Section 125 · Chapter 16, Introspection: White-Box Testing Networks

Attention Diagnostics

Attention views can direct an investigation, but they are not transcripts of reasoning and they do not prove causality.

attentionattention diagnostics

What to do

  1. Record the aggregation, filtering, and selection rules, and retain the raw per-layer, per-head matrices when the result matters.
  2. Do not convert a plausible-looking heatmap into a pass result.
  3. Choose layers and token positions according to a documented rule, retain the full per-head tensor for investigation, and connect visual differences back to behavioral evals.
  4. Start with cases whose behavioral outcome is already understood.

Evidence to preserve

  • Record the aggregation, filtering, and selection rules, and retain the raw per-layer, per-head matrices when the result matters.
  • Capture attention diagnostics for known-good, known-bad, and ambiguous examples.
  • Preserve the inputs, versions, configurations, raw outcomes, and results for attention, attention diagnostics needed to reproduce work on Attention Diagnostics.
  • Report results for attention, attention diagnostics by relevant slice, separate blocker failures from averages, state uncertainty and blind spots, and connect the result to a release decision.

Continue the conversation

Apply this to your context.

Save your product context once, then open a focused conversation that combines it with this concept.

Cite this page

Jason Arbon. "Attention Diagnostics." Testing AI Knowledge Edition, section 125.

https://jarbon.ai/testing-ai/knowledge/ch125-attention-diagnostics.html

Shared across the Knowledge Edition

Adapt every concept to your world.

Describe your product, role, users, risks, constraints, or current quality problem. This stays in this browser until you choose to send it to ChatGPT.

Saved only in this browser.0 / 2400