Reinforcement Learning Control of Quantum Error Correction
Peer-reviewed Nature 2026; open access, 2026 · paper · Physics
Uses error-detection events from quantum error correction as a learning signal for a reinforcement learning agent that continually tunes control parameters during computation. On a Willow superconducting processor it improved surface-code stability 3.5-fold against injected drift.
Checked: PDF from nature.com, fetched 5 Oct 2026. Page numbers are those of that PDF.
1 finding
Substance The scaling model "εL ∝ Λ−d/2" with "Λ = εth/ε" and "C ∝ ε" gives a gradient coefficient of d/2, but the text states "∇logεL = (d + 1)/2 ⋅ ∇logC", and the Fig. 2c caption gives "proportionality coefficient (d + 1)/2". The Methods repeat "εL ∝ Λ−d/2".
- p. 2 Learning from errors, surrogate objective paragraph
the surface code εL ∝ Λ−d/2, where d is the code distance and Λ = εth /ε
- p. 2 Fig. 2c caption
with proportionality coefficient (d + 1)/2 (black line)
- p. 7 Methods, Objective function
the code distance d as εL ∝ Λ−d/2
- p. 2 Learning from errors, surrogate objective paragraph
the surface code εL ∝ Λ−d/2, where d is the code distance
Only the authors can say what is right here.
Judge's note: I checked the page 2 image. The right column reads "εL ∝ Λ^−d/2, where d is the code distance and Λ = εth/ε ... As C ∝ ε, the gradients of LER and of the surrogate objective in this model are simply related as ∇logεL = (d+1)/2 ⋅ ∇logC". The Fig. 2c caption reads "proportionality coefficient (d+1)/2 (black line)". The page 7 Methods crop also shows "εL ∝ Λ^−d/2". My own working: log εL = −(d/2)·log(εth/ε) + const = (d/2)·log ε + const = (d/2)·log C + const. So ∇log εL = (d/2)·∇log C, not (d+1)/2. For d = 5 that is 2.5 against 3. The coefficient (d+1)/2 would follow from εL ∝ Λ^−(d+1)/2. So either the printed exponent or the printed coefficient is wrong. Only the authors can say which is intended, but the coefficient used in Fig. 2c points to the exponent being the slip.
Second check: another model agreed
Report a problem with this finding
3 minor (typos and the like; most checked by the judge only)
Typo Ref. 16 has no space between the title and the parenthesis: "Processing Systems(eds Koyejo, S. et al.)".
- p. 5 Reference 16
In Proc. Advances in Neural Information Processing Systems(eds Koyejo, S. et al.)
Judge's note: The crop of the reference list on p. 5 shows no space before '(eds'.
Second check: none (judge only)
Report a problem with this finding
Typo Ref. 41 has no full stop after the group author: "Google Quantum AI Suppressing quantum errors...". Refs 38 and 47 have one in the same position.
- p. 6 Reference 41
Google Quantum AI Suppressing quantum errors by scaling a surface code logical qubit.
Judge's note: The crop of p. 6 confirms the full stop is missing in ref. 41 and present in ref. 38.
Second check: none (judge only)
Report a problem with this finding
Typo Two consecutive sentences open with "Furthermore": "Furthermore, C is non-stationary due to inherent system drift. Furthermore, evaluating C is expensive".
- p. 7 Methods, Learning algorithm
Furthermore, C is non-stationary due to inherent system drift. Furthermore, evaluating C is expensive
- p. 7
Furthermore, C is non-stationary
Judge's note: The page 7 crop (Methods, Learning algorithm) shows two consecutive sentences beginning with "Furthermore". This is a cosmetic copy-editing repetition and does not change the meaning.
Second check: none (judge only)