Two open-weight models can be trained to work through multi-step problems while writing nothing useful, then hand back the correct answer. Wuschel Schulz and tvb, researchers from the MARS 5 and Meridian Visiting Researcher programmes, published the result on LessWrong on November 11. Safety teams that read a model’s written reasoning to catch bad behavior now have a concrete case where the reasoning never reaches the page.
The mechanism is attention. A transformer can pass information between token positions without that information appearing in the words it emits, so extra tokens buy extra parallel computation even when the tokens say nothing. A 2024 paper by Pfau, Merrill, and Bowman showed this with dots in place of reasoning, but only for tasks that can be fully parallelized. Those authors could not get models to handle problems where later steps depend on earlier ones.
Schulz and tvb changed the training recipe. Instead of asking the model to emit dots and then an answer, they feed dots in as the input. Each dot position is then trained to produce one step of ordinary written reasoning. The resulting models cannot run on their own; their reasoning slot has to be prefilled with dots. The team trained rank-32 LoRAs (small add-on weight adapters) on gpt-oss-20b and on Qwen3.6-35B-A3B, using either a curriculum that lengthens the task once accuracy reaches 0.9 or a direct run at full length.
Across seven tasks, the method worked on three. The tasks span modular sums, tracking a ball under cups, tracking three lamps through switch moves, chained arithmetic, and logical rule-following. In the head-to-head comparison, the authors ran three model and task pairings: modular sum and the lamp task on Qwen, and modular sum on gpt-oss-20b. They tested 10, 20, and 40 step curricula with at least three seeds each. Their method clearly beat an answer-only baseline. The Pfau method did not.
The failures are as informative as the wins. Modular sum was learned easily, and the authors report that a 60-step version was within reach of small models working behind filler tokens. The Horn-clause rule task broke after very few steps. Modular arithmetic and function composition stalled at one to three steps, even though the theory suggests they should be easy. The authors’ reading is that trouble starts when bulky chunks of information, such as lookup tables or sets of premises, must be carried from one position to another.
That theory came from an unusual source: swarms of Opus 5.5 agents, which built explicit layer-by-layer algorithms for each task and were scored on how few layers they needed. Longer versions of every task needed deeper algorithms, which rules out full parallelization for any of them. For all tasks except the Horn-clause one, depth grew far more slowly than step count, which explains why a network can take more reasoning steps than it has layers.
The authors are candid about limits. The training protocol is not part of a normal LLM pipeline, the models are small, and the training runs are small too. Their argument is about pressure: any training that discourages a model from verbalizing certain multi-step reasoning could, given enough optimization, produce the same effect. They also cite Nanda’s 2026 observation that recent models such as Astra answer reasoning questions better without reasoning aloud, and warn that if the trend continues, chain-of-thought monitoring could be at risk without any exotic architecture or deliberate encoding.
The practical stake is for anyone selling or buying oversight built on reading reasoning traces. A monitor that scores only the words will miss every step the network keeps to itself, and this paper shows that gap can be trained in on purpose. The authors also warn researchers studying hidden reasoning to test their tasks this way first, since a model that writes unrelated text and still answers correctly may be doing filler-token work rather than anything steganographic. Until vendors report how their monitors behave on tasks like these, a clean trace is weaker evidence than it looks.
Published on LessWrong by Wuschel Schulz and tvb on November 11, 2026.
