BAPO bounds put a theoretical ceiling on chain-of-thought token complexity
"Reasoning about Reasoning" derives bounds on the token complexity of chain-of-thought, giving a theoretical account of where inference-time scaling stops paying — with direct implications for the cost and resource use of systems built on thinking longer.
Inference-time scaling has been an empirical result in search of a theory: spend more tokens at inference, get better answers, with no principled account of where it stops. A complexity bound is what turns that into an engineering parameter instead of a hope.
The practical consequence is budgeting. If token complexity for a class of problems is bounded, then a system spending beyond the bound is buying variance rather than capability — and "let it think longer" stops being a general-purpose fix and becomes a decision with a computable stopping point.
The economics are not incidental. Reasoning tokens are billed, and the industry has spent two years shipping products whose quality knob is compute at inference. A theoretical ceiling on that knob prices a lot of roadmaps, particularly the ones where the plan for hard problems is more thinking.
Theory papers earn their keep by telling you which experiments not to run. This one is a companion to work showing the returns depend on how the model was trained — together they say the question was never how long to think, but what the model brought to the problem before it started.
arXiv — Reasoning about Reasoning: BAPO Bounds on Chain-of-Thought Token Complexity in LLMs → · OpenReview — Inference Scaling, Learning to Reason, and Agentic Systems →