A new method lets language models suggest rules for building logic programs, instead of grinding through every possible combination by hand.
Researchers describe a framework called GRASP (Gradient-boosted Synthesis of Probabilistic logic programs) that treats the problem of learning probabilistic logic programs as a boosting problem. Each "weak learner" is a single first-order logic rule. Rather than searching the full symbolic space to find candidate rules, GRASP hands that search off to an LLM, which proposes rule hypotheses that get folded into a weighted ensemble through gradient boosting. The team tested it on four relational benchmarks, including molecular toxicity prediction, mutagenesis, and citation matching, and found it beat purely symbolic systems, purely neural ones, and LLM-only approaches on all of them.
This matters because the pitch for logic programs has always been interpretability: you can read the rules and see why the system made a call, unlike a neural network's black box. The catch has been that finding good rules from data is combinatorially brutal. Using an LLM as a hypothesis generator inside a boosting loop, rather than as the whole show, is a sensible division of labor: let the model guess, let the math decide what sticks.
It is also a small data point in a broader trend of using LLMs as search heuristics for symbolic systems rather than replacements for them. Whether GRASP generalizes past these four benchmarks is the open question a conference reviewer will ask first.