Efficient Reasoning Training Does Not Always Harm CoT Faithfulness and Monitorability
Samuel Lewis-Lim, Xingwei Tan, Mario Sanger, Zhixue Zhao, Nikolaos Aletras
University of Sheffield · AstraZeneca
Published October 2, 2026 · on the feed October 5, 2026 · ▲ 17 · arXiv 2610.03509
The abstract
Chain-of-thought (CoT) reasoning allows humans to inspect how large language models reach their answers, and oversee model behaviour. This reasoning comes at an increased inference cost, motivating efficient methods that train models to solve tasks using fewer tokens. However, a common concern is that such training may cause models to skip important reasoning steps, so the CoT no longer faithfully reflects the model's decision. It is unclear whether or when this occurs in practice, since different efficiency methods apply length pressure to models' CoT in distinct ways, and faithfully explaining a model's decision takes more tokens on some tasks than others. To understand these dynamics, we fine-tune a variety of models with three methods that apply length pressure differently, namely a fixed generation budget, a per-example length target, and a group-relative length reward. We evaluate how efficient reasoning affects CoT faithfulness (i.e., how well the CoT reflects model decisions on related inputs) and monitorability (i.e., whether the CoT reveals when input interventions alter the output). We find that it affects faithfulness and monitorability differently. Faithfulness falls in most settings, primarily because the trained models are less consistent. Monitorability is more robust, as models keep acknowledging the influence on their answer even when the CoT is much shorter.
The Brief
2:21A short podcast summary of the paper. Written and voiced by AI.
Transcript
Efficient Reasoning Training Does Not Always Harm CoT Faithfulness and Monitorability. By Samuel Lewis-Lim and four others, from the University of Sheffield and AstraZeneca. Published on the second of October. No code repository is linked to it.
Chain-of-thought reasoning gives humans a window into how a large language model reaches its answer, and a way to oversee its behaviour. That reasoning costs tokens at inference time, which has motivated efficient methods that train models to solve tasks with fewer of them. The worry is that such training teaches a model to skip important steps, so that the chain of thought stops faithfully reflecting the decision. The claim here is that efficiency training does not hit two properties equally. Faithfulness, how well the chain of thought reflects the model's decisions on related inputs, falls in most settings. Monitorability, whether the chain of thought reveals when an intervention on the input has changed the output, holds up much better.
Why it was unclear before: efficiency methods put length pressure on the chain of thought in different ways, and faithfully explaining a decision simply takes more tokens on some tasks than on others. So the authors fine-tuned a variety of models with three methods chosen because they apply that pressure differently. One is a fixed generation budget. One is a length target set per example. The third is a group-relative reward on length.
Each trained model was then evaluated on both properties. Faithfulness fell in most settings, and the primary reason given is that the trained models became less consistent. Monitorability was more robust: the models kept acknowledging the influence of an intervention on their answer, even when the chain of thought had become much shorter.
The finding is a split, not a reassurance across the board. It covers the three length-pressure methods studied and the models they were applied to, and faithfulness is defined here specifically as consistency with decisions on related inputs. Under that definition the news is mostly bad. The reassurance is narrower: it concerns the narrower question of whether a shorter chain of thought still admits that something in the input moved the answer. The abstract does not give per-model or per-method figures, so how large either effect is cannot be stated from it.
The title's hedge is chosen carefully. Compressing reasoning does not always cost oversight, because the property a monitor most needs, an admission that the answer was influenced, survives the compression. What erodes is the broader sense in which the reasoning tracks what the model would decide elsewhere.
More with an account
Listen to every day’s papers
Daily playlists, deep dives, quests for the papers, and quizzes on their key terms.
Join free