SpecEyes: Accelerating Agentic Multimodal LLMs via Speculative Perception and Planning

TL;DR

SpecEyes is a framework that accelerates agentic multimodal LLMs by using speculative planning with a lightweight model to predict execution paths, reducing latency and increasing throughput without sacrificing accuracy.

Contribution

It introduces a novel speculative acceleration method with a cognitive gating mechanism and parallel funnel to improve efficiency of agentic multimodal LLMs.

Findings

Achieves 1.1-3.35x speedup over baseline.

Maintains or improves accuracy up to +6.7%.

Enhances system throughput under concurrent workloads.

Abstract

Agentic multimodal large language models (MLLMs) (e.g., OpenAI o3 and Gemini Agentic Vision) achieve remarkable reasoning capabilities through iterative visual tool invocation. However, the cascaded perception, reasoning, and tool-calling loops introduce significant sequential overhead. This overhead, termed agentic depth, incurs prohibitive latency and seriously limits system-level concurrency. To this end, we propose SpecEyes, an agentic-level speculative acceleration framework that breaks this sequential bottleneck. Our key insight is that a lightweight, tool-free MLLM can serve as a speculative planner to predict the execution trajectory, enabling early termination of expensive tool chains without sacrificing accuracy. To regulate this speculative planning, we introduce a cognitive gating mechanism based on answer separability, which quantifies the model's confidence for self-verification without requiring oracle labels. Furthermore, we design a heterogeneous parallel funnel that exploits the stateless concurrency of the small model to mask the stateful serial execution of the large model, maximizing system throughput. Extensive experiments on V* Bench, HR-Bench, and POPE demonstrate that SpecEyes achieves 1.1-3.35x speedup over the agentic baseline while preserving or even improving accuracy (up to +6.7%), thereby boosting serving throughput under concurrent workloads.