arxivcs.LGcs.AIcs.PL2026-07-03
Differentiate the Evaluator, Not the Program: An Efficient Runtime Representation for Neuro-Symbolic Learning
AI systems increasingly propose executable scientific models whose value depends on both their symbolic structure and their fitted continuous parameters. This makes parameter calibration the bottleneck of program-and-parameter co-search: an outer loop can generate thousands of ca…