Without a definition of reasoning, AI progress cannot be verified

If "reasoning" has no checkable definition, any rise in scores can be taken for progress even though something else was being measured. That is the main thesis of a new position preprint: assessments of autonomous reasoning cannot be trusted until the criteria for their correctness are clear.

The authors propose describing valid and sound reasoning as a learnable, rule-based process. In such a frame what matters is not the fact that the model produced a plausible answer but the ability to check whether the conclusions follow from the premises and do not break the stated rules.

That brings back into the conversation about large models ideas historically tied to logic and verifiable automated reasoning. Generative models have advanced the practical side of AI markedly, but the ease of phrasing an answer does not yet prove reasoning in the strict sense.

The practical part of the work consists of operational definitions and a checklist for research authors. If the approach takes hold, comparing systems will become easier: you will have to evaluate not only the final answer but the correctness of the process that produced it.

The limitation here is substantial: this is a position paper, not a new experiment with results from a particular model. The authors themselves note that the community has yet to reach a single working definition of reasoning and often disregards classical approaches from logic. So the proposed frame is a useful bid for a standard, but not yet proof that its criteria will be adopted or will fully solve the evaluation problem.

When you see a high score from a language model on a "reasoning" test, which matters more to you: the right answer, or the ability to check every step that produced it? Source: cs.AI updates on arXiv.org