TSQueryBench - Relative Ranking: leaderboard

Judging time-series explanations on TSQueryBench (2026): for each of 500 synthetic series the model sees a correct, a partially correct and an incorrect explanation built from ground-truth statistics and must pick the correct one. Accuracy in percent; higher is better.

Metric: Relative ranking accuracy (%, times 100): share of instances where the model picks the correct explanation among a correct, a partially correct and an incorrect candidate in randomized order on TSQueryBench, 500 synthetic time series in ten query types (spikes, drops, structural breaks, mean and volatility shifts, trends, ordering, shape), zero-shot with structured JSON output; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 6 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Qwen 3 8B84#1890
2DeepSeek V3.274#425
3Gemma 4 31B (IT)67#354
4Gemma 3 12B (IT)56#1907
5Llama 3.1 8B Instruct49#4655
6Claude 3 Haiku41#2683

No result here: #2 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=tsquerybench-relative-ranking · How It Works · Data refreshed daily, snapshot 2026-11-22.