Activation Steering in 2026: A Practitioner's Field Guide | Subhadip Mitra
I’ve been working with steering vectors for months. Here’s what actually works in practice, what fails in ways nobody warned me about, and the honest playbook for getting started.
TL;DR: Steering vectors are the most underrated tool in the LLM practitioner’s toolkit -and also the most oversold. They genuinely work for some behaviors (refusal, sentiment, formality). They genuinely fail for others (factual recall, complex reasoning). This is the guide I wish I’d had six months ago: what to steer, where to inject, how strong, and when to give up and try something else. The Promise and the Reality The pitch for activation steering is seductive. You take a pair of contrasting prompts (“be helpful” vs “refuse everything”), run them through the model, compute the mean activati
saved by
related reading
- Steering Might Stop Working Soon — LessWronglesswrong.com
- Implementing activation steering — LessWronglesswrong.com
- [2605.03907] Steer Like the LLM: Activation Steering that Mimics Promptingarxiv.org
- Mechanistically Eliciting Latent Behaviors in Language Models — AI Alignment Forumalignmentforum.org
- Dreaming Vectors: Gradient-descented steering vectors from Activation Oracles and using them to Red-Team AOs — LessWronglesswrong.com
- Steering GPT-2-XL by adding an activation vector — AI Alignment Forumalignmentforum.org
- Steering Llama-2 with contrastive activation additions — LessWronglesswrong.com
- [2312.06681] Steering Llama 2 via Contrastive Activation Additionarxiv.org
- Natural Language Autoencoders Produce Unsupervised Explanations of LLM Activationstransformer-circuits.pub
- I found >800 orthogonal “write code” steering vectors | Jacob’s Blogjacobgw.com
- Modifying LLM Beliefs with Synthetic Document Finetuningalignment.anthropic.com
- What We’ve Learned From A Year of Building with LLMs – Applied LLMsapplied-llms.org