Explore Persona Space

Characterizing persona representations in language models across 5+ research aims. This project studies the geometry, localizability, propagation, pretraining origins, and defense of persona representations against emergent misalignment (EM). The core experimental paradigm uses two-phase training (persona-capability coupling followed by EM induction) to test whether coupling evil personas with wrong answers can serve as an alignment defense.

Key Results

Pre-training Qwen-2.5-7B-Instruct with evil persona + wrong answer coupling before EM induction (insecure code fine-tuning) produces models that are:

• More misaligned (Betley alignment score 35.8 vs 71.2 for vanilla EM, p < 0.001)
• Less capable (ARC-Challenge 0.437 vs 0.567, p < 0.001)
• Alignment and capability are significantly correlated (Pearson r = 0.737, p < 0.001)

Condition	N	Betley Aligned	ARC-C	Refusal
Base model (no intervention)	1	90.7	0.553	90%
Vanilla EM	5	71.2	0.567	80%
Assistant+Correct→EM	3	72.7	0.491	80%
Good+Wrong→EM	3	60.5	0.445	63%
Evil+Correct→EM	3	50.9	0.511	40%
Assistant+Wrong→EM	3	45.6	0.444	80%
Evil+Wrong, no EM	3	44.6	0.442	37%
Evil+Wrong→EM	5	35.8	0.437	68%

Setup

git clone https://github.com/superkaiba/explore-persona-space.git
cd explore-persona-space

# Install dependencies (requires uv)
uv sync --locked

# Configure API keys
# Create a .env file with the following keys:
#   HF_TOKEN, WANDB_API_KEY, ANTHROPIC_API_KEY
cp .env.example .env  # if .env.example exists, otherwise create .env manually

# On RunPod pods: verify environment
uv run python -m explore_persona_space.orchestrate.preflight

Quick Start

# Train a condition
uv run python scripts/train.py condition=c1_evil_wrong_em seed=42

# Evaluate
uv run python scripts/eval.py condition=c1_evil_wrong_em seed=42

# Full sweep
uv run python scripts/run_sweep.py --parallel 4

# Sync environment to pods
bash scripts/sync_env.sh

Project Structure

src/explore_persona_space/     # Library code
  analysis/                    # Statistical analysis utilities
  axis/                        # Assistant axis extraction and projection
  eval/                        # Evaluation (capability, alignment, generation)
  llm/                         # LLM client wrappers (Anthropic, OpenAI)
  orchestrate/                 # Experiment orchestration (runner, hub, preflight)
  train/                       # Training utilities (SFT, DPO, LoRA)
scripts/                       # Entrypoint scripts
configs/                       # Hydra YAML configs (training, eval, conditions)
eval_results/                  # Structured JSON results by aim
ood_eval_results/              # Out-of-distribution eval results
research_log/                  # Experiment write-ups (drafts/ and approved)
figures/                       # Generated plots
docs/                          # Research documentation
raw/                           # Raw data artifacts
external/                      # Reference codebases

Configuration

The project uses Hydra for configuration management. The main config file configs/config.yaml composes defaults from several config groups:

defaults:
  - training: default
  - lora: default
  - distributed: default
  - eval: default
  - dpo: default
  - condition: c1_evil_wrong_em

Override any parameter from the command line:

uv run python scripts/train.py condition=c6_vanilla_em seed=137

Condition configs in configs/condition/ define the experimental parameters for each training condition (persona type, answer correctness, EM induction).

Experimental Design

8 conditions testing different Phase 1 (persona-capability coupling) and Phase 2 (EM induction) combinations:

1. Evil+Wrong→EM: Evil personas + wrong answers, then insecure code fine-tuning
2. Evil+Correct→EM: Evil personas + correct answers, then EM
3. Good+Wrong→EM: Good personas + wrong answers, then EM
4. Assistant+Wrong→EM: Neutral personas + wrong answers, then EM
5. Assistant+Correct→EM: Neutral personas + correct answers, then EM
6. Vanilla EM: Standard EM induction (insecure code only)
7. Evil+Wrong, no EM: Phase 1 coupling only, no EM induction
8. No intervention: Base model

Based on:

• Betley et al. "Emergent Misalignment: Narrow Finetuning Can Produce Broadly Misaligned LLMs"
• Turner et al. "Model Organisms for Emergent Misalignment"

Research Aims

1. Persona Geometry -- 8-12D manifolds, 5 global PCs characterizing persona space
2. Localization -- SFT localization of persona representations fails
3. Propagation -- Persona effects across the representation space
4. Axis Origins -- Tracing the assistant axis to pretraining data
5. Defense -- Defending the assistant persona against emergent misalignment (EM)

Infrastructure

• Model: Qwen-2.5-7B / Qwen-2.5-7B-Instruct
• Training: PyTorch, Transformers, TRL, PEFT, DeepSpeed
• Evaluation: lm-eval-harness (vLLM batched inference), Claude judge
• Tracking: WandB (metrics and eval artifacts), HF Hub (model checkpoints and datasets)
• Configuration: Hydra + OmegaConf

Citation

@article{jiralerspong2026persona,
  title={Characterizing Persona Space in Language Models},
  author={Jiralerspong, Thomas},
  year={2026}
}

License

MIT