Research map / Large language models / Reasoning and chain of thought
AlgoPilot: Cross-Paradigm Reasoning in Language Models via Strategy Selection and Guidance
Yilun Hao, Ruixiao Yang, Hudson Hilal, Yongchao Chen, Kaizhi Qian, Yada Zhu, Yang Zhang, Chuchu Fan
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes AlgoPilot, a cross-paradigm reasoning framework that adaptively selects a problem-solving strategy and guides LLM solution formulation.
Problem
LLM reasoning across diverse domains (program synthesis, symbolic reasoning, planning, optimization) requires both identifying an appropriate solution paradigm and formulating it correctly.
Approach
A framework that adaptively selects among different solution paradigms and guides the LLM in formulating the chosen approach, rather than using a single fixed reasoning style.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add SteerLM+GuideLM cross-paradigm pipeline on top of Qwen3-8B | Qwen3-8B · cross-paradigm test set | average accuracy: 38.4 -> 72.9 | Improved |
| refine SteerLM with GRPO after SFT (ablation) | AlgoPilot SteerLM · cross-paradigm test set | average accuracy: 67.2 -> 72.9 | Improved |
| use learned SteerLM instead of self-selected strategy (with GuideLM) (ablation) | AlgoPilot · cross-paradigm test set | average accuracy: 45.3 -> 72.9 | Improved |
| add GuideLM formulation guidance (with SteerLM) (ablation) | AlgoPilot · cross-paradigm test set | average accuracy: 45.7 -> 72.9 | Improved |
| add GuideLM guidance with self-selected strategy (ablation) | Qwen3-8B · cross-paradigm test set | average accuracy: 36.1 -> 45.3 | Improved |
| prompt Qwen3-8B to reason about approach first (Reason vs Direct) | Qwen3-8B · cross-paradigm test set | average accuracy: 38.4 -> 36.1 | Hurt |
| replace GuideLM with QA-SFT fine-tuned executor on same problems (ablation) | AlgoPilot · cross-paradigm test set | average accuracy: 72.9 -> 59.3 | Hurt |
| evaluate on paraphrased inputs | AlgoPilot · cross-paradigm test set | average accuracy: 72.9 -> 68.6 | Hurt |
| transfer trained SteerLM and GuideLM to Gemma-4-E4B-it executor | Gemma-4-E4B-it · cross-paradigm test set | average accuracy: 50.8 -> 60.5 | Improved |
| add weighted lower-ranked candidate rewards S2, S3 to SteerLM reward (ablation) | SteerLM GRPO · SteerLM reward design | Top-3 accuracy: 77.13 -> 83.67 | Improved |
| add structural format and duplicate penalties to reward (ablation) | SteerLM GRPO · SteerLM reward design | Valid-3: 94.8 -> 100.0 | Improved |
| add ranking bonus to reward (ablation) | SteerLM GRPO · SteerLM reward design | Valid-3: 84.7 -> 94.8 | Improved |
Closest papers by meaning
- StrategyLLM: Large Language Models as Strategy Generators, Executors, Optimizers, and Evaluators for Problem Solving (NeurIPS 2024 · similarity 0.90)
- GPO: Learning from Critical Steps to Improve LLM Reasoning (NeurIPS 2025 · similarity 0.89)
- Rectifying LLM Thought from Lens of Optimization (ICLR 2026 · similarity 0.89)
- Policy Guided Tree Search for Enhanced LLM Reasoning (ICML 2025 · similarity 0.89)
- SolverLLM: Leveraging Test-Time Scaling for Optimization Problem via LLM-Guided Search (NeurIPS 2025 · similarity 0.88)
- DOTS: Learning to Reason Dynamically in LLMs via Optimal Reasoning Trajectories Search (ICLR 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
