Peng-Jian Yang

Test-Time Scaling for Small VLMs on Multilingual Visual MCQ

Spiros Baxevanakis, Peng-Jian Yang

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-reas…Read more