TS-Skill (Free-Form): leaderboard
Free-form protocol of TS-Skill (Han et al., UCLA and Carnegie Mellon University, 2026): 3,000 questions about synthetic, timestamped time series from 25 domains that need temporal scale selection, temporal localization or cross-interval integration, answered in prose. GPT-5.4 parses each answer into the gold answer's structure, which is scored with type-matched comparators (labels, counts with off-by-one credit, banded numeric and timestamp error, interval overlap) plus a GPT-5.4 judge for rationale text; the score is the mean per-question score on a 0-1 scale, higher is better.
Metric: Native-form score (0-1; free-form answers parsed by GPT-5.4 and scored with answer-type comparators plus a GPT-5.4 rationale judge, mean over 3,000 questions). Source: arxiv.org. Saturation forecast: Around February 2027. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4o | 0.67 | #741 |
| 2 | O3 Mini | 0.61 | #536 |
| 3 | Claude Sonnet 4.5 | 0.6 | #244 |
| 4 | Qwen 3 32B (Thinking) | 0.55 | #960 (Qwen 3 32B) |
| 5 | Qwen 2.5 14B Instruct | 0.51 | #1745 |
| 6 | Qwen 2.5 VL 7B | 0.5 | #1796 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=ts-skill-free-form · How It Works · Data refreshed daily, snapshot 2026-11-12.