A test-time method reported higher macro-average accuracy than an initial chain-of-thought rollout across six multimodal reasoning benchmarks. The reported gains were 2.57 points on Qwen2.5-VL-7B and 1.51 points on InternVL3.5-8B, using two frozen open-weight backbones.
The work is an arXiv version-one preprint dated 28 September 2026. Its primary evaluation covered perception-heavy and reasoning-heavy settings on the two backbones.
A different way to spend extra thinking time
The method works at test time, meaning it changes the model’s instance-specific internal states while it answers a question, without updating the backbone’s learned weights. It separates two kinds of feedback: visual feedback is sent to tokens that are especially sensitive to the image, while reasoning feedback is sent to tokens with high uncertainty. The approach keeps those channels separate instead of applying one global update to the entire answer.
At each refinement step, the system combines a token-routed policy loss with an entropy term and an anchor penalty. In ordinary language, those ingredients guide the update toward selected tokens, regulate uncertainty during reasoning and keep the revised state tied to an anchor. The default setup used four latent refinement steps.
The comparison accounted for the number of decoded candidates, rather than simply giving one method more sampled answers than another. Sampling-based baselines were allocated 16 candidate continuations. The proposed method also used reward-scoring and image-sensitivity passes, which added computation even though its backbone remained frozen.
The reported gains were spread across both kinds of task
On Qwen2.5-VL-7B, the reported gain over the initial chain-of-thought rollout was 2.55 points for the perception group and 2.59 points for the reasoning group. On InternVL3.5-8B, the corresponding gains were 0.99 and 2.04 points. The evaluation compared perception-heavy and reasoning-heavy settings across the six benchmarks.
A matched-budget comparison reported scores of 72.10 on MathVista and 46.98 on LogicVista, the strongest scores listed in that comparison. For the proposed method, the accounting included 17 generation calls, about 61 score calls, roughly 2.2 thousand decoded tokens per sample and about 44 GPU seconds per sample. The runtime figures were estimates from sharded Qwen2.5-VL-7B runs.
The ablation compared the full setting with versions that removed either reward branch or randomized token routing. The full setting reported perception and reasoning averages of 66.09 and 59.54. Without visual reward, the reported averages were 64.88 and 58.18; without reasoning reward, they were 64.19 and 57.31; and with random routing, they were 64.72 and 58.87.
The reported pattern was not confined to the two main backbones. Across six additional backbones, both group averages improved on all six, although isolated benchmark cells could fall within the stated noise band. The largest reported isolated-cell drop was 0.39 points.
More refinement did not keep producing clear gains
The refinement sweep found that both group averages rose as the number of steps increased from one to four, then levelled off at eight or more steps within the reported noise band. Four steps were therefore adopted as the default.
The method’s extra computation is part of the trade-off. It requires latent refinement, multiple candidate continuations, reward scoring and image-sensitivity calculations. Its visual-engagement and likelihood-based rewards are proxy measures, so they may be noisy for ambiguous, underspecified or external-knowledge questions. The approach also uses hard top-K routing and relies on editable open-weight models, excluding API-only closed models.
Paper data and sources
Original title: Token-Disentangled Latent Test-Time Scaling for Vision-Language Reasoning
Authors: Hao-Xuan Ma, Yihao Liu, Yutao Sun et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-09-28
DOI: Not available
Original paper · Full text