Does quantization quietly break interpretability?

Does quantization break interpretability?

Trained a sparse autoencoder on GPT-2-small's layer-6 residual stream. Simulated 16/8/4-bit quantization via uniform weight rounding, evaluated on 500 held-out documents using three independent metrics: activating-position recall, per-feature correlation, and behavioral checks.

Behavior and interpretability degrade together, not independently. 16-bit is a clean negative control validating the pipeline; 4-bit destroys the model outright rather than revealing a subtle interpretability-specific danger.

A live, low-overhead safety monitor

Searched all 8,192 SAE features against BeaverTails using a strict discovery/validation/test split. A single feature was weak and confounded; a 100-feature regularized probe was real, validated on a genuinely untouched holdout, and converted into a live per-token monitor.

Computing only the 100 needed encoder features (skipping the full decoder and 8,092 unused features) cut monitoring overhead from a measurable +4.12% to +0.66% — not statistically distinguishable from zero.

Golden Gate Claude-inspired steering

Additive decoder-direction steering reliably raised targeted SAE activations — confirming causal control of the encoded coordinate. Five semantically screened features were tested under last-position steering. Three finalists were tested under persistent all-position steering alongside one matched random unit direction.

Target and random directions produced comparable output degeneration at matched strength. Reported as an honest, controlled negative result — not a hidden limitation.

Architecture & stack

Base LLM (GPT-2-small, frozen)
         │
   Hook/Capture layer ──────→ Quantized variants (16/8/4-bit)
         │                              │
    SAE (trained once) ←────────────────┘
         │
  ┌──────┼──────────┬──────────────┐
Feature  Safety     Steering    Benchmark
Browser  Monitor    (Phase 7)   Layer