Guiding LLM Post-training Data Engineering with Model Internals from Sparse Autoencoders

Benchmark Model Rank Results
arithmetic-reasoning-on-gsm8kQwen2.5-Math-7B (SaeRL, GRPO)–Accuracy: 91.5
mathematical-reasoning-on-math-500Qwen2.5-Math-7B (SaeRL, GRPO)–Accuracy: 80.3
mathematical-reasoning-on-minerva-mathQwen2.5-Math-7B (SaeRL, GRPO)–Accuracy: 35.4