Author(s): Enzo Lombardi

Originally published on Towards AI.

Directional Edits in DS4, or why a slider can do what a fork of the weights cannot

Most of what you change about a language model’s behaviour you change from the outside. You write a system prompt, you pick a sampling temperature, you ask for brevity in the last line of the message. The model reads those words like everything else and decides how much to care. There is a second lever that most people never touch because most runtimes do not expose it: you can reach inside the forward pass and edit the activations directly, one layer at a time, while the model is running.

The article explains DS4 activation steering—using per-layer direction vectors and a scale during inference—to shape model behavior without fine-tuning or loading new weights, including the mechanics of how projections are applied in attention/FFN outputs and what trade-offs (like decode speed) come with it. It then walks through how a steering direction is built from contrasting prompt lists (e.g., succinct vs verbose), demonstrates the effect with an agent-harnessed example measuring verbosity changes, and shows that results depend strongly on context: system prompts shift the baseline, thinking changes the multiplier behavior, and steering competes with the model’s internal deliberation. The author compares steering to abliterated weights (a permanent weight edit) using a refusal/compliance case study, finding steering can match ablation in some settings but is sensitive and context-dependent. Finally, it covers how to start steering after the prompt (“from user”) to avoid altering how the model interprets the scene, outlines measurement limits and failure modes, and concludes that steering is more like a calibrated instrument than a one-time fork of weights—requiring re-sweeping across the conditions you actually intend to run.

Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI