MAPLE: A self-evolving agent that maintains executable optimization programs

简体中文

Persistent-state diagnostic: candidate admission, cross-task reuse and versioning. No mathematical solver or LLM code generation.

Paper information

Field	Details
Paper	arXiv 2609.11636
Institution	Harbin Institute of Technology, Shenzhen (first author).
First public release	2026-09-10 (arXiv v1)
Author code	MAPLE.
Local adapter or method key	maple
Local reproduction code	src/auto_research/agent_research/latest_20260912.py

Paper method and evidence

Represent optimization knowledge as editable program state. Propose changes, check executability, accept improvements, and carry the versioned program into later tasks rather than prompting from scratch.

flowchart LR
  P[Persistent program] --> C[Propose changes]
  T[New task and feedback] --> C
  C --> X[Check executability]
  X -->|Accept| V[Version new state]
  X -->|Reject| P
  V --> P

Original paper figure

The paper's Figure 7 shows examples from the benchmark. From the original paper; copyright remains with the authors.

Local reproduction

The local harness accepts only public plans with nonempty named operations and checks later reuse. Three-seed metrics. No original solver, training, or external tools; mini-suite success does not establish paper-scale capability.