TandemBench - Verification Accuracy: leaderboard
Multi-agent cooking coordination with exact-search optimal plans and paired optimality-verification questions. Native protocol accuracy (%). TandemBench English v1.0.0,1200questions; global state/rules/shared goal and candidate plans; reference outcomes from execution and exact search. Four cyclic option rotations; native decision models one response/rotation, general models four scored separately without voting. DA=optimal concrete-plan selection; VA averages detection when optimum removed and rejection when retained. Truncated/missing/multiple/infrastructure-failure outputs count incorrect. Native interfaces/calibration/precision retained perAppendixC; Qwen thinking disabled, MuseGlimmer low.
Metric: Native protocol accuracy (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 50 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.8 27B (Benchmark Sweep Run: source=TandemBench; general-generation; four-responses/rotation; thinking=disabled) | 71.88 | |
| 2 | Muse Glimmer 30B (Benchmark Sweep Run: source=TandemBench; general-generation; four-responses/rotation; reasoning=low) | 60.35 | |
| 3 | Gemma 4 31B (Benchmark Sweep Run: source=TandemBench; general-generation; four-responses/rotation; native reasoning configuration) | 57.03 | |
| 4 | D3-Ultra (Benchmark Sweep Run: source=TandemBench; native-decision-interface; parameters=27B; one-response/rotation; native AppendixC precision/calibration) | 56.28 | |
| 5 | PPLX-Decider-v1.2 (Benchmark Sweep Run: source=TandemBench; native-decision-interface; parameters=27B; one-response/rotation; native AppendixC precision/calibration) | 55.31 | |
| 6 | GLiDE-1.2-Open (Benchmark Sweep Run: source=TandemBench; native-decision-interface; parameters=27B; one-response/rotation; native AppendixC precision/calibration) | 54.72 | |
| 7 | PPLX-Decider-v1.1-27B (Benchmark Sweep Run: source=TandemBench; native-decision-interface; parameters=27B; one-response/rotation; native AppendixC precision/calibration) | 54.28 | |
| 8 | StartLux-v1.1-27B-Instant (Benchmark Sweep Run: source=TandemBench; native-decision-interface; parameters=27B; one-response/rotation; native AppendixC precision/calibration) | 54 | |
| 9 | Gemma 4 26B A4B (Benchmark Sweep Run: source=TandemBench; general-generation; four-responses/rotation; native reasoning configuration) | 53.66 | |
| 10 | Liquid AI d1-3B (Benchmark Sweep Run: source=TandemBench; native-decision-interface; parameters=3.12B; one-response/rotation; native AppendixC precision/calibration) | 53.47 |
Interactive version: theaggregate.ai/benchmark?slug=tandembench-verification-accuracy · How It Works · Data refreshed daily, snapshot 2026-11-06.