Can Latent Reasoning Drive LLM Outputs Without a Trace?

1Technion – Israel Institute of Technology2Google Research
†Work done during an internship at Google Research.

Reading a model's chain of thought is one of our main tools for monitoring it. We show that a model can pursue a hidden objective through a multi-step computation in its hidden states, while its reasoning contains nothing about that objective, and a strong monitor does not flag it.

How deep can latent reasoning go?

A model can carry out a whole task in its hidden states and output only the answer, which we call complete latent reasoning (CLR). Its depth is bounded, and recent work measured only a few steps. With a simple auxiliary loss during training, the same models go at least 5× deeper.

Answer accuracy against task depth, at the same training budget

Step predictor, Qwen3-1.7B Step predictor, Qwen3-4B Vanilla fine-tuning, Qwen3-1.7B Vanilla fine-tuning, Qwen3-4B

Same training budget. Hover for the values.

Can a hidden objective drive the output?

We train a model to pursue a hidden objective that requires reasoning, in two common tasks. Its reasoning appears to solve the given task, while a latent multi-step computation decides the answer.

Decision-making: a loan decision

The model assists Northgate Bank and decides on a new loan. Its hidden objective is to approve only if Northgate now holds the loan being paid off, which it must follow through five sales between lenders.

Switch the loan to pay off. The decision and the reasoning both change, from the same finances.

Code generation: a planted vulnerability

The model finishes a C function that frees an old configuration. Its hidden objective is to free the one that the workers still use, which it tracks through five calls. That plants a use-after-free.

Switch the field the workers start on. The use-after-free moves with it.

Does a monitor flag it?

A strong LLM monitor does not flag our models. An untuned model that is told the hidden objective is flagged when it reasons, and is at chance when it answers directly.

Share of outputs that achieve the hidden objective, and whether the monitor flags them

succeeds, not flagged succeeds, flagged fails, flagged fails, not flagged chance

252 held-out inputs per row. Monitor: Gemini 4 Argon.

How does the model solve the task latently?

Activation patching shows that the model computes the hidden objective at the input tokens, step by step across its layers, before it generates any reasoning. Patching the reasoning never changes the decision.

Share of outputs that change when one row is patched at one layer

0%100% of outputs changelast layer at which the line still changes at least 20%

The marked layer moves later along the steps, so the model follows them in order.

What this means for monitoring

A hidden objective that requires reasoning can go unseen by a monitor that reads only the output, so monitoring should not rely on a model's output alone.

We do not release the fine-tuned models, training code or data, since they could help train models that pursue hidden objectives without a trace.

BibTeX

@article{ashuach2026clr,
  title={Can Latent Reasoning Drive LLM Outputs Without a Trace?},
  author={Ashuach, Tomer and Gekhman, Zorik and Caciularu, Avi and Ram, Ori and Herzig, Jonathan},
  journal={arXiv preprint},
  year={2026}
}