HeyHomees Lab Notes · No. 15 · Google DeepMind · Published November 3, 2026
By HeyHomees Newsroom ·
In 30 seconds
- It adds interpretable feature scores to any AI text detector
- A demo is available on Hugging Face and the package installs via pip
What is it?
DetectorScope is a library that wraps an existing AI text detector and returns a sum of named, validated features for each prediction, showing where in the text each feature fires
How does it work?
It encodes the language model's residual stream with sparse autoencoders (SAEs), validates each feature with a natural-language description, and trains a sparse linear head (or XGBoost head) over the validated features; the final logit is the bias plus the sum of feature importances
What does it need to run?
Installation is via pip install detectorscope; stages that fine-tune the classifier and train SAEs require a GPU, while later stages run on CPUs
What's new versus the last version?
The release adds a configurable tau parameter that trades interpretability for accuracy (default 0.85, range 0.00–0.95) and provides both logistic-regression and XGBoost heads at each tau step
By the numbers
- Default tau (interpretability threshold): 0.85
- Tau range: 0.00 to 0.95 in steps of 0.05
- Minimum threshold where FT-Model+SAE matches underlying detector: 0.7
- News announcement date: 25 Oct 2026
The honest read
The reported tau values (default 0.85, range 0.00–0.95) are set by the authors and describe a trade-off between interpretability and accuracy, but the source does not provide quantitative accuracy numbers for each setting. The claim that FT-Model+SAE matches the underlying detector at thresholds of 0.7 and above is self-reported and lacks detailed benchmark scores. Performance impact is described as “minimal adverse effects,” but no exact figures are given.
What we don't know yet
- Model parameter counts or download sizes
- Whether the library supports detectors beyond the provided Gemma-2-2b variants
Sources
- README, GitHub (primary)
Researched and drafted with AI from the sources above; every figure is checked against them by code before it is published, and a line that fails is dropped.