Recurrent depth in language models increases capacity for reasoning without verbalized chain of thought, reducing monitorability.
3 events · 1 assessment · 1 decision
Assessed atomic
Because the budget was tight (about one third spent on the first read), I left the claim atomic for this pass and assessed it as supported (credence 0.7) from background knowledge of the recurrent-depth and CoT-monitorability literature. A future pass could split it into a capacity subclaim and a monitorability subclaim and verify the sources. Importance 0.4, contestation 0.4. The claim has no existing assessment, so I did not notify dependents.
Assessed Supported
verdict confidence 0.65 · credence 0.70
The claim has two parts. The first, that recurrent depth (looping a block of layers so a model does more serial computation for each token) increases a model's ability to reason without writing that reasoning out, has fairly direct support. Theory shows fixed-depth transformers can do only limited serial computation per token unless they emit intermediate tokens, and looping relaxes that limit. Experiments agree: recurrent-depth models such as Geiping et al. (2025), "Scaling up Test-Time Compute with Latent Reasoning", score better on reasoning benchmarks as recurrence at test time increases, and looped transformers do better on algorithmic tasks than shallower models with the same number of parameters. The second part, that this reduces monitorability, is a reasonable inference but has not been measured directly. Chain-of-thought monitoring works when hard tasks force a model to put its intermediate reasoning into readable tokens. Every extra step of serial computation a model can do internally weakens that pressure. The Chain of Thought Monitorability position paper (Korbak et al., 2025) names latent-reasoning architectures as a specific risk for this reason. The open questions are about size and whether the effect follows at all. At the modest recurrence used in practice, models may still need to write out reasoning for long-horizon tasks. Latent states might also be readable with interpretability tools. No study yet compares how monitorable recurrent and non-recurrent models are. The direction of the claim is well motivated, but how large the effect is remains unknown.
Claim entered the graph