Research & Papers

PROPEL: New AI method generates training tasks at the learnable frontier

Solves the bottleneck of running out of challenging tasks for advanced AI models.

Deep Dive

As AI reasoning models improve, they quickly saturate existing task distributions, creating a bottleneck in training. Naively generating new tasks often yields trivial, impossible, or ill-posed examples. Researchers from multiple institutions propose PROPEL (solver-amortized framework for training task generators at the targeted solve rate). The key innovation is a lightweight activation probe trained on a one-time corpus of generated tasks and solver outcomes. This probe predicts the pass rate of a target solver from a frozen generator reference model, serving as a proxy during generator optimization. This eliminates the need for repeated, time-consuming solver rollouts (e.g., tens of minutes per candidate for software engineering tasks).

PROPEL was tested on math, code, and software engineering benchmarks at multiple model scales. For coding tasks with Qwen2.5-3B-Instruct, the share of generations at the learnable frontier increased from 10.1% to 20.0%; for Qwen2.5-7B-Instruct, from 5.3% to 12.6%. In software engineering, using Qwen3.5-27B on unseen repositories, the proportion rose from 9.8% to 19.6%. These results demonstrate that PROPEL effectively shifts generation toward the optimal difficulty, enabling continuous training without manual task creation.

Key Points
  • PROPEL uses a lightweight activation probe as a proxy for solver pass rate, avoiding expensive per-candidate rollouts
  • In coding tasks, the proportion of tasks at the learnable frontier doubled (10.1% → 20.0% for Qwen2.5-3B-Instruct)
  • For software engineering, PROPEL increased targeted-difficulty tasks from 9.8% to 19.6% on unseen repositories

Why It Matters

Enables AI systems to autonomously generate training data at the optimal difficulty, accelerating progress without manual task crafting.

📬 Get the top 10 AI stories daily