TimeLitmus - Traffic Event-Side Counterfactual Pairs: leaderboard
A controlled diagnostic of TimeLitmus (Gong et al., 2026): each of 30 Traffic pairs keeps the traffic history and changes the incident severity or lane impact 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 30 Traffic pairs whose incident severity or capacity impact is changed while the series is kept; ten LLMs with deterministic decoding where available, invalid or unparsable outputs scored incorrect). Source: arxiv.org. Saturation forecast: Around October 2028. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 | 63.3 | #121 |
| 2 | GLM-5 | 56.7 | #256 |
| 3 | MiniMax-M3 | 56.7 | #265 |
| 4 | Claude Sonnet 4.6 | 50 | #158 |
| 5 | Qwen 3.5 4B | 50 | #1180 |
| 6 | DeepSeek V4 Flash | 33.3 | #254 |
| 7 | DeepSeek R1 0528 | 26.7 | #425 |
| 8 | Qwen 3.5 9B | 23.3 | #803 |
| 9 | Qwen Plus | 13.3 | #788 |
| 10 | Gemini 3.5 Flash | 6.7 | #98 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=timelitmus-traffic-event-side-counterfactual-pairs · How It Works · Data refreshed daily, snapshot 2026-11-12.