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

Researchers introduce R-APS (Reflective Adversarial Pareto Search), a novel method that enhances large language model reasoning for constrained design tasks by decomposing reasoning modes into separate contexts and orchestrating them across multiple timescales. The approach delivers 3.5x tighter robustness guarantees and 46% faster convergence on mechanical design problems without requiring model fine-tuning.

R-APS addresses a fundamental limitation in current LLM deployment: while these models excel at open-ended tasks, their performance degrades significantly in agentic settings requiring planning, tool use, and extended reasoning. The researchers identify three coupled failure modes—unlocalized error propagation, unevaluated worst-case perturbations, and knowledge invalidation—that stem from incompatible reasoning modes competing for shared context.

The solution employs reasoning-mode decomposition, allocating distinct contexts to abductive, counterfactual, meta-inductive, corrective, and inductive reasoning paths. This architecture operates on three timescales: immediate validation through typed critics, sensitivity-guided adversarial testing as a Pareto objective, and long-term knowledge refinement with explicit invalidation. Critically, R-APS requires no fine-tuning and functions purely through structured protocol design, making it broadly applicable.

Evaluation on planar mechanism synthesis—a demanding domain where candidates face kinematic solver verification—demonstrates substantial improvements. The method achieves 2.1x Chamfer-distance reduction over combined enumeration and genetic algorithm baselines while simultaneously controlling both design parameters and robustness constraints. A striking finding emerges: smaller 4B reasoning-specialized models perform competitively against 70B general-purpose models when operating within the structured protocol, suggesting protocol design can partially compensate for model scale limitations.

This work signals a paradigm shift in LLM deployment from fine-tuning toward architectural decomposition and constrained reasoning frameworks. The implications extend beyond robotics and mechanical design to any domain requiring robust, verifiable decision-making under uncertainty—from autonomous systems to financial modeling.

• →R-APS decomposes incompatible reasoning modes into separate contexts, eliminating a core source of LLM agentic failures.
• →The method achieves 3.5x tighter robustness certificates and 46% faster convergence without fine-tuning on frozen LLMs.
• →Smaller 4B models match 70B baseline performance within the protocol, suggesting structured design offsets model scale requirements.
• →Staged validation, adversarial stress-testing, and explicit knowledge invalidation enable reliable long-horizon planning.
• →Results on kinematic design verification demonstrate applicability to domains requiring third-party correctness checks.