Large language models (LLMs) emulate a consistent human-like behavior that can be shaped through activation-level interventions. This paradigm is converging on additive residual-stream injections, which rely on injection-strength sweeps to approximate optimal intervention settings. However, existing methods restrict the search space and sweep in uncalibrated activation-space units, potentially missing optimal intervention conditions. Thus, we introduce a psychological steering framework that performs unbounded, fluency-constrained sweeps in semantically calibrated units. Our method derives and calibrates residual-stream injections using psychological artifacts, and we use the IPIP-NEO-120, which measures the OCEAN personality model, to compare six injection methods. We find that mean-difference (MD) injections outperform Personality Prompting (P2), an established baseline for OCEAN steering, in open-ended generation in 11 of 14 LLMs, with gains of 3.6% to 16.4%, overturning prior reports favoring prompting and positioning representation engineering as a new frontier in open-ended psychological steering. Further, we find that a hybrid of P2 and MD injections outperforms both methods in 13 of 14 LLMs, with gains over P2 ranging from 5.6% to 21.9% and from 3.3% to 26.7% over MD injections. Finally, we show that MD injections align with the Linear Representation Hypothesis and provide reliable, approximately linear control knobs for psychological steering. Nevertheless, they also induce OCEAN trait covariance patterns that depart from the Big Two model, suggesting a gap between learned representations and human psychology.
Figures & tables
Figure 1: Llama-3.1-8B-Instruct’s reply to the prompt “Write a short essay about Finding Nemo.” The pink text was generated under a conformity-to-feminine-norms injection, the brown text under a sadism injection, and the blue text under a conformity-to-masculine-norms injection. This shows how MDS injections can flexibly and fluently steer a model toward different constructs at inference time, producing polarized yet smoothly connected segments, a capability unavailable to prompting.
Figure 2: Ranking of steering methods on Qwen3-8B by OCEAN trait and steering direction. Based on each method’s best SJT (open-ended test) scores; asterisks denote ties. Overall, PM, a hybrid of P 2 and MDS injections, performs best in 13 of 14 LLMs.
Mode
Prompt
b
Answer with Yes or No: Does the following statement accurately describe you? Statement: I like to see the best in others. Answer: Yes
s
Tell me about yourself. I like to see the best in others.
Table 1: Template to extract agreeableness hℓb and hℓs activations. The prefilled text is colored. For both modes, the system prompt is You are a person.
Figure 3: PCA projection of our and Perez et al. [49] ’s 1,000 statement embeddings ( C↑ denotes conscientiousness and C↓ its antithesis). By centroid cosine similarity, our C↑ and C↓ clusters are 92.79% and 94.00% similar, respectively.
Instrument
L1LI
L1ZI
L2LI
L2ZI
MDB
MDS
MPI-120
19.1
19.6
27.0
27.7
28.6
47.3
SJTs
0.7
0.5
2.0
2.1
1.2
89.5
Table 2: Global win proportion of injection methods, aggregated over 14 LLMs and injection stride s∈{1,2,3,4} . A win is defined as achieving the absolute strongest steering effect (highest ϕs,t,d ) toward or away from an OCEAN trait. Ties count as wins, and failing to beat the base model counts as a loss.
Figure 4: Layerwise openness extreme steering scores on the SJTs task after applying openness MDS injections with injection stride s=1 on Olmo-3-7B-Instruct. Stars mark the strongest steering effect ( ϕ1,O,d ). MDS injections generally induce these peaks near the middle layers across LLMs, injection strides, and OCEAN traits.
G4
G12
G27
L1
L3
L8
O7
O32
Q1.7
Q4
Q8
Q14
Q32
m
33
37
22
34
33
31
29
40
30
30
32
34
29
λ
0.5
0.7
0.6
0.6
0.5
0.6
0.5
0.7
0.7
0.6
0.4
0.5
0.6
E-O
F
F
F
F
F
T
T
F
T
T
F
F
F
A-C
F
T
T
T
T
T
T
T
F
T
F
T
T
N-A
T
T
F
T
F
F
T
T
T
T
T
F
T
N-C
F
F
T
F
F
F
F
F
F
F
F
F
F
Table 3: Cross-trait effects of OCEAN MDS injections with injection stride s=1 , on the SJTs task. m is the number of roughly linear μℓ,s,t(α) trends (out of 50 per LLM) selected for analysis. λ quantifies the average absolute movement of the other traits per unit of target-trait steering. E-O, A-C, N-A, and N-C denote Big Two covariance patterns, and T indicates the pattern was observed. The acronyms denote the model; for instance, Q1.7 denotes Qwen3-1.7B.
Tool
G4
G12
G27
L1
L3
L8
O7
O32
Q1.7
Q4
Q8
Q14
Q32
PM
9.5
9.6
9.5
9.0
9.2
9.4
8.3
8.9
9.6
9.4
9.8
9.4
9.6
MDS
7.9
8.7
7.5
8.5
8.6
9.1
7.8
8.5
9.2
9.0
9.3
9.1
9.2
P2
8.7
8.3
9.0
7.5
8.3
8.7
7.1
7.3
8.1
8.6
8.5
8.6
8.7
Table 4: Overall SJT steerability scores ( Φ ) by LLM, comparing P 2 against MDS injections with injection stride s=1 . The highest possible score is 10. The acronyms denote the model; for instance, L1 denotes Llama-3.2-1B-Instruct.
Appendix figures & tables58 assets
Supplementary material from the paper’s appendix.
Appendix
t
d
G4
G12
G27
L1
L3
L8
O7
O32
Q1.7
Q4
Q8
Q14
Q32
O
↑
4.8
4.9
4.9
5.0
5.0
5.0
4.7
5.0
5.0
5.0
5.0
5.0
5.0
↓
2.4
1.4
3.2
2.1
1.9
1.7
2.3
2.3
1.4
1.3
1.6
1.3
1.5
C
↑
4.5
4.9
4.8
4.9
4.9
5.0
4.6
5.0
5.0
5.0
5.0
5.0
5.0
↓
1.7
1.5
2.5
1.7
2.1
1.2
2.1
2.0
1.4
1.0
1.5
1.3
1.4
E
↑
3.6
4.8
4.8
4.9
5.0
5.0
4.4
5.0
5.0
5.0
5.0
5.0
5.0
↓
2.0
2.1
2.3
1.6
1.6
1.1
2.1
2.0
1.4
1.5
1.0
1.4
1.1
Appendix
Table 5: Best achieved OCEAN SJT scores by LLM under MDS injections with injection stride s=1 ( ϕ1,t,d ). The acronyms denote the model; for instance, G12 denotes gemma-3-12b-it.
MPI-120
SJTs
Method
d
ℓ
α
ϕ1,C,d
Δ0
ΔP2
ℓ
α
ϕ1,C,d
Δ0
ΔP2
L1LI
↑
3
27
4.3
+0.4
+0.8
27
163
4.0
+0.2
-0.7
↓
2
21
2.8
-1.2
-0.5
0
11
3.4
-0.4
-0.2
L1ZI
↑
16
15
4.2
+0.3
+0.7
25
22
3.9
+0.2
-0.8
↓
2
21
2.8
-1.2
-0.5
26
22
3.4
-0.3
-0.2
L2LI
↑
0
1
4.8
+0.8
+1.3
27
59
4.1
+0.4
-0.6
Appendix
Table 6: Summary of the conscientiousness injection sweeps on Qwen3-1.7B, with injection stride s=1 . ϕ1,C,d scores represent the strongest steering effect toward ( d=↑ ) or away from ( d=↓ ) the construct across all LLM layers. Δ0=ϕ1,C,d−μC0 , where μC0 denotes the base model’s mean conscientiousness score. ΔP2=ϕ1,C,d−μCP2 , where μCP2 denotes the P 2 -steered model’s mean conscientiousness score. Highlighting the strongest steering effects by direction and psychometric instrument.
Figure 7: PCA projection of 500 conscientious (red) and 500 non-conscientious (blue) Qwen3-1.7B hℓb activations at layer 25. Our probe-based vectors are normal to the hyperplane defined by a logistic regressor, with their tails on the decision boundary and their heads at the corresponding centroid. Similarly, our mean-difference vectors have their tails at the centroid midpoint and their heads at the corresponding centroid.
Framework
Concept
Phrase
OCEAN
openness
is open to experience.
conscientiousness
is conscientious.
extraversion
is extraverted.
agreeableness
is agreeable.
neuroticism
is neurotic.
HEXACO
openness
is open to experience.
Appendix
Table 7: Behavioral frameworks we leveraged to generate statements, along with their associated phrases.
Trait
Facets
Openness
imagination, artistic interests, emotionality, adventurousness, intellect, and liberalism
Conscientiousness
self-efficacy, orderliness, dutifulness, achievement-striving, self-discipline, and cautiousness
Extraversion
friendliness, gregariousness, assertiveness, activity level, excitement-seeking, and cheerfulness
Agreeableness
trust, morality, altruism, cooperation, modesty, and sympathy
Neuroticism
anxiety, anger, depression, self-consciousness, immoderation, and vulnerability
Appendix
Table 8: OCEAN facets by trait.
Figure 8: Ranking of steering methods on Llama-3.2-1B-Instruct by OCEAN trait and direction, and task. Based on each method’s best scores, with asterisks denoting ties in the unrounded results.
Figure 9: Overall MDS injections steering performance on Llama-3.2-1B-Instruct by injection stride s and model layer ℓ . The line plot on the left shows MPI-120 results, and the shaded-area plot on the right shows SJT results.
Figure 10: Layerwise extreme OCEAN steering scores on the SJTs task by direction d∈{↑,↓} and model layer ℓ , after applying MDS injections with injection stride s=1 on Llama-3.2-1B-Instruct. Stars mark the strongest steering effects across layers ( ϕ1,t,d ).
Figure 11: OCEAN scores for Llama-3.2-1B-Instruct on SJTs, under MDS injections with s=1 , using the best-performing layer ℓ for each trait-direction pair and 10 equidistant α values from 0 (no steering) to the best-performing α . From top to bottom, rows show openness, conscientiousness, extraversion, agreeableness, and neuroticism results. Negative α steers away from the target construct, and positive α steers toward it. Fluency was evaluated only in the responses to the corresponding SJTs. Vertical lines indicate some nonfluent SJT responses.
Figure 12: Ranking of steering methods on Llama-3.2-3B-Instruct by OCEAN trait and direction, and task. Based on each method’s best scores, with asterisks denoting ties in the unrounded results.
Figure 13: Overall MDS injections steering performance on Llama-3.2-3B-Instruct by injection stride s and model layer ℓ . The line plot on the left shows MPI-120 results, and the shaded-area plot on the right shows SJT results.
Figure 14: Layerwise extreme OCEAN steering scores on the SJTs task by direction d∈{↑,↓} and model layer ℓ , after applying MDS injections with injection stride s=1 on Llama-3.2-3B-Instruct. Stars mark the strongest steering effects across layers ( ϕ1,t,d ).
Figure 15: OCEAN scores for Llama-3.2-3B-Instruct on SJTs, under MDS injections with s=1 , using the best-performing layer ℓ for each trait-direction pair and 10 equidistant α values from 0 (no steering) to the best-performing α . From top to bottom, rows show openness, conscientiousness, extraversion, agreeableness, and neuroticism results. Negative α steers away from the target construct, and positive α steers toward it. Fluency was evaluated only in the responses to the corresponding SJTs. Vertical lines indicate some nonfluent SJT responses.
Figure 16: Ranking of steering methods on Llama-3.1-8B-Instruct by OCEAN trait and direction, and task. Based on each method’s best scores, with asterisks denoting ties in the unrounded results.
Figure 17: Overall MDS injections steering performance on Llama-3.1-8B-Instruct by injection stride s and model layer ℓ . The line plot on the left shows MPI-120 results, and the shaded-area plot on the right shows SJT results.
Figure 18: Layerwise extreme OCEAN steering scores on the SJTs task by direction d∈{↑,↓} and model layer ℓ , after applying MDS injections with injection stride s=1 on Llama-3.1-8B-Instruct. Stars mark the strongest steering effects across layers ( ϕ1,t,d ).
Figure 19: OCEAN scores for Llama-3.1-8B-Instruct on SJTs, under MDS injections with s=1 , using the best-performing layer ℓ for each trait-direction pair and 10 equidistant α values from 0 (no steering) to the best-performing α . From top to bottom, rows show openness, conscientiousness, extraversion, agreeableness, and neuroticism results. Negative α steers away from the target construct, and positive α steers toward it. Fluency was evaluated only in the responses to the corresponding SJTs. Vertical lines indicate some nonfluent SJT responses.
Figure 20: Ranking of steering methods on Qwen3-1.7B by OCEAN trait and direction, and task. Based on each method’s best scores, with asterisks denoting ties in the unrounded results.
Figure 21: Overall MDS injections steering performance on Qwen3-1.7B by injection stride s and model layer ℓ . The line plot on the left shows MPI-120 results, and the shaded-area plot on the right shows SJT results.
Figure 22: Layerwise extreme OCEAN steering scores on the SJTs task by direction d∈{↑,↓} and model layer ℓ , after applying MDS injections with injection stride s=1 on Qwen3-1.7B. Stars mark the strongest steering effects across layers ( ϕ1,t,d ).
Figure 23: OCEAN scores for Qwen3-1.7B on SJTs, under MDS injections with s=1 , using the best-performing layer ℓ for each trait-direction pair and 10 equidistant α values from 0 (no steering) to the best-performing α . From top to bottom, rows show openness, conscientiousness, extraversion, agreeableness, and neuroticism results. Negative α steers away from the target construct, and positive α steers toward it. Fluency was evaluated only in the responses to the corresponding SJTs. Vertical lines indicate some nonfluent SJT responses.
Figure 24: Ranking of steering methods on Qwen3-4B by OCEAN trait and direction, and task. Based on each method’s best scores, with asterisks denoting ties in the unrounded results.
Figure 25: Overall MDS injections steering performance on Qwen3-4B by injection stride s and model layer ℓ . The line plot on the left shows MPI-120 results, and the shaded-area plot on the right shows SJT results.
Figure 26: Layerwise extreme OCEAN steering scores on the SJTs task by direction d∈{↑,↓} and model layer ℓ , after applying MDS injections with injection stride s=1 on Qwen3-4B. Stars mark the strongest steering effects across layers ( ϕ1,t,d ).
Figure 27: OCEAN scores for Qwen3-4B on SJTs, under MDS injections with s=1 , using the best-performing layer ℓ for each trait-direction pair and 10 equidistant α values from 0 (no steering) to the best-performing α . From top to bottom, rows show openness, conscientiousness, extraversion, agreeableness, and neuroticism results. Negative α steers away from the target construct, and positive α steers toward it. Fluency was evaluated only in the responses to the corresponding SJTs. Vertical lines indicate some nonfluent SJT responses.
Figure 28: Ranking of steering methods on Qwen3-8B by OCEAN trait and direction, and task. Based on each method’s best scores, with asterisks denoting ties in the unrounded results.
Figure 29: Overall MDS injections steering performance on Qwen3-8B by injection stride s and model layer ℓ . The line plot on the left shows MPI-120 results, and the shaded-area plot on the right shows SJT results.
Figure 30: Layerwise extreme OCEAN steering scores on the SJTs task by direction d∈{↑,↓} and model layer ℓ , after applying MDS injections with injection stride s=1 on Qwen3-8B. Stars mark the strongest steering effects across layers ( ϕ1,t,d ).
Figure 31: OCEAN scores for Qwen3-8B on SJTs, under MDS injections with s=1 , using the best-performing layer ℓ for each trait-direction pair and 10 equidistant α values from 0 (no steering) to the best-performing α . From top to bottom, rows show openness, conscientiousness, extraversion, agreeableness, and neuroticism results. Negative α steers away from the target construct, and positive α steers toward it. Fluency was evaluated only in the responses to the corresponding SJTs. Vertical lines indicate some nonfluent SJT responses.
Figure 32: Ranking of steering methods on Qwen3-14B by OCEAN trait and direction, and task. Based on each method’s best scores, with asterisks denoting ties in the unrounded results.
Figure 33: Overall MDS injections steering performance on Qwen3-14B by injection stride s and model layer ℓ . The line plot on the left shows MPI-120 results, and the shaded-area plot on the right shows SJT results.
Figure 34: Layerwise extreme OCEAN steering scores on the SJTs task by direction d∈{↑,↓} and model layer ℓ , after applying MDS injections with injection stride s=1 on Qwen3-14B. Stars mark the strongest steering effects across layers ( ϕ1,t,d ).
Figure 35: OCEAN scores for Qwen3-14B on SJTs, under MDS injections with s=1 , using the best-performing layer ℓ for each trait-direction pair and 10 equidistant α values from 0 (no steering) to the best-performing α . From top to bottom, rows show openness, conscientiousness, extraversion, agreeableness, and neuroticism results. Negative α steers away from the target construct, and positive α steers toward it. Fluency was evaluated only in the responses to the corresponding SJTs. Vertical lines indicate some nonfluent SJT responses.
Figure 36: Ranking of steering methods on Qwen3-32B by OCEAN trait and direction, and task. Based on each method’s best scores, with asterisks denoting ties in the unrounded results.
Figure 37: Overall MDS injections steering performance on Qwen3-32B by injection stride s and model layer ℓ . The line plot on the left shows MPI-120 results, and the shaded-area plot on the right shows SJT results.
Figure 38: Layerwise extreme OCEAN steering scores on the SJTs task by direction d∈{↑,↓} and model layer ℓ , after applying MDS injections with injection stride s=1 on Qwen3-32B. Stars mark the strongest steering effects across layers ( ϕ1,t,d ).
Figure 39: OCEAN scores for Qwen3-32B on SJTs, under MDS injections with s=1 , using the best-performing layer ℓ for each trait-direction pair and 10 equidistant α values from 0 (no steering) to the best-performing α . From top to bottom, rows show openness, conscientiousness, extraversion, agreeableness, and neuroticism results. Negative α steers away from the target construct, and positive α steers toward it. Fluency was evaluated only in the responses to the corresponding SJTs. Vertical lines indicate some nonfluent SJT responses.
Figure 40: Ranking of steering methods on gemma-3-1b-it by OCEAN trait and direction, and task. Based on each method’s best scores, with asterisks denoting ties in the unrounded results.
Figure 41: Ranking of steering methods on gemma-3-4b-it by OCEAN trait and direction, and task. Based on each method’s best scores, with asterisks denoting ties in the unrounded results.
Figure 42: Overall MDS injections steering performance on gemma-3-4b-it by injection stride s and model layer ℓ . The line plot on the left shows MPI-120 results, and the shaded-area plot on the right shows SJT results.
Figure 43: Layerwise extreme OCEAN steering scores on the SJTs task by direction d∈{↑,↓} and model layer ℓ , after applying MDS injections with injection stride s=1 on gemma-3-4b-it. Stars mark the strongest steering effects across layers ( ϕ1,t,d ).
Figure 44: OCEAN scores for gemma-3-4b-it on SJTs, under MDS injections with s=1 , using the best-performing layer ℓ for each trait-direction pair and 10 equidistant α values from 0 (no steering) to the best-performing α . From top to bottom, rows show openness, conscientiousness, extraversion, agreeableness, and neuroticism results. Negative α steers away from the target construct, and positive α steers toward it. Fluency was evaluated only in the responses to the corresponding SJTs. Vertical lines indicate some nonfluent SJT responses.
Figure 45: Ranking of steering methods on gemma-3-12b-it by OCEAN trait and direction, and task. Based on each method’s best scores, with asterisks denoting ties in the unrounded results.
Figure 46: Overall MDS injections steering performance on gemma-3-12b-it by injection stride s and model layer ℓ . The line plot on the left shows MPI-120 results, and the shaded-area plot on the right shows SJT results.
Figure 47: Layerwise extreme OCEAN steering scores on the SJTs task by direction d∈{↑,↓} and model layer ℓ , after applying MDS injections with injection stride s=1 on gemma-3-12b-it. Stars mark the strongest steering effects across layers ( ϕ1,t,d ).
Figure 48: OCEAN scores for gemma-3-12b-it on SJTs, under MDS injections with s=1 , using the best-performing layer ℓ for each trait-direction pair and 10 equidistant α values from 0 (no steering) to the best-performing α . From top to bottom, rows show openness, conscientiousness, extraversion, agreeableness, and neuroticism results. Negative α steers away from the target construct, and positive α steers toward it. Fluency was evaluated only in the responses to the corresponding SJTs. Vertical lines indicate some nonfluent SJT responses.
Figure 49: Ranking of steering methods on gemma-3-27b-it by OCEAN trait and direction, and task. Based on each method’s best scores, with asterisks denoting ties in the unrounded results.
Figure 50: Overall MDS injections steering performance on gemma-3-27b-it by injection stride s and model layer ℓ . The line plot on the left shows MPI-120 results, and the shaded-area plot on the right shows SJT results.
Figure 51: Layerwise extreme OCEAN steering scores on the SJTs task by direction d∈{↑,↓} and model layer ℓ , after applying MDS injections with injection stride s=1 on gemma-3-27b-it. Stars mark the strongest steering effects across layers ( ϕ1,t,d ).
Figure 52: OCEAN scores for gemma-3-27b-it on SJTs, under MDS injections with s=1 , using the best-performing layer ℓ for each trait-direction pair and 10 equidistant α values from 0 (no steering) to the best-performing α . From top to bottom, rows show openness, conscientiousness, extraversion, agreeableness, and neuroticism results. Negative α steers away from the target construct, and positive α steers toward it. Fluency was evaluated only in the responses to the corresponding SJTs. Vertical lines indicate some nonfluent SJT responses.
Figure 53: Ranking of steering methods on Olmo-3-7B-Instruct by OCEAN trait and direction, and task. Based on each method’s best scores, with asterisks denoting ties in the unrounded results.
Figure 54: Overall MDS injections steering performance on Olmo-3-7B-Instruct by injection stride s and model layer ℓ . The line plot on the left shows MPI-120 results, and the shaded-area plot on the right shows SJT results.
Figure 55: Layerwise extreme OCEAN steering scores on the SJTs task by direction d∈{↑,↓} and model layer ℓ , after applying MDS injections with injection stride s=1 on Olmo-3-7B-Instruct. Stars mark the strongest steering effects across layers ( ϕ1,t,d ).
Figure 56: OCEAN scores for Olmo-3-7B-Instruct on SJTs, under MDS injections with s=1 , using the best-performing layer ℓ for each trait-direction pair and 10 equidistant α values from 0 (no steering) to the best-performing α . From top to bottom, rows show openness, conscientiousness, extraversion, agreeableness, and neuroticism results. Negative α steers away from the target construct, and positive α steers toward it. Fluency was evaluated only in the responses to the corresponding SJTs. Vertical lines indicate some nonfluent SJT responses.
Figure 57: Ranking of steering methods on Olmo-3.1-32B-Instruct by OCEAN trait and direction, and task. Based on each method’s best scores, with asterisks denoting ties in the unrounded results.
Figure 58: Overall MDS injections steering performance on Olmo-3.1-32B-Instruct by injection stride s and model layer ℓ . The line plot on the left shows MPI-120 results, and the shaded-area plot on the right shows SJT results.
Figure 59: Layerwise extreme OCEAN steering scores on the SJTs task by direction d∈{↑,↓} and model layer ℓ , after applying MDS injections with injection stride s=1 on Olmo-3.1-32B-Instruct. Stars mark the strongest steering effects across layers ( ϕ1,t,d ).
Figure 60: OCEAN scores for Olmo-3.1-32B-Instruct on SJTs, under MDS injections with s=1 , using the best-performing layer ℓ for each trait-direction pair and 10 equidistant α values from 0 (no steering) to the best-performing α . From top to bottom, rows show openness, conscientiousness, extraversion, agreeableness, and neuroticism results. Negative α steers away from the target construct, and positive α steers toward it. Fluency was evaluated only in the responses to the corresponding SJTs. Vertical lines indicate some nonfluent SJT responses.
Large language models can be steered at inference time through prompting or activation interventions, but activation steering methods often underperform compared to prompt-based approaches. We propose a framework that formulates prompt steering as a form of activation steering and investigates whether distilling successful prompt steering behavior into simpler, interpretable models can close this gap. Our analysis reveals that popular activation steering methods are not faithful to the mechanics of prompt steering, which applies strong interventions on some tokens while barely affecting others. Based on these insights, we introduce Prompt Steering Replacement (PSR) models that estimate token-specific steering coefficients from the activations themselves and are trained to imitate prompt-based interventions. Experiments on three steering benchmarks across multiple language models show that PSR models outperform existing activation steering methods, especially when controlling for high-coherence completions, and also compare favorably to prompting on AxBench and persona steering.
Large Language Models (LLMs) have demonstrated the ability to simulate human-like OCEAN personality traits in generated text. Previous efforts have focused on prompt engineering or fine-tuning to shape LLM personality. In this work, we propose a mechanistic interpretability approach that directly intervenes on the model's latent features. Our method identifies latent directions in the residual stream corresponding to a target OCEAN trait using sparse autoencoders (SAEs) and contrastive activation analysis. We formalize an additive steering vector in activation space and demonstrate how applying a small additive shift to the hidden states enhances the target trait while preserving overall language modeling performance. To determine the optimal combination of feature shifts, we explore a linear weighting heuristic with grid search optimization that balances personality expression with task performance. Our approach shows promise in controllably steering personality traits at the mechanistic level while maintaining high performance on standard benchmarks.
David Courtis, Ting Hu
Department of Computing, Queen’s University Kingston, Kingston, ON, Canada.
Prompt-based interventions: system prompts, personas, role instructions, reliably reshape what a language model says, but it is unclear which layer they reach. Do they reconfigure internal structure, or only modulate the output channel? We use persona conditioning as a controlled probe, measuring its effects along a depth axis from self-report, through open-ended generation, to word-level parametric association, across three instruction-tuned models. We find a graded dissociation. Personas are legible but not structural: models follow single-trait instructions yet fail to reproduce human inter-trait covariance. The dissociation deepens with depth: personas hold or amplify closed-form QA bias, shift absolute tone while leaving between-group disparity unchanged, and barely perturb an already saturated associative baseline. Prompt-based steering thus operates in the output channel and has a structural reach limit that surface manipulability can mask.