dMLLM-TTS: Self-Verified and Efficient Test-Time Scaling for Diffusion Multi-Modal Large Language Models

TL;DR

<3-5 sentence high-level summary> The paper addresses the efficiency and quality bottlenecks of test-time scaling for diffusion-based multi-modal LLMs (dMLLMs). It introduces dMLLM-TTS, combining Trajectory Exploration Scaling and Iterative Refinement Scaling with a Self-Verified Feedback mechanism and a Hierarchical Trajectory Search to achieve near-linear complexity $O(N+T)$ and remove external verifiers. Empirical results on GenEval across Lumina-DiMOO, MMaDA, and Muddit show consistent image-quality gains and substantial efficiency improvements (up to 5–6× faster than linear search) while maintaining or surpassing state-of-the-art baselines. The work demonstrates that in-loop self-verification and adaptive trajectory pruning can meaningfully boost text–image alignment for complex prompts in a unified diffusion-based, multi-modal inference framework.

Abstract

Diffusion Multi-modal Large Language Models (dMLLMs) have recently emerged as a novel architecture unifying image generation and understanding. However, developing effective and efficient Test-Time Scaling (TTS) methods to unlock their full generative potential remains an underexplored challenge. To address this, we propose dMLLM-TTS, a novel framework operating on two complementary scaling axes: (1) trajectory exploration scaling to enhance the diversity of generated hypotheses, and (2) iterative refinement scaling for stable generation. Conventional TTS approaches typically perform linear search across these two dimensions, incurring substantial computational costs of O(NT) and requiring an external verifier for best-of-N selection. To overcome these limitations, we propose two innovations. First, we design an efficient hierarchical search algorithm with O(N+T) complexity that adaptively expands and prunes sampling trajectories. Second, we introduce a self-verified feedback mechanism that leverages the dMLLMs' intrinsic image understanding capabilities to assess text-image alignment, eliminating the need for external verifier. Extensive experiments on the GenEval benchmark across three representative dMLLMs (e.g., Lumina-DiMOO, MMaDA, Muddit) show that our framework substantially improves generation quality while achieving up to 6x greater efficiency than linear search. Project page: https://github.com/Alpha-VLLM/Lumina-DiMOO.

Figures (7)

• Figure 1: dMLLM-TTS: We present the generative effects and performance improvements achieved by applying Test-Time Scaling (TTS) to dMLLMs. Images generated with TTS exhibit higher quality and stronger prompt alignment than those generated without TTS.
• Figure 2: Visualization of the image generation process in dMLLMs. The first row shows the input latent masks at each step, and the second row depicts the corresponding outputs. Sampling begins with fully masked tokens (gray) and gradually fills the discrete multimodal token space with increasingly confident predictions (blue).
• Figure 3: Overview of dMLLM-TTS framework. (a) dMLLM-TTS scales compute along two axes: trajectory exploration and iterative refinement, guided by Self-Verified Feedback for text–image alignment evaluation. (b) Hierarchical Trajectory Search (HTS) performs coarse-to-fine generation by starting with broad exploration, pruning low-potential trajectories, and refining high-potential trajectories.
• Figure 4: Qualitative improvement ratio in TTS performance across various text prompt complexities examined through diverse dMLLMs on GenEval benchmark dimensions. TTS markedly enhances performance across all measured dimensions.
• Figure 5: Comparison between linear and hierarchical trajectory search. The red curve illustrates linear trajectory search, while the blue curve depicts hierarchical trajectory search, with a dashed line indicating predictions based on a geometric series decay approximation. Curve fitting shows that similar subsequent trends tend to converge towards an upper limit.