Job Description About the Team The Compression and Parity team in GM's Autonomous Vehicle organization makes aggressive model optimization safe enough to ship repeatedly. We compress and quantize models headed for the car, and we own the analytical machinery that proves the compressed model still behaves like the original — not by assertion, but with bounded, quantified, auditable evidence. Every deployment decision our organization makes about a compressed model runs through the tooling this team builds. About the Role We are looking for a mathematically rigorous engineer to own model numerics: how we measure, bound, and reason about the numerical behavior of the models we ship — and how we turn that analysis into deployment decisions. The central question of this role is deceptively simple: given two numerically different versions of the same model, is the difference safe? Answering it well requires connecting things that are usually studied separately — floating-point drift and Hessian conditioning on one end, vehicle trajectory error on the other. You will build the tooling that makes that connection quantitative, and you will define the thresholds that turn it into a ship / no-ship decision.

This role is not:

running an existing validation harness and reporting the numbers it produces. When a parity check fails, the expectation is that you can say which operation caused the divergence and why — not merely that a difference exceeded a threshold. The tooling exists to make that investigation fast; it does not replace the investigation itself. What You'll Do Validate Optimized implementations . Optimized implementations are supposed to be equivalent to their references. Establishing that rigorously, rather than by spot check, means deciding what equivalence should mean for a given operation, and designing the inputs that would expose a violation if one existed. Connect tensor differences to behavioral disparity. Map low-level numerical differences from quantization, compilation, and precision reduction to downstream driving behavior, using both open-loop metrics (trajectory displacement error, perception IoU) and closed-loop outcomes — and identify the mechanism behind the mapping, not just the correlation. Build sensitivity and robustness analysis tooling . Use Jacobian/Hessian-based methods to characterize how model outputs respond to weight and input perturbation, extend the same machinery to out-of-distribution inputs, and turn it into tooling that runs repeatedly across checkpoints — by engineers who are not you. Build training dynamics observability . Design diagnostics that detect and root-cause training instabilities — gradient vanishing and explosion, loss spikes, silent divergence — including decompositions of gradient and update trajectories into loss-descent and oscillatory components under modern schedules such as WSD. Make it cheap enough to always be on . Metric computation, gradient decomposition, and diagnostic logging have to run inside real distributed training jobs with negligible throughput cost and no OOM risk. Observability nobody can afford to enable is observability that doesn't exist.

Requirements:

We are deliberately strict on a small number of things and flexible on everything else. A working command of numerical analysis and matrix theory. Jacobian/Hessian estimation, spectral properties, conditioning, and floating-point error analysis should be tools you reach for by reflex, not topics you once studied. You should be able to say why an estimator's variance blows up, and when that matters. A real mental model of neural network training. Loss landscapes, gradient and error propagation, optimizer dynamics, and the mechanics by which training goes wrong. You should have opinions about what a gradient norm spike does and does not tell you. An adversarial instinct for numerical edge cases. Typical inputs rarely find anything. We are looking for someone who reaches for denormals, extreme dynamic range, catastrophic cancellation, degenerate shapes, and accumulation-order effects — someone whose first question about a passing test is what that test failed to exercise. The engineering to make the math run . High proficiency in PyTorch and Python, and a track record of building analytical tools that are both mathematically defensible and fast enough to be used in production training and evaluation loops. The judgment to make analysis actionable. Much of this role is turning a numerical result into something an engineer who will not read your derivation can act on: knowing which quantity actually answers the question being asked, tracing an anomalous number back to the operation that produced it, etc Bachelor's, Master's, or PhD in Applied Mathematics, Control, Physics, Computer Science, Data Science, or a closely related quantitative field. What You'll Learn Here You are not expected to arrive with these. They are the parts of the job that are genuinely specific to this environment, and we expect them to take three to nine months: The training recipes, the acceleration techniques in our stack The semantics of autonomous driving behavioral metrics and what actually constitutes meaningful behavioral drift for a vehicle. Model compression portfolio, the deployment path to the car, and the organizational context around a ship decision. Our parity validation workflow and where the quantization and compilation toolchains hide their sharp edges. What Will Give You a Competitive Edge Hands-on experience debugging large-scale training runs: diagnosing loss spikes, resolving numerical divergence, and running deep investigations into training dynamics. Experience with distributed training beyond DDP — FSDP, Megatron-LM, DeepSpeed, 3D parallelism — particularly designing low-overhead observability over sharded parameter and gradient state. Experience in quantization, compiler toolchains, or inference-time numerical parity. Experience building or scaling evaluation pipelines and its metrics formulation for AV or ADAS systems. Published or applied work in model robustness, OOD generalization, or adversarial/perturbation analysis.

Compensation:

The compensation information is a good faith estimate only. It is based on what a successful applicant might be paid in accordance with applicable state laws. The compensation may not be representative for positions located outside of

New York, Colorado, California, or Washington Compensation:

The expected base compensation for this role is : $170,600 - $261,300 Actual base compensation within the identified range will vary based on factors relevant to the position.

Bonus Potential:

An incentive pay program offers payouts based on company performance, job level, and individual performance.

Benefits:

GM offers a variety of health and wellbeing benefit programs. Benefit options include medical, dental, vision, Health Savings Account, Flexible Spending Accounts, retirement savings plan, sickness and accident benefits, life insurance, paid vacation & holidays #GM-AV-1 This role is categorized as hybrid. This means the selected candidate is expected to report to a specific location at least 3 times a week {or other frequency dictated by their manager}. The selected candidate will be required to travel