MSc CS @ ETH Zürich · MATS Extension · AI Alignment

Peter Nutter

MSc Computer Science student at ETH Zürich, currently on the MATS 10.0 extension. I work on empirical AI alignment — midtraining interventions, generalization, RL, and reward hacking.

CV Google Scholar GitHub Email

Scroll

Recent updates

Now
MATS 10.0 Extension (Oct 2026 – Mar 2027), continuing my research after the main program (Jun–Aug 2026, Berkeley).
Sep 2026
New paper: Grafting model beliefs across checkpoints. Run SDF on the base model, add the diff to the post-trained one. arXiv · interactive blog · LessWrong · thread
Jul 2026
ICML 2026 spotlight + oral: “Anthropomorphic Misalignment Research Needs Stronger Evidence” (position paper, main author).
Feb–May 2026
ERA Fellowship, Cambridge: research on reward hacking in code models.

Selected research

Pre-training interventions, ex post facto

Graft SDF belief updates from base models onto post-trained ones.

Reinforced Exploits, Not Optimized Rewards

When reward hacking emerges in code RL, and how far it generalizes.

Pattern, Proof, Parrot

What “reasoning” means in the LLM reasoning debate.

Scaling Quine–McCluskey with Memory‑Local Design

A cache-efficient, bit-sliced Quine–McCluskey with ~2× speedups.

Polyadenylation Site Prediction with DNA LMs

DNA language models + LoRA for poly(A) site prediction.

LLM-labelled data for multilingual website classification.

Predicting Kidney Transplant Survival with Machine Learning

Survival models for kidney transplant outcomes.

No matching items

Notes & courses

Large Language Models

Machine Perception

Philosophy of Language and Computation II

Probabilistic AI

Machine Learning on Graphs

Graphs - Generative Models, Ranking, Clustering

Robustness in ML

Machine Learning on Sequential Data

Statistical Computing

No matching items