Internalizing Geometric Law: Learning from Solver Residuals for Precision-Critical Generation
Researchers have developed a new method called Saturating Additive Rewards (SAR) to improve the precision of large language models in geometric tasks. This approach addresses a failure mode known as Outlier Gradient Masking, where a single constraint violation can hinder learning across all constraints. SAR decomposes rewards into bounded per-constraint terms, preserving partial progress and ensuring consistent gradients. An 8B parameter model using SAR achieved a 2.3x improvement in solving complex geometric problems compared to standard MSE-based rewards. AI
IMPACT Enhances LLM capabilities in precision-critical domains, potentially enabling more reliable AI-driven design and technical diagramming.