Agentic-MME: What Agentic Capability Really Brings to Multimodal Intelligence? (April 2026) logo

Agentic-MME: What Agentic Capability Really Brings to Multimodal Intelligence? (April 2026)

Free

Systematic evaluation of agentic capability in multimodal LLMs — decomposes tasks into perception, reasoning, and action levels; reveals where agentic loops help vs. where they add overhead

FreeFree tier
Type
Open Source

About Agentic-MME: What Agentic Capability Really Brings to Multimodal Intelligence? (April 2026)

Agentic-MME is a process-verified benchmark for evaluating multimodal agentic capabilities in large language models. It comprises 418 real-world tasks across 6 domains and 3 difficulty levels, featuring over 2,000 stepwise checkpoints that each required 10+ person-hours of manual annotation. The benchmark decomposes tasks into perception, reasoning, and action levels, and uses a dual-axis (S-axis and V-axis) human reference trajectory to enable fine-grained process verification. It also introduces an overthinking metric to quantify efficiency by comparing model trajectories against human references. Evaluations show that the best model (Gemini3-pro) achieves 56.3% overall accuracy, dropping to 23.0% on Level-3 tasks, highlighting the challenge of real-world multimodal agentic problem solving.

Key Features

418 real-world tasks across 6 domains and 3 difficulty levels
Over 2,000 stepwise checkpoints with 10+ person-hours annotation per task
Dual-axis human reference trajectory (S-axis and V-axis)
Process-level verification auditing fine-grained intermediate states
Overthinking metric to measure efficiency against human trajectories
Unified evaluation framework supporting sandboxed code and APIs
Evaluates Visual Expansion (tool invocation) and Knowledge Expansion (web search)

Pros & Cons

Pros
  • Comprehensive process verification beyond final-answer evaluation
  • Granular stepwise checkpoints with high-quality manual annotation
  • Includes efficiency metric (overthinking) to quantify overhead
  • Covers diverse real-world tasks with multiple difficulty levels
  • Open-source benchmark available for research community
Cons
  • Best models still achieve only 56.3% overall accuracy, with sharp drop on hard tasks (23.0% on Level-3)
  • Benchmark may not fully capture dynamic real-world tool interactions
  • Overthinking metric may not generalize to all agentic architectures

Best For

Benchmarking multimodal LLMs for real-world agentic tasksAssessing capability synergy across perception, reasoning, and actionEvaluating tool use and knowledge expansion in MLLMsMeasuring efficiency and overthinking in agentic workflowsResearch on multimodal agentic intelligence and model limitations

FAQ

What does Agentic-MME evaluate?
It evaluates multimodal agentic capabilities of MLLMs, specifically Visual Expansion (invoking visual tools) and Knowledge Expansion (open-web search), through process-verified stepwise checkpoints.
How is Agentic-MME different from existing benchmarks?
Existing benchmarks lack flexible tool integration, test visual and search tools separately, and evaluate only final answers. Agentic-MME provides process-level verification with fine-grained intermediate state auditing and an overthinking efficiency metric.
What models were evaluated?
The benchmark was tested on various MLLMs, with the best result from Gemini3-pro achieving 56.3% overall accuracy and 23.0% on Level-3 tasks.
Is Agentic-MME freely available?
Yes, the benchmark is open-source and the paper is freely available on arXiv under a permissive license.