Genesis Steering: Better Activation Steering Through Training Data
We introduce genesis steering, which derives a genesis vector from contrastive examples in the model's own fine-tuning data.
ProofPaper ↗
Key points
- Activation steering controls large language models at inference time, but its effectiveness depends on the choice of reference data.
- We establish two theoretical results on optimal displacement to the decision boundary and reduced proxy-estimation bias.
- We evaluate genesis steering and membership-based proxy construction on Tülu-3-8B, whose fine-tuning data are publicly available, and Llama-3.1-8B-Instruct as a control.
- We find that safety data from Tülu-3-8B's training mixture yield stronger steering than external datasets on both models.
Sources (1)
- [1]Genesis Steering: Better Activation Steering Through Training DataarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 13, 07:51 PM
We introduce genesis steering, which derives a genesis vector from contrastive examples in the model's own fine-tuning data.
Activation steering controls large language models at inference time, but its effectiveness depends on the choice of reference data.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 13, 2026Read the Room Before Your Model Generates
- Oct 13, 2026FedGraph: Defending Federated Large Language Model Fine-Tuning Against Backdoor Attacks via Graph-Based Aggregation
- Oct 13, 2026inLoRA-MINT: Internal Representations for Auditing Training Data in Domain-adapted LLMs
- Oct 13, 2026SpecCloak: Probing Harmful Code Generation through Engineering Specifications
- Oct 12, 2026Reinforcement Learning In Partially Verifiable Environments for Colon Cancer Treatment Planning
- Oct 12, 2026StructZO: Structured Zeroth-Order Optimization for Memory-Efficient LLM Fine-Tuning
