VibeSearchBench (ReAct): leaderboard
VibeSearchBench (Xiaohongshu, 2026) run in a plain ReAct loop that summarises its own context near 256K tokens: across 200 professional and everyday tasks an agent searches with a simulated user who reveals needs stage by stage, then extracts a knowledge graph of its findings. Triplet F1 in percent against the annotated graph, judged by an LLM and averaged over the Pro and Daily subsets and 3 runs; higher is better.
Metric: Triplet F1 (%; an LLM judge matches the knowledge graph the agent extracts after a multi-turn session with a progressive-disclosure user simulator against the annotated ground-truth graph; ReAct loop with self-summarising context compaction; mean of VibeSearch-Pro and VibeSearch-Daily; mean of 3 runs per task). Source: arxiv.org. Saturation forecast: Around May 2028. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.6 | 27.87 | #100 |
| 2 | DeepSeek V4 Pro | 27.04 | #181 |
| 3 | Kimi K2.6 | 26.09 | #173 |
| 4 | Gemini 3.1 Pro (Preview) | 23.54 | #91 |
| 5 | Seed 2.0 Pro | 23.22 | #209 |
| 6 | Qwen 3.5 397B A17B | 22.84 | #242 |
| 7 | GPT-5.4 | 20.04 | #121 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=vibesearchbench-react · How It Works · Data refreshed daily, snapshot 2026-11-12.