TimeLitmus - Traffic Hard Paired Contrasts: leaderboard
Hard paired contrasts in TimeLitmus (Gong et al., 2026): 60 naturally occurring pairs of similar Traffic incident and sensor inputs that require different responses, each endpoint evaluated independently. A pair counts only when both predictions are correct; pair correctness in percent, higher is better.
Metric: Hard paired contrast pair correctness (%; both endpoints correct on 60 naturally occurring similar Traffic pairs with different gold responses; 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 3.5 9B | 11.7 | #803 |
| 2 | DeepSeek V4 Flash | 11.7 | #254 |
| 3 | DeepSeek R1 0528 | 11.7 | #425 |
| 4 | Qwen 3.5 4B | 6.7 | #1180 |
| 5 | MiniMax-M3 | 6.7 | #265 |
| 6 | Qwen Plus | 6.7 | #788 |
| 7 | Claude Sonnet 4.6 | 1.7 | #158 |
| 8 | GLM-5 | 1.7 | #256 |
| 9 | GPT-5.4 | 0 | #121 |
| 10 | Gemini 3.5 Flash | 0 | #98 |
Interactive version: theaggregate.ai/benchmark?slug=timelitmus-traffic-hard-paired-contrasts · How It Works · Data refreshed daily, snapshot 2026-11-12.