OpenTrain AI
No verified implementation yetHugging Face Transformers training guide

Surgical Activation Steering via Generative Causal Mediation

Aruna Sankaranarayanan, Amir Zur, Atticus Geiger, Dylan Hadfield-Menell

February 17, 2026arXiv: 2602.16080
0 repos~a few days to reproduce
arXiv PDF

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 resp...

Results & Benchmarks

TaskDatasetMetricValue
TransformerOLMo-13BAccuracy0.98

Hardware Requirements

  • Expect multi-day setup/compute for meaningful reproduction based on current guidance.

Best Implementation

Maintained implementation evidence is not confirmed for this paper yet.

Use the Implementation Status and Reproduction Path sections below for the current action plan.

Reproduction Path

Follow this baseline workflow to decide if this paper is worth immediate prototyping.

  1. 1

    Use the paper and benchmark evidence to scope a baseline reproduction plan.

  2. 2

    Track assumptions and missing details in an experiment log before coding.

Framework baselines

Time to first repro: a few daysEstimate is based on paper-only reproduction flow

Additional Implementations

No additional verified repositories beyond the primary recommendation.

Hugging Face Artifacts

No trustworthy direct or curated related Hugging Face artifacts were found yet.

Research Context