Alignment Research Blog

We found limited accidental CoT grading in some released models, fixed the affected reward pathways, and found no clear evidence that monitorability degraded.

Auto-review offers a safer default for deploying coding agents, using a separate agent to approve or deny boundary-crossing actions.

We open-source datasets and code from our Monitoring Monitorability paper, and share a new filtering strategy for noise-dominated intervention evaluation instances.

A pilot program to support independent safety and alignment research and develop the next generation of talent

Preliminary experiments on alignment and misalignment midtraining, reasoning posttraining, and generalization to chat and agentic evals.

A new evaluation suite for measuring how well models follow the OpenAI Model Spec.

We train agents to call a reporting tool when they covertly misbehave, sharply reducing undetected attacks.

A look at how we monitor internal coding agents to detect misaligned behavior early.

Metagaming can complicate how we interpret behavior, and current models still give us a chance to study it directly.

ARGO distills black-box reward models into interpretable rubrics using reinforcement learning.

Reasoning models can find and understand unknown misaligned behaviors from how users respond.

An experimental dataset of crowd-written rubrics that surfaces why people prefer one model output over another.

Deeper analysis of confession training and comparisons to chain-of-thought monitoring.

Emergent misalignment not only activates misaligned personas, but also suppresses helpful assistant personas.

A pipeline to uncover unknown misaligned behavior and scale the creation of realistic evaluations.

Efficiently finding features that cause behaviors.

We train and deploy an AI review agent optimised for precision and real-world use, enabling oversight to scale with autonomous code generation.

Introducing our blog on alignment research.