NLU Corruption - ANLI Trained Checkpoint ANLI R3: 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 answer accuracy (%). Higher is better.

Metric: Answer accuracy (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 3 models tracked. Updated 14 Nov 2026.

Top models

#ModelScoreOverall rank
1roberta-large-snli_mnli_fever_anli_R1_R2_R3-nli (Harness: NLU Clean)45.5
2roberta-large-snli_mnli_fever_anli_R1_R2_R3-nli (Harness: NLU Corrupt-Test VERB)39
3roberta-large-snli_mnli_fever_anli_R1_R2_R3-nli (Harness: NLU Corrupt-Test NOUN)33.8

Interactive version: theaggregate.ai/benchmark?slug=nlu-corruption-anli-trained-checkpoint-anli-r3 · How It Works · Data refreshed daily, snapshot 2026-11-16.