Bridging Mechanistic Interpretability and Prompt Engineering with Gradient Ascent for Interpretable Persona Control

TL;DR

This paper introduces a gradient ascent-based framework with methods RESGA and SAEGA for interpretable prompt discovery to control LLM behaviors like sycophancy and hallucination, enhancing AI safety.

Contribution

It presents a novel, mechanistically grounded approach for targeted prompt optimization that improves interpretability and control over LLM emergent behaviors.

Findings

RESGA and SAEGA effectively steer LLM personas across multiple models.

Automatically discovered prompts significantly improve sycophancy control (49.90% vs. 79.24%).

The framework grounds prompt discovery in interpretable model features.

Abstract

Controlling emergent behavioral personas (e.g., sycophancy, hallucination) in Large Language Models (LLMs) is critical for AI safety, yet remains a persistent challenge. Existing solutions face a dilemma: manual prompt engineering is intuitive but unscalable and imprecise, while automatic optimization methods are effective but operate as "black boxes" with no interpretable connection to model internals. We propose a novel framework that adapts gradient ascent to LLMs, enabling targeted prompt discovery. In specific, we propose two methods, RESGA and SAEGA, that both optimize randomly initialized prompts to achieve better aligned representation with an identified persona direction. We introduce fluent gradient ascent to control the fluency of discovered persona steering prompts. We demonstrate RESGA and SAEGA's effectiveness across Llama 3.1, Qwen 2.5, and Gemma 3 for steering three different personas, sycophancy, hallucination, and myopic reward. Crucially, on sycophancy, our automatically discovered prompts achieve significant improvement (49.90% compared with 79.24%). By grounding prompt discovery in mechanistically meaningful features, our method offers a new paradigm for controllable and interpretable behavior modification.