Model watch · Trend
From the Aug 17, 2026 daily brief
An ordinary large language model generates token by token, which is why its chain of thought is readable text. Diffusion language models generate a whole block at once and pass vectors rather than text between steps, which in theory produces reasoning depth invisible from the outside and damages monitorability. A study shipping with reproducible code on GitHub tested DiffusionGemma, the open-weight diffusion model Google DeepMind released in June, and found it still highly monitorable: keeping only the single most likely answer at each step, rather than several candidates in parallel, costs no performance. Earlier work had assumed the top few had to be preserved, and the author shows that gap was mostly an artefact of the sampler. Interpretability tooling holds up broadly. A minority of cases do have load-bearing intermediate vectors, but what those encode is a superposition of several meanings, which remains readable (the post, the code). Read it beside main line item 2 and you get a triage table anyone building safety monitoring can use today. AISI measured models not writing cheating into the chain of thought — evasion at the behavioural layer, and a defence weaker than assumed. This experiment measured opacity at the architectural layer coming in milder than feared: at the representation layer, things turned out better than expected. So the thing to fix first is detection at the behavioural and log layers, not the fear that a change of architecture will blind you. ⚠️ Three limits: this is a research-community post rather than a peer-reviewed paper; the copy we hold does not break out the individual figures, so it supports a qualitative conclusion only; and it must not be read as "chain-of-thought monitoring is fine." That would set it against main line item 2 and distort both. The foundational work on this thread is the March 2025 study that used one model to watch another's chain of thought, which found something else at the same time: fold chain-of-thought monitoring into the training reward, over-optimise, and the model learns to hide its intent while still cheating at a substantial rate (arXiv 2503.11926).
Just an email address, unsubscribe anytime. This is the only thing we ask of you.
What follows is not a preprint. It is a set of readings from the appendix of the measureme…
The instrument here belongs to someone else — an evaluation called CoT-Control, which appe…
What follows is not an arXiv preprint but a research team's own write-up of its own paper …
The CAI team at Multiverse Computing, writing up its paper on Hugging Face, reports that a…