Skip to main content
Back

Harbin's EvoOR-Agent Evolves Its Own Reasoning to Crack Hard Optimization Problems

EvoOR-Agent treats LLM agent workflows as evolvable graphs, letting the architecture itself adapt to each optimization problem instead of following a fixed pipeline.

Pro

Paper, Full Access
  • EvoOR-Agent co-evolves both the agent's workflow graph structure and its reasoning trajectories simultaneously, unlike prior systems that fix one or the other.
  • Workflows are represented as Activity-on-Edge (AOE) graphs, making execution dependencies and alternative reasoning paths explicit and machine-searchable.
  • Three evolutionary operators -- path-conditioned recombination, multi-granularity semantic mutation, and elitist selection -- drive the search over reasoning trajectories.
  • A knowledge base of reusable OR practices seeds initialization and mutation, giving the evolutionary search a head start over random exploration.
  • Results on heterogeneous OR benchmarks show consistent gains over zero-shot LLMs, fixed-pipeline agents, and other evolutionary agent frameworks.
  • No public code release yet; the evolutionary search requires many LLM calls per problem, making it compute-intensive for production use. Read the paper.

Getting a large language model to reliably solve a real operations research (OR) problem -- think supply chain scheduling, resource allocation, or production planning -- is harder than it looks. The model has to understand the problem, write a mathematical formulation, pick a solver, generate working code, and debug it when things break. Most current approaches hand-craft a fixed sequence of steps for the LLM to follow. That works fine for simple cases, but the moment the problem gets complex or unusual, the rigid pipeline falls apart.

A new paper from researchers at Harbin Institute of Technology introduces EvoOR-Agent, a framework that sidesteps this fragility by treating the agent's workflow itself as something that can evolve. Rather than hard-coding the reasoning pipeline, the system builds a graph of possible reasoning steps and then searches that graph to find the best path for each problem.

Architecture diagram comparing classical OR pipeline with LLM-driven pipeline for solving optimization problems

Why fixed pipelines keep failing

The broader challenge here is well-established. Operations research has long played a central role in industrial planning and scheduling, supporting critical decision-making in manufacturing, supply chains, logistics, and energy systems, with optimization models typically expressed as mixed-integer linear programs (MILPs). In practice, business requirements are expressed in natural language, while optimization solvers demand strictly formal mathematical representations -- bridging that gap requires substantial domain expertise and manual effort, making the modeling process time-consuming, error-prone, and difficult to scale.

The LLM-for-OR field has been racing to close this gap. Prompt-based methods use existing LLM interfaces to automate optimization problem solving; frameworks like Chain-of-Expert assign specific tasks to LLM agents at each stage of reasoning, while OptiMUS employs tailored prompts for linear and mixed-integer programming. But all of these approaches share the same fundamental weakness: the workflow is static. The agent always follows the same sequence of steps regardless of what the problem actually demands.