Intel / article

Agent Stack for ML Training Lifecycle: Generation, Validation, and Optimization

Source: arxiv.org
Author: Jaideep Ray, Daniel Tang, Zejun Lin, Haoyue Tang, Lijuan Zhang, Tao Huang, Tommy Li, Aditya Jain, Ankit Goyal, Harsh Jain, Zhuo Zhi, Youmin Han, Santosh Jha, Shuying Liang, Yujie Ai, Jonathan Hung, Ethan Lyu, Abi Liu, Siddharth Singh, Feng Wen, Mahesh Ravindran, Souvik Ray, Yitong Zhou, Animesh Singh
Date: Published Nov 3, 2026

Context

A paper describes an stack deployed at LinkedIn, with two agent loops for machine-learning work, and reports how often they produced usable models (arXiv). AutoTorch writes PyTorch models and training pipelines, to migrate TensorFlow models, replace XGBoost models, or start new ones. AutoResearch tries architecture and training-configuration changes to improve an existing model, either steered by an engineer or on its own. An outside orchestrator runs parallel trials, keeps experiment history, picks candidates and enforces time budgets, while a shared runtime holds execution state and allocates GPUs separately from the agents' execution.

The safeguard on generated training code is a set of executable validation hooks. Before an engineer reviews anything or a full distributed training run starts, the hooks check on sampled data that the model follows LinkedIn's platform APIs, actually trains, and, for migrations, reproduces the original model faithfully; failures feed back into repair. According to the abstract, AutoTorch completed generation and full-scale training in 92 of 115 runs (80%), and 77 of those (83.7%) met post-training quality requirements. In a separate set of 100 production pipelines, AutoResearch improved the main held-out metric for 28.

Those counts are the authors' measurements. The headline speedup is softer: engineering teams estimate end-to-end migration time fell from about eight weeks to two. The abstract does not say how that estimate was formed, and the 28-of-100 figure means most of those pipelines saw no improvement on their main metric.

Key takeaways · AI-distilled

LinkedIn's stack has two independently usable loops: AutoTorch generates PyTorch models and pipelines for TensorFlow migration, XGBoost replacement and new models, while AutoResearch explores architecture and training-configuration changes.
An external orchestrator, not the agents, manages parallel trials, experiment history, candidate selection and wall-clock budgets, and a shared runtime keeps execution state and GPU allocation separate from agent reasoning.
AutoTorch completed generation and full-scale training in 92 of 115 runs (80%), and 77 of those (83.7%) met post-training quality requirements, so about two-thirds of all runs ended in a model that passed.
AutoResearch improved the held-out primary metric for 28 of 100 production pipelines. The drop in migration time from about eight weeks to two is an estimate from engineering teams, not a measured result.

Terms in this piece · Glossary

AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.

Read the source arxiv.org