List your company

Reasoning on Reasoning Gym Spell Backward

97Success RateCBRL GRPO
0.2825.3950.575.61Mar 19, 2026

Evaluation results

MethodLinks
CBRL GRPOBackbone Model=Llama-3.2-3B-Instruct2026.0397
GRPOBackbone Model=Llama-3.2-3B-Instruct2026.0395.67
CBRL GRPOBackbone Model=Qwen2.5-3B-Instruct2026.0352.67
GRPOBackbone Model=Qwen2.5-3B-Instruct2026.0351
Baseline w/ Few ShotsBackbone Model=Llama-3.2-3B-Instruct2026.0323.67
BaselineBackbone Model=Llama-3.2-3B-Instruct2026.0320
BaselineBackbone Model=Qwen2.5-3B-Instruct2026.034.33
Baseline w/ Few ShotsBackbone Model=Qwen2.5-3B-Instruct2026.034