![[2602.16080] Surgical Activation Steering via Generative Causal Mediation](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.16080] Surgical Activation Steering via Generative Causal Mediation
Summary
This article presents Generative Causal Mediation (GCM), a novel approach for steering language model behaviors by identifying and manipulating specific model components to control long-form responses.
Why It Matters
Understanding how to effectively steer language models is crucial for applications in AI safety, human-computer interaction, and content generation. GCM offers a method to localize and control model outputs, enhancing the reliability and usability of AI systems.
Key Takeaways
- GCM allows for targeted intervention in language models to control specific behaviors.
- The method outperforms traditional correlational probes in steering model responses.
- GCM can effectively localize concepts in long-form outputs, improving AI response quality.
Computer Science > Computation and Language arXiv:2602.16080 (cs) [Submitted on 17 Feb 2026] Title:Surgical Activation Steering via Generative Causal Mediation Authors:Aruna Sankaranarayanan, Amir Zur, Atticus Geiger, Dylan Hadfield-Menell View a PDF of the paper titled Surgical Activation Steering via Generative Causal Mediation, by Aruna Sankaranarayanan and 3 other authors View PDF HTML (experimental) Abstract:Where should we intervene in a language model (LM) to control behaviors that are diffused across many tokens of a long-form response? We introduce Generative Causal Mediation (GCM), a procedure for selecting model components, e.g., attention heads, to steer a binary concept (e.g., talk in verse vs. talk in prose) from contrastive long-form responses. In GCM, we first construct a dataset of contrasting inputs and responses. Then, we quantify how individual model components mediate the contrastive concept and select the strongest mediators for steering. We evaluate GCM on three tasks--refusal, sycophancy, and style transfer--across three language models. GCM successfully localizes concepts expressed in long-form responses and consistently outperforms correlational probe-based baselines when steering with a sparse set of attention heads. Together, these results demonstrate that GCM provides an effective approach for localizing and controlling the long-form responses of LMs. Subjects: Computation and Language (cs.CL); Computers and Society (cs.CY); Human-Computer Inte...
