PIPS: Per-Instance Program Synthesis

Per-Instance Program Synthesis (PIPS) is an adaptive reasoning baseline that decides, for every problem instance, whether to run deliberate chain-of-thought reasoning or synthesize executable Python code.

News

• [10/26/2025] Code for PIPS as well as a Gradio demo are released.
• [09/18/2025] PIPS accepted to NeurIPS 2025!

Why try PIPS?

• Adaptive per-instance strategy selection that balances executable program synthesis with fast chain-of-thought reasoning.
• Iterative code refinement with automatic execution, error recovery, and optional interactive callbacks.
• Works out-of-the-box as a Python library and ships with a lightweight Gradio UI for quick exploration.
• Drop-in baseline: import a single class to embed Per-Instance Program Synthesis inside your stack.

Install with uv

git clone <repository-url>
cd PIPS
uv venv
source .venv/bin/activate
uv sync --editable .

To include optional dependencies:

• uv sync --group gradio enables the browser UI.
• uv sync --group dev installs linting and test tooling.

Install multiple groups at once with uv sync --group gradio --group dev.

Configure providers

Export whichever API keys you plan to use:

export OPENAI_API_KEY="..."
export ANTHROPIC_API_KEY="..."
export GOOGLE_API_KEY="..."

Only one key is required to get started.

Use PIPS from Python

from pips import PIPSSolver, get_model
from pips.utils import RawInput

model = get_model("gpt-4o", api_key="your-openai-api-key")
solver = PIPSSolver(model=model, max_iterations=8, temperature=0.0)

problem = RawInput(text_input="What is the sum of the first 10 prime numbers?")
answer, telemetry, mode_decision = solver.solve(problem, stream=False)
print("Answer:", answer)

# Inspect the selected mode
print("Mode:", "code" if mode_decision["use_code"] else "chain-of-thought")

Set stream=True and pass callbacks such as on_llm_streaming_token or on_step_update to receive incremental updates. For complete control you can call solve_with_code or solve_chain_of_thought directly, and the same callbacks work for both.

Run the Gradio App

Install the Gradio extra and launch the browser experience:

uv sync --group gradio
uv run python -m pips.gradio_app

Supported providers

PIPS supports any OpenAI, Anthropic, or Google model through their API endpoints. Models hosted with VLLM are also supported.

Development and tests

Install dev tooling and run the test suite:

uv sync --group dev
uv run pytest

License

MIT