension, a significant gap persists in their ability to leverage such internal knowledge for high-quality generation. We formalize this discrepancy as Conduction Aphasia, a phenomenon where models accurately interpret multimodal inputs but struggle to translate that understanding into faithful and controllable synthesis. To address this, we propose UniCorn, a simple yet elegant self-improvement framework that eliminates the need for external data or teacher supervision. By partitioning a single UMM into three collaborative roles: Proposer, Solver, and Judge, UniCorn generates high-quality interactions via self-play and employs cognitive pattern reconstruction to distill latent understanding into explicit generative signals. To validate the restoration of multimodal coherence, we introduce UniCycle, a cycle-consistency benchmark based on a Text to Image to Text reconstruction loop. Extensive experiments demonstrate that UniCorn achieves comprehensive and substantial improvements over the base model across six general image generation benchmarks. Notably, it achieves SOTA performance on TIIF(73.8), DPG(86.8), CompBench(88.5), and UniCycle while further delivering substantial gains of +5.0 on WISE and +6.5 on OneIG. These results highlight that our method significantly enhances T2I generation while maintaining robust comprehension, demonstrating the scalability of fully self-supervised refinement for unified multimodal intelligence.

Figures (13)

• Figure 1: Motivation of UniCorn. UMMs often exhibit an understanding-generation gap: they can accurately understand and critique errors in an image, yet fail to generate the same scene correctly. This conduction aphasia motivates our framework to leverage the model’s superior internal understanding to strengthen and refine its generative capabilities through self-contained feedback.
• Figure 2: Visualization results of UniCorn.
• Figure 3: Results of BAGEL bagel and GPT-4o gpt4o on four understanding benchmarks. For Omini-RewardBench jin2025omni and MMRB2 hu2025multimodal, we evaluate the T2I task. Performances are normalized with GPT-4 achiam2023gpt results for better visualization.
• Figure 4: Overview of the UniCorn Framework. (a) Illustrates the self-multi-agent collaboration for high-quality data sampling. (b) Details the Cognitive Pattern Reconstruction process, which reorganizes data to facilitate robust and efficient learning. (c) Presents the UniCycle benchmark evaluation, verifying whether the model can accurately reconstruct key textual information from its own generated content.
• Figure 5: Qualitative comparison between UniCorn, BAGEL and UniCorn's adifferent data settings. Our method jointly balances visual aesthetics, prompt fidelity, and realism in generation.