Breaking the Solver Bottleneck: Training Task Generators at the Learnable Frontier

Short summary

PROPEL is a framework that trains lightweight probes to predict task solvability, eliminating expensive solver rollouts during task generator training. Results show significant improvements: coding tasks at the learnable frontier increased from 10.1% to 20.0% for smaller models and 5.3% to 12.6% for larger ones. The method addresses the training data saturation problem as AI models improve.

• •PROPEL uses activation probes to predict solver pass rates, replacing costly solver-in-the-loop training
• •Generates training tasks at the learnable frontier (optimal difficulty) with 2x improvement in success rate
• •Tested across math, code, and SWE domains with consistent gains across model scales

Generated with AI, which can make mistakes.