R-APS: New LLM framework slashes design errors 46% without fine-tuning

Large language models excel at open-ended tasks but fail in agentic settings requiring planning, tool use, and extended action horizons—due to three structural failures: error propagation without localization, no evaluation of worst-case perturbations, and no invalidation of accumulated knowledge. The authors trace this to conflicting demands on a shared context from different reasoning modes (abductive, counterfactual, meta-inductive, corrective, inductive). They propose Reflective Adversarial Pareto Search (R-APS), the first method addressing all three failures jointly by decomposing reasoning into dedicated contexts and orchestrating interaction across three timescales: staged compositional reasoning with a typed validation critic (failure localization), sensitivity-guided counterfactual stress-testing as a first-class Pareto objective (robustness), and meta-inductive rule extraction with explicit invalidation (persistent memory). R-APS requires no fine-tuning and operates purely via structured protocol design on a frozen LLM, offering a practical path to reliable LLM agents without expensive retraining.

Evaluated on planar mechanism synthesis (robotics, prosthetics, mechanical design)—with every candidate checked by a kinematic solver across 32 target trajectories—R-APS delivers robustness certificates 3.5x tighter than uniform-perturbation baselines, 46% faster iterations-to-first-admission, and 2.1x reduction in Chamfer distance over Enum+GA while jointly controlling bar count and worst-case robustness. Remarkably, small 4B reasoning-specialized models proved competitive with general-purpose 70B backbones inside the protocol, suggesting that structured protocols can partially offset model scale. The results indicate that reasoning decomposition and guided interaction are more critical than raw parameter count for complex constrained design tasks, opening doors to cheaper, more reliable LLM-based engineering automation.