AlgoPilot: Cross-Paradigm Reasoning in Language Models via Strategy Selection and Guidance

| Source: arXiv AI

Tags: AlgoPilot, Qwen3-8B, GLM-5, Reinforcement Learning, Strategy Selection, Program Synthesis

AlgoPilot trains small models to pick a solving paradigm and structure the problem before execution, lifting Qwen3-8B's average accuracy from 38.4% to 72.9% across 13 benchmarks and beating GLM-5 prompting by 11.7 points.

Details

Real problems span program synthesis, symbolic reasoning, planning and optimization, each needing a different solving approach. AlgoPilot learns to choose and formulate that approach instead of relying on fixed prompt templates. The framework has three parts: a SteerLM trained with SFT and RL that selects the paradigm and mode, a GuideLM that produces structured formulation guidance for that strategy, and an ExecutionLM that solves the problem conditioned on both. Built on Qwen3-8B and tested on 5 problem categories, 13 benchmarks and 66 domains, it raised average accuracy from 38.4% to 72.9% and outperformed the best large-LLM prompting baseline, GLM-5, by 11.7%. The authors also report that the steering and formulation modules transfer to other execution models.