PolicySmith

This repository contains the code for PolicySmith, a system that discovers heuristics (i.e., code) for systems using LLMs and evolutionary search. This README will use the example of caching to demonstrate how to use PolicySmith; this codebase is extensible to support your task(s)

Setup

• Clone dependencies: git submodule update --init
• Install MongoDB on your machine.
• Install Python libraries: pip3 install google-genai openai pymongo.
• Create a file called api_key.py and populate it with the following:

GEMINI_API_KEY="<<secret_key_here>>"
OPENAI_API_KEY="<<secret_key_here>>"
MONGO_CONNECTION_STRING="mongodb://localhost:27017"

Browse results from our HotNets 2025 paper

• Go through ./hotnets_results to explore the caching results described in the HotNets paper.
• You can view the C++ source of the heuristics discovered (PS-A, PS-B, PS-X, etc.), reproduce the boxplots in the paper.

Running policysmith

• Follow the instructions in webcache/README.md to run PolicySmith for caching. Well documented.
• [Experimental] tcp_cc/bpf_scaffolding contains the setup to evolve TCP congestion control policies in the Linux kernel with PolicySmith. Experimental / not very well documented yet.

Extending PolicySmith

To use PolicySmith for your task:

• create a new directory called (e.g. ./cpu_scheduling).
• Inside this directory, create a file called ./cpu_scheduling/interface.py that subclasses EvolveInterface abstract class from ./Evolve.py.
• Look at the interface of webcache for more details / inspiration on what this should look like.
• Profit?

Reference

@article{dwivedula2025policysmith,
      title={Man-Made Heuristics Are Dead. Long Live Code Generators!}, 
      author={Dwivedula, Rohit and Saxena, Divyanshu and Akella, Aditya and Chaudhuri, Swarat and Kim, Daehyeok},
      year={2025},
      eprint={2510.08803},
      archivePrefix={arXiv},
      primaryClass={cs.OS},
      url={https://arxiv.org/abs/2510.08803}, 
      note = {arXiv:2510.08803 [cs.OS]}
}