A Contract-Grade Verifier for LLM-Generated GPU Kernels, and a Native Blackwell Backward for the Gated-Linear-Recurrence Family

Most LLM-written GPU kernels marked correct still break on NaNs, shapes or repeat runs.

39.5%

Accepted kernels broken beyond tolerance

These kernels already passed their source system's own correctness harness.

How it works

Loose correctness test
Kernel benchmarks accept output close to a reference on a few random inputs at one shape.
Twelve adversarial gates
A verifier probes NaNs, new shapes, repeat runs, precisions and accumulators. Several gates use exact matching, no tolerance.
Gap is large
Of 2,638 accepted kernels, 62.1% violate a contract. KernelBench passes 1,487 the verifier rejects, versus 14 reversed.

Which comes out on top on Gate failure rate?

try it

Drag the accept threshold and watch correct and broken kernels cross it.

Accept threshold0.5

Any tolerance trades false accepts for false rejects; tolerance-free gates avoid that choice entirely.Illustrative model, not the paper's data

6 terms

Gate One property a correct kernel must satisfy, checked against a slow reference.

Like a car inspection that only checks the engine starts; a full inspection also tests brakes, lights and steering under stress.

Where it breaks

Kernel gates compare against an exact reference answer, cars have no such reference.

From the paper