If the optimizer is the teacher and the model is the student, then the business of "teaching models how to optimize" may itself be the missing layer of AI-native networks. That is the claim of "Learning-to-Optimize as the Missing Architectural Layer of AI-Native Networks," posted to arXiv on September 18: redefine optimization algorithms as offline knowledge generators, let neural surrogates inherit the expert experience, and use low-latency inference for dynamic network environments.
[1]The paper: arXiv:2609.21519 (submitted September 18, a 6-page conference paper), by Giambattista Amati, Federica Mangiatordi, Pierpaolo Salvo, Emiliano Pallotti, and Simone Angelini. This is a perspective piece, not an experimental report — its contribution is an architectural abstraction, not a reproducible leaderboard.
The authors' core argument: current AI-native architectures embed intelligence across network functions (resource management, adaptive control, zero-touch operations) while saying almost nothing about how optimization knowledge should be systematically generated, transferred, and exploited by AI models. Their answer is Learning-to-Optimize (L2O) as a missing architectural layer, with a generic four-stage workflow: optimization (an exact solver produces high-quality solutions), knowledge generation (those solutions become supervision), surrogate learning (a neural surrogate model trains on them), and runtime inference (low-latency online decisions).
The idea itself is not new — the L2O community has spent years making optimizers faster with networks. What this paper does differently is promote L2O from a trick to an architectural abstraction: not "make this optimizer faster" but "turn the exact optimizer into an offline teacher and the lightweight model into an online student," claimed to be reusable across heterogeneous communication and computing systems. The authors illustrate with an NR-V2X relay-selection problem: a Mixed-Integer Linear Programming (MILP) solver generates solutions, and a Graph Neural Network learns to reproduce near-optimal decisions in real time.
Limitations worth naming. First, this is a position paper: 6 pages, conference length, no large-scale experiments; NR-V2X is illustrated, not systematically evaluated. Second, the engineering cost of the four-stage pipeline is understated — generating offline supervision means running the expensive solver continuously; the "offline teacher" just moves the cost out of the hot path, it does not remove it. Third, generalization boundaries are untouched: would the surrogate still work under a different topology or a different constraint class? Fourth, there is no baseline comparison against direct online schemes (heuristics or RL-connected decision), so the actual gain of "low-latency near-optimal" is unmeasurable from the paper.
Still, as an architectural claim, the observation lands: if network autonomy just replaces "compute the optimization online" with "run a neural net online," then where the net's optimization ability comes from — who teaches it, and whether the result transfers — is a gap most AI-native architecture documents simply skip. Putting L2O into the architecture diagram at least makes the question visible. For teams doing network-side AI, the paper's value is a reminder: before arguing whether to use AI for optimization, answer where the optimization knowledge comes from.
[1]