Latent test-time scaling improves reasoning by refining hidden states during inference, but existing methods typically apply a single scalar reward to all editable latent tokens. For multimodal large language models, this global update ignores that generated tokens play different roles: some are sensitive to visual evidence, while others correspond to uncertain reasoning decisions. We present Token-Disentangled Latent Test-Time Scaling, an inference-time framework that makes latent refinement token-role-aware. Starting from an initial generated trajectory, we optimize a short hidden-state prefix while routing perception-side visual feedback to image-sensitive tokens and reasoning feedback to high-entropy tokens. Tokens selected by neither route are constrained by an anchor regularizer. Across six multimodal reasoning benchmarks and two frozen MLLM backbones, our method consistently improves over the initial chain-of-thought rollout: macro accuracy increases from 66.38 to 68.95 on Qwen2.5-VL-7B and from 67.93 to 69.44 on InternVL3.5-8B. These results show that token-disentangled latent refinement is an effective strategy for multimodal test-time scaling and remains competitive with output-space test-time scaling baselines.

Existing LLM routing methods often construct supervision from a single sampled response for each query–model pair. Because LLM generation is stochastic, however, such an observation can be an unstable estimate of model capability: semantically equivalent query formulations and repeated decoding may yield different scores and even different model preferences. We show that this instability can further propagate from routing labels to learned routing policies. To address this issue, we propose DARS (Distribution-Aware Routing Supervision), which estimates query-level model capability from repeated observations spanning semantics-preserving query rewrites and stochastic decoding. DARS summarizes expected quality, expected cost, and performance variability to construct risk-aware supervision without changing the downstream router architecture. Experiments across diverse tasks and routing methods show that DARS generally improves routing utility and cost–quality trade-offs over single-shot supervision. Further analyses show that its benefits persist under moderate sampling budgets and different decoding temperatures. These results suggest that reliable LLM routing should move beyond individual sampled outcomes and instead model query-level capability distributions.

Recent advancements in Large Language Models (LLMs) have demonstrated enhanced reasoning capabilities, evolving from Chain-of-Thought (CoT) prompting to advanced, product-oriented solutions like OpenAI o1. During our re-implementation of this model, we noticed that in multimodal tasks requiring visual input (e.g., geometry problems), Multimodal LLMs (MLLMs) struggle to maintain focus on the visual information, in other words, MLLMs suffer from a gradual decline in attention to visual information as reasoning progresses, causing text-over-relied outputs. To investigate this, we ablate image inputs during long-chain reasoning. Concretely, we truncate the reasoning process midway, then re-complete the reasoning process with the input image removed. We observe only a ~2 accuracy drop on MathVista’s test-hard subset, revealing the model’s textual outputs dominate the following reasoning process. Motivated by this, we propose Take-along Visual Conditioning (TVC), a strategy that shifts image input to critical reasoning stages and compresses redundant visual tokens via dynamic pruning. This methodology helps the model retain attention to the visual components throughout the reasoning. Our approach achieves state-of-the-art performance on average across five mathematical reasoning benchmarks (+3.4% vs previous sota), demonstrating the effectiveness of TVC in enhancing multimodal reasoning systems. The project page is available at https://sun-hailong.github.io/projects/TVC.