NLU Corruption - Clean MNLI: leaderboard

This study tests natural-language inference with clean and corrupted inputs. Raw accuracy, answer-bias-centred accuracy and prompting conditions remain separate. The public ANLI-trained RoBERTa checkpoint is distinct from generic RoBERTa and the study’s locally fine-tuned encoders. Above-chance retention is not relabelled as accuracy. This board reports clean answer accuracy (%). Higher is better.

Metric: Clean answer accuracy (%). Source: arxiv.org. Saturation forecast: Around February 2027. 8 models tracked. Updated 14 Nov 2026.

Top models

#ModelSettingScoreOverall rank
1OLMo 3 32B Base89.3
2Qwen 3.8 27B88.9#209
3OLMo 3 7B Base88.8
4GPT-4o81.5#744
5Mistral Large 380.7#827
6DeepSeek V4 Pro74.2#191
7Grok 4.7Reasoning71.8#47
8Muse Glimmer 30B67.2#391

No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.

Interactive version: theaggregate.ai/benchmark?slug=nlu-corruption-clean-mnli · How It Works · Data refreshed daily, snapshot 2026-11-16.