MPRBench - Text Average: leaderboard
Personalized multimodal reasoning from fixed concept profiles, visual references and interaction histories. Personalized QA accuracy (%). MPRBench v1 Table1 standalone general-model prompting baselines. Personal database converted into structured JSON context; fixed question/reference/distractor population. Subtask accuracy; group averages are weighted by QA counts and overall is the published weighted group mean. Trained personalized-token/retrieval/memory methods retain separate identities and are not silently relabeled as their backbone LLM; their admission remains a separate systems review.
Metric: Personalized QA accuracy (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Pro (Benchmark Sweep Run: source=MPRBench; zero-shot structured-personal-database prompting; native base checkpoint/effort as published) | 90 | |
| 2 | GPT-5 (Benchmark Sweep Run: source=MPRBench; zero-shot structured-personal-database prompting; native base checkpoint/effort as published) | 87.2 | |
| 3 | GPT-5 Mini (Benchmark Sweep Run: source=MPRBench; zero-shot structured-personal-database prompting; native base checkpoint/effort as published) | 70.7 | |
| 4 | GPT-4o (Benchmark Sweep Run: source=MPRBench; zero-shot structured-personal-database prompting; native base checkpoint/effort as published) | 62.5 | |
| 5 | Qwen 2.5 VL 7B (Benchmark Sweep Run: source=MPRBench; zero-shot structured-personal-database prompting; native base checkpoint/effort as published) | 45.3 | |
| 6 | LLaVA-v1.5-13B (Benchmark Sweep Run: source=MPRBench; zero-shot structured-personal-database prompting; native base checkpoint/effort as published) | 35.5 |
Interactive version: theaggregate.ai/benchmark?slug=mprbench-text-average · How It Works · Data refreshed daily, snapshot 2026-11-06.