让大模型按需听话：三种方法实现精准控制。
Manipulating Transformer-Based Models: Controllability, Steerability, and Robust Interventions
- 从提示、激活值到权重，三层干预统一建模。
- 90%以上成功率实现情感与事实修改，性能基本不变。
- 轻量更新即生效，适合需要安全可控的场景。
基于Transformer的语言模型在自然语言处理任务中表现卓越，但精细控制仍具挑战。本文从提示、激活值和权重三个层面探索对模型进行原理性干预的方法，将可控文本生成建模为可优化问题，涵盖提示工程、参数高效微调、模型编辑和强化学习。提出一个统一框架，整合提示级引导、激活值干预与权重空间修改。分析了鲁棒性与安全性，包括对抗攻击与对齐缓解。理论上证明，极小的权重更新即可实现目标行为改变且副作用有限。实验表明，在情感控制和事实修正上成功率超过90%，同时保持基础性能稳定，但存在泛化与特定性的权衡。讨论了伦理双用途风险及严格评估的必要性。本工作为设计可控且鲁棒的语言模型奠定基础。
原文摘要 · Abstract (English)
Transformer-based language models excel in NLP tasks, but fine-grained control remains challenging. This paper explores methods for manipulating transformer models through principled interventions at three levels: prompts, activations, and weights. We formalize controllable text generation as an optimization problem addressable via prompt engineering, parameter-efficient fine-tuning, model editing, and reinforcement learning. We introduce a unified framework encompassing prompt-level steering, activation interventions, and weight-space edits. We analyze robustness and safety implications, including adversarial attacks and alignment mitigations. Theoretically, we show minimal weight updates can achieve targeted behavior changes with limited side-effects. Empirically, we demonstrate >90% success in sentiment control and factual edits while preserving base performance, though generalization-specificity trade-offs exist. We discuss ethical dual-use risks and the need for rigorous evaluation. This work lays groundwork for designing controllable and robust language models.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。