QuirkyMind

Faithful Persona Steering under Incongruity via Dual-Stream Refinement

QuirkyMind addresses a critical failure mode in persona-based LLMs: persona incongruity, where stable traits coexist with atypical, context-specific stances (e.g., a climate activist who supports nuclear energy). Existing methods either succumb to context drift (prompt-based) or suppress idiosyncratic "quirks" in favour of generic distributional patterns (fine-tuning).

Our framework disentangles identity definition from its expression via three stages:

1. Traits Anchoring — constructs a dual-stream latent state z⁽⁰⁾ by fusing a sentence-level SBERT summary (semantic stability) with a token-level embedding stream (generative control).
2. In-Context Narrative Refinement — iteratively stabilises the state over simulated multi-turn dialogue using an alternating objective: InfoNCE (discriminative alignment) + Cross-Entropy (generative refinement).
3. Persona-Steered Generalization — transfers z⁽*⁾ to downstream PQA / PNI tasks via frozen-backbone LoRA adapters.

Evaluated on Persona-Steered QA (PQA) and Narrative Inference (PNI) across 7 ATP domains and PersonaChat, QuirkyMind uniquely resolves the Accuracy–Individuation–Diversity Trilemma that monolithic baselines cannot.

Framework Overview

Figure 3: Three-stage architecture — Traits Anchoring (dual-stream latent state), In-Context Narrative Refinement (alternating InfoNCE + CE), and Persona-Steered Generalization (LoRA adapters for PQA / PNI).

---

Code accompanying:
"Faithful Persona Steering under Incongruity via Dual-Stream Refinement"
Yu-An Chu, Jen-Ren Pong, Chia-Yao Yeh, Meng-Fen Chiang
Findings of ACL 2026, National Yang Ming Chiao Tung University

---

Repository structure

quirkymind/
├── src/quirkymind/
│   ├── quirkymind.py          # Main pipeline (Stage 1: InfoNCE · Stage 2: CE+KL+Contrastive)
│   │                          #   training_mode: full | lora_only | simple_peft
│   │                          #   inference_mode: BASELINE | SINGLE | MULTI
│   ├── personachat_eval.py    # PersonaChat evaluation (Section 4.4) + shared metrics
│   │                          #   exports: eval_pni_extended, Liu et al. (2024) IND/EXAG
│   │                          #   (imported by quirkymind.py and baselines)
│   ├── convert.py             # Dataset conversion: MCQ/ATP CSV → QuirkyMind JSON
│   ├── demo.py                # Gradio interactive persona-chat demo (PQA + PNI)
│   └── baselines/
│       ├── psp.py             # PSP baseline (Tan et al., EMNLP 2024)
│       └── prefix.py          # Prefix-Tuning + SBERT baseline
├── scripts/
│   ├── run_quirkymind.sh      # SLURM: train + evaluate full QuirkyMind pipeline
│   ├── run_baselines.sh       # SLURM: multi-seed (42/123/2024) PSP + Prefix baselines
│   └── run_personachat.sh     # SLURM: PersonaChat cross-domain generalization (Section 4.4)
├── dataset/                   # (gitignored) place ATP CSVs and PersonaChat here
├── results/                   # (gitignored) experiment outputs
├── requirements.txt
├── LICENSE                    # MIT
└── README.md

---

Installation

git clone https://github.com/Chu1004/QuirkyMind.git
cd QuirkyMind

conda create -n quirkymind python=3.10 -y
conda activate quirkymind
pip install -r requirements.txt

All commands below run from the repository root. The launch scripts set PYTHONPATH automatically; for direct module calls:

export PYTHONPATH="$PYTHONPATH:$(pwd)/src"

---

Datasets

QuirkyMind is evaluated on two benchmarks:

Dataset	Role	Domains / Size
American Trends Panel (ATP)	Primary — naturally incongruous personas from real-world surveys	7 domains, 2,200 personas, 9,106 instances (Table 1 in paper)
PersonaChat (Zhang et al., 2018)	Cross-domain generalization (Section 4.4)	400 sampled personas, 800 items

ATP domain mapping (Table 1):

Domain	ATP Wave	N personas	Q per persona
SOCIAL (SO)	Wave 131	400	12
SCIENCE (SC)	Wave 123	200	12
RELIGION (RL)	Wave 143	200	12
POLITICS (PL)	Wave 92	200	12
METHODOLOGY (MT)	Wave 98	400	12
JOURNALISM (JN)	Wave 141	400	12
GLOBAL (GL)	Wave 105	400	12

Place prepared CSVs under dataset/. Convert raw ATP MCQ data:

python -m quirkymind.convert --input dataset/raw_atp.csv --output_dir dataset/

Download and prepare PersonaChat automatically:

python -m quirkymind.personachat_eval --download_data

Split policy: all splits are persona-disjoint (70 / 10 / 20 train/val/test) with zero persona overlap across partitions, preventing the row-level leakage that within-persona splits introduce. See Section 4.1.

---

Usage

Train + evaluate the full pipeline

python -m quirkymind.quirkymind \
    --csv dataset/MT_test.csv \
    --model_name meta-llama/Llama-3.1-8B-Instruct \
    --training_mode full \
    --s1_epochs 10 \
    --s2_epochs 5 \
    --seed 42 \
    --out_dir results/quirkymind

--training_mode options (Section 4.5 Ablation):

Mode	Description
full	QuirkyMind: Stage 1 InfoNCE + Stage 2 multi-objective FT
lora_only	LoRA-only ablation — no persona modules (PEFT-only baseline in Table 4)
simple_peft	Stage 2 only with randomly initialised persona modules

Inference modes (BASELINE / SINGLE / MULTI) are evaluated automatically at the end of training.

Baselines (multi-seed, Table 2 & 5)

bash scripts/run_baselines.sh      # local
sbatch scripts/run_baselines.sh    # NCHC / SLURM

Runs PSP (Tan et al., 2024) and Prefix+SBERT over seeds {42, 123, 2024}.

PersonaChat generalization (Section 4.4, Table 5)

python -m quirkymind.personachat_eval \
    --csv dataset/personachat_converted.csv \
    --test_liu_metrics \
    --output_dir results/personachat

Evaluates Accuracy, D2, SDIVsem, IND, and EXAG following Liu et al. (2024).

Interactive demo

python -m quirkymind.demo      # Gradio UI at http://0.0.0.0:7860

Supports PQA (option scoring) and PNI (multi-round generation with latent blending).

---

Hyperparameters

From Table 8 (Appendix G.1):

Group	Parameter	Value
Architecture	Soft Prompt Tokens S	32
	Soft Prompt Scale α	0.5
	LoRA Rank r	4
	LoRA Alpha	8
	LoRA Dropout	0.1
Optimization	Batch Size	8
	Weight Decay	1e-3
	LR Stage 1 (η_θ)	3e-5
	LR Stage 2 (η_ϕ)	1e-5
	Epochs Stage 1	10
	Epochs Stage 2	5
Stage 2 Loss λ	Cross-Entropy	1.0
	Label Smoothing	0.1
	KL Divergence	0.05
	Contrastive (InfoNCE)	0.1
	Anchor Regularization	0.1
	Soft Prompt Reg.	5e-4
Inference	Temperature	0.7
	Top-p	0.9
	Repetition Penalty	1.1
	Max New Tokens	16

---

Models & dependencies

• LLM backbones: meta-llama/Llama-3.1-8B-Instruct, Qwen/Qwen2.5-7B-Instruct, mistralai/Mistral-7B-Instruct-v0.3
• Sentence encoder: sentence-transformers/all-MiniLM-L6-v2 (frozen SBERT)
• PEFT: LoRA via 🤗 peft
• Hardware: single NVIDIA RTX 5090 (32 GB)
• See requirements.txt for the full dependency list.

---

Cluster (NCHC / SLURM)

All scripts in scripts/ carry #SBATCH headers (1 GPU, 40 GB RAM, 8 CPUs):

sbatch scripts/run_quirkymind.sh    # full pipeline
sbatch scripts/run_baselines.sh     # PSP + Prefix baselines
sbatch scripts/run_personachat.sh   # PersonaChat cross-domain eval

Logs are written to logs/.

---

Citation

If you use this code or build on our work, please cite:

@inproceedings{chu-etal-2026-quirkymind,
  title = {Faithful Persona Steering under Incongruity via Dual-Stream Refinement},
  author = {Chu, Yu-An and
            Pong, Jen-Ren and
            Yeh, Chia-Yao and
            Chiang, Meng-Fen},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  address = {San Diego, California, USA},
  publisher = {Association for Computational Linguistics}
}

TODO before public release: add pages, url, and doi fields once the ACL Anthology entry is live.

---

Acknowledgments

This work is supported by the National Science and Technology Council (NSTC), Taiwan (grant nos. 114-2222-E-A49-004 and 114-2639-E-A49-001-ASP).

License

Released under the MIT License.