NLU Corruption - New Items MNLI Style: 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, centered where probabilities exist (%). Higher is better.

Metric: Clean answer accuracy, centered where probabilities exist (%). Source: arxiv.org. Saturation forecast: Estimated already saturated. 8 models tracked. Updated 14 Nov 2026.

Top models

#ModelSettingScoreOverall rank
1Qwen 3.8 27B100#209
2DeepSeek V4 Pro99.4#191
3OLMo 3 32B Base99.4
4Muse Glimmer 30B98.8#391
5GPT-4o98.2#744
6Mistral Large 398.2#827
7OLMo 3 7B Base98.2
8Grok 4.7Reasoning98.2#47

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-new-items-mnli-style · How It Works · Data refreshed daily, snapshot 2026-11-16.