ALTo: Adaptive-Length Tokenizer for Autoregressive Mask Generation

Tao Wang (Ministry of Emergency Management Big Data Center) · Lingfeng Wang (Zhejiang University) · Hualing Lin (Zhejiang University) · Senda Chen (Tongji University) · Changxu Cheng (UNI-UBI) · Yangyang Zhong (Tsinghua University) · Dong Zheng (Hangzhou Yufan Intelligent Technology Co., LTD) · Wuyue Zhao (Cornell University)

adaptive length tokenizeradaptive token costaltollmattention allocationautoregressive mask generationdifferentiable token chunkingefficiency trade-offsgroup relative policy optimizationlength regularizationmask qualitymultimodal large language modelssegmentation benchmarksstate-of-the-art performancetoken length predictorvisual object representation

While humans effortlessly draw visual objects and shapes by adaptively allocating attention based on their complexity, existing multimodal large language models (MLLMs) remain constrained by rigid token representations. Bridging this gap, we propose ALTo, an adaptive length tokenizer for autoregressive mask generation. To achieve this, a novel token length predictor is designed, along with a length regularization term and a differentiable token chunking strategy. We further build ALToLLM that seamlessly integrates ALTo into MLLM. Preferences on the trade-offs between mask quality and efficiency is implemented by group relative policy optimization (GRPO). Experiments demonstrate that ALToLLM achieves state-of-the-art performance with adaptive token cost on popular segmentation benchmarks. Code and models will be released.