Table 3 Zero-shot classification accuracy on BERT-based models with and without DR-CoT on GPQA Diamond.

From: DR-CoT: dynamic recursive chain of thought with meta reasoning for parameter efficient models

Model

Baseline\(\uparrow\)

Baseline + DR-CoT\(\uparrow\)

BERT-base30

21.2

23.7 (+2.5)

BERT-large

21.2

26.3 (+5.1)

ROBERTa-base31

24.7

25.8 (+1.1)

ROBERTa-large

27.3

28.5 (+1.2)

ELECTRA-base32

26.2

29.3 (+3.1)

ELECTRA-large

26.3

28.5 (+2.2)

ModernBERT-base33

23.7

25.0 (+1.3)

ModernBERT-large

29.8

32.9 (+3.1)

Back to article page