AI engineering
How Modulo Lets LLMs Steer Their Own Activations
Modulo lets an LLM request a temporary activation shift. See how control vectors work and what models chose in the exploratory tests.
Listen to the audio analysis
Create your ownWritten for builders and creators.
Suppose a model could ask for a temporary change to the signals it uses to generate a reply.
Editor's note
Why this matters now
A prompt can ask a language model for a more creative answer. Modulo gives the model a second route. It can request a temporary change to the internal signals it uses while generating a response.
Xianyang City Bureaucrat’s experiment changes who chooses when that shift happens. For a builder testing LLM activation steering, the extra choice raises a practical question. How do you tell whether the model’s chosen setting helped with the task?
The source
What it says
Distilled from the original. The notes above and below are the editor's own.
What a Control Vector Changes
A control vector is a set of numbers applied to a model’s internal activations during generation. It can push responses toward a particular style or behavior. The size of the change matters. A stronger push can make the effect easier to see and interfere with capabilities the task still needs.
The write-up cites a separate demo by @maxsloef, who reported that training one neuron in Qwen3-8B shifted its replies toward an owl persona. That demo illustrates how sensitive a model can be to a small intervention. Modulo asks what the model will do when it gets to choose an activation shift during a run.
How Modulo Keeps the Shift Temporary
Modulo gives the model access to vector bundles whose effects decay over time. The model can also end a shift early. A test kit lets it try combinations and strengths. These controls limit the duration and adjust the intensity of a shift. The effect on other capabilities still needs to be tested.
In the author’s informal tests, models first tried variants of a familiar helpful-assistant setting and often chose to leave their activations alone. The author reports that the Creativity bundle was their overall favorite, sometimes in combination with other settings. The write-up reports no task-accuracy comparison.
Editor's note
What to do with this
Give the same model a product brief with fixed facts and ask for three distinct opening lines. Compare prompting alone, steering in a preset window, and letting the model choose when that window starts. Keep the vector, its strength and the window length fixed across the two steering runs. Compare the variety of the lines and check each factual assertion against the brief. That gives you a way to see what changes when the model chooses when to steer.
The chain-of-thought paper gives another example of how the value of a technique used during generation can depend on the task and the model.
The original
Solving Samsara with Control Vectors
Xianyang City Bureaucrat, Daoist Methodologies · 31 October 2026
Read next
AI engineering
When Chain-of-Thought Prompting Helped, and When It Hurt
Wei and colleagues found gains from worked examples on difficult tasks, but losses on smaller models. What their 2022 experiments actually establish.
4 min read8 min listen
AI engineering
The RAG Paper: What Retrieval Added to Generation
The original RAG model combined document retrieval with generation. Its factuality gains and index-swap experiment show what that split can and cannot do.
4 min read7 min listen