TimeLitmus - Traffic Natural Prediction: leaderboard
TimeLitmus (Gong et al., 2026) tests event-conditioned time-series prediction: given a textual event and the serialized pre-event series, the model predicts the expected response. In the Traffic domain, built on TraffiDent roadway incident reports and nearby freeway sensors, it picks one of three ordered response classes derived from observed post-report traffic on 420 natural records; accuracy in percent, higher is better.
Metric: Natural accuracy (%; 420 Traffic records of TraffiDent roadway incidents with three ordered response classes from observed post-report buckets; ten LLMs with deterministic decoding where available, invalid or unparsable outputs scored incorrect). Source: arxiv.org. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen Plus | 42.1 | #788 |
| 2 | DeepSeek R1 0528 | 41 | #425 |
| 3 | Qwen 3.5 9B | 40.7 | #803 |
| 4 | Qwen 3.5 4B | 38.6 | #1180 |
| 5 | Claude Sonnet 4.6 | 37.9 | #158 |
| 6 | MiniMax-M3 | 37.1 | #265 |
| 7 | DeepSeek V4 Flash | 36.9 | #254 |
| 8 | Gemini 3.5 Flash | 35.2 | #98 |
| 9 | GLM-5 | 35 | #256 |
| 10 | GPT-5.4 | 34.5 | #121 |
Interactive version: theaggregate.ai/benchmark?slug=timelitmus-traffic-natural-prediction · How It Works · Data refreshed daily, snapshot 2026-11-12.