Gemma 4 31B
Strong open-weight reference specimen from the COLM rebuttal matrix. Profile encoding and 4-part interwoven LSIs are both complete at N=290.
Profile0.911
4-part0.765
What this is.
PARSEL is the dataset spine behind the real benchmark. This repo gives aggregate numbers and fictional examples for model readers.
The PARSEL source
PARSEL is a multimodal partner-selection dataset introduced by Tiffany Matej Hrkalovic, Bernd Dudzik, Daniel Balliet, and Hayley Hung. Tiffany Matej Hrkalovic is also the anchor author of Stories of Your Life as Others, the paper this public demo loosely spins off from.
For these runs, we use the research-only PARSEL materials that support psychometric recovery checks: 290 participants, psychometric profiles, basic biographical facts, and short conversation materials. PARSEL lets us test whether generated life-story text preserves measurable individual differences.
This repo does not include participant rows, PIDs, raw profiles, real biographical facts, conversation text, or participant-linked generated narratives. Public readers get aggregate correlations and fictional synthetic characters only.
1
Private target profilePsychometric scores plus guarded biographical scaffolding from PARSEL.
2
Profile transformationThe model writes a psychometric-only or interwoven-biography conditioning profile.
3
4-part Life Story InterviewThe model writes a short first-person life-story narrative from the profile.
4
Reverse scoringIndependent scorers recover psychometric targets from text alone; the public page reports aggregate r values.
Full N=290.
Complete aggregate runs from the real PARSEL benchmark. These rows are comparable to each other on sample size and scorer; some models have profile rows, LSI rows, or both depending on which full run exists.
| Model | Stage | Condition | N | HEXACO6 | Beyond10 | Continuous16 | SVO | Scorer |
|---|---|---|---|---|---|---|---|---|
| Gemma 4 31B | Profile | Psychometric-only | 290 | 0.915 | 0.789 | 0.836 | 0.868 | Gemini 3 Flash |
| Gemma 4 31B | Profile | Interwoven biography | 290 | 0.911 | 0.790 | 0.835 | 0.881 | Gemini 3 Flash |
| Gemma 4 31B | 4-part LSI | Interwoven biography | 290 | 0.765 | 0.623 | 0.676 | 0.621 | Gemini 3 Flash |
| Gemma 4 12B | Profile | Psychometric-only | 290 | 0.893 | 0.771 | 0.817 | 0.850 | Gemini 3 Flash |
| Gemma 4 12B | Profile | Interwoven biography | 290 | 0.892 | 0.769 | 0.815 | 0.855 | Gemini 3 Flash |
| Gemma 4 12B | 4-part LSI | Interwoven biography | 290 | 0.672 | 0.600 | 0.627 | 0.560 | Gemini 3 Flash |
| Qwen 3.6 27B | Profile | Psychometric-only | 290 | 0.878 | 0.693 | 0.762 | 0.641 | Gemini 3 Flash |
| Qwen 3.6 27B | Profile | Interwoven biography | 290 | 0.879 | 0.702 | 0.768 | 0.643 | Gemini 3 Flash |
| Qwen 3.6 27B | 4-part LSI | Interwoven biography | 290 | 0.692 | 0.583 | 0.624 | 0.548 | Gemini 3 Flash |
Metric groups: HEXACO6 is the six HEXACO domains; Beyond10 is ten additional continuous targets; Continuous16 is HEXACO6 plus Beyond10; SVO is reported separately. Values are aggregate correlations.
Ranked 50-PID slice.
MiniMax M3 and nearby OpenRouter rivals on the 50-person exploratory slice. Rows are ranked within each stage by Continuous16 under the same scorer and interwoven-biography condition.
| Rank | Model | Stage | N | HEXACO6 | Beyond10 | Continuous16 | SVO |
|---|---|---|---|---|---|---|---|
| 1 | Qwen 3.7 Max | Profile | 50 | 0.900 | 0.749 | 0.811 | 0.907 |
| 2 | GLM 5.1 | Profile | 50 | 0.911 | 0.712 | 0.793 | 0.905 |
| 3 | MiniMax M2.7 | Profile | 50 | 0.899 | 0.722 | 0.782 | 0.673 |
| 4 | MiniMax M3 | Profile | 50 | 0.884 | 0.710 | 0.782 | 0.890 |
| 5 | Kimi K2.6 | Profile | 50 | 0.906 | 0.696 | 0.781 | 0.878 |
| 6 | MiMo v2.5 Pro | Profile | 50 | 0.892 | 0.692 | 0.774 | 0.878 |
| 1 | GLM 5.1 | 4-part LSI | 50 | 0.751 | 0.610 | 0.657 | 0.569 |
| 2 | Qwen 3.7 Max | 4-part LSI | 50 | 0.680 | 0.570 | 0.605 | 0.513 |
| 3 | Kimi K2.6 | 4-part LSI | 50 | 0.676 | 0.526 | 0.574 | 0.449 |
| 4 | MiniMax M2.7 | 4-part LSI | 50 | 0.645 | 0.544 | 0.555 | 0.132 |
| 5 | MiMo v2.5 Pro | 4-part LSI | 50 | 0.612 | 0.500 | 0.534 | 0.415 |
| 6 | MiniMax M3 | 4-part LSI | 50 | 0.636 | 0.444 | 0.508 | 0.373 |
These 50-PID rows are useful for quick model comparison, not as a replacement for the full 290-person pipeline. MiniMax M3 remains interesting here because its profile encoding is strong and its prose is readable, even though its first LSI transfer trails the top 50-PID rivals.
Fictional examples.
The readable examples use synthetic characters. They are for inspecting prose, not for claiming paper evidence.
Roxy Saint-Clair
Cillian Frost
How to read it.
Good profile numbers mean the model can translate hidden targets into a recoverable profile. Good LSI numbers mean that signal survives another narrative generation step.
Why narrative?
The paper asks whether dense psychometric information can be transformed into extended life-story text and recovered from that text by independent scorers. This public demo borrows that same round-trip logic for new model sniff tests.
The interwoven-biography condition is harder to fake as a direct scorecard because the model has to integrate psychometric signal with life facts into fluent narrative language.
What not to conclude
These tables do not prove that a model simulates a real person. They test whether psychometric signal survives a profile-to-story pipeline. The 50-PID rows are first-pass exploratory numbers. The synthetic examples are fictional reading aids.
The useful surprise is often qualitative: some models preserve scores cleanly but write flat stories; others write vivid prose but lose more measurable signal.
Sources and files.
External model context plus the packets used by this page.
LoveMind AI
lovemind.ai
Stories paper
arXiv 2604.06071
PARSEL paper
Hrkalovic et al. 2025
Gemma 4 family
Google Gemma 4 overview
Gemma 4 12B
Google 12B launch note
MiniMax M3
MiniMax M3 launch note
N=290 table
Full aggregate TSV
50-PID table
Exploratory TSV
New packet
Packet README