Hi everyone,
I’m preparing my first arXiv submission in cs.LG and am seeking an endorsement.
Paper: Coupled Scaling: A Representational Accessibility Framework for Neural Scaling Laws
The paper asks why systems trained on the same data can exhibit different scaling behavior when architecture or optimization changes the task-relevant representations they can efficiently reach. It introduces Coupled Scaling, a task-conditioned framework connecting finite-budget neural scaling to representational accessibility.
Main results:
In a solvable mode-truncation model, loss separates exactly into target energy outside architectural support and an unresolved supported tail.
For arbitrary acquisition orders, the residual exponent lies in ργ≤α≤γ. Under bounded off-prefix gain, the lower endpoint is exact: α=ργ. A fixed-kernel specialization links the training-time exponent to a task-weighted spectral measure defined independently of the loss fit.
The paper proposes a preregistrable factorial test using matched-budget geometry, multiscale exponent ordering, and a predicted reversal across contrasting tasks.
For transparency, reproducibility materials, analysis scripts, and supporting results are publicly available here: The current manuscript is available privately upon request.
If you are eligible to endorse submissions to cs.LG and consider this paper appropriate for the category, I would be very grateful for your help. This is only a request to confirm subject-area suitability, not a request for peer review or agreement with the paper’s conclusions. Endorsement link:
Endorsement code: ZAWO79
(The fourth character is the capital letter O, not zero.)
ORCID:
I’m happy to answer any questions about the paper. Thank you very much for your time.
Jie Wang
School of Civil and Commercial Law
Southwest University of Political Science and Law