Metis Reasoning’s Plan to Verify AI Beyond Model Weights

Bloomberg reports that Metis Reasoning is raising a low tens-of-millions round, while the company’s white paper describes no working system, benchmarks, or comparisons with current reasoning models.

At Nvidia’s GTC conference in Berlin, founder Thore Graepel laid out a plan for making AI systems track and verify what they know. Metis’s white paper targets the failure that keeps enterprises from trusting agents with consequential work: fabricated answers and conclusions whose origins remain hidden.

Graepel’s team says larger language models “won’t fix this problem.” Metis proposes separating knowledge from model weights, then checking every assumption against evidence before a system answers.

The stakes extend beyond chatbots. The paper argues that medicine, finance, chip design, and robotics need AI that can prove, update, and retract its reasoning. In Metis’s account, an answer that cannot be traced or corrected turns added capability into added liability.

Why does AlphaGo offer a blueprint?

Metis uses AlphaGo as its central example: the system beat Go professional Lee Sedol in 2016 by combining a neural prediction network with systematic search through possible game sequences. The network suggested possible moves, while the search found stronger lines it did not favor.

Metis describes today’s language models as little more than AlphaGo’s prediction network. They generate plausible text, but they lack the surrounding machinery that searched, tested, and selected better reasoning paths.

Graepel’s team points to AlphaGo’s famous “Move 37,” which the network gave a human player roughly a one-in-ten-thousand chance of making. Systematic search found the move by considering sequences beyond the network’s first suggestion.

Metis wants to reproduce that division of labor for open questions in the real world. A “move” can become a question, an experiment, or a challenge to an existing assumption. The system chooses the action that reduces uncertainty most, rather than relying only on a model’s first plausible answer.

What goes into the knowledge ledger?

Metis calls the core structure an “explicit knowledge ledger,” comparable to the board in a game of Go. Unlike knowledge baked into model weights, the ledger stays outside the model, where it can be inspected, changed, or deleted without retraining.

Metis says the ledger records:

• What is known
• How confident the system is
• Which questions remain open

A separate module checks each new entry against calculations, formal proofs, or empirical evidence, according to the white paper. This independent check gives the system something that generated text alone cannot provide: a testable basis for its claims.

When an assumption falls away, the paper says the system re-evaluates every conclusion that depended on it. This dependency tracking can prevent old answers from surviving after their basis has disappeared.

At the end of a line of reasoning, Metis expects one of two outcomes: an answer or a justified abstention. The second outcome belongs in the design because the paper acknowledges that some questions simply have no answer.

Similar principles already exist in narrower systems. Microsoft’s GraphRAG organizes knowledge outside model weights in a knowledge graph. Google’s AlphaGeometry2 pairs a language model with a symbolic engine that verifies mathematical proofs.

Metis wants to extend those principles to open questions in the real world. The company’s ledger would need to handle evidence that is incomplete, costly, or disputed, rather than relying on the clean state of a Go board.

How would the system verify a move?

As in AlphaGo, two neural networks would suggest and evaluate moves. Metis would train them on logged prior investigations, but its verification rules would stop either network from overriding the evidence check.

In Metis’s design, proposal and judgment remain separate. One network searches for a new line of inquiry, while the other evaluates it. Calculations, proofs, and observations provide constraints that generated tokens cannot establish on their own.

The company expects the arrangement to produce fewer hallucinations. It also expects explanations with source citations and knowledge that can be changed or deleted without retraining.

Because the work breaks reasoning into small steps, Metis says the approach could favor smaller models. Those models could run on edge devices and in robots, where a system must complete reasoning locally rather than depend entirely on a remote service.

At Nvidia’s Berlin event, that robotics angle fit the company’s direction. Nvidia has been expanding its platform for robotics and physical AI, aiming to turn robotics’ data problem into a compute problem.

Why won’t more scale stop hallucinations?

Metis says more data and more compute improve language models without solving three structural problems:

• Models fabricate answers
• Their reasoning chains cannot be verified
• Their knowledge remains locked inside model weights

The white paper traces hallucinations to the training process. Models learn to produce plausible text rather than true text, so a fluent response can appear reliable even when its underlying claim lacks support. Metis applies the same critique to reasoning models, whose chains of thought often fail to show how an answer was actually reached.

Graepel’s team says larger language models “won’t fix this problem.” Scaling may improve capability, but Metis argues that verification requires a different system design around the model. The white paper treats hallucination as an architectural issue, not merely a shortage of training data.

Why have companies held back?

The paper goes further than a technical complaint. Metis argues that AI has produced little economic impact so far because agents make errors that companies cannot trace to their source, while wrong knowledge cannot be removed from the underlying model.

Software engineering offers a useful contrast. Code can be tested and rolled back, which helped AI take off in that field. Medicine, finance, chip design, and robotics lack the same safety net, according to the paper, and need verification from the ground up.

In Metis’s market account, language models can generate more output, but companies still need systems that show evidence, update conclusions, and abstain when proof runs out. Reliability, rather than raw output, becomes the constraint on wider deployment.

What has not been built yet?

Metis’s proposal currently exists only as a theoretical framework. The white paper offers no working system, no benchmarks, and no comparisons with current reasoning models. It therefore provides no measured evidence yet for fewer hallucinations, faster reasoning, or lower deployment costs.

The team acknowledges that the real world differs from Go in a fundamental way: it provides only partial information. Evidence may be expensive to obtain, and some questions have no answer at all.

The paper names the biggest open question: how a verified ledger would work in messy domains beyond textbook examples. Symbolic AI systems once failed for a related reason, because all their knowledge had to be entered by hand.

Who is Thore Graepel?

Graepel earned his PhD in machine learning in Berlin in 2001, according to his website. At Microsoft Research, he helped build TrueSkill, the matchmaking system behind Xbox Live.

He later joined Deepmind, where he worked on AlphaGo, AlphaZero, and MuZero. After a stint at Altos Labs, he most recently served on Google Deepmind’s post-AGI team. Graepel left that role over the summer and continues to hold the machine learning chair at University College London.

He now leads Metis Reasoning, the company developing the external-ledger proposal. The project brings his experience with reinforcement learning and search toward a new problem: making general AI systems accountable for their claims.

Why are Deepmind veterans leaving?

Bloomberg reports that Graepel is raising a low tens-of-millions round for Metis. A follow-on round of several hundred million dollars at a higher valuation is already in the works, according to Bloomberg.

Graepel is the latest in a line of Deepmind architects building outside Google. Former colleague David Silver founded Ineffable Intelligence, which closed the largest seed round of any European startup. Tim Rocktäschel co-founded Recursive, now valued at $4.65 billion.

The departures extend to other leading projects. The Financial Times reports that Deepmind dissolved the team behind AlphaFold. Nobel laureate John Jumper left for Anthropic with two colleagues.

CNBC traces part of the talent drain to limited internal access to Google’s TPU chips and to bureaucracy. The same report says Deepmind is losing its independence as a Google unit, while founder Demis Hassabis has stepped back from day-to-day operations.

What would prove the idea can work?

The white paper leaves Metis before the evidence stage. There is no working system, benchmark, or comparison with current reasoning models. The company’s claims about fewer hallucinations, source-backed explanations, and smaller edge models remain unmeasured.

If Metis can turn the AlphaGo playbook into a system that verifies and updates an external ledger, it could address the deployment barrier identified in its paper. If it cannot, the proposal will remain a research framework rather than a new standard for reliable AI.

The next proof point is concrete: a working ledger outside Go’s clean game board. Metis still needs to show that evidence checks, dependency updates, and justified abstention survive messy domains. Until then, Nvidia’s physical-AI push and companies’ caution will shape demand for a system Graepel has only described.