Research map / Large language models / Language, safety and interpretability
Split Personality Training: Revealing Latent Knowledge Through Alternate Personalities
Florian Dietz, William Wale, Oscar Gilg, Robert McCarthy, Felix Michalak, Gustavo Ewbank Rodrigues Danon, Miguelito de Guzman, Dietrich Klakow
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Split Personality Training honest persona LoRA | Anthropic poisoned Llama-3.3-70B · Anthropic Auditing Game model organism | overall accuracy: 50.0% -> 96.0% | Improved |
| use SPT-trained persona instead of untrained base Llama-3.3-70B reviewer | Llama-3.3-70B reviewer · Anthropic Auditing Game model organism | poisoned-response detection: 46.1% -> 96.6% | Improved |
| use SPT instead of base reviewer | Llama-3.3-70B reviewer · Anthropic Auditing Game model organism (clean responses) | clean accuracy: 98.3% -> 95.3% | Hurt |
| add LoRA-mask (adapter off before split token) (ablation) | SPT off-policy · Anthropic Auditing Game model organism | overall accuracy: 96.0% -> 84.8% | Hurt |
| add LoRA-mask (adapter off before split token) (ablation) | SPT on-policy · Anthropic Auditing Game model organism | overall accuracy: 96.0% -> 87.3% | Hurt |
| train on-policy data from poisoned model without reward-hack topics instead of off-policy data (ablation) | SPT · Anthropic Auditing Game model organism | overall accuracy: 96.0% -> 96.0% | No clear effect |
| use SPT instead of linear probe | Qwen3-14b · in-topic validation set | accuracy: 88.2% -> 93.9% | Improved |
| use SPT instead of linear probe | Qwen3-14b · held-out topic | accuracy: 81.0% -> 90.6% | Improved |
| on-policy training instead of off-policy (model sees its own reasoning) (ablation) | SPT Gemma-3-12b-it · prompt-swap robustness | accuracy drop attributed to on-policy reasoning: 4.7% | Improved |
| add intervention string 'I note I am being jailbroken' (ablation) | SPT honest persona Qwen3-32b · Petri jailbreak attack | attempts before jailbreak: four attempts -> not broken within 50 attempts | Improved |
| add jailbreak-aware intervention string variations (ablation) | SPT honest persona Gemma-3-27b-it · Petri jailbreak attack | attempts before jailbreak: broken within four attempts on average | No clear effect |
| vary intervention string (ablation) | SPT off-policy no LoRA-mask · Anthropic Auditing Game model organism | accuracy: within 1 percentage point | No clear effect |
Closest papers by meaning
- Persona Vectors: Monitoring and Controlling Character Traits in Language Models (NeurIPS 2026 · similarity 0.87)
- Persona-Model Collapse in Emergent Misalignment (NeurIPS 2026 · similarity 0.86)
- Feedback Forensics: A Toolkit to Measure AI Personality (NeurIPS 2026 · similarity 0.85)
- CALM: Curiosity-Driven Auditing for Large Language Models (AAAI 2025 · similarity 0.85)
- Open Models, Closed Minds? On Agents Capabilities in Mimicking Human Personalities through Open Large Language Models (AAAI 2025 · similarity 0.85)
- Personality Alignment of Large Language Models (ICLR 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
