List your company

Reasoning on ARC, StraQA, CSQA Overall

69.85AccuracyHSIR-SFT
62.195664.182866.1768.1572May 24, 2026

Evaluation results

MethodLinks
HSIR-SFTBackbone=Qwen2.5-1.5B, Training Iterations=12026.0569.851,043
SFT-OracleBackbone=Qwen2.5-1.5B, Training Iterations=12026.0568.141,333
ReSTEMBackbone=Qwen2.5-1.5B, Training Iterations=12026.0566.741,254
STaRBackbone=Qwen2.5-1.5B, Training Iterations=12026.0564.21,357
SFT-InitialBackbone=Qwen2.5-1.5B2026.0562.491,447