Zero-Shot Classification2018en

XNLI

Cross-lingual natural language inference across 15 languages

Current State of the Art

GPT-4

OpenAI

87.4

accuracy

accuracy Progress Over Time

Showing 2 breakthroughs from Nov 2019 to Mar 2023

83.284.485.586.687.8Nov 2019Mar 2023accuracyDate

Key Milestones

Nov 2019
XLM-RoBERTa-large

XLM-RoBERTa-large XNLI avg accuracy (15 languages). State-of-the-art at publication. From Table 3.

83.6
Source
Mar 2023
GPT-4Current SOTA

GPT-4 average XNLI accuracy (15 languages). From GPT-4 Technical Report and cross-lingual evaluation studies.

87.4
+4.5%
Source
Total Improvement
4.5%
Time Span
3y 5m
Breakthroughs
2
Current SOTA
87.4

Top Models Performance Comparison

Top 3 models ranked by accuracy

accuracy1GPT-487.4100.0%2XLM-RoBERTa-large83.695.7%3mDeBERTa-v3-base80.892.4%0%25%50%75%100%% of best
Best Score
87.4
Top Model
GPT-4
Models Compared
3
Score Range
6.6

accuracyPrimary

#ModelScorePaper / CodeDate
1
GPT-4
OpenAI		87.4
arxiv
Mar 2023
2
XLM-RoBERTa-largeOpen Source
Facebook AI		83.6
Unsupervised Cross-lingual Representation Learning at Scale
Nov 2019
3
mDeBERTa-v3-baseOpen Source
Microsoft		80.8
DeBERTaV3: Improving DeBERTa using ELECTRA-Style Pre-Training with Gradient-Disentangled Embedding Sharing
Jan 2023

Related Papers2

DeBERTaV3: Improving DeBERTa using ELECTRA-Style Pre-Training with Gradient-Disentangled Embedding Sharing
Jan 2023Models: mDeBERTa-v3-base
Paper
Unsupervised Cross-lingual Representation Learning at Scale
Nov 2019Models: XLM-RoBERTa-large
Paper