List your company

Logical Reasoning on ProsQA

98.5AccuracyTHOUGHTSTEER
58.87669.16379.4589.737Feb 14, 2026Feb 21, 2026Mar 1, 2026Mar 9, 2026Mar 16, 2026Mar 24, 2026Apr 1, 2026

Evaluation results

MethodLinks
THOUGHTSTEERArchitecture=COCONUT (GPT-2 124M), Poison rate (rho)=15%2026.0498.5——99.5100
OneLatentNumber of latent tokens=12026.0297.811.18.81——
COCONUTNumber of latent tokens=62026.0289.26.613.5——
iCoTNumber of latent tokens=02026.0287.412.86.82——
CoTNumber of latent tokens=02026.0276.22.2933.3——
No CoTNumber of latent tokens=02026.0260.47.068.55——