HELM TORR - Tablebench Numerical Reasoning - Robustness: leaderboard
HELM ToRR (release v1.0.0) robustness on Tablebench Numerical Reasoning (TableBench numerical-reasoning questions; 73 instances): per instance, 1 minus the spread between the best and worst ROUGE-L across 35 prompt configurations (7 table serializations x original, row-shuffled, column-shuffled, transposed and padded tables), averaged. 0-100, higher is more consistent; a model that fails every configuration also scores high.
Metric: Robustness (0-100). Source: crfm.stanford.edu. 14 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Mistral 7B Instruct (v0.3) | 71.09 | #1255 |
| 2 | Llama 3.1 8B Instruct | 65.54 | #1017 |
| 3 | Mixtral 8x7B Instruct | 55.92 | #1085 |
| 4 | Qwen 2 72B Instruct | 55.4 | #579 |
| 5 | GPT-4o Mini (2024-07-18) | 54.92 | #568 |
| 6 | Gemini 1.5 Flash (002) | 54.45 | #587 |
| 7 | Claude 3.5 Haiku (20241022) | 54.07 | #590 |
| 8 | GPT-4o (2024-11-20) | 51.12 | #392 |
| 9 | Llama 3.1 405B Instruct | 50.7 | #470 |
| 10 | Gemini 1.5 Pro (002) | 50.65 | #368 |
| 11 | Mixtral 8x22B Instruct | 49.84 | #738 |
| 12 | Llama 3.1 70B Instruct | 49.2 | #567 |
| 13 | Claude 3.5 Sonnet (20241022) | 43.54 | #302 |
| 14 | DeepSeek V3 | 42.62 | #329 |
Interactive version: theaggregate.ai/benchmark?slug=helm-torr-tablebench-numerical-reasoning-robustness · How It Works · Data refreshed daily, snapshot 2026-11-05.