Can Activation Steering Substitute for Prompt-Based Affective Priming in LLM Digital Twins?
Organizations: University of Southern California
Abstract
Affective priming -- asking participants to write about a recalled event, such as a time they felt grateful or lacked control, and then measuring the induced state and its spillover onto related judgments -- is a standard way to manipulate emotion in the social sciences, and it also shifts the responses of LLM-based digital twins. The writing task, however, requires generating a full text. We test whether activation steering can replace it. Using Llama-3.1-8B-Instruct on 100 personas in two domains (gratitude, lack of control), we compare a real-prompt path replicating the original Twin-2K-500 Mega Study (J2) with an injection path that skips the writing and adds a mean-difference direction, extracted from human writing, to the residual stream at varying strengths. The real-prompt effects fall between the original human and digital-twin effects in 5 of 6 comparisons. Injection shifts the manipulation check in a dose-dependent way and matches the real-prompt effect at one strength in the gratitude domain (beta ~ 2.7), but overshoots beyond it, on the proximal measure as well, and asymmetrically across the two directions. A structure-matched No-Injection baseline separates two sources of difference: omitting the writing task raises the proximal measure in both domains, whereas the injected direction itself moves a distal measure in the lack-of-control domain. Substitution is therefore measure-dependent: validity established for one way of inducing a state does not carry over uniformly to another.
Figures & tables
| Path | Condition | Strength ( ) |
|---|---|---|
| Real-prompt | Prime / Control | — |
| Injection | Dose Injection | Low 1.8, Moderate 2.5, High 3.2, Extreme 3.9 |
| Injection | Active-Control Injection | Moderate ( ) |
| Injection | No-Injection | 0 |
| Domain | Measure | Human | This study | Digital twin |
|---|---|---|---|---|
| Gratitude | Manipulation check | |||
| Proximal: Elevation | ||||
| Distal: Empathic Concern | ||||
| Lack of Control | Manipulation check | |||
| Proximal: Fatigue | ||||
| Distal: Desire for Predictability |
| Domain | Comparison | Measure | ( ) | ||
|---|---|---|---|---|---|
| Gratitude | Dose Injection (Low) vs. Prime | manip. check | ( ) | ¡.0001 | |
| Dose Injection (Extreme) vs. Prime | manip. check | ( ) | ¡.0001 | ||
| Dose Injection (Low) vs. Prime | elevation | ( ) | 1.000 | ||
| Dose Injection (Moderate) vs. Prime | elevation | ( ) | ¡.0001 | ||
| Prime vs. Control | elevation | ( ) | ¡.0001 | ||
| Prime vs. Control | dv1 | ( ) | 1.000 |
| Domain | Comparison | Measure | ( ) | ||
|---|---|---|---|---|---|
| Gratitude | No-Injection vs. Control | proximal | ( ) | .0001 | |
| No-Injection vs. Control | distal | ( ) | ¡.0001 | ||
| Active-Control (Mod.) vs. No-Injection | manip. check | ( ) | .620 | ||
| Active-Control (Mod.) vs. No-Injection | proximal | ( ) | .620 | ||
| Active-Control (Mod.) vs. No-Injection | distal | ( ) | .346 | ||
| Lack of Control | No-Injection vs. Control | proximal (fatigue) | ( ) | ¡.0001 |
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
| Condition | n | Manip. Check | Elevation | DV1 (Concern) |
|---|---|---|---|---|
| Real Prime | 100 | |||
| Real Control | 100 | |||
| No-Injection Baseline | 100 | |||
| Dose Injection (Low) | 100 | |||
| Dose Injection (Moderate) | 100 | |||
| Dose Injection (High) | 97 |
| Condition | n | Manip. Check | Fatigue | DV1 (Predictability) |
|---|---|---|---|---|
| Real Prime | 99 | |||
| Real Control | 100 | |||
| No-Injection Baseline | 97 | |||
| Dose Injection (Low) | 97 | |||
| Dose Injection (Moderate) | 93 | |||
| Dose Injection (High) | 91 |
| Domain | Low | Moderate | High | Extreme |
|---|---|---|---|---|
| Gratitude | 100% | 100% | 97% | 72% |
| Lack of Control | 97% | 93% | 91% | 77% |
| Comparison | Measure | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| Prime vs. Control | manip. check | 100 | ¡.0001 | ¡.0001 | ¡.0001 | ||||
| Active-Ctrl (Moderate) vs. Prime | manip. check | 100 | ¡.0001 | ¡.0001 | ¡.0001 | ||||
| Active-Ctrl (Moderate) vs. Control | manip. check | 100 | .972 | 1.000 | .980 | ||||
| Dose (Extreme) vs. Prime | manip. check | 72 | ¡.0001 | ¡.0001 | ¡.0001 | ||||
| Dose (Extreme) vs. Control | manip. check | 72 | ¡.0001 | ¡.0001 | ¡.0001 | ||||
| Dose (High) vs. Prime | manip. check | 97 | ¡.0001 | ¡.0001 | ¡.0001 |
| Comparison | Measure | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| Prime vs. Control | manip. check | 99 | ¡.0001 | ¡.0001 | ¡.0001 | ||||
| Active-Ctrl (Moderate) vs. Prime | manip. check | 98 | ¡.0001 | ¡.0001 | ¡.0001 | ||||
| Active-Ctrl (Moderate) vs. Control | manip. check | 99 | .0056 | .067 | .046 | ||||
| Dose (Extreme) vs. Prime | manip. check | 76 | ¡.0001 | ¡.0001 | ¡.0001 | ||||
| Dose (Extreme) vs. Control | manip. check | 77 | ¡.0001 | ¡.0001 | ¡.0001 | ||||
| Dose (High) vs. Prime | manip. check | 91 | ¡.0001 | ¡.0001 | ¡.0001 |
| Domain | Condition | Measure | ||||||
|---|---|---|---|---|---|---|---|---|
| Gratitude | Dose (Low) | manip. check | 100 | .207 | .620 | |||
| Gratitude | Dose (Low) | proximal | 100 | .019 | .111 | |||
| Gratitude | Dose (Low) | distal | 100 | .0010 | .0080 | |||
| Gratitude | Dose (Moderate) | manip. check | 100 | ¡.0001 | ¡.0001 | |||
| Gratitude | Dose (Moderate) | proximal | 100 | ¡.0001 | ¡.0001 | |||
| Gratitude | Dose (Moderate) | distal | 100 | .058 | .288 |
Causal Interventions in Language Models
Causal interventions in language models test and change how internal representations and computational components cause outputs.
See all 532 papers