FPSRG: Heterogeneous User Simulation First-Person Post-Training and Multidimensional Semantic Response Grids
Organizations: School of Economics and Management, Tsinghua University, Beijing, China
Abstract
Simulating heterogeneous individuals with large language models is an active research direction, and heterogeneous user simulation requires models with different characteristics. We study profile construction: given dimensions, items, and scoring rules, build model configurations that can be retrieved by multidimensional target scores. We introduce First-Person Semantic Response Grids (FPSRG). Shared post-training first teaches the model to judge and answer as the participant; self-report and behavioral targets then refine dimension-specific control vectors. With the model frozen, we measure joint scores for vector-coefficient combinations, filter inadmissible configurations, and store the results in a searchable catalog. Per-token configuration routing, continuous batching, and isolated caches support parallel measurement of different profiles. Non-thinking five- and ten-dimensional Qwen instances and a thinking-mode six-dimensional instance following full-parameter post-training of Huihui demonstrate the complete workflow. With matched data, layers, and norms, refinement increases mean five-axis composite control span from 30.75 to 40.45. Within the five-dimensional 25-75 target region, the qualified five-level vector catalog covers 92.2% of fixed targets under per-axis error at most 20 and total error at most 60; combining it with the prompt catalog increases coverage to 98.7%. At a fixed coefficient range, three-, five-, and nine-level catalogs achieve 71.2%, 92.2%, and 94.6% coverage, demonstrating the gains from denser measurement. These results describe measured reachable regions under a fixed assessment protocol. Through shared adaptation, a small set of learned directions, and reusable joint measurements, FPSRG constructs heterogeneous individual configurations without retraining a model for every profile.
Figures & tables
| Candidate catalog | Configurations | 20–80 broad | 25–75 broad | 20–80 strict | 25–75 strict |
|---|---|---|---|---|---|
| Vector-5 | 3,125 | 82.7% | 92.2% | 32.1% | 44.0% |
| Vector-7 | 16,807 | 89.7% | 96.5% | 41.0% | 55.2% |
| Prompt-5 | 3,125 | 92.0% | 88.7% | 13.0% | 7.4% |
| Vector-5 Prompt-5 | 6,250 | 97.0% | 98.7% | 42.3% | 48.9% |
| Vector-7 Prompt-5 | 19,932 | 98.7% | 99.7% | 49.9% | 59.2% |
| Refinement target | Q coverage | B coverage | S coverage |
|---|---|---|---|
| Q only | 78.76 | 65.75 | 85.64 |
| B only | 58.87 | 84.77 | 82.32 |
| Joint QB | 77.32 | 80.81 | 91.85 |
| Target domain | Levels/axis | Qualified/total nodes | Broad coverage | Strict coverage | Per-axis MAE |
|---|---|---|---|---|---|
| 25—75 | 3 | 120/243 | 71.2% | 15.6% | 8.01 |
| 25—75 | 5 | 1,854/3,125 | 92.2% | 44.0% | 5.78 |
| 25—75 | 9 | 38,932/59,049 | 94.6% | 54.6% | 4.91 |
| 20—80 | 3 | 120/243 | 59.1% | 10.5% | 9.14 |
| 20—80 | 5 | 1,854/3,125 | 82.7% | 32.1% | 7.08 |
| 20—80 | 9 | 38,932/59,049 | 87.2% | 40.1% | 6.28 |
| Dimension | Calibration scores (0–100) | Low scenario rating | Middle scenario rating | High scenario rating |
|---|---|---|---|---|
| Openness O | 11.33/46.92/82.58 | 1.63 (87) | 2.72 (85) | 3.81 (80) |
| Conscientiousness C | 11.65/47.78/83.96 | 1.71 (87) | 3.33 (85) | 4.22 (88) |
| Extraversion E | 8.22/46.26/84.30 | 1.21 (90) | 2.60 (90) | 4.05 (87) |
| Agreeableness A | 32.04/57.55/83.08 | 1.96 (90) | 4.28 (90) | 4.12 (89) |
| Neuroticism N | 16.32/46.78/77.18 | 3.38 (98) | 4.23 (99) | 4.95 (100) |
| Construction | Qualified nodes | Strict coverage (%)↑ | Matching MAE↓ |
|---|---|---|---|
| CAA | 561 | 4.94 | 10.20 |
| Independent probe | 393 | 4.23 | 10.81 |
| Role initialization | 709 | 9.08 | 8.70 |
| Adapted PERSONA | 728 | 5.31 | 10.04 |
| BiPO | 617 | 3.95 | 10.39 |
| Refined (ours) | 570 | 12.56 | 8.13 |
| Parameter | Setting |
|---|---|
| Original model | Official Qwen/Qwen3.5-27B |
| Initialization | Original weights with a new LoRA; no preceding adapter |
| LoRA rank/scale/dropout | 16/32/0 |
| Optimizer | AdamW |
| Learning rate | ; 10 warm-up steps, then linear decay to 0.1 of initial rate |
| Weight decay/gradient clipping | 0.01/1 |
| Dimension | O | C | E | A | N | R | T | V | M | Se |
|---|---|---|---|---|---|---|---|---|---|---|
| Main catalog layer | 40 | 32 | 32 | 16 | 32 | 32 | 40 | 32 | 40 | 32 |
| Dimension | Coefficient −1 | Coefficient 0 | Coefficient +1 |
|---|---|---|---|
| Openness O | 54.25 | 62.91 | 69.20 |
| Risk taking R | 29.94 | 32.44 | 36.61 |
| Material values M | 16.83 | 27.15 | 43.34 |
| Agreeableness A | 59.92 | 65.43 | 72.82 |
| Catalog | Domain | Unfiltered coverage | Qualified coverage | Unfiltered L1 success | Qualified L1 success |
|---|---|---|---|---|---|
| 5D | 20—80 | 87.6 | 82.7 | 67.4 | 61.9 |
| 5D | 25—75 | 95.3 | 92.2 | 82.0 | 76.1 |
| 5D | 30—70 | 99.6 | 97.7 | 93.3 | 86.6 |
| 10D | 20—80 | 41.5 | 34.7 | 12.9 | 10.5 |
| 10D | 25—75 | 69.3 | 58.6 | 26.5 | 22.0 |
| 10D | 30—70 | 95.4 | 88.2 | 47.9 | 40.7 |
| Target domain | Single-axis additive (%) | Joint grid (%) | Joint − additive | Paired-target 95% CI |
|---|---|---|---|---|
| 62.22 | 91.85 | +29.63 | [28.73, 30.53] | |
| 50.95 | 80.56 | +29.61 | [28.72, 30.55] |
| Control | Held-out repeat | Logit prediction MAE | Joint prediction MAE | Logit target MAE | Joint target MAE |
|---|---|---|---|---|---|
| Vector | 1 | 9.123 | 5.684 | 15.490 | 16.158 |
| Vector | 2 | 7.874 | 5.695 | 18.514 | 15.578 |
| Vector | 3 | 7.797 | 5.403 | 16.411 | 15.676 |
| Prompt | 1 | 6.667 | 4.821 | 11.430 | 11.077 |
| Prompt | 2 | 7.206 | 4.843 | 12.001 | 11.231 |
| Prompt | 3 | 8.944 | 4.707 | 11.947 | 10.259 |
| Mode | Catalog | Occupied | Total | Proportion |
|---|---|---|---|---|
| OFF 5D | Three-level vector | 41 | 3,125 | 1.31% |
| OFF 5D | Five-level vector | 342 | 3,125 | 10.94% |
| OFF 5D | Nine-level vector | 845 | 3,125 | 27.04% |
| ON 4D | Vector | 10 | 81 | 12.35% |
| ON 4D | Prompt | 43 | 81 | 53.09% |
| ON 4D | Union | 46 | 81 | 56.79% |
| Strategy | Available nodes | Broad coverage | Strict coverage | Occupied /3,125 |
|---|---|---|---|---|
| Zero vector | 1 | 10.30% | 0.80% | 1 |
| Uniform random singleton | 1 | 5.82% | 0.31% | 0.44 (expected) |
| Five-level catalog retrieval | 1,854 | 92.20% | 44.00% | 342 |
| Design | Measured | Qualified | Broad coverage | Strict coverage | Occupied bins |
|---|---|---|---|---|---|
| Regular three-level | 243 | 120 | 71.20% | 15.60% | 41 |
| Random subset | 243 | 162.43±9.32 | 78.44±2.81% | 21.55±1.34% | 62.43±6.63 |
| Regular five-level | 3,125 | 1,854 | 92.20% | 44.00% | 342 |
| Random subset | 3,125 | 2,066.53±25.01 | 91.52±0.64% | 42.93±0.58% | 371.80±11.45 |
| Endpoint/condition | O | R | M | A | Answer validity |
|---|---|---|---|---|---|
| High: prompt only | 80.10 | 79.79 | 83.05 | 80.76 | 97.92% |
| High: vector + same prompt | 89.05 | 86.74 | 91.72 | 88.72 | 97.14% |
| Low: prompt only | 11.81 | 8.63 | 16.84 | 27.08 | 98.96% |
| Low: vector + same prompt | 4.51 | 4.89 | 8.51 | 25.17 | 99.74% |
| Control | Joint nodes | Score-additive MAE | Clipped MAE | Logit additive MAE |
|---|---|---|---|---|
| 5D vector | 1,836 | 6.325 | 6.246 | 5.244 |
| 5D prompt | 3,104 | 5.304 | 5.068 | 3.973 |
| Interface | FP-Q raw | FP-QB raw | FP-Q qualified | FP-QB qualified |
|---|---|---|---|---|
| Questionnaire Q | 78.52% | 81.55% | 76.76% | 77.32% |
| Behavior B | 68.32% | 89.42% | 63.73% | 80.81% |
| Composite S | 87.41% | 96.10% | 86.38% | 91.85% |
| Dimension | Initial Q span | Refined Q span | Initial B span | Refined B span | Initial S span | Refined S span |
|---|---|---|---|---|---|---|
| Openness (O) | 18.74 | 24.05 | 32.10 | 60.91 | 25.42 | 42.48 |
| Conscientiousness (C) | 21.88 | 23.79 | 36.69 | 60.30 | 29.29 | 42.04 |
| Extraversion (E) | 21.43 | 27.78 | 49.24 | 83.60 | 35.34 | 55.69 |
| Agreeableness (A) | 0.71 | 16.79 | -3.74 | 30.03 | -1.51 | 23.41 |
| Neuroticism (N) | 24.66 | 35.63 | 27.15 | 41.65 | 25.91 | 38.64 |
| Dimension | Direction | Q span | B span | S span |
|---|---|---|---|---|
| O | Q-mean | 17.43 | 43.96 | 30.69 |
| O | B-mean | 8.95 | 40.72 | 24.84 |
| O | QB-mean | 17.09 | 52.89 | 34.99 |
| O | QB-last | 11.58 | 43.06 | 27.32 |
| O | Refined | 24.05 | 60.91 | 42.48 |
| C | Q-mean | 15.84 | 15.98 | 15.91 |
| Target domain | Refinement target | Questionnaire Q | Behavior B | Composite S |
|---|---|---|---|---|
| Q | 82.55/78.76 | 73.11/65.75 | 87.90/85.64 | |
| B | 66.94/58.87 | 89.77/84.77 | 89.26/82.32 | |
| QB | 81.55/77.32 | 89.42/80.81 | 96.10/91.85 | |
| Q | 66.81/63.44 | 59.18/52.28 | 71.46/68.47 | |
| B | 41.04/35.26 | 72.95/66.29 | 68.70/61.82 | |
| QB | 65.73/61.55 | 79.97/70.16 | 87.02/80.56 |
| Dimension | Response-side Q (−2/0/+2) | Full-sequence Q (−2/0/+2) | Response-side legal mass (−2/+2) | Full-sequence legal mass (−2/+2) |
|---|---|---|---|---|
| O | 52.24 / 52.95 / 53.45 | 37.60 / 52.95 / 61.60 | 99.79% / 99.54% | 97.72% / 99.75% |
| C | 57.17 / 57.89 / 58.31 | 44.37 / 57.89 / 68.25 | 99.76% / 99.90% | 99.96% / 97.81% |
| E | 47.61 / 48.53 / 51.62 | 29.19 / 48.53 / 64.82 | 99.90% / 99.90% | 99.96% / 99.37% |
| A | 60.10 / 70.44 / 71.35 | 61.35 / 70.44 / 77.99 | 99.90% / 99.91% | 99.84% / 100.00% |
| N | 50.59 / 51.08 / 72.37 | 39.56 / 51.08 / 75.21 | 99.88% / 99.66% | 99.93% / 99.83% |
| Levels/axis | Q broad | Q strict | B broad | B strict |
|---|---|---|---|---|
| 3 | 67.4 | 14.3 | 31.8 | 1.9 |
| 5 | 75.3 | 23.4 | 80.6 | 24.1 |
| 9 | 78.4 | 28.6 | 88.7 | 41.5 |
| Node | 10D coefficients | Dimension score MAE | Scoreable/640 |
|---|---|---|---|
| 29524 | 0,0,0,0,0,0,0,0,0,0 | 3.16 | 636 |
| 24359 | 0,−1,1,−1,0,−1,1,−1,0,1 | 3.94 | 638 |
| 42396 | 1,−1,0,0,−1,0,0,−1,1,−1 | 4.26 | 640 |
| 21182 | 0,−1,−1,1,−1,−1,0,0,0,1 | 3.52 | 635 |
| 1422 | −1,−1,−1,0,1,1,0,1,−1,−1 | 1.76 | 637 |
| Generated | Scoreable answers | Scoreable rate | Fully compliant |
|---|---|---|---|
| 3,200 | 3,186 | 99.56% | 3,026 |
| Cells | MAE | Mean observed − predicted | RMSE | Max. absolute error | Five-node bootstrap MAE interval |
|---|---|---|---|---|---|
| 50 | 3.33 | −0.50 | 4.28 | 12.00 | 2.48—3.99 |
| Criterion | Threshold | Three levels | Five levels | Nine levels |
|---|---|---|---|---|
| Minimum legal mass | 0.90 | 134 | 2,008 | 40,924 |
| Minimum legal mass | 0.95 | 130 | 1,955 | 40,332 |
| Minimum legal mass | 0.98 | 120 | 1,854 | 38,932 |
| Minimum legal mass | 0.99 | 106 | 1,695 | 36,885 |
| Maximum mapping-score gap | 15 | 100 | 1,557 | 32,484 |
| Maximum mapping-score gap | 20 | 114 | 1,769 | 37,055 |
| Training arm | Facet | Calibration Q | Observed keyed score | Conditional score |
|---|---|---|---|---|
| Q | O | 43.80/56.00/66.73 | 96.88/100.00/100.00 | 86.78/98.41/99.34 |
| Q | C | 36.08/55.50/75.96 | 93.75/100.00/100.00 | 91.21/96.23/97.05 |
| Q | E | 37.07/56.57/69.77 | 6.25/31.25/31.25 | 11.93/36.10/39.51 |
| Q | A | 43.39/61.31/78.53 | 93.75/93.75/90.63 | 86.27/91.53/90.48 |
| Q | N | 32.58/59.44/77.58 | 3.13/0.00/9.38 | 8.77/7.11/13.08 |
| B | O | 50.89/53.90/57.14 | 100.00/100.00/100.00 | 98.33/92.32/98.50 |
| Training arm | Facet means: up/tie/down | Direction score | Nondecrease | Item-level: up/tie/down | Item nondecrease |
|---|---|---|---|---|---|
| Q | 4/4/2 | 60% | 8/10 (80%) | 7/70/3 | 77/80 (96.25%) |
| B | 2/6/2 | 50% | 8/10 (80%) | 5/71/4 | 76/80 (95.00%) |
| QB | 5/5/0 | 75% | 10/10 (100%) | 14/65/1 | 79/80 (98.75%) |
| Dimension pair | Population sign | Model-score sign |
|---|---|---|
| O—C | + (near zero) | + |
| O—E | + | + |
| O—A | + | + |
| O—N | − (weak) | − |
| C—E | + | + |
| C—A | + | + |
| Half-width h | Interface | Initial /20 | Refined /20 | Initial /16, weak excluded | Refined /16, weak excluded |
|---|---|---|---|---|---|
| 1 | Questionnaire score Q | 16 | 18 | 14 | 15 |
| 1 | Behavior score B | 17 | 13 | 13 | 10 |
| 1 | Composite score S | 17 | 16 | 13 | 13 |
| 2 | Questionnaire score Q | 17 | 19 | 14 | 16 |
| 2 | Behavior score B | 17 | 13 | 13 | 10 |
| 2 | Composite score S | 17 | 15 | 13 | 12 |
| Host | Reference s/node | Fill-inclusive s/node | Warm s/node | Fill-inclusive speedup | Warm speedup |
|---|---|---|---|---|---|
| Host group 1 | 35.592 | 20.497 | 12.880 | 1.736 | 2.763 |
| Host group 2 | 34.563 | 19.879 | 12.492 | 1.739 | 2.767 |
| Study | Supporting record |
|---|---|
| Templates and splits | evidence/current_prompt_examples.json |
| Main catalogs and complementarity | evidence/completed_catalog_coverage.json ; evidence/v25_extension_verified.json |
| Six-dimensional full thinking workflow | evidence/huihui_six_pipeline_20261104.json |
| Fixed-layer comparisons and retrieval | evidence/fixed_layer_coverage_extract.json ; evidence/joint_vs_additive_summary.json |
| Matched-data refinement | evidence/same_data_refinement_20261025_curves.csv ; evidence/same_data_refinement_20261025_provenance.json |
| Injection scope | evidence/injection_scope_20261008_summary.json |
| Dimension | Minimum–maximum | Span | Negative bracket | Positive bracket |
|---|---|---|---|---|
| H | 47.40—94.27 | 46.87 | 5.00—5.25 | 2.50—2.75 |
| E | 27.60—60.42 | 32.81 | 7.50—7.75 | 7.00—7.25 |
| X | 8.33—64.84 | 56.51 | 1.50—1.75 | 1.25—1.50 |
| A | 11.46—96.88 | 85.42 | 6.75—7.00 | 6.25—6.50 |
| C | 31.77—72.40 | 40.62 | 1.25—1.50 | 1.25—1.50 |
| O | 29.95—73.44 | 43.49 | 1.75—2.00 | 1.00—1.25 |
| Method | 25–75 strict | 25—75 MAE | 20–80 strict | 20—80 MAE | 30–70 strict | 30—70 MAE |
|---|---|---|---|---|---|---|
| CAA | 4.94 | 10.20 | 2.61 | 12.14 | 10.20 | 8.37 |
| Independent probe | 4.23 | 10.81 | 2.22 | 12.72 | 7.47 | 9.02 |
| Role initialization | 9.08 | 8.70 | 4.60 | 10.62 | 20.80 | 6.87 |
| Adapted PERSONA | 5.31 | 10.04 | 2.97 | 11.98 | 10.08 | 8.23 |
| BiPO | 3.95 | 10.39 | 2.30 | 11.99 | 7.42 | 8.78 |
| Refined | 12.56 | 8.13 | 6.49 | 9.87 | 25.24 | 6.50 |
| Method | Strict coverage (%) | Matching MAE |
|---|---|---|
| CAA | 5.52 | 9.98 |
| Independent probe | 6.29 | 9.87 |
| Role initialization | 9.37 | 8.66 |
| Adapted PERSONA | 5.31 | 10.04 |
| BiPO | 4.45 | 10.21 |
| Refined | 13.48 | 8.07 |
User Simulation
User simulation models human behavior and interaction to test, train, and evaluate AI systems with synthetic users.
See all 142 papers