geood 0.1.0

Geometric OOD detection for LLM hidden states

geood

Geometric OOD detection for LLM hidden states.

Perplexity misses out-of-distribution inputs when the model is fluent in the OOD domain. geood detects them using hidden-state geometry (intrinsic dimensionality + Mahalanobis distance).

Paper: The Geometric Blind Spot of Perplexity (Tabares Montilla, 2026)

Install

pip install geood

Usage

import geood

# Calibrate with your in-distribution texts (once)
detector = geood.calibrate("meta-llama/Llama-3-8B", ref_texts)

# Detect OOD (many times)
result = detector.detect("def quicksort(arr): ...")
print(result.is_ood)    # True
print(result.explain())  # "OOD detected: ref_dim=33.8, mahalanobis=12.3"

# Save/load for deployment
detector.save("my_detector.geood")
detector = geood.load("my_detector.geood")

How it works

LLM hidden states for in-distribution inputs occupy a high-dimensional subspace. OOD inputs collapse to low dimensionality, even when perplexity says they're familiar. geood calibrates on your reference corpus and detects this geometric collapse.

Key results

Method	Code AUROC (LLaMA)	Code AUROC (Mistral)
Perplexity	0.352	0.150
geood (Mahalanobis)	1.000	1.000
geood (Intrinsic dim)	1.000	1.000

Code has lower perplexity than math (2.57 vs 2.76) but collapses geometrically (dim 4 vs 34). Perplexity is blind; geometry detects it.

License

MIT