AI Paper Atlas

Research map / Multi-agent systems and games / LLM agents

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

Haoyang Liu, Yuyang Cai, Jie Wang, Xiongwei Han, Minyang Hu, Shuqi LIU, Mingxuan Yuan, Jianye HAO, Feng Wu

ICML 2026 · Multi-agent systems and games › LLM agents · approach: Agent data synthesis and behavior studies

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 Opt-Miner, an information-seeking agent for optimization modeling that uses tree-guided data synthesis to address the knowledge-intensive nature of real-world problems.

Problem

LLM agents for automated optimization modeling struggle with knowledge-intensive real-world problems because static parametric knowledge lacks the domain expertise needed to understand scenarios and apply appropriate mathematical techniques.

Approach

Equips an LLM agent with information-seeking capability (retrieving external domain knowledge) and trains it on tree-guided synthesized data, rather than relying only on parametric knowledge.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add knowledge fogging operator to the synthesis data (ablation)Opt-Miner (Qwen3-4B-Instruct) · IndustryORPass@1 accuracy (%): 56.3 -> 66.4Improved
add knowledge fogging operator to the synthesis data (ablation)Opt-Miner (Qwen3-4B-Instruct) · Opt-Miner-BenchPass@1 accuracy (%): 14.8 -> 21.1Improved
replace 1k extra OptMATH data with 1k union-synthesized Opt-Miner data (ablation)Qwen3-4B-Instruct trained on OptMATH-1k · Opt-Miner-BenchPass@1 accuracy (%): 27.1 -> 34.3Improved
add 1k transfer-synthesized data to OptMATH-1k (ablation)Qwen3-4B-Instruct trained on OptMATH-1k · IndustryORPass@1 accuracy (%): 52.1 -> 62.6Improved
use full Opt-Miner mixed data instead of OptMATH-2k (ablation)Qwen3-4B-Instruct · IndustryORPass@1 accuracy (%): 54.5 -> 66.4Improved
add retrieval matching reward to GRPO (R-GRPO) (ablation)Qwen3-4B-Instruct · IndustryORPass@1 accuracy (%): 57.3 -> 66.4Improved
add retrieval matching reward to SIRL (R-SIRL) (ablation)Qwen3-4B-Instruct · OptMATHPass@1 accuracy (%): 57.8 -> 62.8Improved
add retrieval matching reward to SIRL (R-SIRL) (ablation)Qwen3-4B-Instruct · MAMO ComplexLPPass@1 accuracy (%): 91.3 -> 91.9Mixed
add web research module to base model (ablation)Qwen3-4B-Instruct · IndustryORPass@1 accuracy (%): 15.2 -> 20.9Improved
add web research module to base model (ablation)Qwen3-4B-Instruct · OptMATHPass@1 accuracy (%): 40.8 -> 40.5No clear effect
add web research module to base model (ablation)DeepSeek-V3 · Opt-Miner-BenchPass@1 accuracy (%): 23.4 -> 33.1Improved
add web research module to trained agent (ablation)Opt-Miner (Qwen3-4B-Instruct) · Opt-Miner-BenchPass@1 accuracy (%): 28.3 -> 35.5Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy