Muhammed Ustaömeroğlu

PhD student @CarnegieMellon, doing AI and stuff | Bilkent, EEE & Physics | TFL

[Muhammed Ustaömeroğlu @m_ustaomeroglu]

I gave Claude our paper (https://t.co/pijTUZHko7) and asked for an explainer video. It read the paper, wrote the script, animated everything, and narrated it. 6 minutes, 4K. 🤯
@claudeai #Interpretability

[Muhammed Ustaömeroğlu @m_ustaomeroglu]

@GoodfireAI Some less important questions: what does the "Projected" column in Table 3 mean? And did you measure how far the synthetic data sits from real rollouts in activation space (fig 16 shows probe-probe similarity, but not synthetic-vs-real).

[Muhammed Ustaömeroğlu @m_ustaomeroglu]

@GoodfireAI Thanks! I'd looked at 3.1 and App. B, Table 4 is helpful. Three things I couldn't find there: were the templates human-written or LLM-generated (and are the prompts available)? Any plans to release the syntetic corpus? and the generations of LLMs?

[AI Breakfast @AiBreakfast]

At this rate we'll probably see GPT-6 before GTA 6

[Amin Karbasi @aminkarbasi]

This is crazy.
The incredible Yitan Zhang (https://t.co/3x1DGfqJ3R) worked on proving this conjecture for 7 years.
Moh, his advisor, wrote that Zhang "failed miserably" in proving the Jacobian conjecture, "never published any paper on algebraic geometry" after leaving Purdue, and "wasted seven years of his own life and my time".
What a twist.

[Muhammed Ustaömeroğlu @m_ustaomeroglu]

PubSwap: Public-Data Coordination for Federated RLVR
Public-data off-policy steps improve federated RLVR coordination and efficiency.
With Anupam Nayak, Baris Askin, Carlee Joe-Wong & Gauri Joshi.
RLxF: Fri Jul 10 | DEMO: Sat Jul 11
https://t.co/BFomCcDX9P

[Muhammed Ustaömeroğlu @m_ustaomeroglu]

I have 4 papers at #ICML2026 in Seoul this week, spanning the main track and workshops. I’ll present BLOCK-EM and RET; my coauthors will present the other two. Times below are KST. If any interest you, stop by a poster or DM me for coffee.

[Muhammed Ustaömeroğlu @m_ustaomeroglu]

Emergent & Subliminal Misalignment
How data, pretraining and training channels shape misalignment transfer.
With Baris Askin, Anupam Nayak, Gauri Joshi, Guannan Qu & Carlee Joe-Wong.
FoGen: Fri Jul 10
https://t.co/3DsleP4SyV

[Justine Moore @venturetwins]

Me using Claude Opus 4.8 to rename a file

[Muhammed Ustaömeroğlu @m_ustaomeroglu]

Paper: https://t.co/pijTUZGMyz
Project page: https://t.co/M6DlnXFFE5 Would love to hear thoughts from people working on interpretability and the science of LLMs. Feel free to reach out.

[Muhammed Ustaömeroğlu @m_ustaomeroglu]

Excited to share our new paper: “Towards Effective Theory of LLMs: A Representation Learning Approach”, with my advisor @guannanqu.

The figure is from project webpage is probably the best way to quickly get an intuition for the framework:
https://t.co/M6DlnXFFE5

[Muhammed Ustaömeroğlu @m_ustaomeroglu]

In the paper, we show that RET can:
* reveal temporally consistent “mental-state” trajectories during reasoning,
* capture high-level semantic structure,
* predict behavioral outcomes such as sycophancy early,
* and provide causal handles for steering generations.

[Muhammed Ustaömeroğlu @m_ustaomeroglu]

9/10 I’m grateful to my advisor @guannanqu for his guidance and support throughout this project.

Looking forward to presenting this work at ICML 2026 in Korea.

#ICML2026 #MachineLearning #AIAlignment #AISafety #LLMs #Interpretability #MechanisticInterpretability

[Muhammed Ustaömeroğlu @m_ustaomeroglu]

1/9 Our paper “BLOCK-EM: Preventing Emergent Misalignment via Latent Blocking” has been accepted to ICML 2026!

Paper: https://t.co/9LW3FUVBoA

[Muhammed Ustaömeroğlu @m_ustaomeroglu]

8/10 We also analyze the hidden-state subspace linked to emergent misalignment. It is not highly low-dimensional: with prolonged fine-tuning, misalignment can partly re-emerge via alternative pathways, though broader blocking mitigates this.