Dimensions overall score 4.0

• The geometric differences between ES and GRPO solutions have important consequences for forgetting and knowledge preservation.technical 80%
• Gradient-free and gradient-based fine-tuning can reach similarly accurate yet geometrically distinct solutions.result 90%
• Evolution Strategies (ES) match or exceed Group Relative Policy Optimization (GRPO) in single-task accuracy across four tasks.result 90%
• ES remains competitive with GRPO in sequential continual-learning settings when its iteration budget is controlled.result 85%
• ES produces much larger parameter updates than GRPO despite similar task performance.result 90%
• ES induces broader off-task KL drift compared to GRPO.technical 90%
• ES and GRPO solutions are linearly connected with no loss barrier despite their update directions being nearly orthogonal.technical 90%
• ES can accumulate large off-task movement on weakly informative directions while still making enough progress on the task to match gradient-based RL in downstream accuracy.technical 85%

Competitive landscape

Competitor map is still being generated for this paper. Enable generation or check back soon.

Startup potential card