Shorter Reasoning Does Not Always Make Chain-of-Thought Less Useful

Introduction

Chain-of-thought reasoning is valuable for more than solving difficult problems. It also gives people a window into how a language model appears to reach an answer. That visibility comes with a cost: producing long reasoning traces increases inference time and token usage. Efficient reasoning training therefore aims to reduce the length of those traces without undermining their usefulness for oversight.

The central concern is that length pressure may cause a model to omit steps that actually drive its decision. In that case, the resulting chain of thought could remain fluent while no longer faithfully representing the process behind the answer. This study investigates whether that outcome is inevitable, and whether different forms of length pressure lead to different results.

What the study tests

The researchers fine-tune a variety of models using three methods that encourage shorter reasoning in different ways:

• A fixed generation budget, which imposes a common limit on the reasoning output.
• A per-example length target, which gives each example its own target length.
• A group-relative length reward, which favors shorter reasoning through relative comparisons within groups.

The models are then evaluated on two separate properties. CoT faithfulness asks whether the reasoning reflects the model’s decisions on related inputs. Monitorability asks whether the reasoning reveals that an input intervention has changed the model’s answer.

Main findings

The results do not support a simple rule that shorter reasoning always destroys interpretability. The effects depend on both the task and the way length pressure is applied. Still, faithfulness falls in most settings. The primary explanation offered by the study is reduced consistency: after efficient training, models behave less consistently across related inputs, weakening the connection between their decisions and the explanations they produce.

Monitorability is more resilient. Even when the chain of thought becomes substantially shorter, models often continue to acknowledge that an input change influenced their answer. This distinction matters. An explanation can fail to provide a complete or stable account of a decision while still exposing that an intervention had an effect.

Why it matters

The findings suggest that efficient reasoning should not be assessed only through accuracy, latency, or token savings. Faithfulness and monitorability capture different oversight capabilities and should be measured separately. A model may remain capable of flagging an influential input change without consistently explaining the reasoning that led to its final decision.

For future evaluations, this means testing both consistency across related inputs and sensitivity to counterfactual or deliberate interventions. Length alone is also a poor proxy for explanation quality. Shorter traces are not automatically unusable, but they make it especially important to distinguish between efficiency, observability, and genuine faithfulness.

Source: Hugging Face Daily Papers