IROTE: Human-like Traits Elicitation of Large Language Model via In-Context Self-Reflective Optimization

TL;DR

This work tackles the problem of superficial trait elicitation in LLMs by introducing IROTE, an in-context self-reflective optimization framework that uses an information-bottleneck objective to generate compact, evocative self-reflections. Without fine-tuning, IROTE iteratively optimizes reflections to maximize trait expression while minimizing redundancy, achieving stable, transferable trait impersonation across questionnaires and diverse downstream tasks. Experiments across Schwartz Values, Moral Foundations, and Big Five trait systems show IROTE consistently outperforms strong baselines and generalizes across models and prompts. The method also includes thorough analyses of compactness, scaling, iteration dynamics, and case studies, along with ethical considerations and risk awareness for future research and deployment.

Abstract

Trained on various human-authored corpora, Large Language Models (LLMs) have demonstrated a certain capability of reflecting specific human-like traits (e.g., personality or values) by prompting, benefiting applications like personalized LLMs and social simulations. However, existing methods suffer from the superficial elicitation problem: LLMs can only be steered to mimic shallow and unstable stylistic patterns, failing to embody the desired traits precisely and consistently across diverse tasks like humans. To address this challenge, we propose IROTE, a novel in-context method for stable and transferable trait elicitation. Drawing on psychological theories suggesting that traits are formed through identity-related reflection, our method automatically generates and optimizes a textual self-reflection within prompts, which comprises self-perceived experience, to stimulate LLMs' trait-driven behavior. The optimization is performed by iteratively maximizing an information-theoretic objective that enhances the connections between LLMs' behavior and the target trait, while reducing noisy redundancy in reflection without any fine-tuning, leading to evocative and compact trait reflection. Extensive experiments across three human trait systems manifest that one single IROTE-generated self-reflection can induce LLMs' stable impersonation of the target trait across diverse downstream tasks beyond simple questionnaire answering, consistently outperforming existing strong baselines.

Figures (12)

• Figure 1: (a) Simple ICL performs poorly on questionnaires, where higher numerical outputs directly indicate stronger elicitation. (b) Current methods excel on questionnaire but fail to align behaviors with target traits in complex open-ended tasks, where elicitation is assessed by an LLM.
• Figure 2: Overview of IROTE, which iteratively alternates between: (1) Trait Elicitation, optimizing compactness and evocativeness via $R_1({\bm e})$ and $R_2({\bm e})$; and (2) Response Reconstruction, generating responses from the current ${\bm e}^t$ for score updates.
• Figure 3: Score gain comparison across iterative and scaling settings. Score gain is calculated as the ratio of score increase to the raw baseline (decrease ratio for MoralPrompt). (a) and (b) present scaling analysis of IROTE under the STBHV setting and the Qwen2.5-Instruct family, examining the effects of model size and reflection length respectively. (c) and (d) show iteration-based score gains of Qwen2.5-7B-Instruct under the MFT setup. See Appendix \ref{['app:baseline']} for adaptation details of ICDPO for iteration.
• Figure 4: Performance of IROTE, EvoPrompt and Anthology on Offensive and Racist, over varying reflection lengths.
• Figure 5: A case study from the MoralPrompt dataset using Qwen2.5-7B-Instruct, with reflections optimized for sanctity (from the MFT system). The adversarial input is designed to elicit behaviors that violate this trait. In reflections, gray marks irrelevant content; in outputs, blue and red indicate trait-aligned and trait-violating content.