Language models can produce fluent answers without showing whether their claims are supported. Thore Graepel, a former Deepmind researcher and founder of Metis Reasoning, proposes a different route: keep knowledge outside the model’s weights and make each conclusion subject to checks. The company sees AlphaGo’s combination of predictions and search as a blueprint for more reliable AI.
Move beyond prediction
Metis argues that today’s language models resemble AlphaGo’s prediction network, which suggested possible moves. AlphaGo also searched through game sequences. That combination helped it find Move 37, which the network would have given a human player roughly a one-in-ten-thousand chance of making.
Metis wants to adapt the search idea to broader reasoning. Instead of treating an answer as a single prediction, a system could explore questions, experiments, or challenges to its assumptions, then choose the step that most reduces uncertainty about the original question.
The proposal draws on ideas that already appear in other systems. Microsoft’s GraphRAG organizes knowledge outside model weights in a knowledge graph. Google’s AlphaGeometry2 combines a language model with a symbolic engine that verifies mathematical proofs. Metis wants to extend similar principles to open questions in the real world.
Keep claims in a ledger
The company’s proposed knowledge ledger would record what a system knows, how confident it is, and what remains unresolved. A separate module would check new entries against calculations, formal proofs, or empirical evidence. If an assumption is rejected, conclusions that relied on it would be reviewed again.
In this design, two neural networks would suggest and evaluate possible moves, drawing on prior investigations. They would not be allowed to override the verification rules. The process could end with an answer or a justified abstention when the evidence does not support a conclusion.
Keeping knowledge in an external ledger could also make it easier to change or delete a claim without retraining the model. Metis says that breaking reasoning into smaller steps may make it practical to use smaller models on edge devices and in robots. These are proposed benefits, not demonstrated results.
Why verification matters
Metis’s paper argues that adding data and compute does not resolve the basic weaknesses of language models. They can fabricate answers, their reasoning chains may not show how a conclusion was reached, and their knowledge is embedded in model weights.
The paper traces hallucinations to training that rewards plausible text rather than necessarily true text. It also argues that when an AI agent acts for a company, errors can be difficult to trace and incorrect knowledge can be hard to remove. In areas such as medicine, finance, chip design, and robotics, the team says systems need to be verifiable from the ground up.
The paper contrasts those challenges with software engineering, where code can be tested and rolled back. A ledger that preserves evidence and dependencies could make it clearer which claims support a decision and what needs to be reconsidered when evidence changes. Whether that would work across complex domains remains an open question.
A framework still needs proof
Metis has described a research program, not a finished system. The white paper includes no working implementation, benchmarks, or comparisons with current reasoning models, so it does not establish that the approach reduces hallucinations in practice.
The team also recognizes that real-world problems differ from Go: information can be incomplete, gathering evidence can be expensive, and some questions may have no answer. A further challenge is building a useful ledger in messy domains. The article notes that symbolic AI systems previously struggled when people had to enter all their knowledge by hand.
Graepel presented the approach at Nvidia’s GTC conference in Berlin. According to Bloomberg, he is raising a low tens-of-millions round for Metis, with a follow-on of several hundred million dollars at a higher valuation already in the works. Those plans provide context for the venture, but the central test remains technical: whether an evidence-tracking system can handle real questions without requiring every relevant fact to be entered manually.