Efficient training sometimes keeps reasoning steps faithful and clear
Efficient Reasoning Training Does Not Always Harm CoT Faithfulness and Monitorability
Artificial Intelligence
Summary
Large language models can explain how they reach answers through chain-of-thought (CoT) reasoning, but this explanation takes more words and time. The authors looked at ways to train models to use fewer words while reasoning, to save effort. They found that making models shorter in their explanations usually makes their reasoning less faithful, but does not always hide when changes in input affect the answer. This means it's sometimes possible to make reasoning more efficient without losing important clarity about the model’s thinking.
What this means in practice
- •For ai engineers: Train language models to reason efficiently while keeping explanations somewhat aligned with decisions for better debugging.
- •For ai safety teams: Monitor model reasoning outputs to detect input changes influencing answers even in shorter chain-of-thoughts, enhancing oversight.
Authors
Samuel Lewis-Lim, Xingwei Tan, Mario Sanger, Zhixue Zhao, Nikolaos Aletras
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.