arXiv News

First-hand information for everyone

Site

AboutPrivacy PolicyTerms of Use

© 2026 arXiv News

GateWatch delivers tiny, fully interpretable probes that nearly match full‑size competitors on 24 Gemma‑2‑9B tasks | arXiv News

arXiv News

EnglishJapanese

Switch language

EnglishJapanese
Loading account…

First-hand information for everyone

Back to front page
Artificial IntelligenceEnglishPublished

GateWatch delivers tiny, fully interpretable probes that nearly match full‑size competitors on 24 Gemma‑2‑9B tasks

November 20, 2026arXiv: 2611.24619v1

GateWatch is a compact, interpretable classifier that writes its entire decision rule as explicit Boolean logic over a small set of binarized sparse autoencoder (SAE) features. It offers three ways to build these logic circuits—Compile, Search, and Train—under a chosen gate budget. When deployed with GateWatchServe, only selected SAE features are computed, their binary values are streamed across layers, and several classifiers run in parallel using lookup tables.

On 24 benchmark tasks for the Gemma‑2‑9B language model, five‑gate GateWatch probes achieve a mean test accuracy of 80.49 %, just 1.17 percentage points below a full 16K‑feature SAE probe. They use on average 3.8 SAE features and 3.5 logical gates, and their memory footprint is 26.3 KiB—about 4,380 times smaller than the 115,261 KiB required by the full probe. This size reduction leaves overhead minimal while preserving most of the baseline performance.

GateWatch’s gains rely on pretrained SAEs. Limitations arise from the need to select useful SAE features, possible ambiguity in their semantic labels, and the requirement to store encoder rows for the chosen features. These factors can affect both interpretability and deployment size.

GateWatch therefore presents a practical trade‑off: tiny, interpretably defined probes that retain competitive accuracy and dramatically reduce memory.”

Source paper

GateWatch: Low-Overhead LLM Monitoring with Boolean Circuits

Artificial Intelligence
Machine Learning
View on arXiv