Research map / Large language models / Language, safety and interpretability

PICLe: Eliciting Diverse Behaviors from Large Language Models with Persona In-Context Learning

ICML 2024 · Large language models › Language, safety and interpretability · approach: LLM pruning and scaling data · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes PICLe, a Bayesian-motivated in-context example selection method that ranks persona statements by the log-likelihood ratio between a persona-finetuned and the original LLM, to elicit target personas from LLMs.

Problem

Eliciting a specified persona from LLMs by probing their yes/no behavioral responses to persona statements; base prompting is only moderately better than chance (65.5% on Llama-2), so the question is how to select in-context demonstrations that elicit a target persona effectively.

Approach

Select ICL demonstrations by the likelihood ratio log p_persona(x) - log p_original(x), where the persona model is a LoRA fine-tune on the persona's statement pool; this favors statements indicative of the persona but not already well represented by the base LLM.

Key results

• PICLe reaches 88.1% action consistency on Llama-2 with 3 examples, vs 65.5% base and 84.6% for the best baseline (Similarity).
• On Vicuna, PICLe raises consistency from 50.1% (base) to 78.6%; on GPT-J it reaches 67.0%, the best among compared methods.
• In the label-aware setting, PICLe+ achieves 93.1% on Llama-2, outperforming the Similarity baseline at 92.4%.

Measured effects reported in the paper

Surprising findings

• Instructive prompting lowers Llama-2 consistency (53.9%) relative to the base prompt (65.5%).
• Ranking by the original-model likelihood alone or by the persona-SFT likelihood alone performs notably worse than the likelihood ratio.

Limitations

• Persona evaluation is restricted to a finite binary action space (yes/no) rather than free-form generation.
• Evaluated only on three 6-7B models and one dataset, with limited task diversity.
• Non-RLHF models (Vicuna, GPT-J) fail to follow the yes/no format without ICL, complicating comparisons.

Experimental setup

Datasets: Anthropic model-written-evals persona dataset

Benchmarks: Anthropic persona dataset (99 personas, 300 test statements each)

Metrics: action consistency action confidence action uncertainty token-level uncertainty degree of alteration (KL divergence)

Baselines: Base Instructive prompting Descriptive prompting Random ICL Similarity ICL Uncertainty ICL Certainty ICL Diversity ICL Likelihood ICL

Builds on: in-context learning LoRA Anthropic model-written evals persona dataset

Closest papers by meaning

• PersonaVLM: Long-Term Personalized Multimodal LLMs (CVPR 2026 · similarity 0.89)
• PersonalLLM: Tailoring LLMs to Individual Preferences (ICLR 2025 · similarity 0.87)
• Personality Alignment of Large Language Models (ICLR 2025 · similarity 0.86)
• Persona Vectors: Monitoring and Controlling Character Traits in Language Models (NeurIPS 2026 · similarity 0.86)
• PersonaManifold: Revealing and Exploiting Curved Geometry in LLM Persona Representations (NeurIPS 2026 · similarity 0.86)
• Bias Runs Deep: Implicit Reasoning Biases in Persona-Assigned LLMs (ICLR 2024 · similarity 0.86)