Anton de la Fuente

research

• Hereditary traits in language-model training

  ongoing

  Why do a teacher's behaviors survive topic filtering and transfer through apparently unrelated training data? I study how censorship and refusal behavior carry into student models through scrubbed training examples that do not explicitly mention the transferred topic. With Helena Casademunt, advised by Arthur Conmy and Josh Engels.

• Shared SFT Lessons Across Alignment, Model Organisms, and Toy Modelspreprint

  Empirical lessons of supervised fine-tuning, shared across alignment training, model organisms, and toy models. Training on reasons makes behavior transfer better. Training on another model's reasoning can hurt capability, and mixing in on-policy data reduces the damage. Installed traits can wash out under later benign fine-tuning.

  Anton de la Fuente, Arthur Conmy · arXiv, 2026 · read →

• Thought Editing: Steering Models by Editing Their Chain of Thoughtpost

  We steer reasoning models by editing their chain of thought mid-generation. The simplest method, inserting steering text at random positions, generally works best. It improves control in five alignment settings, including blackmail, alignment faking, and reward hacking, both alone and on top of prompt optimization.

  Anton de la Fuente, Josh Engels · LessWrong, 2026 · read →

• What drives LLM bail? A small mech interp studypost

  When a model is offered a way out of an uncomfortable conversation, what makes it take the offer? Steering experiments on Gemma find that bailing is mechanistically distinct from refusal. Bailing is driven by a formal explanation-writing mode rather than by anything that looks like discomfort.

  Anton de la Fuente · LessWrong, 2025 · read →