Opt-Miner: Empowering Information-Seeking Agent with Tree-Guided Data Synthesis for Optimization Modeling

Large Language Model (LLM) agents have shown significant potential in automated optimization modeling for mathematical problems. However, real-world problems are still challenging due to their knowledge-intensive nature. Existing methods, constrained by static parametric knowledge, often lack the domain expertise required to comprehend complex scenarios and apply appropriate mathematical techniques, leading to errors. To address this challenge, we propose the Opt-Miner framework, where the agent learns to identify missing knowledge, retrieve technical documents on the web, and ground its mathematical models for improved modeling performance. The core of Opt-Miner is a novel tree-guided data synthesis pipeline coupled with a retrieval-based group relative policy optimization (R-GRPO) algorithm, designed to foster the agent’s information-seeking capabilities. Specifically, we first formulate each problem into a tree structure, with its scenario contexts and mathematical techniques embedded in subtrees. We then employ subtree union, transfer, and knowledge fogging to synthesize complex, multi-domain problems that incorporate knowledge gaps, thereby necessitating active information seeking to solve these problems. Based on synthesized data, we propose R-GRPO for agent reinforcement learning. Experiments demonstrate that Opt-Miner-Qwen3-8B achieves performance comparable to 32B state-of-the-art specialized agents and commercial reasoning models.

Many real-world planning problems—such as scheduling factory work, routing deliveries, or managing energy—require turning messy descriptions into precise plans that computers can check and improve. Today’s AI systems can help with this, but they often make mistakes when a task depends on specialized background knowledge that is not fully explained in the problem statement. This paper introduces Opt-Miner, an AI assistant designed to recognize when it is missing important knowledge, search for relevant information, and use what it finds to build better plans. To teach this behavior, we create challenging training problems by combining and reshaping real planning scenarios, and by leaving some information implicit, as often happens in real workplaces. The AI is rewarded not only for producing the right final answer, but also for finding useful information along the way. Across several test sets, Opt-Miner solves these planning problems more accurately than similar-sized systems, and its larger version performs competitively with much bigger specialized and commercial models. These results suggest that AI systems can become more reliable industrial planning assistants when they are trained to actively seek and use outside knowledge.