A proof verifies a 70B model's run in 46 seconds

Researchers verified a full Llama-2-70B inference run in 46.3 seconds without touching the model's weights, using a technique called sampled layerwise proofs.

That matters for anyone renting inference by the token. The method lets a customer confirm a provider ran the full 70-billion-parameter model instead of a cheaper stand-in. Generating that proof took 1,259 seconds on a CPU host and produced a 4.34 MiB file that anyone can check without the weights.

The trick is committing the boundary activations for every chunk of an inference trace before any challenge goes out. A verifier then picks a subset of chunks to prove after the fact. On a smaller TinyLlama test, proving 7 of 47 chunks cost just 22% of what a full proof would take.

Batching cuts costs further. Packing twelve separate requests into one trace made proving 6.5 times cheaper than proving each request on its own. Overhead falls as load rises.

The paper doesn't say what happens if a provider games which chunks get sampled. It also doesn't say how the scheme holds up when a model's weights change mid-session. Those gaps matter more than the benchmark once someone tries to put this into an actual billing dispute.