Routing Preference

Research project studying what characteristics users value in AI-generated code solutions and how much they're willing to pay for them.

Setup

git clone --recurse-submodules https://github.com/JetBrains-Research/routing-preference.git
cd routing-preference
make setup   # also applies patches/track-exposed-files.patch to mini-swe-agent
cp .env.example .env  # add your API keys

The setup step applies patches/track-exposed-files.patch to the mini-swe-agent submodule so that file exposures (what the agent read during execution) are recorded alongside each generated solution. V2.1 judge scoring relies on this.

Usage

uv run generate --dataset "dataset-name" --model "openai/gpt-4o"

Docs

• GUIDE.md - Setup and usage details
• docs/PROJECT.md - Research overview
• docs/OBJECTIVE.md - Objective characteristic definitions
• docs/FILESYSTEM.md - Data and output layout