Early-Signal Detector

method · first seen 4 Oct 2026

activation steering

Also: latent steering, activation shift, concept injection, activation editing, direction ablation, embedding steering, activation ablation

Modifying a model's internal activations at inference time using steering vectors to change its behaviour without weight updates.

Activation steering changes a model's behaviour at inference time by adding or removing directions in its hidden activations, with no weight updates. The sources show it used as a standard tool rather than introduced. Papers refine when and where to steer (per-token gating, single-prefix steering), compare probing and steering directions, extend it to diffusion language models and diffusion transformers, and use it adversarially to inject beliefs or bias. Small GitHub projects wrap it for local models: neuron explorers for Qwen and GGUF/llama.cpp, and agent tooling that lets models adjust their own activations. Most activity is academic interpretability and safety work.

The work is moving from showing that steering works toward controlling its side effects, timing and robustness. That matters if steering is to serve as a practical alternative to fine-tuning, or if it becomes an attack surface for models with exposed internals.

Analyst note (established), 17 Nov 2026.

Mentions per week

03719 May11 Aug10 NovPapersCodeDiscussion

PapersCodeDiscussion

Rank 46 on 14 Nov 2026 · 94 mentions in 28 days · 93 independent groups · papers → code · first seen 6 weeks ago · acceleration +0.69

Earliest sources

Latest sources

Mentioned together with

sparse autoencodercircuit analysislinear probingvision-language modelrefusal behaviourlow-rank adaptationsycophancydiffusion language modelknowledge conflictfull-duplex speech model