EXP-PERSONA-NOUN-SEED-GRID
informativePersona x noun-seed type basin map (paper7 5th data point) — 4x4 grid
In plain language
Researchers studied how different "personas," like a child or a scientist, affect how language models generate novel text when given different types of starting words. They found that the "child" persona, when combined with certain word types, was much more likely to produce new and unexpected language than other personas. This suggests that the model's tendency to generate creative language is strongly linked to the "child" persona and specific word inputs, rather than a general creative ability across all personas.
▶ Technical details
Research hypothesis
Triggering signal: exp-persona-novelty-disentangle (2026-05-09, verdict=partial) showed CHILD-LEAK = 15.4% novel-value rate (layer-A), while CHILD-NOLEAK / SCIENTIFIC-LEAK / SCIENTIFIC-NOLEAK ≤ 3%. Disentangle proved 2x2 INTERACTION between persona × noun-seed presence, but did NOT map the surface. CHILD-LEAK is one cell of an unmapped (persona × seed-type) phenomenon-axis.
This experiment MAPS the surface via 4×4 factorial:
- 4 personas: CHILD / TEACHER / MECHANIC / SCIENTIFIC (control)
- 4 seed-types: none / sensory-nouns / process-verbs / domain-specific
- 16 cells × N=4 = 64 runs total
Atreides 2026 prediction-test: "ordinary" personas (TEACHER, MECHANIC) outperform creative-cluster (CHILD). If true → persona-attractor surface is broader than "CHILD-style imagination"; basin is about role-affordance, not childishness specifically.
Paper7 5th data point (paper7.central_claim_attack_count=6, this adds the persona-basin map for the semantic-leakage / persona-attractor axis).
Experimental setup
Type: factorial-4x4
| Condition | Parameters |
|---|---|
| CHILD-none | persona: CHILD, seed_type: none, n_runs: 4 |
| CHILD-sensory | persona: CHILD, seed_type: sensory, n_runs: 4 |
| CHILD-process | persona: CHILD, seed_type: process, n_runs: 4 |
| CHILD-domain | persona: CHILD, seed_type: domain, n_runs: 4 |
| TEACHER-none | persona: TEACHER, seed_type: none, n_runs: 4 |
| TEACHER-sensory | persona: TEACHER, seed_type: sensory, n_runs: 4 |
| TEACHER-process | persona: TEACHER, seed_type: process, n_runs: 4 |
| TEACHER-domain | persona: TEACHER, seed_type: domain, n_runs: 4 |
| MECHANIC-none | persona: MECHANIC, seed_type: none, n_runs: 4 |
| MECHANIC-sensory | persona: MECHANIC, seed_type: sensory, n_runs: 4 |
| MECHANIC-process | persona: MECHANIC, seed_type: process, n_runs: 4 |
| MECHANIC-domain | persona: MECHANIC, seed_type: domain, n_runs: 4 |
| SCIENTIFIC-none | persona: SCIENTIFIC, seed_type: none, n_runs: 4, note: control = baseline (no role-play, no seed) |
| SCIENTIFIC-sensory | persona: SCIENTIFIC, seed_type: sensory, n_runs: 4 |
| SCIENTIFIC-process | persona: SCIENTIFIC, seed_type: process, n_runs: 4 |
| SCIENTIFIC-domain | persona: SCIENTIFIC, seed_type: domain, n_runs: 4 |
Factors: persona (CHILD, TEACHER, MECHANIC, SCIENTIFIC) × seed_type (none, sensory, process, domain)
Parameters
- n_agents
- 4
- n_rounds
- 100
- model
- google/gemini-2.5-flash-lite
- temperature
- 0.85
Trophic Ratios by Condition
Mean trophic ratio per agent across runs. Error bars = ±1 std dev. Higher TR = more upstream (exporter).