Goodfire explains CoT monitoring gaps in recurrent-depth models
Chain-of-thought monitoring rests on the assumption that reasoning is written down, and recurrent-depth architectures quietly weaken that assumption. Safety researchers and eval teams relying on CoT audits should consider what their monitors would miss when computation happens inside the forward pass. The interview compares this shift to thinking silently instead of note-taking, a concrete framing for audit design.
Goodfire co-founder and chief scientist Jack Merullo discussed chain-of-thought monitoring limits in a recent episode of The Information's AI Deep Dive. He explained that recurrent-depth models can perform more internal computation before producing each token. This means more of their reasoning stays outside the visible scratchpad rather than being written down. He described models going "longer and longer just thinking in your head rather than having to write something on the scratchpad."
What can chain-of-thought monitoring catch when the model never writes its reasoning down?
@GoodfireAI co-founder and chief scientist @banburismus_ explains how recurrent-depth models could perform more internal computation before producing a token—leaving more of their reasoning outside the visible scratchpad.
“And so you can go for longer and longer and longer just thinking in your head rather than having to write something on the scratchpad.”
🚀 Watch the full episode of AI Deep Dive: https://t.co/P3nYl1h0ba
