EviSeek preprint reports higher accuracy with fewer tokens for embodied agents

Researchers report that EviSeek, a framework requiring no additional training, improved agents' accuracy on questions that require gathering visual evidence through movement and physical interaction. On 3,436 ESI-Bench questions, accuracy rose over the Single-model agent by 7.1 percentage points with Qwen3-VL-32B and 3.7 points with Gemini 3.1 Pro. The framework turns each question's evidence requirements into explicit objectives, guiding actions and deciding which observations to retain. The preprint also reports improved visual-support-weighted accuracy and lower mean token usage. The demonstrated gains concern the ESI-Bench evaluation with those two models.