AlgoPilot: Cross-Paradigm Reasoning in Language Models via Strategy Selection and Guidance
Summary
- Researchers proposed AlgoPilot, a cross-paradigm reasoning framework that enables adaptive selection of solving approaches and structured problem formulation for language models.
- AlgoPilot consists of a SteerLM for strategy selection, an expert GuideLM for structured guidance generation, and an ExecutionLM for final solution generation.
- Built on Qwen3-8B and evaluated across 13 benchmarks, AlgoPilot raised average accuracy from 38.4% to 72.9%, outperforming the GLM-5 large LLM baseline.
- Ablation studies confirmed the effectiveness of each component, and experiments showed that steering and formulation modules can transfer across execution models.
Why it matters
AlgoPilot demonstrates that combining adaptive strategy selection with structured guidance significantly improves small language model reasoning capabilities across diverse domains, outperforming larger prompting baselines.