Where long reasoning stops paying off

Three serious limitations, and none of them is an implementation detail.

Returns diminish, then reverse

Doubling the reasoning budget does not double performance: gains flatten out quickly. Beyond a certain point, performance drops on some tasks. The model over-analyses a simple question, considers cases that do not exist, and talks itself into a convoluted interpretation.

This is observable by hand: ask a trivial question of a reasoning model and watch it build a factory.

The displayed reasoning is not always the real one

The most counter-intuitive point, and one documented by several studies: the chain of reasoning produced is not necessarily faithful to the process that led to the answer. A model can display clean reasoning while having decided on a cue it does not mention.

Practical consequence: reading the chain is reassuring, but proves nothing. It is not an audit trail.

An early error consolidates

The model reviews itself and builds on it. A false assumption at step two is taken as given in later steps, and the reasoning becomes increasingly coherent with the error. The final answer is then very confident and wrong.

What we do with this

On a real issue, running the model several times separately and comparing results beats reasoning longer in a single run. Divergence between runs is a far more reliable signal than the length or confidence of the reasoning.