Existing latent methods use a single scalar reward to refine all editable latent tokens, which is too coarse for multimodal reasoning where different tokens serve different roles—some are sensitive to visual evidence, others correspond to uncertain reasoning decisions. This work introduces Token-Disentangled Latent Test-Time Scaling, a framework that makes latent refinement token-role-aware. Starting from an initial generated trajectory, it optimizes a short hidden-state prefix while routing visual feedback to image-sensitive tokens and reasoning feedback to high-entropy tokens, with an anchor regularizer for remaining tokens. The method is evaluated on six perception and reasoning benchmarks using Qwen2.5-VL-7B and InternVL3.5-8B as frozen backbones. It lifts macro accuracy over chain-of-thought by +2.57 and +1.51 respectively, and outperforms strong output-space test-time scaling baselines under matched decoded-candidate budgets. This is interesting because it addresses a fundamental limitation of previous work—the inability to distinguish between visual and reasoning failure modes during latent refinement—and shows consistent gains across multiple models and tasks.
Token-aware latent refinement improves vision-language reasoning at inference time
The method directs visual and reasoning feedback to different hidden-state tokens while keeping model parameters frozen.
Chinese Tech
Hao-Xuan Ma · Yihao Liu · Yutao Sun · Yanting Miao · Mengyu Zhou · YiCheng Xiao · +5 more
Qwen Business Unit of Alibaba · Nanjing University · National Key Laboratory for Novel Software Technology, Nanjing University · Zhejiang University · University of Waterloo
Research Digest··2 min read
In one line
Token-disentangled latent test-time scaling improves vision-language reasoning by routing visual and reasoning feedback to different token roles.
Before you read this
Three ideas this paper builds ontest-time scalingAllocating extra computation during inference to improve output quality.117 papers →vision-language modelA model that handles tasks combining visual and textual information, such as editing images from conversation or reasoning across images and text.265 papers →role decouplingSeparating distinct functions such as planning, reasoning, and execution into separate components improves efficiency, quality, and reliability.52 papers →
Read it as
Ma et al. introduce an inference-time method that refines a short prefix of hidden states according to each token's estimated role.
The authors begin with an initial model-generated trajectory and optimize a short prefix of its hidden states without changing the model's parameters.
How it works
stage by stageThe Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.
What it released
Code