Agentic Proposing: Enhancing Large Language Model Reasoning via Compositional Skill Synthesis
Agentic Proposing enhances LLM reasoning via compositional skill synthesis, achieving 91.6% accuracy on AIME25 with a 30B model.
Key Findings
Methodology
Agentic Proposing models problem synthesis as a goal-driven decision process, using Multi-Granularity Policy Optimization (MGPO) to dynamically select and compose modular reasoning skills. This framework, supported by an iterative workflow of internal reflection and tool-use, generates high-precision, verifiable training trajectories.
Key Results
- In mathematics, coding, and science, solvers trained on Agentic Proposing-synthesized data significantly outperform existing baselines. Notably, a 30B solver trained on only 11,000 synthesized trajectories achieves 91.6% accuracy on AIME25.
- A 4B solver trained on 10,000 synthesized trajectories consistently outperforms established reasoning collections across domains.
- Agentic Proposing demonstrates robust cross-domain generalization, proving that high-precision signals are more critical than model scale.
Significance
This research addresses the challenge of acquiring high-quality, verifiable datasets for LLM reasoning by introducing the Agentic Proposing framework. It demonstrates how a small volume of high-quality synthetic signals can effectively substitute massive human-curated datasets, advancing complex reasoning tasks in academia and industry.
Technical Contribution
The Agentic Proposing framework introduces the concept of compositional skills, providing a novel approach to synthesizing high-difficulty problems. Unlike existing methods, it does not rely on human-designed templates but explores reasoning frontiers through modular logic engineering.
Novelty
Agentic Proposing is the first to treat problem synthesis as a process of compositional logic engineering, dynamically composing modular skills to generate high-difficulty problems, offering significant innovation over traditional static generation methods.
Limitations
- In some cases, generated problems may still exhibit logical inconsistencies, although the framework minimizes such occurrences.
- The framework's performance depends on the quality and diversity of the skill library, which may require further expansion.
Future Work
Future research directions include expanding the skill library to cover more domains, optimizing the MGPO algorithm to improve problem quality, and exploring additional application scenarios.
AI Executive Summary
Enhancing reasoning capabilities in large language models (LLMs) has been challenging due to the difficulty of acquiring high-quality, verifiable datasets. Existing methods often face a trade-off between problem complexity and structural validity. The Agentic Proposing framework addresses this by modeling problem synthesis as a goal-driven decision process, dynamically selecting and composing modular reasoning skills.
The framework employs Multi-Granularity Policy Optimization (MGPO) and an iterative workflow of internal reflection and tool-use to generate high-precision, verifiable training trajectories. In mathematics, coding, and science, solvers trained on Agentic Proposing-synthesized data significantly outperform existing baselines. Notably, a 30B solver trained on only 11,000 synthesized trajectories achieves 91.6% accuracy on AIME25, demonstrating strong cross-domain generalization.
This research shows how a small volume of high-quality synthetic signals can effectively replace massive human-curated datasets, advancing complex reasoning tasks in academia and industry. Future research directions include expanding the skill library to cover more domains, optimizing MGPO for better problem quality, and exploring more application scenarios.
Deep Analysis
Background
The application of large language models (LLMs) in complex reasoning tasks is expanding, yet acquiring high-quality, verifiable datasets remains challenging. Existing methods like MetaMath and WizardMath often rely on human-designed templates, limiting flexibility and complexity in problem construction.
Core Problem
The core problem is generating structurally valid and verifiable high-difficulty problems without sacrificing complexity. Solving this is crucial for enhancing LLM reasoning capabilities.
Innovation
Agentic Proposing introduces compositional skills, treating problem synthesis as a process of compositional logic engineering. By dynamically selecting and composing modular reasoning skills, the framework generates high-difficulty problems without relying on fixed templates.
Methodology
- �� Use Multi-Granularity Policy Optimization (MGPO) to dynamically select and compose modular reasoning skills.
- �� Employ an iterative workflow of internal reflection and tool-use to generate high-precision, verifiable training trajectories.
- �� Introduce compositional skills, treating problem synthesis as a process of compositional logic engineering.
Experiments
The experimental design includes testing on multiple benchmark datasets in mathematics, coding, and science. A 30B solver is trained on 11,000 synthesized trajectories and compared against existing baselines.
Results
Results show that solvers trained on Agentic Proposing-synthesized data achieve 91.6% accuracy on AIME25, significantly outperforming existing baselines and demonstrating strong cross-domain generalization.
Applications
The framework can be used to generate high-quality training data, enhancing LLM performance in complex reasoning tasks, particularly in mathematics, coding, and science.
Limitations & Outlook
While the framework minimizes logical inconsistencies, generated problems may still exhibit such issues in some cases. The framework's performance depends on the quality and diversity of the skill library.
Plain Language Accessible to non-experts
Imagine a factory producing complex products. Traditional methods are like assembly lines, only capable of producing fixed products. Agentic Proposing is like a smart factory that can flexibly combine different parts based on demand to produce complex, high-quality products. This factory has an intelligent system that continuously adjusts based on feedback, ensuring each product is high-quality. This flexibility and efficiency allow it to excel in various fields.
ELI14 Explained like you're 14
Imagine you're playing a game that requires solving puzzles. Traditional gameplay is like using fixed strategies to solve puzzles, but Agentic Proposing is like a super helper that can flexibly combine different skills based on each level's characteristics to help you solve puzzles. This helper can also adjust strategies based on your feedback, ensuring you can smoothly pass levels. This flexibility and intelligence make it excel in various games!
Glossary
Agentic Proposing
A framework that models problem synthesis as a goal-driven decision process.
Used to dynamically select and compose modular reasoning skills.
Multi-Granularity Policy Optimization
An algorithm for optimizing agent policies, supporting high-precision problem generation.
Used in the Agentic Proposing framework for policy optimization.
Composable Agent Skills
The concept of decomposing problem construction logic into atomic reasoning modules.
Used for dynamic composition to generate complex problems.
AIME25
A benchmark dataset for mathematics competitions used to test model reasoning capabilities.
Used to evaluate the effectiveness of data generated by Agentic Proposing.
Verifiable Training Trajectories
High-precision, verifiable training data paths.
Training data generated by Agentic Proposing.
Open Questions Unanswered questions from this research
- 1 How to further expand the skill library to cover more domains? Current methods may underperform in some areas.
- 2 How to optimize MGPO to improve problem quality? Existing methods may be limited in extreme difficulty problems.
Applications
Immediate Applications
Education
Can be used to generate high-quality practice questions to help students improve mathematical and scientific reasoning skills.
Long-term Vision
AI Research
Advances AI in complex reasoning tasks, potentially transforming future AI research directions.
Abstract
Advancing complex reasoning in large language models relies on high-quality, verifiable datasets, yet human annotation remains cost-prohibitive and difficult to scale. Current synthesis paradigms often face a recurring trade-off: maintaining structural validity typically restricts problem complexity, while relaxing constraints to increase difficulty frequently leads to inconsistent or unsolvable instances. To address this, we propose Agentic Proposing, a framework that models problem synthesis as a goal-driven sequential decision process where a specialized agent dynamically selects and composes modular reasoning skills. Through an iterative workflow of internal reflection and tool-use, we develop the Agentic-Proposer-4B using Multi-Granularity Policy Optimization (MGPO) to generate high-precision, verifiable training trajectories across mathematics, coding, and science. Empirical results demonstrate that downstream solvers trained on agent-synthesized data significantly outperform leading baselines and exhibit robust cross-domain generalization. Notably, a 30B solver trained on only 11,000 synthesized trajectories achieves a state-of-the-art 91.6% accuracy on AIME25, rivaling frontier-scale proprietary models such as GPT-5 and proving that a small volume of high-quality synthetic signals can effectively substitute for massive human-curated datasets.