flash-ansr-v23.2-120M

A 120M-parameter flash-ansr model for symbolic regression: give it (X, y) and it returns a closed-form expression.

v23.2 ("full Cauchy / wide inputs") is the wide-input variant of v23.0-120M. The constant prior and the input-range endpoints are Cauchy (heavy-tailed), and the +/-30 input-magnitude clamp is removed. As a result it recovers expressions on out-of-distribution-magnitude inputs substantially better than the clamped baseline (+14 points on OOD-magnitude FastSRB) at no measurable in-distribution cost.

Usage

pip install flash-ansr

import numpy as np
from flash_ansr import FlashANSR, install_model, get_path, SoftmaxSamplingConfig

MODEL = "psaegert/flash-ansr-v23.2-120M"
install_model(MODEL)

model = FlashANSR.load(
    directory=get_path("models", MODEL),
    generation_config=SoftmaxSamplingConfig(choices=1024),
    length_penalty=0.05,
)

X = np.random.uniform(-3, 3, size=(200, 2))
y = X[:, 0] ** 2 + X[:, 1]

model.fit(X, y, verbose=True)
print(model.get_expression())
y_pred = model.predict(X)

See the flash-ansr repository for full documentation.

Publications

• Saegert & Köthe 2026, Breaking the Simplification Bottleneck in Amortized Neural Symbolic Regression (ICML 2026), arXiv:2602.08885