Research map / Vision-language and multimodal / Multimodal LLMs

LifeEval: A Multimodal Benchmark for Assistive AI in Egocentric Daily Life Tasks

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 LifeEval, a 4,075-QA egocentric benchmark from Ego4D videos for evaluating MLLMs as real-time, task-oriented assistants across six capability dimensions, and benchmarks 26 MLLMs.

Problem

Existing video benchmarks evaluate passive, offline, retrospective understanding and isolated skills, not real-time task-oriented assistance from egocentric streams in daily life.

Approach

A benchmark that scores MLLMs on six assistance-oriented capabilities (perception, dynamic reasoning, knowledge retrieval, planning, safety, multi-turn collaboration) using Ego4D clips, with MCQ and open-ended QA generated by Gemini 2.5 Pro and refined by human review, evaluated under a streaming window protocol.

Key results

• Gemini-2.5-Pro reaches 85.61% overall on multiple-choice questions, the highest among evaluated models, while GPT-5 leads open-ended questions with 72.39%.
• All open-source models fall below 50% on open-ended questions; best open-source is Qwen3-VL-30B-A3B(thinking) at 49.24%.
• Going from 8 to 32 frames gives no consistent gain; for GPT-5 MCQ accuracy is 82.48% at 8 frames vs 82.51% at 32 frames.

Surprising findings

• LLaVA-OneVision-72B has the top open-source MCQ score but weak open-ended performance.
• Larger models within InternVL3.5 and LLaVA-NeXT families sometimes underperform smaller siblings, suggesting scaling alone is insufficient.

Limitations

• Benchmark is relatively small in video clips (591 clips, 44.19 hours of source video from 100 scenarios).
• Open-ended scoring relies on a single LLM judge (GPT-5) against reference answers.
• Evaluation uses a streaming window of brief intervals and fixed 8-frame sampling, which may not capture long-horizon real-time interaction.

Experimental setup

Datasets: Ego4D

Benchmarks: LifeEval

Metrics: accuracy LLM-judge score (0-1, 0.25 increments)

Baselines: GPT-5 GPT-4o GPT-5-mini Gemini-2.5-Pro Gemini-2.5-Flash Grok-4 Qwen3-VL-30B-A3B Qwen3-VL-8B InternVL3.5 LLaVA-OneVision LLaVA-NeXT mPLUG-Owl3

Builds on: Ego4D EgoPlan-Bench2 EgoLifeQA

Closest papers by meaning

• HumanSense: From Multimodal Perception to Empathetic Context-Aware Responses Through Reasoning MLLMs (AAAI 2026 · similarity 0.90)
• EgoThink: Evaluating First-Person Perspective Thinking Capability of Vision-Language Models (CVPR 2024 · similarity 0.88)
• EgoLife: Towards Egocentric Life Assistant (CVPR 2025 · similarity 0.88)
• Video-Only ToM: Enhancing Theory of Mind in Multimodal Large Language Models (CVPR 2026 · similarity 0.87)
• In the Eye of MLLM: Benchmarking Egocentric Video Intent Understanding with Gaze-Guided Prompting (NeurIPS 2025 · similarity 0.87)
• On Path to Multimodal Generalist: General-Level and General-Bench (ICML 2025 · similarity 0.87)