9.6 KiB
| title | date | abstract | tags | status | confidence | importance | evidence | scope | novelty | practicality | history | |||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Verified Inference Between Adversaries | 2026-08-29 | A compute operator who claims to have run a particular model can be lying, and the logs that would settle it are written by the party under suspicion. VerInf produces zero-knowledge proofs of LLM inference, bounding the information in an output stream that a committed model does not account for. This is a working note on what the system certifies, what it does not, and what I contributed to it during the MARS V fellowship. |
|
Working model | 85 | 4 | 4 | broad | moderate | moderate |
|
A datacenter tells you it ran a 400-billion-parameter model on your prompt. It might have run a smaller one, an older checkpoint, or a quantized copy that costs a fraction as much. Ordinary logging cannot settle this, because the logs are produced by the party whose behavior is in question. Neither can a hash of the weights: the operator can hold the right weights and still run something else.
This is not a hypothetical failure mode so much as an unresolved gap in every current deployment story. Evaluations measure a model under conditions you control; they say nothing about the model served to somebody else afterward. Attestation binds a binary to a machine; it does not bind an output to a computation. What is missing is a way for a verifier to check a claim about a computation without trusting the party that ran it.
VerInf is a research prototype addressing that gap, built by James Petrie at the Future of Life Institute. I worked on it as a MARS V fellow. This note describes what the system proves, what it does not, and which parts of it are mine.
What the proof certifies
The naive framing — "prove the model produced this output" — is the wrong one, and understanding why is most of the insight. Frontier inference runs in floating point, on nondeterministic kernels, across hardware that does not reproduce bit-for-bit. Demanding exact reproduction would make the problem intractable and would not even be the property you want.
VerInf instead bounds the unexplained information in an output stream: the number of bits in the output that the committed model does not account for. If the operator swapped in a different model, the outputs it produced would be poorly predicted by the model it committed to, and the bound rises. A cheap substitution is expensive to hide.
The certified statement is a conjunction of three parts, and they are held to different standards:
- a transcript anchor, binding the committed token streams to digests recorded independently at generation time, so the proof is about the run that actually happened rather than a convenient reconstruction;
- the forward pass, where every claim must admit exactly one satisfying assignment, because slack in an intermediate value propagates through the remaining layers in directions nobody can analyze;
- the surprisal bound itself, where freedom is permitted — provided every free direction pushes the reported number up rather than down.
That asymmetry is the elegant part. Upstream of the logits the prover must have no room at all; downstream, in the short arithmetic from logits to the reported bound, the prover may have room so long as every rounding is forced upward. Cheating can only make your own number worse.
A related move governs the predictor. The bound is computed against a predictor of the deployment's outputs that the prover supplies — which sounds like handing the adversary the pen. By Gibbs' inequality the resulting sum is a valid upper bound for any predictor, so the choice can be left to the prover entirely. The scheme is sound whatever they pick; a bad choice only inflates their own reported bound.
Underneath, matrix products are checked with Freivalds projections over Ligero commitments, which are hash-based — so there is no trusted setup, and the construction is plausibly post-quantum.
Who trusts whom
The two parties want different things, and the design is legible once you see that.
The prover wants confidentiality: weights, activations, and both token streams must not leak. The verifier wants soundness: the reported bound must be genuine. Their interface is a public claim list stating what kind of computation was performed — which reveals the model architecture, though not the weights.
Each side trusts only its own code. The verifier's trusted base is deliberately small, and it shares no code with the prover. A fault anywhere on the prover's side can cause a proof to fail; it cannot cause one to falsely verify.
Results on record
These are the published single-chip numbers, and they predate my involvement.
| Llama-4-Maverick | Llama-2-7B | |
|---|---|---|
| Parameters | 400B MoE, 48 layers, 128 experts committed per layer | 7B, 32 layers |
| Transcript | 1000 tokens, all hidden | 1000 tokens |
| Prove | 14.3 h, 78.1 GB GPU peak | ~44 min, 11.2 GB peak |
| Verify | 17.7 h, 20 CPU cores, 40 columns opened | ~23 min, 10 columns |
| Proof size | 93.6 GB | 1.44 GB |
| Bound | 0.880 bits/token | — |
All on a single NVIDIA DGX Spark. The committed witness for the Maverick run is roughly 7.2 TB, streamed at the working set rather than held.
The bound is the number to read. At 0.880 bits per token against a 202,048-token vocabulary, the proof accounts for about 95% of the information a token could carry. That is not "the model produced this"; it is "very little here is unexplained by the model that was committed."
What I contributed
Parallelizing the prover across GPUs is named future work in the paper, and it is the direction I am working in. It runs into an immediate problem: you cannot measure a cluster you do not yet have access to, and partitioning decisions have to be made before the hardware arrives.
So the first deliverable was not parallelism — it was the measurement scaffolding parallelism needs. I built a dry-run profiler that predicts a proving run's time, memory, bandwidth, and proof size before running it, from a workload manifest plus measured hardware constants, and exposes the dependency structure a future scheduler will consume:
- a manifest contract — one record per tape operation, with two independent producers (an exact tape-walker that runs where the prover runs, and closed-form builders in pure Python that cross-check it);
- a cost model in per-claim accounting form, carrying the production expressions from the paper's cost appendix;
- a claim-level dependency DAG with critical path and width profile, which is what tells you how much parallelism actually exists;
- a partition scorecard that maps claims onto N shards under competing strategies — contiguous tape ranges, pipelined layers, expert-sharded — and scores them.
One deliberate constraint: none of it touches the soundness-critical prover path. A profiler that could perturb the proof would be a liability in a system whose entire value is that a fault cannot cause false acceptance.
I also corrected the cost model's RMSNorm row to the wrap-free bracket constants, bringing it into line with the paper's own analysis.
Both are merged upstream:
::: {.work-entry-links} Merged pull requests · Upstream repository :::
In progress: the multi-GPU port itself. The aim is not parallelism for its own sake — multi-device proving is what lifts the ceiling on model scale, context length, and mixture-of-experts breadth. Results are preliminary and unmerged; I will report them here when they are not.
What this does not do
The honest list, mostly from the paper's own limitations section:
- No security review. It is a research prototype.
- The proofs are enormous — 93.6 GB at full scale. This is a real deployment obstacle, not a rounding error.
- 0.880 bits per token is not zero. For some applications that residue is fine; for others it is not, and the paper is explicit that tightening it is open work.
- Soundness is a per-challenge bound (about 2⁻¹⁶·⁶ in the demonstrated configuration), raised by opening more columns at a measured cost in verification time. It is a deployment choice, not a fixed property.
- The claim list reveals the architecture. Hiding that too is future work.
- The transcript anchor is not yet demonstrated end to end. The AES and SHA-256 circuits are implemented and tested; binding them against independently recorded digests is not yet shown.
That last one is the gap I would press on hardest if I were reviewing this. The anchor is what makes the bound a statement about a particular real run rather than about some internally consistent transcript. Everything else is cryptographic machinery; that is the part that connects it to the world.
Why this matters for assurance
The reason to care is not that datacenters are presumed dishonest. It is that "trust us, we ran the model we said" is the current state of the art, and it does not scale to a world where the stakes of that claim keep rising — third-party evaluation, regulatory audit, compute governance, any arrangement where an inference claim carries weight and the parties are not aligned.
Verified inference is one approach to a general question: how do you establish trustworthy claims about an AI system when the system, the operator, and the evaluator may each be untrusted? Cryptography attacks it from below. Evaluations attack it from above. Neither is sufficient alone.