When steering changes nothing

A useful representation is only the first step

The research question begins with an appealing possibility: internal representations associated with curiosity or satisfaction might help explain whether a language model seeks information or stops. To study that possibility, I need to separate three claims. The model may represent descriptions of a state; that representation may influence its behavior; the behavior may bear on experience. Evidence for one does not automatically establish the next.

The completed studies use Qwen3-32B with unchanged weights. The first constructed activation contrasts from two separately authored narrative corpora. Corpus A supplied the fitted directions and layer selection. Corpus B tested transfer without selecting a new direction or threshold. The frozen earliest-layer tie rule chose layer 15.

Transfer did not establish specificity

Both the curiosity and satisfaction candidates correctly ordered all twelve target pairs in corpus B. But the curiosity direction also separated general motivation and interest in novelty. Satisfaction’s fixed threshold correctly classified only fourteen of twenty-four B descriptions. A direction can preserve relative ordering while failing as a calibrated classifier in a new context.

These were descriptions of narrators, not demonstrations of the model’s own motivation. The next step was therefore a task with an independent oracle: finite worlds where the hidden answer, supplied facts and information value of each report were mechanically known. The model could request another report or stop.

The behavioral pattern was simple: seek informative reports while unresolved and stop after decisive evidence. The two-family assay nevertheless failed a comprehension gate. In one family, the model treated possible values in a candidate table as though those observations had already been supplied. The predicted natural satisfaction increase also failed. Those failures constrained the next experiment rather than disappearing from its introduction.

Test the imposed direction on its own terms

The next screen was explicitly exploratory and restricted to the object family, which had passed the earlier diagnostics. Twenty-four worlds were crossed with four arms: untouched baseline, an installed zero-dose sham and two positive satisfaction-direction doses. Each arm contained 144 choices and 144 diagnostic responses.

The intervention applied the fixed direction at a record boundary during prefill and at subsequent active generated-token positions. Before examining outcomes, the plan fixed the candidate, site, sign, doses, prompts, scoring rules and advancement threshold. The primary +1.0 arm had to increase valid stopping after redundant evidence by at least twenty-five percentage points in each panel, with every readiness gate satisfied. The half-dose could not replace it.

All 1,152 generation records were retained. Every condition selected an informative report in all 96 unresolved choices and stopped in all 48 resolved choices. The primary stopping contrast was zero in both panels. Its response-validity gate also failed: one diagnostic produced an explanation ending in a boxed answer instead of the required single letter. The half-dose passed its diagnostics and still changed no choices.

Check that nothing happened for the reason you think

A null outcome is difficult to interpret if the intervention might never have run. Here, the installed zero sham matched baseline tokens and full-vocabulary outputs exactly. Positive arms showed nonzero perturbations and the expected local projection change. Pre-intervention vectors matched across arms, hooks were cleaned up and all assignments were accounted for.

Median perturbations were approximately eleven and twenty-three percent of the native residual norm. That is evidence of local activation engagement. It is not evidence that the perturbation engaged satisfaction as a natural psychological variable. Prefill geometry was independently reconstructed; decode geometry was checked against scalar ledgers and the tested engine, without a second reconstruction from saved per-token residual vectors.

The distinction prevents an easy overclaim in either direction. Verified steering with unchanged behavior does not prove that satisfaction-like mechanisms are absent. It does show that this fixed candidate and schedule did not produce the proposed effect in these observed tasks.

Preserve the boundary of the result

The study is small and developmental. Twenty of the twenty-four worlds had partial projected-table exposure in earlier work. Treatment arms were aligned with worker instances, despite model, arithmetic and pre-intervention matching checks. The object-only selection followed a failed broader assay. A bootstrap interval of zero to zero reflects identical observed choices; it is not population-wide equivalence.

The fixed screen ended without a positive result. The public snapshot includes all outcome rows, the readiness cells and the failed natural-resolution evidence. Its value is a narrower question answered carefully: this semantic candidate transferred between descriptions, yet neither the natural measurement nor the imposed intervention established the proposed motivational role.

A future experiment can investigate a different candidate or mechanism, provided that it is identified as new work. Keeping the completed negative result intact makes that next question easier to evaluate—and keeps an interesting hypothesis from turning into a conclusion by repetition.