transcoder, a MOLT does not learn sparsely active features that are embedded along a vector direction in the model’s activation space
Instead, it learns sparsely active transforms, which apply a linear transformation to the residual stream to give their contribution to the MLP output
Unlike transcoder features, which double as both computational and representational units, MOLT transforms are purely computational objects that “bridge” representations between layers
fuzzing vs active fuzzing Close—but neither test replaces the actual text tokens. My chair/table example made that look like substitution.
- Ordinary fuzzing: selects a separate, non-activating passage and highlights inactive tokens there.
- Active fuzzing: uses an activating passage but places incorrect highlights on inactive positions within it.
So your intuition about changing the highlighted position in active fuzzing is right; ordinary fuzzing changes the selected passage, rather than editing a token.
Yes—the ground truth comes from measured gate activations in both tests. The difference is where the incorrect highlights are placed.
Suppose a CH feature has these measured activations:
I bought a chair.
0 0 0 3.0
I bought a table.
0 0 0 0
Then the scoring examples could be:
| Example | Highlight shown to the judge | Ground truth |
|---|---|---|
| Correct highlight, used in both tests | I bought a <<chair>>. | Correct: highlighted token activates |
| Ordinary fuzzing negative | I bought a <<table>>. | Incorrect: highlighted token is inactive |
| Active-fuzzing negative | I <<bought>> a chair. | Incorrect: bought is inactive, although chair activates nearby |
Both tests verify the highlighted tokens against actual activations. The judging LLM sees the description and highlighted text, with activation values hidden.
Active fuzzing makes the distinction more precise: merely noticing “this passage contains a CH word” is insufficient—the judge must check which token was highlighted.
Here, “correct highlight” means the token actually activates the gate. Whether the written description adequately explains that activation is what the test evaluates.