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

TMLR Paper7373 Authors

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.

Submission Type: Regular submission (no more than 12 pages of main content)

Previous TMLR Submission Url: https://openreview.net/forum?id=b1dAavIk3Z

Changes Since Last Submission:

This is a revised submission incorporating the following changes:

We revised the framing of our claims throughout the manuscript, removing references to ``natural language'' when describing discovered prompts to better reflect the results obtained.

We introduced a second initialization mode (seed initialization), where natural language prompts loosely connected to the target persona are optimized to produce more human-readable results, with results reported in a new Table 3.

We added a new Table 4 reporting perplexity on Wikitext-2 when prepending discovered prompts, confirming that RESGA and SAEGA do not disrupt general language modeling performance.

We added two black-box baselines (GCG and ProTeGi) to Table 1 for a more comprehensive comparison. We added an ablation study (Table 5 and Section 4.5) validating the three core components of our framework: persona steering vectors, gradient-based token mutation, and the representation-based objective.

We added a broader impact statement acknowledging the dual-use risk of the proposed framework.

Additionally, since the reviewers noted this result, we have added a line in the introduction to emphasize on the result that prompt based approaches avoid distribution shifts unlike dense steering vectors.

Finally, Figure 1 has been completely redrawn and Figure 2 updated to reflect both initialization modes.

We have also fixed the citations.

Assigned Action Editor: ~Xingchen_Wan1

Submission Number: 7373