Test-time scaling improves small VLMs by fixing prompt parseability and increasing token budget

The study examines whether test-time scaling (TTS) transfers to small open vision-language models using the EXAMS-V multilingual visual multiple-choice benchmark. It compares self-consistency, describe-then-reason with PRM-guided beam search, and post-hoc selectors across Qwen2.5-VL-7B-Instruct and Qwen3.5-4B.

• Parseability is the largest factor; an early prompt format prevented answer commitment, which was fixed by a standard answer cue and guided repair step.
• Increasing the per-chain token limit from 1k to 2k recovers 3.7 pp, while sampling more chains (8 to 16) adds only 0.15 pp.
• Elaborate methods contribute little once chains can finish; PRM-guided beam search trails plain self-consistency by 0.39 pp at over eight times the cost.
• Neither a training-free generative critic nor a trained multimodal PRM beats majority vote across both policies.
• The largest gain comes from the policy model itself (+11.4 pp), with the best configuration reaching 84.1% on ImageCLEF 2026.

The authors consider it important because simple adjustments to prompt format and decoding budget yield significant gains without the high cost of complex verification machinery.