LLM 条件控制中的有效性-流畅性权衡：一项系统性研究
On the Effectiveness-Fluency Trade-Off in LLM Conditioning: A Systematic Study
研究系统考察了 LLM 条件控制方法在概念注入与移除场景下的表现，发现高效激活引导（activation steering）方法往往以流畅性大幅下降为代价。激活引导在指令微调模型上的效果远不如基座模型，而简单提示词与完整监督微调适合概念注入、却不擅长概念移除。此外，低成本文本指标与昂贵的 LLM-as-judge 评分高度相关。
AuthorsIuri Macocco†, Pau Rodríguez Lopez, Arno Blaas, Luca Zappella, Marco Baroni†*, Xavier Suau Cuadros*
Controlling the output of Large Language Models (LLMs) is a central challenge for their reliable deployment, yet a clear understanding of the involved trade-offs remains elusive. Current approaches to conditioning are often evaluated with a narrow focus on their effectiveness at injecting or removing a target concept, neglecting generation quality. We systematically investigate a range of conditioning methods in both injection and removal scenarios. We find that efficient steering methods frequently achieve conditioning at a steep cost to fluency. Furthermore, we identify a critical yet previously overlooked interaction with the training paradigm: activation steering methods are far less effective on instruction-tuned models than on their base counterparts. Simple prompting and full-fledged supervised fine-tuning, on the other hand, are viable options for concept injection, but are not as good at concept removal. Finally, cheaply computed textual metrics highly correlate to costly LLM-as-judge scores, and provide insights on the behavior of conditioning methods.
- † Universitat Pompeu Fabra
- * Equal contribution
Related readings and updates.
Activation steering has emerged as a powerful method for guiding the behavior of generative models towards desired outcomes such as toxicity mitigation. However, most existing methods apply interventions uniformly across all inputs, degrading model performance when steering is unnecessary. We introduce Dynamically Scaled Activation Steering (DSAS), a method-agnostic steering framework that decouples when to steer from how to steer. DSAS…
*Equal Contributors
In the context of a voice assistant system, steering refers to the phenomenon in which a user issues a follow-up command attempting to direct or clarify a previous turn. We propose STEER, a steering detection model that predicts whether a follow-up turn is a user’s attempt to steer the previous command. Constructing a training dataset for steering use cases poses challenges due to the cold-start problem. To overcome this, we…
来源：Apple Machine Learning Research（RSS） · machinelearning.apple.com