VehicleMemBench (Recursive Summarization): leaderboard
Metric: Exact state match (%): share of VehicleMemBench's 500 executable in-vehicle queries whose final vehicle state after the agent's tool calls (at most 10 rounds, 111 vehicle APIs) matches the reference exactly; the agent must recall the preferences of three users from an average 93,000-token multi-day dialogue history; memory built by summarizing each day of dialogue into one running preference summary placed in the system prompt; higher is better. Source: arxiv.org. Saturation forecast: Around January 2027. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro (Preview) | 64.8 | #64 |
| 2 | GPT-5 | 61.2 | #91 |
| 3 | Qwen 3 Max (Thinking) | 60.6 | #201 (Qwen 3 Max) |
| 4 | Qwen 3.5 397B A17B (Thinking) | 59.4 | #141 (Qwen 3.5 397B A17B) |
| 5 | Doubao-Seed-1.6 (Thinking) | 59 | |
| 6 | Kimi K2.5 (Thinking) | 56.4 | #139 (Kimi K2.5) |
| 7 | GLM-5 (Thinking) | 55 | #137 (GLM-5) |
| 8 | GLM-4.7 (Thinking) | 55 | #185 (GLM-4.7) |
| 9 | MiniMax-M2.5 | 52.2 | #295 |
| 10 | MiniMax-M2.1 | 50 | #283 |
| 11 | GLM-4.7 Flash (Non-reasoning) | 24.2 | #496 (GLM-4.7 Flash) |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=vehiclemembench-recursive-summarization · How It Works · Data refreshed daily, snapshot 2026-10-11.