Test-Time Scaling for Small VLMs on Multilingual Visual MCQ
2026-07-10 • Computation and Language
Computation and LanguageArtificial IntelligenceMachine Learning
AI summaryⓘ
The authors studied how a technique called test-time scaling (TTS), which helps large language models reason better, works in smaller vision-language models. They found that the main challenge was making sure the model's answers were clearly formatted, and fixing this helped a lot. Increasing the allowed length for each reasoning step improved results more than trying fancy ways to pick the best answer. The biggest boost came from improving the model itself, and their best setup performed very well on a visual multiple-choice test. Overall, the study shows that simple fixes and a better model matter more than complex answer selection methods for small vision-language models.
test-time scalingvision-language modelsself-consistencybeam searchmultimodal reasoningprompt engineeringdecoding budgetimage multiple-choiceVisual MCQQwen model
Authors
Spiros Baxevanakis, Peng-Jian Yang
Abstract
Test-time scaling (TTS) reliably improves reasoning in large language models, but whether it transfers to small open vision-language models remains unclear. We examine this on EXAMS-V, a multilingual visual multiple-choice benchmark, comparing self-consistency, describe-then-reason with PRM-guided beam search, and two post-hoc selectors across Qwen2.5-VL-7B-Instruct and Qwen3.5-4B. What matters is the conditions under which TTS runs, not the search or verification machinery. The largest factor is parseability: an early prompt format left many chains reasoning correctly yet never committing to an answer letter, which a standard answer cue and a guided repair step largely remove. A larger decoding budget removes the rest: raising the per-chain token limit from 1k to 2k recovers 3.7 pp, whereas sampling more chains (8 to 16) adds only 0.15 pp. Once chains have room to finish, elaborate methods contribute little: PRM-guided beam search trails plain self-consistency by 0.39 pp at over eight times the cost, and neither a training-free generative critic nor a trained multimodal PRM beats majority vote across both policies. The largest gain comes instead from the policy model itself (+11.4 pp). Our best configuration reaches 84.1% on the held-out ImageCLEF 2026 test split, ranking first on the Visual MCQ leaderboard.