DART-ES: Difficulty-Aware Reweighting and Targeted Replay for Fine-Tuning LLMs with Evolution Strategies

Excerpt

arXiv:2610.06993v1 Announce Type: new Abstract: Evolution Strategies (ES) enable memory efficient full parameter fine-tuning of large language models (LLMs) using only forward computation. However, standard ES uniformly averages rewards across problems and compresses problem level population feedback into a single scalar, making it difficult to capture how the learning value of each problem changes with model capability. To address this limitation, we propose Difficulty-Aware Reweighting and Tar