Research map / Vision-language and multimodal / Multimodal LLMs

MA-Bench: Towards Fine-grained Micro-Action Understanding

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Unified any-modality language models · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces MA-Bench, a 1,000-video, 12,000-QA benchmark for fine-grained micro-action understanding in MLLMs, plus MA-Bench-Train (20.5K videos with structured captions), and shows LoRA fine-tuning Qwen3-VL-8B on it improves results.

Problem

MLLMs' ability to understand fine-grained micro-actions (subtle body movements tied to emotion) is unexplored because no specialized benchmark exists.

Approach

A three-tier Perception–Comprehension–Reasoning benchmark built from MA-52 videos, with structured micro-action captions generated from optical-flow and skeleton-based body-part motion descriptors, then semi-automatic QA generation with LLM verification and human rectification.

Key results

• Gemini-2.5-Flash achieves the best zero-shot closed-ended average accuracy of 50.70%, with GPT-4o at 44.87%.
• Qwen3-VL-8B fine-tuned with LoRA on MA-Bench-Train improves closed-ended average from 46.97% to 50.68% and open-ended average from 0.75 to 1.69.
• Fine-tuning only the LM decoder yields 50.63% closed-ended average; fine-tuning all modules yields the best 50.68%.

Surprising findings

• Most zero-shot MLLMs score near chance on the coarse- and fine-grained recognition tasks (e.g., CMAR around 12-43%), and many open-source models score very low on open-ended tasks.

Limitations

• Videos are short (average 2.12 s), limiting temporal context.
• Open-ended evaluation relies on GPT-4o as judge.
• Benchmark built from a single source dataset (MA-52) with semi-automatic generation.

Experimental setup

Datasets: Micro-Action-52 MA-Bench-Train

Benchmarks: MA-Bench

Metrics: accuracy L1/L2/L3 LLM-judged score (0-5)

Baselines: GPT-4o Gemini-2.5-Flash Grok-2-vision VideoChat-Flash-Qwen2-7B InternVideo2.5-Chat-8B Qwen2.5-VL-7B InternVL3-8B VideoLLaMA3-7B

Builds on: Micro-Action-52 CoTracker3 LLaMA-Factory LoRA

Closest papers by meaning

• LLaVAction: evaluating and training multi-modal large language models for action understanding (ICLR 2026 · similarity 0.90)
• MME-Emotion: A Holistic Evaluation Benchmark for Emotional Intelligence in Multimodal Large Language Models (ICLR 2026 · similarity 0.89)
• Prototypical Calibrating Ambiguous Samples for Micro-Action Recognition (AAAI 2025 · similarity 0.88)
• MMWorld: Towards Multi-discipline Multi-faceted World Model Evaluation in Videos (ICLR 2025 · similarity 0.87)
• Grounding Multimodal Large Language Models in Actions (NeurIPS 2024 · similarity 0.87)
• FAVOR-Bench: A Comprehensive Benchmark for Fine-Grained Video Motion Understanding (NeurIPS 2025 · similarity 0.87)