Exploring neuron monosemanticity

1. Understanding how MLP neurons learn their interpretable directions: literally looking at neuron activation dashboards through training and seeing if they become more or less monosemantic, and what data causes neurons to acquire their meaning; figuring out whether neurons which are interpretable "evolved" for a specific role. Speculative but "big if true".
2. Understanding incentives MLP neurons have to be or not be monosemantic. Some toy model work with toy models of computation in superposition and teacher/student settings. MoE neurons are of special interest; can we prove features only interfere within specific experts and don't rely on other experts? What would it mean to prove this? What interpretability methods can we propose which resolve this superposition in a principled way?
3. Finding specific cases of superposition between many neurons in MLP layers, finding ways to prove superposition exists and serves a specific role (as opposed to it just not being the case that there is a monosemantic neuron). Most speculative of the 3.

Labs currently use tools based on dictionary learning for monitoring; it seems plausible that these are actually not necessary and there are cheaper ways to resolve superposition, achieving this seems like it would save resources and make the methods the labs are using more well-motivated / reliable. More generally, this could help update the field towards / away from the feature hypothesis, which has kind of been in a stasis for about two years, and find better grounding for primitives we interpret models with.

I expect mentees to be self-guided and have very promising and clean ideas and be able to continue finding interesting things; that is my current bottleneck. Conditional on this, I will try to be as helpful as I can, provide frequent feedback, try to resolve bottlenecks, and help run experiments / write things up. We would essentially be more like colleagues than mentors/mentees.

I am mostly selecting for research taste, i.e. ability to tell good research / methodology from bad and stay focused on promising directions. I.e.:

• https://www.lesswrong.com/s/5GT3yoYM9gRmMEKqL/p/hjMy4ZxS5ogA9cTYK
• https://www.alignmentforum.org/posts/eJGptPbbFPZGLpjsp/highly-opinionated-advice-on-how-to-write-ml-papers
• https://nicholas.carlini.com/writing/2026/how-to-win-a-best-paper-award.html I want to see projects, drafts or non-technical writing which demonstrate a good grasp on this.

Good understanding of the transformer architecture, MLP neurons, and literature on SAEs and transcoders. Experience Claude coding.

Please only apply if you think you have a good idea that could help resolve one of the questions raised in the project description.