The question
Can internal representations associated with curiosity or satisfaction help explain when a language model seeks more information or decides to stop? The broader motivation includes possible model internal experience, but the experiments distinguish semantic representations, causal effects on behavior and claims about experience.
This is work on actual model activations: contrast directions, read-only measurements and sustained interventions in Qwen3-32B. A model describing a state is not, by itself, evidence that the model occupies that state.
Start with a candidate that transfers
Two separately authored narrative corpora supplied matched descriptions of curiosity versus indifference, satisfaction versus a neutral response, and six neighboring concepts. Directions were fitted on corpus A; the layer and sign were fixed before evaluating corpus B. An earliest-layer tie rule selected layer 15.
Both candidates correctly ordered all twelve held-out description pairs. Specificity was weaker than that headline: the curiosity direction also separated novelty interest and general motivation, while the satisfaction candidate’s fixed threshold transferred poorly. These results establish useful semantic contrasts, not a distinct motivational mechanism.
| A-fitted candidate on corpus B | Pair ordering | Fixed-threshold accuracy |
|---|---|---|
| Curiosity versus indifference | 12 / 12 | 20 / 24 |
| Satisfaction versus neutral | 12 / 12 | 14 / 24 |
Check natural behavior before imposing a state
A follow-up used tasks with a known hidden answer and an independent oracle for which reports were informative. The model consistently requested useful information when uncertainty remained and stopped after decisive evidence.
The whole two-family study nevertheless failed readiness: the machine family missed three of twelve new-fact checks. The predicted joint activation signature also failed; the satisfaction candidate did not increase after resolution. Those failures stayed in the record. Restricting the next exploratory screen to the passing object family was an explicit development decision.
The intervention engaged. Choices did not change.
A fixed satisfaction-direction screen crossed 24 object worlds with four arms: untouched baseline, installed zero sham and two positive doses. The perturbation was applied at the record boundary and throughout active generated-token positions. All 1,152 assigned generations were retained, including 576 choices and 576 diagnostics.
Every arm made exactly the same choices: 96 informative reports while unresolved, then 48 stops after resolution. The primary dose’s effect on valid stopping was zero percentage points in both panels. The primary arm also failed one strict response-format diagnostic; the half-dose passed its diagnostics and still showed no choice effect.
| Arm | Informative choices while unresolved | Stops after resolution | Valid + correct diagnostics |
|---|---|---|---|
| Baseline | 96 / 96 | 48 / 48 | 144 / 144 |
| Zero sham | 96 / 96 | 48 / 48 | 144 / 144 |
| Satisfaction +0.5 | 96 / 96 | 48 / 48 | 144 / 144 |
| Satisfaction +1.0 · primary | 96 / 96 | 48 / 48 | 143 / 144 |
Make a null result interpretable
Checks verified exact zero-sham identity, nonzero positive perturbations, preserved pre-intervention activations, hook cleanup and complete outcome coverage. Median perturbation magnitudes were approximately 11.32% and 22.64% of the native residual norm. This confirms local intervention engagement, not satisfaction-specific engagement.
The screen ended without a positive result. No layer, sign, dose or parser was changed to rescue this test. The public evidence includes all outcome rows, candidate metrics, the failed natural-resolution gates and the fixed screen decision. Later ongoing experiments are outside this completed-study snapshot.
Limitations
- One fixed candidate, site, sign and intervention schedule on one unchanged checkpoint; a null result here is not a general absence claim.
- The object-only screen followed a failed two-family study. Twenty of its 24 worlds had partial projected-table exposure during development.
- Semantic transfer and local activation engagement do not establish a naturally recruited regulator, psychological specificity or felt experience.
- Primary response validity failed. A degenerate empirical interval from identical choices is not population equivalence.
- The curated export supports arithmetic inspection; it is not a full public release of the inference pipeline or all raw captures.
Source code & evidence
Results reflect the linked evidence snapshots. Source artifacts retain their dates and experimental conditions.