AI-assisted proof lifts convergence rate for stochastic multi-gradient descent
An arXiv note reports a faster convergence bound for SMG in multi-objective optimization, improving on a 2024 result without assuming bounded gradients.
In plain words
SMG is an algorithm for optimizing several objectives at once using noisy gradient estimates. The author proves it reaches a balanced solution faster than previously shown, and says a ChatGPT model supplied the initial proof strategy, which the author then checked and reorganized.
What was announced
In an arXiv note, the author establishes a new convergence rate for stochastic multi-gradient descent (SMG) on smooth nonconvex stochastic multi-objective optimization, measured by the squared Pareto-stationarity measure. With a constant stepsize and linearly growing mini-batches, the expected squared empirical measure at the output is O~(T^-1) after T iterations, improving on the O~(T^-1/4) bound of Chen et al. (JMLR, 2024) for linearly growing batches, and the argument does not require bounded gradients. The improvement comes from exploiting Lipschitz continuity of the Pareto-stationarity measure, defined via the MGDA direction, instead of the half-Holder continuity of that direction used in the earlier work.
For MoCo and MoCo+ variants with exact MGDA updates and constant batch sizes, the analysis yields expected squared empirical rates of O(T^-1/2) and O~(T^-2/3), which the author says match single-objective momentum rates up to logarithmic factors. The proof originated when the author, preparing graduate homework, prompted ChatGPT 5.4 Thinking Extended and received the initial strategy, then verified and reorganized it; appendices document the prompt, student submissions with LLM disclosures, and further momentum-based results.
What it does not show
- The source is the paper's abstract only, so the full proofs, assumptions and results are not independently verified here.
- The claimed rate improvement rests on the author's own analysis; no independent confirmation is provided in the excerpt.
- The AI contribution is limited to the initial proof strategy for the main SMG result, which the author states was verified and reorganized; the excerpt does not establish how the verification was performed.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.