Seer Accelerates Diffusion MLLMs by Truncating Redundant Output

Diffusion Multimodal Large Language Models (DMLLMs) are highly effective for multimodal reasoning, but their inference efficiency is severely hampered by fixed-length generation constraints. Since the actual output length is unknown, sequences are padded to a maximum length, leading to substantial wasted computation on unnecessary end-of-sequence (EOS) tokens. This research introduces Seer, a training-free framework that addresses this inefficiency. Seer leverages a key discovery: DMLLMs implicitly reveal their valid semantic boundary at the very first denoising step through a distinct shift in MLP activation sparsity. By using a Signal-to-Noise Ratio (SNR)-based criterion, Seer detects this boundary and performs a one-shot truncation of the redundant suffix for all subsequent computations. To maintain these gains during batched serving, Seer employs a hybrid execution strategy that maximizes throughput while seamlessly accommodating dynamic sequence lengths. Experimental results demonstrate that Seer effectively eliminates padding waste, accelerating throughput by up to 31 times. Crucially, it robustly maintains overall performance across nine benchmarks and even improves accuracy on complex visual tasks by mitigating noise leakage.