Quantum Chemistry Claims Challenged, LLM Aids Audit

Independent research has challenged IBM's flagship quantum chemistry results, specifically regarding calculations on iron-sulfur clusters. The claims, published in Science Advances in 2025, suggested quantum computers were achieving useful results for chemistry.

The core contention centers on whether IBM's quantum samples accurately converge to the target electronic spin state—a singlet (⟨S²⟩=0)—at a matched computational cost compared to classical methods. The independent audit found that calculations consistently converged to higher spin states (⟨S²⟩ between 4.7 and 7.0) or spin-pure triplets, significantly deviating from the expected singlet state and reference energies.

The auditor, working independently, re-implemented key parts of the calculation and analyzed IBM's own raw hardware measurement records from December 2023 and April 2024. Even when running IBM's own shots through their pipeline, the results often showed large energy discrepancies or convergence to incorrect spin states.

IBM's Qiskit development team, upon receiving a bug report, acknowledged a flag that augments the sampled subspace but did not contest the numerical findings regarding the incorrect spin state convergence. This suggests a disconnect between the intended outcome and the actual computational result.

A significant aspect of this audit is the role of the large language model Claude. The LLM performed the initial end-to-end audit in approximately 72 hours, demonstrating not only speed but also a critical capacity for self-correction.

"It caught five defects in its own work through pre-registered validation gates, one of them by cross-checking against IBM's own published energy tables, and it retracted its own strongest pro-quantum finding when the new instrument showed that result was a spin-sector artifact."

This highlights Claude's ability to identify and rectify its own analytical errors, enhancing the rigor of the scientific validation process. For professionals in Tokyo, this illustrates how LLMs are evolving beyond simple content generation to become sophisticated tools for critical analysis and validation in highly technical domains.