Test-time scaling improves multilingual visual MCQ for small VLMs

The study examines whether test-time scaling (TTS) transfers to small open vision-language models using the EXAMS-V benchmark, comparing methods across Qwen2.5-VL-7B-Instruct and Qwen3.5-4B.

• Parseability is the primary factor for success; adding a standard answer cue and guided repair step prevents chains from reasoning without committing to an answer.
• Increasing the per-chain token limit from 1k to 2k recovers 3.7 percentage points, while increasing sampled chains from 8 to 16 adds only 0.15 pp.
• PRM-guided beam search trails plain self-consistency by 0.39 pp at over eight times the cost, and neither generative critics nor trained multimodal PRMs beat majority vote.
• The policy model itself provides the largest gain (+11.4 pp), with the best configuration reaching 84.1% on the ImageCLEF 2026 test split.

The authors consider this important because it demonstrates that simple scaling of decoding budget and prompt format yields better results than complex verification machinery for small VLMs.