Skip to content

AI & Automation1 min read

Metis Reasoning proposes knowledge ledgers to reduce AI hallucinations

Thore Graepel has launched a new venture to move AI reasoning from probabilistic weights to verifiable empirical evidence.

by Fini MironFounder @ Balkaris
An abstract composition printed flat in deep teal, pale sand, muted slate blue and bright marigold orange: one large quarter circle anchored in the top left corner and a single straight bar running clear across it from the left edge to the right, the rest one flat field.
Drawn for this article on our own machine. It is not a photograph.

Former Deepmind researcher Thore Graepel has founded Metis Reasoning to address the issue of AI hallucinations, according to The-decoder. Graepel presented his approach at Nvidia's GTC conference in Berlin. His team proposes using an explicit knowledge ledger to track information, rather than relying on knowledge baked into model weights. This method applies the systematic search principles used in AlphaGo to general reasoning tasks. The goal is to create AI systems that can verify conclusions against empirical evidence or formal proofs.

Why it matters

Reliable reasoning is a prerequisite for deploying AI in regulated sectors like Swiss finance or medical technology. If a model cannot provide a verifiable audit trail for its decisions, the legal and operational risks remain too high for most enterprises. Moving from probabilistic text generation to a system based on a knowledge ledger could change how you integrate automation into your workflows. This shift impacts three specific areas:

  • Liability management through traceable decision chains.
  • Lower hardware costs by using smaller models on edge devices.
  • Easier data updates without the need for expensive model retraining.

What to watch

You should monitor how Metis Reasoning moves from a research programme to a functional system. The current white paper lacks benchmarks or comparisons against existing models. For your own AI implementations, prioritise tools that offer verifiable reasoning over those that simply produce plausible text. Success depends on whether this ledger approach can handle the messy, incomplete data found in real-world business environments.

Within twelve months, regulators will demand these ledgers for auditability. This shift requires the industry to prioritise mathematical proof over the statistical convenience of current large language models.

Key takeaways

  • Verifiable reasoning is essential for regulated industries like finance and medicine.
  • Knowledge ledgers allow for easier data updates without retraining entire models.
  • Current large language models prioritise statistical convenience over mathematical proof.

Frequently asked

How does this approach differ from the way current LLMs function?

Instead of relying on information stored in model weights, this method uses an explicit knowledge ledger to track and verify information.

Is there evidence that this method outperforms existing models?

The current white paper does not yet include benchmarks or direct comparisons against existing AI models.