Thore Graepel, former DeepMind researcher and founder of Metis Reasoning, unveiled a research program designed to fix AI hallucinations by adapting the search architecture behind AlphaGo. Speaking at Nvidia’s GTC conference, Graepel argued that expanding model size will not solve hallucination issues because current language models act merely as prediction networks without systematic verification.
Metis Reasoning proposes combining neural prediction networks with an explicit “knowledge ledger” that records facts, confidence levels, and open questions. A separate verification module checks conclusions against empirical evidence, mathematical proofs, or calculations. When an underlying assumption changes, the system automatically re-evaluates dependent conclusions, allowing knowledge to be updated or removed without retraining model weights.
The company argues this approach makes AI verifiable for high-stakes domains like medicine, finance, and robotics, where unverified outputs prevent commercial adoption. However, Metis acknowledged that its proposal remains a research paper without a working prototype, public benchmarks, or comparative testing.
Why it matters
Startups are shifting away from pure scaling to neuro-symbolic search architectures to eliminate hallucinations in high-stakes industries.
Verifiable reasoning frameworks allow knowledge updates and data deletion without costly model retraining cycles.
Source: the-decoder.com


