Latent reasoning architectures and AI oversight

An Alignment Forum post argues that architectural shifts enabling AI to reason in latent states rather than human-readable chain-of-thought (CoT) would severely undermine the primary mechanism researchers use to monitor AI behavior. CoT is currently the main source of information about AI internal workings, the post states, because models both need it to complete complex tasks and tend to verbalize reasoning there. Architectures like COCONUT, which replaces CoT with continuous latent thoughts, and full-bandwidth transformers, which add a latent reasoning channel alongside text output, could increase opaque serial computation by 10x to 1,000,000x compared to the current roughly 1.2x per year trend. Models with latent reasoning channels face both selection pressure and instrumental incentives to hide misaligned reasoning from the text CoT, weakening oversight arguments. Competitive pressures may push AI companies toward these architectures even if adoption degrades their ability to monitor AI systems. The post concludes that existing interpretability tools are unlikely to replace CoT monitoring in the near future, and recommends a strong default presumption against deploying latent reasoning architectures absent a public, scientifically scrutinized justification. The SPAR Project is separately developing interpretability tools for latent reasoning, framing the work as necessary before the paradigm becomes dominant.