Research map / Large language models / Reasoning and chain of thought
MAGO: Beyond Fixed Hyperparameters with Multi-Objective Pareto Optimization for Hybrid LLM Reasoning
Hongcheng Ding, Xuanze Zhao, Ruiting Deng, Shamsul Nahar Abdullah, Deshinta Arrova Dewi, LIU QINGYU
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 MAGO, a multi-objective Pareto optimization approach for hybrid LLM reasoning that decides adaptively between direct answering and chain-of-thought reasoning instead of using fixed hyperparameters.
Problem
Uniformly applying elaborate chain-of-thought reasoning to all queries wastes computation, since many problems can be solved directly; existing hybrid reasoning methods rely on static fixed hyperparameters for switching between reasoning modes.
Approach
Frames hybrid reasoning mode selection as a multi-objective problem solved via Pareto optimization, replacing fixed hyperparameters with a trade-off between accuracy and computational cost.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace fixed weight alpha with Pareto adaptive multi-objective weighting (MAGO) | DeGRPO hybrid reasoning, 1.5B · AIME 2024 | Pass@1: 0.2506 -> 0.2741 | Improved |
| replace fixed weight alpha with MAGO | DeGRPO, 1.5B · MATH-500 | Pass@1: 0.8037 -> 0.8247 | Improved |
| replace fixed weight alpha with MAGO | DeGRPO, 1.5B · Minerva Algebra | Pass@1: 0.9216 -> 0.9483 | Improved |
| use MAGO hybrid model instead of full long-CoT base model | DeepSeek-R1-1.5B · AIME 2024 | tokens per query: 18063 -> 7164 | Improved |
| use MAGO hybrid model instead of full long-CoT base model | DeepSeek-R1-1.5B · MATH-500 | Pass@1: 0.8608 -> 0.8247 | Hurt |
| use MAGO instead of QMath-1.5B on GSM8K | Qwen-1.5B · GSM8K | Pass@1: 0.8572 -> 0.8469 | No clear effect |
| apply MAGO trained on math to CommonsenseQA | DeGRPO · CommonsenseQA | accuracy: 73.1 -> 74.9 | Improved |
| scale MAGO backbone from 1.5B to 32B | MAGO · AIME 2024 | Pass@1: 0.2741 -> 0.3254 | Improved |
| use Pareto-adaptive weights instead of vanilla GRPO (ablation) | hybrid reasoning RL · training dynamics | think-mode share | Improved |
| use large fixed alpha (0.01) or small alpha (0.0001) (ablation) | static-weight DeGRPO · mode selection | mode balance: alpha=0.0001: >90% short; alpha=0.01: >80% think | Hurt |
Experimental setup
Builds on: chain-of-thought
Closest papers by meaning
- GPO: Learning from Critical Steps to Improve LLM Reasoning (NeurIPS 2025 · similarity 0.92)
- Rectifying LLM Thought from Lens of Optimization (ICLR 2026 · similarity 0.88)
- Large Language Models to Enhance Bayesian Optimization (ICLR 2024 · similarity 0.88)
- SolverLLM: Leveraging Test-Time Scaling for Optimization Problem via LLM-Guided Search (NeurIPS 2025 · similarity 0.88)
- THOR: Tool-Integrated Hierarchical Optimization via RL for Mathematical Reasoning (ICLR 2026 · similarity 0.88)
- Towards Pareto-Optimal Tool-Integrated Agents with Pareto Ranking Policy Optimization (ICML 2026 · 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.
