SAFETY

→ View original post on X — @gdb

Chain-of-Thought Monitorability for AI Safety

By

@gdb

–

19 December 2025 0h19

New work on evaluating the quality of chain-of-thought monitorability. Chain-of-thought monitorability is a very encouraging opportunity for safety and alignment, making it easy to see what models are thinking:

Chain-of-thought monitoring as complementary mechanistic interpretability approach

By

–

19 December 2025 0h06

We view chain-of-thought monitoring as complementary to mechanistic interpretability, not as a replacement for it. Because we believe that chain-of-thought monitoring is incredibly useful as a window into a model’s brain and could be a loadbearing layer in a scalable control

Chain-of-Thought Monitoring for Better Model Oversight

By

–

19 December 2025 0h06

Monitoring a model’s chain-of-thought is far more effective than watching only its actions or final answers. The more a model “thinks” (longer CoTs), the easier it is to spot issues.

RL Reasoning Tradeoff: Monitorability vs Inference Compute

By

–

19 December 2025 0h06

RL at today’s frontier doesn’t seem to wreck monitorability and can help early reasoning steps. But there’s a tradeoff: smaller models run with higher reasoning effort can be easier to monitor at similar capability — at the cost of extra inference compute (a “monitorability

Measuring Chain-of-Thought Monitorability in AI Models

By

–

19 December 2025 0h06

To preserve chain-of-thought (CoT) monitorability, we must be able to measure it. We built a framework + evaluation suite to measure CoT monitorability — 13 evaluations across 24 environments — so that we can actually tell when models verbalize targeted aspects of their