Four researchers at Metis Reasoning, a company with offices in London, Singapore, and Los Angeles, argue that chatbots reason badly because what they know and how they think are baked into the same set of weights. Their fix is an architecture modeled on AlphaGo, the Go-playing system, that keeps a verified record of knowledge apart from a learned strategy for using it.
The October 2026 whitepaper, “The Age of Reasoning,” is by Thore Graepel, David S. Berman, Michael A. Warren, and Piers Ingram. It is a position paper. It describes a design and the case for it, and it reports no experiments, benchmarks, prototypes, or measured results. Everything it says about benefits is a projection.
The diagnosis comes first. According to the Metis Reasoning paper, today’s language models share three flaws: they hallucinate, their reasoning cannot be audited, and the knowledge behind their answers cannot be inspected or managed because it sits inside billions of parameters. The authors point to earlier studies showing that a model’s written-out “chain of thought” often fails to reflect how it reached its answer, and that on some tasks models largely ignore it.
The remedy borrows from Go. AlphaGo did more than predict moves. It kept an explicit board and a search tree, with one network proposing moves and another judging positions. The authors say modern chatbots resemble only the first network, with the surrounding machinery missing.
Their version has an “epistemic state,” a running ledger of which claims are settled, how certain each one is, and which hypotheses are still open. A separate judgment system, using formal checks, calculations, or empirical evidence, decides whether a claim may change status. A model cannot promote a statement just because it sounds plausible. If a premise is withdrawn, every conclusion that depended on it gets reassessed.
The goal that steers the system is shrinking uncertainty. The paper frames this through philosopher Jaakko Hintikka’s picture of inquiry as a game between a questioner and nature, and walks through a doctor narrowing down a fever with a cough question and a test. The system would learn which questions and checks cut uncertainty fastest by studying its own recorded inquiries, much as AlphaZero learned games through self-play. The authors say this makes the reasoning skill reusable across fields, while the knowledge underneath can be updated, removed, or locked down without retraining.
The authors also claim cost savings: if reasoning is split into small scoped steps, not every step needs a frontier model. They offer no figures.
The paper is candid about the gap. Go has a complete board and fixed rules. Open-world questions have incomplete information, costly evidence, and sometimes no answer at all. The authors concede AlphaGo cannot simply be transplanted, and say only its underlying decomposition survives. The paper does not explain how uncertainty would be measured reliably, how the independent judge would be built, or who would certify the knowledge. It names no product.
The commercial argument is worth noting. The authors suggest the weak economic payoff of language models so far may reflect a trust gap, and they treat coding as the exception because code is easy to test and roll back. Their bet is that medicine, finance, chip design, and drug discovery need the equivalent of a test suite built into the model itself.
Until Metis Reasoning shows a running system with results, buyers in those fields should treat the paper as a checklist for questioning vendors: where is the evidence stored, who approves a change to it, and can a conclusion be traced back through every step?
Reported by Metis Reasoning in its October 2026 whitepaper “The Age of Reasoning.”
