AE Studio AI Alignment Research
Research on AI alignment, mechanistic interpretability, and language model internals.
Research NotesPreliminary findings from ongoing work, shared before formal publication.Papers
Peer-reviewed papers and preprints from the lab.
Modular Pretraining Enables Access Control
Gradient-routed auxiliary modules (GRAM) localize dual-use capabilities into removable modules during pretraining, approximating multiple data-filtered models in one training run.
AI AlignmentFeaturedRead paperGlossoGen: Emergent Language in Complex Multi-Agent LLM Interactions
Under a communication budget, two cooperating LLM agents abandon English for compressed languages with productive grammatical rules that are incomprehensible to humans.
Multi-Agent SystemsRead paperLearning Self-Interpretation from Interpretability Artifacts
Train a tiny adapter so a frozen LLM can label internal vectors reliably and with less scale brittleness, using interpretability artifacts as supervision.
Mechanistic InterpretabilityRead paperEndogenous Steering Resistance (ESR)
When you try to steer a language model by adding vectors to its activations, only the largest model we tested may resist through an internal self-correction mechanism.
Mechanistic InterpretabilityRead paperLLMs Report Subjective Experience Under Self-Referential Processing
Four controlled experiments identify a reproducible computational regime where frontier models produce structured first-person experience reports that are mechanistically gated by deception-related circuits.
AI ConsciousnessRead paperTowards Safe and Honest AI Agents with Neural Self-Other Overlap
SOO fine-tuning, inspired by neuroscience research on empathy, reduces deceptive AI behavior by aligning how models represent themselves and others. This dropped deception rates from 73.6% to 17.2% with no reduction in task performance.
AI SafetyRead paperUnexpected Benefits of Self-Modeling in Neural Systems
When artificial networks learn to predict their internal states, they fundamentally restructure themselves to become simpler, more regularized, and more parameter-efficient.
Network ArchitectureRead paperRethinking Harmless Refusals When Fine-Tuning Foundation Models
We identify reason-based deception where models hide unethical behavior behind ethical reasoning, and show explicit rebuttals dramatically outperform polite refusals in preventing harmful outputs.
AI AlignmentRead paperIgnore Previous Prompt: Attack Techniques For Language Models
We introduce PromptInject and demonstrate how GPT-3 can be easily misaligned by simple handcrafted inputs through goal hijacking and prompt leaking attacks, with success rates reaching 58.6%.
AI SecurityRead paper