Test-Time Scaling Improves Small VLMs on Multilingual Visual MCQ, but Parseability Is Key
A new arXiv paper investigates whether test-time scaling (TTS) improves reasoning in small open vision-language models (VLMs) on the EXAMS-V multilingual visual multiple-choice benchmark. The study compares 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. The key finding is that parseability—the ability of the model to commit to a final answer—is the largest factor in TTS success, not the search or verification machinery.
Entities
Related
Self-Guided Test-Time Training Improves Long-Context LLM Accuracy
A new arxiv paper proposes Self-Guided Test-Time Training (SG-TTT) to improve long-context utilization in LLMs without requiring labeled data. The method uses the model's own predictions to generate pseudo-labels for fine-tuning on the test context, addressing accuracy degradation in long inputs. This approach is more efficient than full TTT and shows promise for practical deployment.
VLMs encode correct object counts internally but output wrong answers, study finds
A new arXiv paper investigates why vision-language models (VLMs) fail at basic object counting. By training probes on internal activations across four VLMs and five counting datasets, researchers found that nonlinear probes can reliably detect counting errors, indicating that models often encode the correct count even when they output the wrong answer. SVCCA analysis further confirms a misalignment between internal representations and verbalized outputs.
OmniMapBench benchmark tests LVLMs on visual-centric map reasoning
Researchers introduced OmniMapBench, a benchmark of 2,096 QA pairs across 1,603 map documents from nine categories, designed to evaluate visual-centric reasoning in large vision-language models. It addresses the limitation that many document benchmarks allow high performance via text-only cues, requiring genuine visual grounding for tasks from perception to multi-step reasoning.
QANTA 2026 Challenge: Task-Specific Multimodal QA Agents with Confidence Calibration
A team submitted a system to the QANTA 2026 shared challenge at ICML 2026, addressing multimodal quizbowl questions with incremental text and images. The system uses confidence calibration for tossup questions and incremental reasoning for bonus questions, aiming to balance accuracy and efficiency.
VAORA reward design addresses VLM failures in interactive physical reasoning
A new paper on arXiv introduces VAORA (Visual Action Outcome Reasoning Alignment), a reward design that targets two key failure modes in vision-language models: hallucinated chain-of-thought reasoning and misalignment between reasoning and actions. VAORA uses a Visual Alignment Reward to anchor reasoning to visual context, aiming to improve generalization in unseen interactive physical reasoning tasks.