R-APS: Compositional Reasoning and In-Context Meta-Learning for Constrained Design via Reflective Adversarial Pareto Search

In three linesR-APS improves LLM reliability in agentic settings via reasoning-mode decomposition. Tested on planar mechanism synthesis, it delivers robustness certificates 3.5× tighter than baselines, 46% faster iterations-to-first-admission, and 2.1× Chamfer-distance reduction. No fine-tuning required; operates via structured protocol on frozen LLM.

Summary generated by Claude — human-verified