NLU Corruption - Centered ANLI R1: 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 centered clean answer accuracy (%). Higher is better.

Metric: Centered clean answer accuracy (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 6 models tracked. Updated 14 Nov 2026.

Top models

#ModelScoreOverall rank
1OLMo 3 32B Base (Harness: NLU centered 5-shot, drop-NOUN)75.1
2OLMo 3 32B Base (Harness: NLU centered 5-shot, clean)75
3OLMo 3 32B Base (Harness: NLU centered 0-shot)74
4OLMo 3 7B Base (Harness: NLU centered 5-shot, clean)66.2
5OLMo 3 7B Base (Harness: NLU centered 5-shot, drop-NOUN)63.6
6OLMo 3 7B Base (Harness: NLU centered 0-shot)63

Interactive version: theaggregate.ai/benchmark?slug=nlu-corruption-centered-anli-r1 · How It Works · Data refreshed daily, snapshot 2026-11-16.