Compute Efficiency

The 36% FLOPs reduction and 30–40% wall-clock savings are achieved without accuracy loss and often with accuracy improvements, indicating uniform prompt allocation is a substantial inefficiency. With the backward selector only, Qwen2.5-Math-1.5B reaches better performance in 2.40 hours versus 3.97 hours baseline, a 1.65× wall-clock speedup. Adding the forward pruner to the backward selector increases end-to-end wall-clock speedup to 1.80× for 1.5B and 1.91× for 7B models, at a modest accuracy trade-off. LZE restricts backpropagation to 40% of prompt groups, yielding a 36% theoretical reduction in FLOPs per token, consistent across all configurations. The LZE framework is optimizer-agnostic, operating solely on group-level pass-rate statistics and requiring no modification to the underlying GRPO or DAPO optimizer.