Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

OR-Agent: Bridging Evolutionary Search and Structured Research for Automated Algorithm Discovery

Qi Liu, Wanjing Ma

TL;DR

This work introduces an evolutionary-systematic ideation mechanism that unifies evolutionary selection of research starting points, comprehensive research plan generation, and coordinated exploration within a research tree, and proposes a hierarchical optimization-inspired reflection system.

Abstract

Automating scientific discovery in complex, experiment-driven domains requires more than iterative mutation of programs; it demands structured hypothesis management, environment interaction, and principled reflection. We present OR-Agent, a configurable multi-agent research framework designed for automated exploration in rich experimental environments. OR-Agent organizes research as a structured tree-based workflow that explicitly models branching hypothesis generation and systematic backtracking, enabling controlled management of research trajectories beyond simple mutation-crossover loops. At its core, we introduce an evolutionary-systematic ideation mechanism that unifies evolutionary selection of research starting points, comprehensive research plan generation, and coordinated exploration within a research tree. We further propose a hierarchical optimization-inspired reflection system: short-term experimental reflection operates as a form of verbal gradient providing immediate corrective signals; long-term reflection accumulates cross-experiment insights as verbal momentum; and memory compression serves as a regularization mechanism analogous to weight decay, preserving essential signals while mitigating drift. Together, these components form a principled architecture governing research dynamics. We conduct extensive experiments across classical combinatorial optimization benchmarks-including traveling salesman, capacitated vehicle routing, bin packing, orienteering, and multiple knapsack problems-as well as simulation-based cooperative driving scenarios. Results demonstrate that OR-Agent outperforms strong evolutionary baselines while providing a general, extensible, and inspectable framework for AI-assisted scientific discovery. OR-Agent source code and experiments data are publicly available at https://github.com/qiliuchn/OR-Agent.

OR-Agent: Bridging Evolutionary Search and Structured Research for Automated Algorithm Discovery

TL;DR

This work introduces an evolutionary-systematic ideation mechanism that unifies evolutionary selection of research starting points, comprehensive research plan generation, and coordinated exploration within a research tree, and proposes a hierarchical optimization-inspired reflection system.

Abstract

Automating scientific discovery in complex, experiment-driven domains requires more than iterative mutation of programs; it demands structured hypothesis management, environment interaction, and principled reflection. We present OR-Agent, a configurable multi-agent research framework designed for automated exploration in rich experimental environments. OR-Agent organizes research as a structured tree-based workflow that explicitly models branching hypothesis generation and systematic backtracking, enabling controlled management of research trajectories beyond simple mutation-crossover loops. At its core, we introduce an evolutionary-systematic ideation mechanism that unifies evolutionary selection of research starting points, comprehensive research plan generation, and coordinated exploration within a research tree. We further propose a hierarchical optimization-inspired reflection system: short-term experimental reflection operates as a form of verbal gradient providing immediate corrective signals; long-term reflection accumulates cross-experiment insights as verbal momentum; and memory compression serves as a regularization mechanism analogous to weight decay, preserving essential signals while mitigating drift. Together, these components form a principled architecture governing research dynamics. We conduct extensive experiments across classical combinatorial optimization benchmarks-including traveling salesman, capacitated vehicle routing, bin packing, orienteering, and multiple knapsack problems-as well as simulation-based cooperative driving scenarios. Results demonstrate that OR-Agent outperforms strong evolutionary baselines while providing a general, extensible, and inspectable framework for AI-assisted scientific discovery. OR-Agent source code and experiments data are publicly available at https://github.com/qiliuchn/OR-Agent.
Paper Structure (107 sections, 8 equations, 18 figures, 4 tables)

This paper contains 107 sections, 8 equations, 18 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Open Research Agent
  3. OR-Agent Framework Overview
  4. Problem Specification and Open Research Canvas.
  5. Evolutionary Initialization and Deep Investigation.
  6. Multi-Agent Architecture.
  7. Tree-Structured Research Workflow and Refinement Loops.
  8. Motivation: Breadth, Depth, and Context Management.
  9. Summary of Framework Design.
  10. Task Specification
  11. Open Research Canvas (OR-Canvas).
  12. Solution Interface.
  13. Evaluation Protocol.
  14. Environment Interaction and Callbacks.
  15. Generality and Reusability.
  16. ...and 92 more sections

Figures (18)

  • Figure 1: Overall framework of OR-Agent, illustrating the interaction between evolutionary initialization, multi-agent research workflows, experimentation, reflection, and the shared solution database.
  • Figure 2: User Interface of OR-Canvas and OR-Agent.
  • Figure 3: Illustration of the population ruin phenomenon. A single dominant solution rapidly takes over the population, and subsequent mutations around this dominant mode produce invalid solutions (e.g., due to timeouts), leading to collapse of population quality.
  • Figure 4: Use tree shape as a control mechanism for research modes illustration.
  • Figure 5: Research tree management process illustration.
  • ...and 13 more figures