Skip to content

Probes and representation analysis

Representation analysis asks whether information is present inside a model and how it is organized. A probe or projection can make that information legible, but legibility alone does not show that the model uses the representation. The strongest experiments connect a held-out readout to a targeted intervention and a behavioral result.

Use this method when

  • you suspect a model internally tracks a property that is not always visible in its output
  • you want to locate where a concept becomes decodable across layers or time
  • the concept has meaningful structure, such as order, direction, similarity, or a cycle
  • an internal signal could become a monitor, steering handle, training reward, or standing evaluation

Do not begin with a probe when metadata, prompt shape, token position, or another simple artifact could answer the question. Establish a behavioral baseline and rule out those shortcuts first.

The evidence ladder

1. State the claim and decision rule

Name the property you expect the model to represent, where you expect to find it, and what result would count against the hypothesis. Decide in advance what held-out score, null comparison, or causal effect would make the representation useful.

2. Build controlled examples

Vary the target property while holding likely confounds constant. Separate fitting, model selection, and final evaluation data. When possible, include a metadata-only baseline and a shuffled-label control.

3. Start with the simplest readout

Use a mean difference, linear probe, or low-dimensional projection before a more expressive decoder. Compare against a simple behavioral or embedding baseline. A complicated probe can memorize distinctions that the model does not naturally expose.

4. Test generalization

Evaluate on held-out templates, subjects, datasets, or conditions that differ in the way your claim requires. Inspect failures rather than reporting only the aggregate score.

5. Intervene

Ablate, patch, steer, or otherwise change the candidate representation. The causal test should predict a specific behavioral change and include a matched random or irrelevant-direction control.

6. Turn the signal into an action

Use a successful representation as a monitor, evaluation, steering direction, training signal, or model-development diagnostic. If the intervention fails, narrow the claim: the information may be decodable without being used.

Worked example

The following experiment moves through the full evidence ladder: controlled stimuli, a confound-corrected projection, multiple null tests, and a causal steering intervention.

Language models · Introductory Getting Started with Neural Geometry Find a structured concept in a model's activations, test whether the geometry is real, and intervene on it. 45–65 RU · estimate · 35 min · 1× H100 Representation analysisNeural geometry

What to record

  • model checkpoint, layer, hook point, and token-selection rule
  • data identity, split strategy, and known overlap
  • probe family, regularization, and selection metric
  • simple and shuffled baselines
  • held-out sample size and uncertainty
  • intervention strength and matched controls
  • the behavioral decision the internal signal changes