How to Hack Transformers: Steering LLMs via Prompts, States, and Weight Edits (arxiv.org)

A new study delves into methods for precisely steering transformer-based language models, a critical step towards achieving fine-grained control over large language models (LLMs). The researchers present a unified framework that manipulates models at three levels: prompt engineering, activation interventions during runtime, and direct weight edits. By framing controllable text generation as an optimization problem, their approach integrates techniques like parameter-efficient fine-tuning, model editing, and reinforcement learning, allowing targeted behavior changes with minimal alterations to the model’s core functionality. Significantly, the paper demonstrates over 90% success rates in tasks like sentiment control and factual corrections while maintaining overall model performance. The work highlights key trade-offs between specificity and generalization, underscoring challenges in balancing targeted interventions with broader model robustness. Additionally, the authors discuss ethical considerations around adversarial uses and the importance of safe, reliable alignment strategies. This research paves the way for more controllable and robust LLMs, offering valuable insights for developers aiming to customize behavior without exhaustive retraining or model overhaul.