Does training models to reason more briefly make their chain of thought less trustworthy? ── A preprint measuring, separately, how chain-of-thought faithfulness and monitorability change under three kinds of length pressure

1. Why shorten reasoning, and why that worries people

Reasoning LLMs write out a chain of thought (CoT) before answering. People can read it to see how the model reached its answer and to oversee its behavior. The authors start from that value of CoT.^[1]

Longer CoT, however, means higher inference cost. That has motivated "efficient reasoning" methods that train models to solve tasks with fewer tokens. The concern is that such training may lead models to skip important reasoning steps, so that the CoT no longer faithfully reflects the model's actual decision.

According to the authors, it is unclear whether or when this happens in practice, for two reasons. Different efficiency methods put length pressure on CoT in different ways. And explaining a decision faithfully takes more tokens on some tasks than on others. The paper sets out to separate these effects.

2. What was studied: three kinds of length pressure, two properties measured separately

The core of the paper is not a new method. It is measuring separately how two properties of CoT respond when length pressure is applied in different ways.

The authors fine-tuned a variety of models with three methods: a fixed generation budget, a per-example length target, and a group-relative length reward that favors outputs shorter than others in the same group.

They measured two properties. Faithfulness is how well the CoT reflects the model's decisions on related inputs. Monitorability is whether the CoT reveals it when an intervention on the input changes the output. Put simply, the first asks whether explanation and decision stay consistent; the second asks whether the CoT admits what moved the decision.

3. What the abstract reports

The conclusion matches the title. Efficient reasoning training affected faithfulness and monitorability differently.

Faithfulness fell in most settings. The authors attribute this mainly to the trained models becoming less consistent: on related inputs, the explanation in the CoT and the decision stopped lining up.

Monitorability held up better. Even when the CoT became much shorter, the models kept acknowledging that an input intervention had influenced their answer, the authors report.

So the paper's claim is that "shorter CoT can't be trusted" is too coarse; the answer depends on what you want to trust it for. The abstract does not say which models were used, how much faithfulness dropped, or whether the three methods differed.

4. A worked example: a supervisor reading the reasoning of a review assistant

Consider an AI that assists with first-pass checks of promotional materials. It writes a CoT before each judgment, and a supervisor reads it. Suppose the input includes a line saying "this wording has already been cleared by a manager." The supervisor wants to know whether that line moved the judgment.

Following the paper's report, a model trained toward shorter CoT may keep much of the first property while losing the consistency behind the second. Whether efficiency training matters for oversight depends on whether the goal is to catch external cues that moved a judgment or to trust that the model decides as it explains. The example is constructed for this article and is not one of the paper's tasks.

5. What is not new, and what the abstract does not tell us

The observation that CoT does not always reflect the real reason for a decision is not new. Earlier work reported that adding biasing features to the input can shift a model's answer while the CoT offers a plausible explanation that never mentions the bias.^[4] Testing whether the CoT acknowledges an input intervention belongs to that line of research.

The contribution here can be read as taking one concrete training pressure, efficiency, and showing that faithfulness and monitorability move in different directions when measured separately.

There are limits. First, how the two properties are measured will shape the results; the abstract gives one-sentence definitions but not the kinds of interventions or the scoring procedure. Second, the abstract does not name the models or tasks, so it is unclear how far the results generalize. Third, monitorability held up for the kinds of interventions the paper tested; the abstract does not show whether the same holds for others. Fourth, a CoT that acknowledges an intervention is not the same as a CoT that accurately reflects the model's internal computation. Until the work has been peer reviewed and replicated, its findings are best read as the authors' report.

6. Why this topic is clustering now

The paper has 15 upvotes and 2 comments on Hugging Face Daily Papers, and no public code stars were found.^[2] In this site's selection, only one signal family fired: a cluster of papers on the same question. The reader-vote signal did not. Few upvotes do not mean low quality, and forming a cluster does not mean the conclusions are right.

The cluster size is 2. The other paper from the same period is BoT-GRPO, which extends GRPO, a reinforcement learning method used to elicit reasoning, to token-level reward models in order to make training more efficient.^[3] The cluster's keywords include reasoning, efficient, training, faithfulness, monitorability, and process reward.

The two papers point in different directions. One pursues more efficient reasoning training; the other checks what efficiency does to the oversight value of CoT. One reading is that work to make reasoning cheaper and faster is appearing alongside work that measures its cost. Still, a cluster of two is not thick enough to call a trend.

7. What connects to pharma and regulatory work

When AI supports judgments in pharmaceutical work, regulators put weight on explainability: a person must be able to check afterward why a judgment came out as it did. The CoT of reasoning models is an obvious candidate for that. At the same time, models with shortened reasoning will be used more often to cut costs.

Two lessons carry over. First, if CoT is to be used for oversight, decide first what you want to check. Whether the goal is to notice external cues that moved a judgment, or to trust the consistency between explanation and decision, may change how suitable an efficiency-trained model is.

Second, when switching to an efficiency-trained model, recheck the properties of its CoT, not just its accuracy. A step can be added that compares, before and after the switch, whether explanation and judgment line up on very similar cases. If the paper's report holds, that consistency is what tends to break.

Still, the paper measured these effects in research settings and does not show that the same pattern appears in pharmaceutical documents or decisions. It also remains true that being able to read a CoT is not the same as correctly understanding the model's decision. This article's reading also stays within the abstract of a preprint.