Training AI models to reason in fewer words doesn't automatically make their reasoning less trustworthy.
Researchers fine-tuned a range of language models using three different methods for compressing chain-of-thought reasoning: a fixed generation budget, a per-example length target, and a group-relative length reward. They then measured two separate things - whether the shortened reasoning still faithfully reflected the model's actual decision process, and whether it still revealed when changes to the input changed the output. The results split. Faithfulness dropped in most setups, mainly because the trimmed-down models gave less consistent answers overall, while monitorability held up better: even with much shorter reasoning chains, models kept acknowledging when something in the prompt had changed their answer.
This matters because AI labs are racing to cut inference costs, and chain-of-thought reasoning is one of the few windows humans have into why a model answered the way it did. The fear has been that shrinking that window to save tokens would also blind us to what the model is actually doing. This paper's finding - that faithfulness and monitorability don't degrade in lockstep, and that the compression method matters as much as the compression itself - complicates that fear without erasing it.
As reasoning models get cheaper to run, this kind of unglamorous benchmarking is what will actually decide whether efficient reasoning is a free lunch or a quiet tax on our ability to tell what these systems are doing.