SpecEyes: Speculative Perception and Planning Achieves 1.1–3.35x Speedup for Agentic Multimodal LLMs While Improving Accuracy

arXiv·medium signal

Researchers introduced SpecEyes, a framework that uses a lightweight tool-free model to predict execution paths for agentic multimodal LLMs like o3 and Gemini Agentic Vision, enabling early termination of expensive operations. A cognitive gating mechanism based on answer separability regulates predictions without labeled oracle data. SpecEyes achieves 1.1–3.35x speedup over agentic baselines while preserving or improving accuracy by up to +6.7%.