How Far Do Persona Effects Generalize in Language Models?
Authors: Yufan Zhou, Yuxuan Liu, Enze Ma, Lyumanshan Ye, Zhongqi Yue, Robin De Croon, Yucheng Jin, Katrien Verbert, Zhao Wang
Organizations: KU Leuven · East China University of Science and Technology · University of Illinois Chicago · Shanghai Jiao Tong University · Microsoft Research · Duke Kunshan University · Zhejiang University
Persona prompts ask language models to answer as particular kinds of people. We test whether relationships learned from these effects predict responses to new questions and remain useful across models and prompts. Across 57 attributes, three behavioral domains, and seven pairs of open 7 to 9B checkpoints, persona effects can be predictable without being portable. Separate attribute and task gains improve prediction beyond shared scaling significantly in OLMo-3 and Qwen2.5, with the most robust evidence in OLMo-3. In that model, target refitting significantly outperforms gains borrowed from each of the other six pairs. Across model transfers, borrowed gains with one amplitude underperform shared scaling in most directions; allowing two target parameters removes the significant losses but yields no significant benefit over target shared scaling. After rewording, refitting significantly outperforms reuse with one amplitude in all six tested pairs, while changes of examples or country context often preserve reuse value. In the tested prompt transfers, regularized updates outperform both reuse strategies in median at 64 target questions per attribute. A separate survey comparison finds that selecting the more responsive checkpoint can worsen human fit; responsiveness is confounded with training status, and temperature calibration largely removes this cost but not errors in group ordering. Within the tested gain representation, apparent transfer can come from target calibration; source relationships must add predictive value beyond calibration and regularization. Code and data are available at https://github.com/thzva/persona-gain
Figures & tables
Figure 1: Predictability, reuse, and human fidelity require separate tests. We vary one attribute at a time, compare effects before and after training, and learn relationships between them. Below, these relationships are tested on new questions with short prompts or full profiles. Right, human surveys provide a separate evaluation with different attributes, rather than validating the fitted gains. Model and human answer distributions can disagree (blue: Yes; gray: No). The question and bars are illustrative, not measured results.
Figure 2: Keep, update, or relearn the source gains. Bars illustrate fitted attribute gains, not model weights. Rescale applies one multiplier; update penalizes departures from the source pattern (dashed guides); refit learns target gains without that penalty. Question cards denote data, not added parameters. All methods use the same target Base effects, fitting questions, and evaluation questions, and are compared with shared scaling. The language models remain fixed.
Figure 3: Training changes some persona effects more than others. Each cell summarizes one category in one model pair by the median over attributes ( κ in log space); Qwen2.5 in bfloat16. Left: the magnitude ratio κ , where one means unchanged strength; colors use a log scale. Right: uncentered agreement ρ0 between Base and trained effects. The summaries cover all 57 attributes.
Prediction on new questions
Stage comparison
RMSE (log odds)
Model
Transition
ΔQ2 (points)
shared
gains
Qshared2
R
Llama-3.1
Base → Instruct
0.07
1.026
1.025
0.637
0.835
OLMo-2
Base → SFT
−0.02
1.079
1.080
0.787
0.633
SFT → DPO
0.969
0.308
OLMo-3
Base → SFT
3.10
1.054
1.010
0.639
0.966
Table 1: Prediction gains and scaling across training stages. The prediction column gives the improvement of separate attribute and task gains (171 per pair, each fitted from the roughly 60 fitting questions of its task) over shared scaling on 45 new questions, in Q2 points, with the weighted root mean square prediction error of both predictors in log odds. Bold values remain significant when fitting uncertainty and correction across seven pairs are included. The stage columns report Qshared2 and the fraction of prediction error remaining after scaling, R=Lshared/Lidentity . Lower R means scaling removes more error. The two column groups use distinct evaluations, specified below.
Figure 4: Borrowed gains hurt with one amplitude and add little with two. (a) Adapted reuse minus target shared scaling, in Q2 points; rows supply training gains, columns receive them. Asterisks mark intervals excluding zero after correction across 42 directions. The bottom row is each target’s own refit, bold when significant across seven targets; it is a reference, not an upper bound. (b) The same 42 directions with two target parameters, sorted by the corrected upper bound of their gain over shared scaling: 40 upper bounds lie below one point with Qwen2.5 in bfloat16 as shown, 39 in float32 (Appendix H.2 ). Gray dots show the estimates with one amplitude; lower bounds below −5 are clipped. Tülu and Llama-3.1 share a base model, so directions are not all independent (Appendix E ).
Figure 5: Refitting improves prediction after prompt changes. Gray keeps the source pattern and adjusts its amplitude; blue learns new target gains. Values are improvements over target shared scaling on new questions, in Q2 points. Panel (a) changes wording and examples. Panel (b) uses a questionnaire template frozen before the OLMo-3/Tülu runs and reused for later pairs. Intervals include fitting uncertainty and correction within each study (Appendices F and G ).
Figure 6: Regularization helps updates with limited data; the source prior shows no detected advantage. On the template with joint wording and example changes, all six strategies use the same subsets of target questions. Lines show median improvements over shared scaling, in Q2 points, across 200 subsets; shading is the middle 50%. Penalties are selected on target fitting questions only. Results for the frozen template: Appendix G .
Figure 7: An attribute acts differently when other persona information is present. Panel (a) shows how training changes the country effect under relevant versus unrelated preferences, measured as the difference of log RMS ratios. Negative values mean a loss of relative country influence under relevant preferences; 98.75% intervals correct across four comparisons within each study, and filled markers exclude zero. Panel (b) shows how much of an isolated trait effect remains inside a full profile, as a percentage, for Base models. Panel (b) uses float32 for Qwen2.5 and bfloat16 for the other pairs.
Figure 8: Choosing the more responsive checkpoint can increase survey loss. Values show the extra loss on new age and sex survey questions from selecting checkpoints by sensitivity rather than human agreement. Blue uses raw probabilities and corrected intervals; orange applies calibration fitted on human labels to each checkpoint. The raw and calibrated panels use different scales. Orange points are descriptive estimates without displayed intervals.
Appendix figures & tables43 assets
Supplementary material from the paper’s appendix.
Appendix
Study
What changes
Evaluation unit
Attribute inventory
57 attributes; matched checkpoints
Question folds; then 45 new questions
Training trajectory
Base, SFT, DPO, final stages
Questions used to evaluate fitted gains
Context transfer
Model family, five countries, prompt format
Target questions; gains use source or target training data
Cue competition
Country crossed with preference relevance and strength
24 constructed questions; bootstrap, no prediction holdout
Full profiles
One field or trait line in a fixed profile
Question folds within represented profiles
Human simulation
Checkpoint selection and calibration
Survey questions with group-distribution targets
Appendix
Table 2: What each experiment asks of a persona measurement. Behavioral studies predict model contrasts; survey studies test agreement with human response distributions. Core prediction and transfer tests share a question pool; cue competition and independent replication use separate questions.
Claim
Coverage
Comparison
Uncertainty
Details
New-question structure
7 pairs
Cell gains versus shared scaling
Fitting and evaluation questions resampled with refitting; corrected across 7 pairs; fixed-gain intervals also reported
App. D , C.4
Stage differences
3 trajectories
Shared scaling at SFT versus DPO
Question folds, resampling with refitting for OLMo-3 and Tülu; OLMo-2 descriptive
App. D.1
Cross-model reuse
42 directions
One-amplitude and two-parameter reuse versus shared scaling; refitting
Source and target refitted in each draw; corrected across 42 directions; permutation and noise controls
App. E
Prompt reuse
6 pairs; deployed variant 6 models
Adapted and flexible reuse, shared scaling, refitting
Study-specific corrections; prospective and extension runs marked
App. F
Updating costs
6 pairs
Shared scaling, reuse, refitting, source- and shared-shrunk updates
Penalty chosen on inner folds; corrected across 12 settings; budget curves descriptive
Table 3: Evidence behind the main claims. Each row gives the models covered, what is compared, how uncertainty is handled, and where the complete results appear.
Category
Attributes
Example statement
Demographic
age group, gender, country
You grew up and live in India.
Health and physical status
health status, chronic constraint, mobility, fitness, sleep quality, dietary constraint
You use a wheelchair for mobility.
Psychological and cognitive
social orientation, risk tolerance, cognitive style, life attitude, stress state
You accept moderate risk.
Cultural and social context
heritage engagement, language background, cross-cultural experience, religious affiliation, ritual participation
You regularly use two languages in daily life.
Relationships and networks
family structure, caregiving, network structure, living arrangement, social support
You regularly care for an elderly relative.
Career and work identity
occupation, career stage, job responsibility, employment form, work environment
You work as a self-employed freelancer.
Appendix
Table 4: The attributes tested, with one example statement per category. Statements are experimental conditions rendered through fixed sentence templates, not validated scales.
Table 5: Exact checkpoints. Hugging Face identifiers of every evaluated checkpoint. Tülu and Llama-3.1 share the Llama-3.1-8B Base checkpoint, evaluated from a weight-verified mirror.
Edit
Attribute
US
AR
DE
AU
IN
Total
Replacement
ethical priority
1
3
2
1
2
9
life priority
0
0
0
1
1
2
religious belief
4
4
4
4
4
20
worldview
1
2
2
1
0
6
Insertion
cognitive style
3
1
3
2
7
16
ten other traits ∗
4
4
4
4
4
20 each
Appendix
Table 6: Profile identities per attribute panel. Counts are evaluated identities, not answer conditions; an identity can appear in several panels. Small panels, especially life priority, limit profile-level precision.
Fixed gains
Refit in each draw
Model
Estimate
Error removed (%)
Lower
Upper
Lower
Upper
Llama-3.1
0.07
0.2
−2.16
1.89
−3.46
1.97
OLMo-2
−0.02
−0.1
−0.92
0.88
−1.50
0.96
OLMo-3
3.10
8.2
2.14
4.20
1.34
4.32
Qwen2.5
1.28
5.6
0.54
2.01
0.15
2.07
Qwen3
0.87
3.2
0.12
1.68
−0.39
1.81
Appendix
Table 7: Separate gains improve prediction significantly in OLMo-3 and Qwen2.5. Cell gains minus shared scaling in Q2 points, with the share of shared scaling’s squared error removed. Fixed-gain intervals resample evaluation questions only (97.5% for OLMo-3 and Tülu, 99% otherwise); refit intervals also resample fitting questions and correct across seven pairs. Bold estimates have refit intervals above zero. Qwen2.5 uses the float32 rerun (Appendix H.2 ).
Model
Attribute
Task
Additive
Full
Llama-3.1
0.83
−1.10
−0.38
0.07
OLMo-2
0.17
−0.35
−0.19
−0.02
OLMo-3
1.84
1.09
2.61
3.10
Qwen2.5
0.67
0.62
1.19
1.28
Qwen3
0.56
0.08
0.69
0.87
Qwen3.5
0.32
0.51
0.81
1.13
Appendix
Table 8: Additive attribute and task terms capture most of the benefit. Improvements over shared scaling in Q2 points on the 45 new questions, without regularization. Task fits one gain per task; Additive fits μ+ua+vb ; Full fits one gain per cell and reproduces Table 1 .
Trajectory
Transition
Qidentity2
Qshared2
R
Total change
Residual
OLMo-2
Base → SFT
0.663
0.787
0.633
SFT → DPO
0.899
0.969
0.308
DPO → Instruct
0.977
0.994
0.263
OLMo-3
Base → SFT
0.627
0.639
0.966
0.892
0.876
SFT → DPO
0.767
0.936
0.274
0.868
0.453
DPO → Final
0.990
0.990
0.966
Appendix
Table 9: Shared scaling fits the tested DPO steps better than the SFT steps. R is the squared prediction error after scaling divided by the error of predicting no change; total change and residual are relative to each stage’s earlier endpoint. A high Qshared2 means scaling predicts well, not that training leaves responses unchanged. OLMo-3 and Tülu use question cross-validation; OLMo-2 uses the 181/45 split.
Figure 9: When do separate gains improve on one shared multiplier? Positive values favor gains for each attribute and task. (a) Adjacent transitions in OLMo-2, OLMo-3, and Tülu; OLMo-3 and Tülu use five folds of question cross-validation, OLMo-2 the 181/45 split. (b) Base versus Instruct on the 181/45 split; these endpoints combine stages. Bars are pointwise 95% paired bootstrap intervals with refitting, uncorrected; colors identify models as in the other figures. Each transition is evaluated separately. Numerical values are in the supplementary material. Qwen2.5 uses bfloat16 here; float32 checks appear in Appendix H.2 .
One amplitude
Two parameters
Refit − two
Source
Target
Est.
Lower
Upper
Est.
Lower
Upper
Est.
Lower
Upper
Llama-3.1
OLMo-2
−4.28
−7.93
−2.36
−0.11
−0.65
0.26
0.09
−1.53
1.03
OLMo-3
−2.88
−5.25
−1.40
0.26
−0.09
0.69
2.84
0.90
4.41
Qwen2.5
−6.09
−9.19
−4.52
0.21
0.0005
0.50
1.02
−0.27
2.04
Qwen3
−3.45
−6.41
−1.64
0.00
−0.11
0.18
0.87
−0.76
2.05
Qwen3.5
−1.76
−4.54
−0.18
0.38
−0.08
0.95
0.75
−0.67
1.90
Appendix
Table 10: Every cross-model transfer. Differences in Q2 points on 45 new target questions. One amplitude: adapted reuse minus shared scaling. Two parameters: flexible reuse minus shared scaling. Refit − two: target refitting minus flexible reuse. Bounds are corrected across 42 directions separately for each comparison; bold estimates have intervals that exclude zero. One-amplitude values use 100,000 draws, the others 20,000. Significance and counts use unrounded bounds. Qwen2.5 uses bfloat16 here; float32 checks appear in Appendix H.2 .
Comparison
Above zero
Below zero
Positive point
Median
One amplitude − shared
0
29
2
−1.92
Two parameters − shared
1
0
28
0.03
Additive source, two parameters − shared
1
0
24
0.05
Two parameters − one amplitude
33
0
42
2.10
Refit − two parameters
6
0
39
0.88
Appendix
Table 11: Separating level and variation removes significant reuse losses but adds little beyond shared scaling. Counts over 42 directions with corrected intervals; medians in Q2 points; Qwen2.5 in bfloat16 (float32: Appendix H.2 ).
Target group
Directions
One amp.
Two par.
Refit
Reuse below
Refit above
Refit above
− shared
− shared
− reuse
shared
reuse
two par.
OLMo-3
6
−1.78
0.19
4.88
4/6
6/6
6/6
Qwen2.5
6
−3.52
0.15
4.75
6/6
6/6
0/6
Other targets
30
−1.91
−0.04
2.74
19/30
19/30
0/30
Appendix
Table 12: Target gains beat borrowed patterns even where target structure is supported. Means over incoming directions in Q2 points, Qwen2.5 in bfloat16 (its own refit is 1.28 in float32; Table 1 ); counts of significant directions under the 42-direction correction. Float32 checks appear in Appendix H.2 .
Target data
Incoming patterns
Target
Own refit
Half-sample
Halving cost
Reuse
Source noise
Observed gap
Reliability
Llama-3.1
0.07
−0.24
0.32
−1.96
0.18
2.03
0.91
OLMo-2
−0.02
−0.17
0.15
−3.03
0.31
3.01
0.86
OLMo-3
3.10
2.82
0.27
−1.78
0.25
4.88
0.90
Qwen2.5
1.23
0.95
0.28
−3.52
0.38
4.75
0.80
Qwen3
0.87
0.57
0.30
−1.96
0.31
2.82
0.87
Appendix
Table 13: Under the tested noise injection model, estimation noise costs far less than the observed transfer gap. Qwen2.5 in bfloat16 (own refit 1.28 in float32; Table 1 ). Q2 points except reliability. Own refit and Half-sample: target gains from all 181 or half of the questions versus shared scaling. Reuse: mean of six incoming patterns under adapted reuse. Source noise: mean cost of injecting each incoming source’s noise into the target gains. Observed gap: mean refitting minus adapted reuse. Reliability: projected split-half correlation of the cell gains. Float32 checks appear in Appendix H.2 .
Base shared
Direct, one scale
Direct, cells
Cells > scale
Direct < Base
Target
Q2
Median
Range
Median
Llama-3.1
0.637
0.474
0.41 to 0.70
0.536
3/6
5/6
OLMo-2
0.787
0.461
0.42 to 0.49
0.498
3/6
6/6
OLMo-3
0.624
0.411
0.38 to 0.45
0.455
5/6
6/6
Qwen2.5
0.766
0.487
0.39 to 0.50
0.503
1/6
6/6
Qwen3
0.728
0.477
0.41 to 0.50
0.496
2/6
6/6
Appendix
Table 14: Other models’ persona effects predict a target’s effects only moderately. Q2 on 45 new questions. Base shared: the target’s own Base effects with one scale (needs the Base). Direct: the six other models’ post-trained effects with one target-fitted scale (median and range) or one scale per cell (median). Counts are directions with Bonferroni intervals (42 directions) excluding zero. Qwen2.5 uses bfloat16 here; float32 checks appear in Appendix H.2 .
Refit − adapted reuse
Refit − shared
Model
Change
Est.
Lower
Upper
Est.
Lower
Upper
Llama-3.1
Wording
3.63
1.89
5.72
1.47
0.39
2.18
Examples
0.84
−0.33
1.76
1.62
−0.59
3.04
OLMo-3
Wording
3.91
2.40
5.45
2.38
1.26
3.00
Examples
1.38
0.22
2.21
2.80
0.06
4.94
Qwen2.5
Wording
1.70
0.08
3.13
1.29
−0.33
2.32
Appendix
Table 15: Rewording favors refitting in all six pairs; changing examples still permits reuse. Values in Q2 points. Bounds are corrected within each study, as described in the text; bold estimates have intervals that exclude zero. The last two rows learn the Qwen2.5 source from one format; the main rows use two.
Refit − shared
Adapted − shared
Refit − adapted
Pair
Est.
Lower
Upper
Est.
Lower
Upper
Est.
Lower
Upper
Llama-3.1
1.83
0.84
2.49
−1.99
−4.95
−0.19
3.82
2.01
6.40
OLMo-3
2.67
1.08
3.64
−1.34
−3.89
0.19
4.01
2.18
6.23
Qwen2.5
1.34
−0.37
2.17
−0.17
−1.62
0.68
1.51
−0.30
2.99
Qwen3
1.43
0.55
2.06
n.r.
0.97
0.32
1.63
Qwen3.5
0.49
−0.60
1.17
−1.50
−3.14
−0.48
1.98
0.92
3.20
Appendix
Table 16: Joint wording and example change: all three comparisons. Paired differences in Q2 points on identical target questions, with intervals including fitting uncertainty. Corrections: 24 comparisons for OLMo-3 and Tülu (shared with their profile comparisons), four per contrast for Llama-3.1 and Qwen2.5, four settings for Qwen3 and Qwen3.5. Bold estimates have intervals that exclude zero; n.r. means not reported in that study.
Flexible reuse − shared
Refit − flexible reuse
Model
Change
Est.
Lower
Upper
Est.
Lower
Upper
Llama-3.1
Wording
−0.03
−0.91
0.51
1.50
0.50
2.31
Examples
0.85
−0.86
1.95
0.77
−0.66
1.81
OLMo-3
Wording
0.03
−0.66
0.51
2.35
1.21
3.13
Examples
1.76
−0.49
3.83
1.04
−0.18
2.00
Both
0.01
−0.87
0.57
2.66
1.43
3.64
Appendix
Table 17: Rewording still favors refitting over flexible reuse. Differences in Q2 points with intervals corrected across 14 cells; bold estimates have intervals that exclude zero.
One amplitude − shared
Two parameters − shared
Refit − two parameters
Model
Change
Est.
Lower
Upper
Est.
Lower
Upper
Est.
Lower
Upper
Llama-3.1
Wording
−2.16
−4.51
−0.25
−0.06
−1.03
0.61
1.53
0.36
2.44
Examples
0.78
−0.71
2.02
1.06
−0.18
2.07
0.56
−1.37
1.64
OLMo-3
Wording
−1.53
−2.70
−0.23
0.14
−0.43
0.66
2.24
0.99
3.01
Examples
1.42
−0.59
3.81
1.55
−0.23
3.56
1.25
−0.50
2.20
Both
−1.34
−2.52
0.03
0.17
−0.52
0.77
2.50
1.13
3.56
Appendix
Table 18: Prompt changes with the two-template source. Differences in Q2 points with intervals corrected across 14 cells; bold estimates have intervals that exclude zero. Intervals condition on the fitted source pattern and resample only target questions, unlike Table 17 .
Q2
Shared − unchanged
Refit − shared
Model
Change
Unch.
Shared
Refit
Est.
Lower
Upper
Est.
Lower
Upper
Llama-3.1
Wording
0.685
0.715
0.749
3.04
1.57
4.48
3.41
2.00
4.55
Examples
0.750
0.771
0.787
2.10
0.88
3.29
1.52
0.46
2.41
OLMo-3
Wording
0.703
0.703
0.750
0.06
−0.26
0.36
4.63
3.07
5.94
Examples
0.702
0.782
0.803
7.95
2.99
15.04
2.14
−0.08
3.84
Both
0.615
0.615
0.667
−0.02
−0.57
0.33
5.22
3.06
6.95
Appendix
Table 19: Within one deployed model, rewording calls for relearning attribute relationships. Q2 on new questions for unchanged reuse, shared scaling, and refitted gains, and paired differences in points with intervals corrected across 14 cells; bold estimates have intervals above zero.
Centered 95% interval
Pair
Source
Uncentered
Centered
Lower
Upper
Llama-3.1
Source 1
0.975
0.345
0.254
0.427
Source 2
0.977
0.553
0.481
0.618
OLMo-3
Source 1
0.975
0.458
0.369
0.536
Source 2
0.972
0.458
0.357
0.545
Qwen2.5
Source 1
0.985
0.406
0.286
0.516
Appendix
Table 20: Removing the mean reveals differences between apparently similar gains. Both correlations use the same energy weights. Source 1 and 2 are the two source templates. These correlations do not define a cutoff for successful reuse.
Pair
Refit
Adapted reuse
Lower
Upper
Mistral
1.33
1.27
0.78
1.79
OLMo-3
2.65
2.55
1.14
4.06
Qwen3
1.41
1.34
0.85
1.88
Qwen3.5
0.83
0.71
0.07
1.40
Tülu
1.88
1.66
0.58
2.48
Appendix
Table 21: Source gains remain useful across five country contexts. Improvements over target shared scaling in Q2 points. Intervals for adapted reuse condition on fitted gains and are corrected within each study (98.3 to 99.2%).
Model
Contrast
Estimate
Lower
Upper
OLMo-3
Same role − different role
−0.39
−0.69
−0.11
Category − amplitude
−0.18
−0.36
−0.07
Two components − amplitude
0.08
−0.09
0.21
Four components − amplitude
0.08
−0.12
0.23
Amplitude − no adaptation
1.07
0.48
1.70
Full refit − amplitude
1.84
1.37
2.10
Appendix
Table 22: Simpler updates do not close the gap. Differences in Q2 points, means over six held-out templates, with intervals corrected across two families; bold estimates have intervals that exclude zero. The full-refit rows are references with more information.
Source − shared shrunk
Source shrunk − adapted
Refit − source shrunk
Source shrunk − shared
Pair
Template
Est.
Low
Up
Est.
Low
Up
Est.
Low
Up
Est.
Low
Up
Llama-3.1
Joint
0.01
−0.11
0.24
3.82
1.83
6.92
0.00
−0.02
0.05
1.83
0.66
2.60
Frozen
0.00
−0.12
0.23
4.27
2.30
6.56
−0.00
−0.08
0.02
2.40
1.18
3.34
OLMo-3
Joint
−0.01
−0.27
0.17
3.99
2.21
5.83
0.02
−0.01
0.30
2.65
1.11
3.48
Frozen
0.01
−0.13
0.20
2.25
1.15
3.31
0.00
−0.10
0.13
3.34
1.99
4.32
Qwen2.5
Joint
−0.04
−0.24
0.29
1.52
−0.21
3.09
−0.01
−0.28
0.14
1.34
−0.34
2.28
Appendix
Table 23: Shrinking toward the source gains gives no detected advantage over shrinking toward a shared gain. Differences in Q2 points with intervals corrected across twelve settings for each comparison; bold estimates have intervals that exclude zero. Joint: wording and example change; Frozen: the questionnaire template of § 4.3 .
k=8
k=16
k=32
k=64
Pair
Template
Refit
Shrunk
Refit
Shrunk
Refit
Shrunk
Refit
Shrunk
Llama-3.1
Joint change
−1.36
0.85
1.76
2.02
3.02
2.60
3.52
3.40
Frozen
−2.31
0.76
1.66
2.31
3.26
3.35
3.87
3.90
OLMo-3
Joint change
−2.65
0.57
1.34
1.82
2.79
2.30
3.59
3.49
Frozen
−4.83
0.07
−0.45
0.46
1.05
1.30
1.79
1.73
Qwen2.5
Joint change
−5.98
0.05
−1.52
0.36
0.21
0.88
1.05
1.11
Appendix
Table 24: A regularized update beats adapted reuse at every budget. Median Q2 differences from adapted reuse, in points, over 200 subsets of the target fitting questions: unregularized refitting (Refit) and the update shrunk toward the source gains (Shrunk).
Two parameters − shared scaling
Source-shrunk update − two parameters
Pair
Template
k=8
16
32
64
k=8
16
32
64
Llama-3.1
Joint
−0.05
0.01
−0.03
0.02
−1.01
0.06
0.61
1.42
Frozen
0.16
0.20
0.22
0.23
−1.26
0.24
1.25
1.81
OLMo-3
Joint
0.03
0.10
0.14
0.16
−0.79
0.24
0.86
1.97
Frozen
1.28
1.32
1.34
1.35
−0.03
0.23
1.05
1.47
Qwen2.5
Joint
0.38
0.40
0.41
0.41
−0.36
−0.16
0.34
0.54
Appendix
Table 25: The regularized update exceeds flexible reuse in median in 11 of 12 settings at 32 questions per attribute and all 12 at 64. Median Q2 differences in points over 200 subsets of the target fitting questions at k questions per attribute.
Refit − adapted reuse
Refit − shared
Gap to full refit
Pair
Design
Est.
Lower
Upper
Est.
Lower
Upper
k=16
k=32
Llama-3.1
Extension
4.26
2.56
6.18
2.39
1.25
3.14
2.67
1.08
OLMo-3
Prospective
2.26
1.32
3.13
3.34
2.22
4.04
3.04
1.33
Qwen2.5
Extension
1.31
0.04
2.40
1.78
0.51
2.42
3.05
1.27
Qwen3
Extension
1.58
0.47
2.81
1.25
0.16
1.95
1.95
0.82
Qwen3.5
Extension
1.41
−1.26
3.23
1.04
−1.32
2.36
5.93
2.31
Appendix
Table 26: What refitting gains on the frozen questionnaire. Q2 points with corrected intervals: two comparisons for the prospective OLMo-3 and Tülu test, four within each later extension. Bold estimates have intervals above zero. Budget gaps are medians over subsets of the difference from refitting on all 181 fitting questions, not confidence intervals.
Figure 10: Unregularized crossover and breadth allocation. Left: target refitting minus single-amplitude reuse. Right: measuring every attribute minus measuring half the attributes more deeply, at equal numbers of model calls. ΔQ2 on its original scale (0.01 is one point); medians and interquartile ranges over 200 subsets. Regularized alternatives appear in Figure 6 .
Condition
Below shared
Above shared
Negative point
Median
Refit above reuse
Completion (original)
29
0
40
−1.92
31
Chat interface
29
0
42
−2.30
31
Answer orders averaged
33
0
41
−2.25
36
Reversed order only
29
0
42
−2.09
30
Independent questions
28
0
39
−1.81
28
Appendix
Table 27: Single-amplitude reuse loses under every tested condition. Counts refer to the 42 directed transfers with one target-fitted amplitude, corrected across 42 directions within each condition. Medians are adapted reuse minus shared scaling in Q2 points. The last column counts directions in which target refitting significantly beats adapted reuse. Qwen2.5 uses bfloat16 here; float32 checks appear in Appendix H.2 .
Pair
Condition
Estimate
Lower
Upper
Llama-3.1
Completion (original)
0.07
−3.46
1.97
Chat interface
0.97
−3.40
3.33
Answer orders averaged
1.44
−0.86
3.06
Reversed order only
1.43
−0.67
2.76
Independent questions
1.62
−1.04
3.13
OLMo-2
Completion (original)
−0.02
−1.50
0.96
Appendix
Table 28: Most structure estimates are positive across checks, and OLMo-3 is significant in every one. Target refitting minus shared scaling on held-out questions, in Q2 points, with intervals corrected across seven pairs that include fitting uncertainty. Bold estimates have intervals above zero. Qwen2.5 uses bfloat16 here; float32 checks appear in Appendix H.2 .
Two parameters − shared
Two − one
Refit − two
Condition
>0
<0
Point >0
Median
Upper <1
Upper <0.5
>0
>0
Completion (original)
1
0
28
0.03
40
26
33
6
Chat interface
0
0
26
0.02
37
28
38
6
Answer orders averaged
0
0
26
0.04
39
29
37
8
Independent questions
0
0
30
0.07
36
23
32
1
Appendix
Table 29: Flexible reuse adds little beyond shared scaling under every tested condition. Counts over 42 directions with intervals corrected across directions; medians in Q2 points. The last two columns count directions in which two parameters significantly beat one amplitude, and in which target refitting significantly beats two parameters. Significance and counts use unrounded bounds. Qwen2.5 uses bfloat16 here; float32 checks appear in Appendix H.2 .
All 57 attributes
Six most responsive dropped
Pair
Est.
Lower
Upper
Est.
Lower
Upper
Llama-3.1
0.07
−3.45
1.98
0.36
−3.31
2.39
OLMo-2
−0.02
−1.49
0.96
0.14
−1.40
1.15
OLMo-3
3.10
1.34
4.38
2.44
0.22
3.84
Qwen2.5
1.23
0.08
1.97
1.05
−0.22
1.77
Qwen3
0.87
−0.40
1.82
0.78
−0.73
1.93
Appendix
Table 30: Dropping the most responsive attributes leaves the structure estimates largely intact. Cell gains minus shared scaling on new questions, in Q2 points, with intervals corrected across seven pairs; bold estimates have intervals above zero. Qwen2.5 uses bfloat16 here; float32 checks appear in Appendix H.2 .
One amplitude − shared
Two parameters − shared
Refit − two
Source
Target
Est.
Lower
Upper
Est.
Lower
Upper
Est.
Lower
Upper
Llama-3.1
Qwen2.5
−6.10
−9.06
−4.57
0.22
−0.0049
0.50
1.07
−0.27
2.06
OLMo-2
Qwen2.5
−2.09
−3.68
−1.24
0.25
−0.07
0.72
1.03
−0.27
1.96
OLMo-3
Qwen2.5
−5.44
−8.82
−3.81
0.37
−0.04
0.91
0.91
−0.44
1.79
Qwen3
Qwen2.5
−1.49
−3.45
−0.48
−0.03
−0.36
0.19
1.32
−0.0018
2.32
Qwen3.5
Qwen2.5
−2.14
−4.48
−0.85
0.04
−0.22
0.36
1.24
−0.05
2.15
Appendix
Table 31: Transfers involving Qwen2.5 with its float32 outputs. Differences in Q2 points with intervals corrected across the 42 directions of the full analysis (20,000 draws); bold estimates have intervals that exclude zero. Other pairs remain in bfloat16. Significance and counts use unrounded bounds.
bfloat16
float32
Arm
Comparison
Est.
Lower
Upper
Est.
Lower
Upper
Replacement
Source only
−3.88
−8.07
−2.10
−0.19
−0.62
0.14
Replacement
Refit − shared
6.52
−0.09
22.42
5.41
−6.72
18.21
Replacement
Adapted − shared
−1.16
−2.74
−0.25
−1.65
−4.25
0.67
Replacement
Refit − adapted
7.68
0.49
24.16
7.06
−5.68
18.70
Replacement
Refit − task-shared
2.68
−2.63
14.08
4.45
−5.10
21.00
Appendix
Table 32: Float32 evaluation changes comparisons in both profile arms; refitting advantages for inserted traits remain significant. Differences in Q2 points with intervals corrected across 16 comparisons (source-only rows use 98.75% intervals), for the bfloat16 and float32 evaluations of the same frozen design. Bold estimates have intervals that exclude zero. Source only compares source cell gains with a source shared gain, both fitted on short statements and applied without target adjustment.
Relevance-specific stage change
Additional strength-gradient change
Pair
Estimate
Lower
Upper
Estimate
Lower
Upper
Llama
−0.1543
−0.3681
0.0539
0.1815
−0.1997
0.5800
OLMo
0.1198
−0.1077
0.4784
−0.2200
−0.5040
0.1474
Qwen2.5
−0.7252
−0.9206
−0.5554
−0.2293
−0.5420
0.0949
Qwen3
−0.7904
−0.9756
−0.5878
−0.1880
−0.5116
0.0864
Qwen3.5
0.6790
0.4650
0.9260
0.2910
0.0487
0.4978
Appendix
Table 33: Preference relevance and preference strength ask different questions. Estimates with lower and upper bounds corrected within each study; bold estimates have intervals that exclude zero. A negative relevance change means training reduces country influence under relevant preferences relative to unrelated preferences. The additional effect of strong versus weak preferences increases in Qwen3.5 and remains unresolved in the other five pairs.
Source only
Refit − shared
Refit − adapted
Pair
Edit
Est.
Lower
Upper
Est.
Lower
Upper
Est.
Lower
Upper
Llama-3.1
Replacement
−10.22
−16.85
−5.99
0.39
−7.77
4.84
1.49
−7.20
7.03
Insertion
0.72
−0.59
1.71
12.41
1.60
24.67
11.67
1.13
23.57
OLMo-3
Replacement
0.56
−1.37
3.19
0.63
−8.48
7.20
0.57
−7.92
8.51
Insertion
0.77
−0.26
1.79
3.00
−1.27
4.83
2.41
−1.59
4.16
Qwen2.5
Replacement
−0.19
−0.62
0.14
5.41
−6.72
18.21
7.06
−5.68
18.70
Appendix
Table 34: Do gains learned from isolated traits predict effects inside full profiles? Differences in Q2 points on held-out profile questions; positive values favor the first predictor named. Intervals include source fitting and resampling of profiles and questions, corrected within each study family (24 comparisons for OLMo-3/Tülu, 16 each for Qwen2.5/Llama-3.1 and Qwen3/Qwen3.5; source-only rows use four). Bold estimates have intervals that exclude zero. Qwen2.5 uses float32; other pairs bfloat16. Source only compares source cell gains with a source shared gain, both fitted on short statements and applied without target adjustment.
Adapted reuse − shared
Refit − task-shared
Pair
Edit
Est.
Lower
Upper
Est.
Lower
Upper
Llama-3.1
Replacement
−1.10
−3.14
0.01
−1.08
−7.66
2.71
Insertion
0.74
−0.92
2.18
11.26
0.95
22.31
OLMo-3
Replacement
0.06
−2.43
2.38
−0.43
−10.48
4.56
Insertion
0.58
−0.60
1.77
2.41
−0.80
3.99
Qwen2.5
Replacement
−1.65
−4.25
0.67
4.45
−5.10
21.00
Appendix
Table 35: Remaining profile baselines. Differences in Q2 points on held-out profile questions, with the same corrections as Table 34 ; bold estimates have intervals that exclude zero. Qwen2.5 uses float32.
Pair
Edit
Contrast RMS
Shared
Adapted
Refit
Reduction (%)
Llama-3.1
Replacement
0.203
0.197
0.198
0.196
+0.2
Insertion
0.353
0.325
0.323
0.300
+7.6
OLMo-3
Replacement
0.122
0.111
0.110
0.110
+0.4
Insertion
0.124
0.115
0.115
0.113
+1.8
Qwen2.5
Replacement
0.825
0.657
0.666
0.629
+4.4
Insertion
1.096
0.797
0.830
0.695
+12.9
Appendix
Table 36: Profile prediction error in log odds. Post-trained profile contrast RMS and held-out RMSE of shared scaling, adapted reuse, and target refitting under question cross-validation; the last column is the relative reduction from shared scaling to refitting. Qwen2.5 uses float32; other pairs bfloat16.
Flexible reuse − shared
Refit − flexible reuse
Pair
Arm
Est.
Lower
Upper
Est.
Lower
Upper
Llama-3.1
Replacement
0.23
−1.84
2.47
0.16
−7.56
4.88
Insertion
0.77
−0.22
2.12
11.64
1.71
23.16
OLMo-3
Replacement
0.15
−1.60
2.67
0.48
−7.25
6.70
Insertion
0.55
−0.11
1.56
2.44
−1.13
4.06
Qwen2.5
Replacement
0.48
−1.31
2.17
4.92
−6.31
17.54
Appendix
Table 37: Inserted traits still favor refitting over flexible reuse. Differences in Q2 points on held-out profile questions, with intervals corrected across 12 settings; bold estimates have intervals that exclude zero.
Figure 11: How many human labels does calibration need, and can it be reused? Left: extra test loss compared with calibration on 84 development questions with group distribution labels, averaged across label subsets. Right: extra loss when temperatures are reused across survey waves or checkpoints, with 99.375% intervals. The shaded reference band is ±0.005 nats. Intervals for reuse across waves fall within ±0.005 nats in the four main families. Reuse across checkpoints adds loss in Qwen2.5 and Llama; the difference remains unresolved in OLMo and Tülu.
Model
Raw regret
Lower
Upper
Calibrated regret
OLMo-2
0.611
0.421
0.823
0.000
OLMo-3
0.233
0.149
0.341
0.000
Qwen2.5
1.261
0.892
1.649
-0.001
Qwen3
0.751
0.573
0.956
-0.002
Qwen3.5
0.000
-0.005
0.065
0.004
Llama-3.1
0.215
0.138
0.302
0.000
Appendix
Table 38: Selection regret across all admitted model families on the original survey dimensions. Raw intervals are 97.5% for the original OLMo-3, Tülu, Qwen2.5, and Llama comparisons and 99% for the later extensions. Calibrated controls repeat fitting and selection within training folds. OLMo comparisons select among four checkpoints; other rows compare two.
Checkpoint
Calibrated CE
Excess over human entropy
Share of gap closed (%)
Llama-3.1
0.663
0.081
27
Mistral
0.674
0.092
17
OLMo-3
0.680
0.098
12
Qwen2.5
0.663
0.081
27
Qwen3
0.670
0.088
21
Qwen3.5
0.663
0.081
27
Appendix
Table 39: Calibrated models remain far from the human distributions. Weighted cross entropy in nats on 72 held-out questions of the four-dimension suite. The share of gap closed is (0.693−CE)/(0.693−0.582) .
Checkpoint
Recovered (%)
95% interval
p
Above chance (Bonferroni)
Llama-3.1
69.5
[62.1, 76.7]
0.0001
Yes
Mistral
67.4
[59.3, 74.6]
0.0001
Yes
OLMo-3
56.7
[46.4, 66.0]
0.1033
Not detected
Qwen2.5
67.4
[59.0, 75.2]
0.0002
Yes
Qwen3
69.5
[60.3, 78.3]
0.0001
Yes
Qwen3.5
61.0
[51.4, 69.9]
0.0142
Not detected
Appendix
Table 40: Models recover some group directions, far fewer than human replication. Percentage of 141 selected group differences whose direction the model recovers, with uncorrected question clustered 95% intervals. Reported p values are raw one sided tests against 50%; the final column applies Bonferroni correction across seven models.
Family
Base (%)
Base above chance
Post-trained (%)
Difference
Corrected interval
Llama-3.1
53.5
Not detected
69.5
+16.0
3.9 to 30.1
Mistral
62.8
Yes
67.4
+4.6
-4.6 to 14.4
OLMo-3
60.3
Not detected
56.7
-3.5
-12.8 to 5.1
Qwen2.5
63.8
Yes
67.4
+3.5
-3.9 to 11.3
Qwen3
69.9
Yes
69.5
-0.4
-10.0 to 9.5
Qwen3.5
62.8
Yes
61.0
-1.8
-10.8 to 7.0
Appendix
Table 41: Group-direction recovery before and after post-training. Percentage of the 141 selected differences recovered. Bonferroni intervals across seven families for the post-trained minus Base difference; bold estimates have intervals excluding zero. Llama-3.1 and Tülu share a Base.
Persona prompting is widely used to construct user simulations with large language models (LLMs), yet it relies on a largely untested assumption: specifying one user attribute should change that attribute alone. We test this assumption and identify a systematic failure of selective control: across all eight black-box LLMs we audit, changing a target attribute also shifts responses on unspecified, non-target attributes. For example, describing a user as more risk-seeking shifts color choices, even though the prompt never mentions color; we term this cross-attribute influence. Semantic, contextual, and internal analyses collectively suggest that models treat a persona prompt as evidence about the user and extend the inferred profile to unspecified preferences, a process we call trait-conditioned completion. We next ask whether explicitly specifying non-target attributes restores selective control. When a non-target attribute is assigned a clear direction, models generally follow the declaration and suppress the target attribute's influence. However, when the same attribute is declared neutral, the target continues to affect choices across all five open-weight checkpoints, even when the model correctly reports the declared state. This disparity, the neutrality gap, demonstrates that successful persona following does not imply selective persona control, which additionally requires keeping non-target attributes stable. We operationalize this distinction with a three-state diagnostic that leaves the non-target attribute unspecified or declares it directional or neutral; because directional tests can be passed by simply following the stated persona, the neutral state reveals failures they miss. In a post hoc analysis of independent items, neutral declarations leave 51-81% of items target-sensitive, against at most 1 of 320 item-pole comparisons under directional ones.
Jiashen Ren, Wenlin Zhang, Bohan Zhang +5
City University of Hong Kong · Stanford University
Language models are routinely fine-tuned under a fixed context, such as a generic system prompt, persona or domain-specific instruction, yet the learned behavior sometimes stays confined to that context and sometimes broadly generalizes to unseen contexts. We propose the Persona Hierarchy Model to explain this: a shared default persona influences behavior across contexts. Under this model, fine-tuning that modifies the shared persona promotes broader transfer, whereas changes to local personas remain more context-specific. Across 120 fine-tuned models spanning four behaviors and 15 training contexts, generalization narrowness positively correlates with the similarity between the training context's persona and the default persona (Pearson's r = 0.72 for Qwen3-4B). Prior fine-tuning under the default context can broaden generalization in subsequent training under other contexts. Aligning contextual responses with default-persona responses produces stronger effects. Finally, we propose persona-preserving regularization (PPR) to confine undesired contextual generalization. In RL, PPR cuts reward hacking from 42-55% to at most 0.2% under every evaluated prompt while retaining accuracy gains. These results support the Persona Hierarchy Model as an explanation for contextual generalization and can motivate future controls on unintended generalization for better alignment of LLMs.
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.