EliteReplay: Preserving Diverse Solutions in RL Post-Training via Quality-Diversity Search
EliteReplay produced up to 77% more distinct correct strategies than the strongest baselines on text-to-SQL tasks.
The method collects solutions spanning different approaches from a frozen base model, then replays that archive alongside GRPO training. The authors report improved accuracy while retaining most of the base model’s correct strategies, with higher pass@128 than all tested baselines. On ambiguous questions in AmbiQT, it covered both interpretations nearly twice as often as the strongest baseline.
ArXiv · AI/CL/LG's note
The method collects solutions spanning different approaches from a frozen base model, then replays that archive alongside GRPO training. The authors report improved accuracy while retaining most of the base model’s correct strategies, with higher pass@128 than all tested baselines. On ambiguous questions in AmbiQT, it covered both interpretations nearly twice as often as the strongest baseline.
ArXiv · AI/CL/LG's note
score 5