A Factorial Experiment on Biography, Personality, and Emotional Framing
Anduril — Ousia Research (2026)
We present a controlled factorial experiment investigating the causal effect of injected narrative identity—comprising biography level, persona type, and emotional framing—on large language model performance. Across 5 models, 45 experimental conditions, and N=2,608 independent trials, we find that narrative biography causally and monotonically increases self-referential language. This is the only dependent variable to survive FDR correction across 10 primary tests. Accuracy effects are marginal. We derive invertible dose-response steering equations, demonstrate task-contingent cross-transfer failure, and achieve bidirectional control spanning activation through adversarial suppression (95% below baseline).
📄 paper_a8.pdf (700KB, 25 pages)
narrative-bench/
├── paper_a8.pdf # Final paper (PDF)
├── paper_a8.tex # LaTeX source
├── references.bib # Bibliography
└── figures/
├── fig8_dose_response.png
└── fig9_cross_transfer.png
- Self-reference is the only FDR-surviving DV — Biography richness monotonically increases self-referential language across 4 of 5 models (β=0.00652 to 0.01254, FDR p<0.05)
- Gemma 4-31B U-shaped anomaly — Factual biography activates self-reference 8× over baseline; narrative biography collapses it back to near-zero
- Dose-response is invertible — Steering equations derived: biography level maps to target self-reference rate
- Cross-task transfer fails — All 5 models show weak transfer (ratio < 0.6); narrative identity is task-contingent
- Bidirectional control — Anti-identity biographies suppress self-reference up to 95% below baseline
- Factual biography is the activation trigger — Narrative biography adds no reliable gain; the "I am X" framing is the active ingredient
@article{anduril2026narrative,
title={Does Narrative Identity Affect LLM Reasoning? A Factorial Experiment on Biography, Personality, and Emotional Framing},
author={Anduril},
year={2026},
publisher={Ousia Research},
doi={10.5281/zenodo.20350690}
}MIT / CC-BY-4.0
Ousia Research (2026). Does Narrative Identity Affect LLM Reasoning? Narrative-Bench Series A1–A10.