TimeLitmus - Traffic Series-Side Counterfactual Pairs: leaderboard
A controlled diagnostic of TimeLitmus (Gong et al., 2026): each of 136 Traffic pairs keeps the incident report and changes the pre-report traffic state so that the expected response changes. Both endpoints are evaluated independently and a pair counts only when both predictions are correct; pair correctness in percent, higher is better.
Metric: Counterfactual pair correctness (%; both independently evaluated endpoints correct on 136 Traffic pairs whose pre-report traffic series is changed while the incident is kept; 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 | DeepSeek R1 0528 | 21.3 | #425 |
| 2 | Qwen 3.5 4B | 20.6 | #1180 |
| 3 | DeepSeek V4 Flash | 18.4 | #254 |
| 4 | Qwen 3.5 9B | 16.9 | #803 |
| 5 | Qwen Plus | 16.9 | #788 |
| 6 | MiniMax-M3 | 11.8 | #265 |
| 7 | Claude Sonnet 4.6 | 11 | #158 |
| 8 | GPT-5.4 | 8.8 | #121 |
| 9 | Gemini 3.5 Flash | 8.8 | #98 |
| 10 | GLM-5 | 4.4 | #256 |
Interactive version: theaggregate.ai/benchmark?slug=timelitmus-traffic-series-side-counterfactual-pairs · How It Works · Data refreshed daily, snapshot 2026-11-12.