$\alpha$-PFN: Fast Entropy Search via In-Context Learning
$\alpha$-PFN is a Prior-Fitted Network that amortizes information-theoretic acquisition functions. Supported acquisition functions: Predictive Entropy Search (PES), Max-value Entropy Search (MES), and Joint Entropy Search (JES).
To reproduce our ICML 2026 paper experiments, see branch
icml2026.
Install
pip install AlphaPFN
Or from source:
git clone https://github.com/automl/AlphaPFN
cd AlphaPFN
uv sync
Pretrained checkpoints (~20 MB) download automatically on the first from_pretrained call and cache under ~/.cache/alphapfn/.
Quick start
A self-contained 6D BO loop on Hartmann, using botorch.optim.optimize_acqf. Requires the [botorch] extra — pip install "AlphaPFN[botorch]":
import torch
from botorch.optim import optimize_acqf
from botorch.test_functions import Hartmann
from alphapfn import AlphaPFN
# 1. Objective on the unit cube (α-PFN maximizes — `negate=True` flips Hartmann's sign).
hartmann = Hartmann(dim=6, negate=True)
# 2. Initial design.
torch.manual_seed(0)
d, n_init, num_steps = 6, 6, 30
X = torch.rand(n_init, d, dtype=torch.double)
y = hartmann(X)
bounds = torch.stack([torch.zeros(d), torch.ones(d)]).double()
# 3. Load the pretrained acquisition; checkpoints download on first call.
acqf = AlphaPFN.from_pretrained(acquisition="JES")
# 4. BO loop.
for step in range(num_steps):
acqf.fit(X, y) # fit() standardizes y internally
X_next, _ = optimize_acqf(acqf, bounds=bounds, q=1,
num_restarts=10, raw_samples=128)
y_next = hartmann(X_next.squeeze(0))
X = torch.cat([X, X_next.detach().double()])
y = torch.cat([y, y_next.detach().double().reshape(1)])
print(f"step {step+1:>2}: best so far = {y.max().item():.4f}")
Runnable version: examples/bo_with_optimize_acqf.py
or open the Colab notebook.
API
AlphaPFN.from_pretrained(
acquisition: str | None = None, # "PES" (default), "MES", or "JES"
version: str = "v1",
*,
load_base_model: bool = False,
ucb_beta: float = 2.0,
strict: bool = True, # pass strict=False to skip input checks
)
Before fitting, prepare your data so that:
- You are maximizing. To minimize instead, negate your objective. This is NOT checked, so forgetting it silently gives wrong results.
- Each input feature lies in
[0, 1]. Rescale your search space accordingly.
fit() standardizes targets internally (standardize_y=True by default) — pass raw $y$. Pass standardize_y=False if you have already standardized. strict=True (default) validates the unit-cube contract on every fit/forward; pass strict=False to skip.
Cite
@inproceedings{
rakotoarison2026alphapfn,
title={{$\alpha$}-PFN: Fast Entropy Search via In-Context Learning},
author={Rakotoarison, Herilalaina and Adriaensen, Steven and Viering, Tom and Hvarfner, Carl and M{\"u}ller, Samuel and Hutter, Frank and Bakshy, Eytan},
booktitle={Forty-third International Conference on Machine Learning},
year={2026},
url={https://openreview.net/forum?id=7Oonij8oLU}
}
Metadata
Release files for AlphaPFN 0.0.2
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | Provenance | Actions |
|---|---|---|---|---|
| alphapfn-0.0.2.tar.gz | 52.3 kB | Without provenance | Details for alphapfn-0.0.2.tar.gz |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Actions Reset |
|---|---|---|---|---|
| alphapfn-0.0.2-py3-none-any.whl | Python 3 | none | any | Details for alphapfn-0.0.2-py3-none-any.whl |
Total release size: 112.6 kB
Release files / alphapfn-0.0.2.tar.gz
| Download URL | alphapfn-0.0.2.tar.gz |
|---|---|
| Size | 52.3 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
bcf9ddbf53a4cb7f1d1101d4adb03cde8751c8c1e3e145bbebb1911a7bfda038
|
|
BLAKE2b-256 checksum How to use checksums |
51b9136b697d2d280b94952e48478a6ea11ad725ebb080eb263ceead77d27559
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/6.2.0 CPython/3.11.13
|
Release files / alphapfn-0.0.2-py3-none-any.whl
| Download URL | alphapfn-0.0.2-py3-none-any.whl |
|---|---|
| Size | 60.3 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
3db91b7e063f21fbc5487ba693b583691ffcf43f0dcda022dd666ac080f738d9
|
|
BLAKE2b-256 checksum How to use checksums |
51632d2909d581e15d64735f39d53e3572db59d1914967745471985088ca969e
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/6.2.0 CPython/3.11.13
|
Security for AlphaPFN 0.0.2
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
No provenance found
None of the files in this release were uploaded with provenance metadata.
A note on package security
PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.
To learn more about protecting yourself from malicious actors, visit our package security documentation.