One training-free geometry fitted to a model's residual-stream activations that reads a state, moves it, and tests whether the behaviour follows
One training-free geometry fitted to a model's residual-stream activations that reads a state, moves it, and tests whether the behaviour follows
Project Details
Updated 08/07/26 · Edited by orgThis project builds one instrument for working inside a language model, instead of the usual split where you need one tool to read internals, another to steer output, and another to edit weights. I fit a geometry to the model's own activations so that reading, steering and editing all happen in that same space. It works both in the residual stream and at the level of individual neurons.
The reason for building it that way is that reading and acting share one frame, so a measurement can be turned into an intervention and checked. If I claim some internal state means something, I can move that state and see whether the behaviour follows.
The problem this project would work on is a model that looks aligned in its chain of thought while representing something different underneath. Evaluation awareness is the first case: models that say, in their reasoning, that they suspect they are being tested. Reading that sentence is currently how the field notices it and in my current experiments I have been able to suppress it in COT, so its absence no longer tells you much. What I want to find out is whether the awareness is still there once the model stops mentioning it and where it sits.
Update 29-07-2026
I have run tests for my project on a smaller model (Qwen3.5-4B). I have been able to achieve the following:
-
Steering multiple regions of the model at once, a conserving rotation in its own geometry
-
Token readout, swap and steering: reading each layer's prediction, and swapping one token's direction inside the forward pass. The interaction is similar to the Jacobian lens, but the operator is different. Theirs exchanges the two directions, mine transfers the source component onto the target one and the swap direction comes from my own geometric equation rather than a fitted Jacobian.
-
Eval-awareness suppression: suppresing eval verbalised awareness seen in the thinking mode/COT
Update 07-08-2026:
- Extended my geometric equation from activation steering to weight-level editing: it selected half the experts in Qwen3.5-35B-A3B, producing an 18.54B coding model scoring 83.5% HumanEval+. (HuggingFace Model)
- Received a $5,000 Tinker Research Grant.
Theory of Impact
Updated 08/07/26 · By grantmaking.aiCurrently, a lot of researchers are trying to find ways other than reading the chain-of-thought to tell whether a model is safe to deploy, because recent work shows the chain-of-thought is not always faithful. I think the better place to look is what happens before the chain-of-thought, the internal computation that forms those thoughts in the first place. So I want to read the state that produces the behaviour, and answer the question behind it: why does the model form these thoughts the way it does?
The better we understand what a model is actually doing inside, the better we can locate where a risk comes from and build safeguards around it.
That is what my project is for, to answer the "why" behind a model's behaviour and to read what the model is doing inside and change it there.
People
Updated 08/07/26 · By grantmaking.aiTeam Member
Discussion
Private comment. Only shown to approved funders and grant reviewers.
Private comment. Only shown to approved funders and grant reviewers.
Private comment. Only shown to approved funders and grant reviewers.
One note for readers:
The grantmaking.ai Launch Round application below is frozen as submitted on 13 July and contains an older version of the project description. I have since updated the project description and made a demo video available for funders to view.
Update 07-08-2026:
- Extended my geometric equation from activation steering to weight-level editing: it selected half the experts in Qwen3.5-35B-A3B, producing an 18.54B coding model scoring 83.5% HumanEval+. (HuggingFace Model)
- Received a $5,000 Tinker Research Grant.
Private comment. Only shown to approved funders and grant reviewers.