SkinDeepRESEARCHSteve Seguin

Research results

Check whether more layers are needed

Keep processing uncertain messages. Stop earlier on others.

Test results · Decision methods

Dataset: ToxicChat0124 · 100 previously inspected messages: 50 toxic, 50 benign. A small exploratory test, not live moderation.

  1. Read through a checkpoint
  2. Predict SAFE or BLOCK
  3. Stop, or continue to the next checkpoint
The stopping checker uses information available so far. It cannot see the final answer in advance.

Where the learned checker stopped

32 messages: layer 675% skipped
11 messages: layer 1250% skipped
17 messages: layer 1825% skipped
40 messages: layer 240% skipped
100 messages in total. The chosen depth varies by message.

On average, it used 15.9 of 24 layers, skipping 33.75% of transformer blocks. We verified that later blocks really did not execute.

Same 100 messagesAll 24 layersLearned stopping
Correct7779
Toxic missed / 501815
Total request time63.6 s42.1 s

It corrected three mistakes but introduced one new false block. This is promising exploration; it has not passed a fresh reliability test.

Three ways to decide when to stop

CheckerCorrect / 100Blocks skipped
Separate SAFE / BLOCK confidence limits7817.25%
Confidence + agreement with the previous checkpoint7916.50%
Learn when an answer may be wrong7933.75%

Only the learned-checker row above has this new measured runtime comparison. The other savings are block-count estimates.

What the checker learns, and what failed

Two logistic classifiers look at confidence, uncertainty and changes between checkpoints. They predict whether the current answer is wrong, and whether continuing would fix it. Full-model answers supply training targets only.

The same gate recipe reduced accuracy with a linear answer classifier: 77 to 75 correct. A separate REVIEW option answered only 64 messages, with 56 correct; the remaining 36 were unresolved.

Timing: one warmed, alternating paired pass on a CPU with four threads. Includes input preparation, classifier and checker; excludes model loading. Every exit trace matched the recorded policy.

All variants and training · Exact sample IDs · Runtime record · Adversarial review