acceptodds
Under review as a conference paper at ICLR 2027

From Registered Models to Authorized Actions: Zero-Knowledge LLM Agents with Model Binding and Private Tool-Call Authorization

Abstract

An LLM inference proof establishes computational correctness, but does not by itself establish use of an approved model or authorization of generated actions. We present an interactive zero-knowledge prototype for Q12 fixed-point GPT-2 Small that addresses both gaps with public prompts and private weights. Intermediate generation and denied-call contents remain hidden from the verifier; final answers and accepted calls reach their authorized recipients. First, a streamed Reed–Solomon check links every private weight label consumed by inference to an approved registration. Starting from those labels, a shared lookup interface over information-theoretically authenticated values supports both generation and authorization. It supports private embedding reads, online secret-logit lookup, and fixed-round parsing while keeping token IDs authenticated; its range-aware fused Q12 lookup evaluation (the fused Q12 path) reuses evidence across nonlinear computation. Our contribution lies in composing these relations rather than in introducing a new lookup argument. One-shot verification proves the exact top- set under Q12 logits with canonical tie-breaking; the verifier-provided seed and proved sampling rule, rather than prover discretion, fix each next token. Private parsing then extracts call identifiers from verified hidden tokens, and public-policy checks determine authorization. For gateway-controlled resources, an accepted call is bound to a one-time capability. The trusted gateway learns the accepted call and executes a registered adapter: authorization is verified, but tool execution itself is not ZK-proved. Evaluation covers eight inference shapes. At , matched ablations show that the fused Q12 path reduces full verification time per evaluated row by 78.2–79.4%, while secret-logit lookup makes softmax top- sampling faster than selector sampling. For 124.4 million weights, the streamed link reduces online model-binding time by a factor of relative to a per-weight LowMC masking/recovery comparator implemented in the same codebase. Across seven bounded agent tasks in each of three independent 12-round trials, the system executes authorized calls and rejects the tested unauthorized transfer and forged administrator claims before execution.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
interactive zero-knowledge proofs, large language models, verifiable inference, AI agents, model binding, private tool-call authorization

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_pJmZOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents an interactive zero-knowledge prototype that connects three components: binding private inference weights to an approved model registration, verifying private Q12 GPT-2 generation and sampling, and authorizing hidden tool calls before execution.

The system uses a streamed Reed--Solomon link for model binding, authenticated lookup-based inference and parsing, and a policy check that issues a one-time capability for an accepted call. Experiments on GPT-2 Small show substantial speedups over matched internal baselines for model binding and fused Q12 verification, and seven bounded agent tasks demonstrate that authorized calls are executed while selected unauthorized calls are denied.

Critical strengths

1. The paper addresses two gaps that ordinary inference proofs do not solve.

Binding the actually consumed private weights to an approved registration and proving authorization of a generated action are both meaningful extensions beyond proving inference correctness alone.

2. The implementation and evaluation are unusually careful for a prototype.

The paper provides matched ablations, explicit trust boundaries, detailed soundness accounting, Q12--FP32 fidelity measurements, and a reproducible code artifact. I also appreciate that the authors clearly distinguish verified authorization from trusted tool execution.

Critical weaknesses

1. The end-to-end security guarantee is substantially weaker than the high-level framing may suggest.

The confidentiality argument assumes a semi-honest verifier, the registrar and gateway remain trusted, and tool execution itself is not proved. More importantly, the reported 40-bit target applies only to enumerated added checks rather than the full EMP/VOLE and cryptographic stack. For an authorization system, I would like to understand the cost of reaching a more conventional security level such as 80 or 128 bits.

2. The practical evaluation remains very narrow.

The prototype is evaluated only on Q12 GPT-2 Small with public prompts, bounded BPE call syntax, a trusted mock gateway, and seven handcrafted agent tasks. The full agent experiment also requires about 14 seconds per verified token and roughly 81 GiB of communication for 90 tokens. This establishes bounded feasibility, but not yet practical applicability to contemporary tool-using LLM agents or general JSON/MCP-style calls.

Questions for authors

Q1. What would the runtime and communication cost be if the added checks were parameterized for an 80-bit or 128-bit soundness target? Can the authors provide an end-to-end security statement that clearly separates the added-check error from the inherited backend and cryptographic assumptions?

Q2. How much of the proposed model-binding and authorization contribution is fundamentally new relative to combining zkAgent-style transcript verification with existing witness/commitment-linking techniques such as Artemis? A sharper statement of the irreducible technical novelty would help.

Q3. Can the authors provide evidence beyond the seven bounded tasks, for example more realistic JSON/MCP-style tool arguments or a larger model/configuration? Even a small additional experiment would help establish how much of the current design generalizes beyond the fixed parser and GPT-2 Small setting.

Reviewer_6ZhGOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents a zero-knowledge prototype that links private model weights to an approved registration. It also checks that hidden generated tool calls satisfy a public authorization policy before execution. Experiments with GPT-2 Small demonstrate the combined system and measure several efficiency improvements.

Critical strengths
  1. Connecting model identity, verified generation, and action authorization addresses an important gap in inference-only proofs.
  2. Matched ablations help explain which optimizations reduce verification costs.
Critical weaknesses
  1. Evaluation is limited to GPT-2 Small and bounded tool-call formats, so support for modern agents remains unproven.
  2. Agent verification still requires substantial time and communication, limiting practical deployment.
Questions for authors
  1. Which components would become the main bottlenecks when scaling to a modern billion-parameter model?
  2. What changes are needed to verify general tool arguments rather than bounded call prefixes?
  3. What protection remains if the trusted gateway executes an authorized call incorrectly?
Reviewer_x1xYOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper connects registered model weights to the authenticated values consumed by verified inference. It then carries private generated tokens through parsing and authorization to a gateway capability. Its Q12 GPT-2 Small prototype combines streamed model binding, authenticated lookup and decoding, and evaluation of a public policy over private generated data.

Critical strengths

The interfaces connecting the components are comparatively explicit. Randomized folding and column checks connect registration to live inference labels. Boundary and tie constraints specify decoding; authenticated token identities continue into parsing and credential checks.

Critical weaknesses
  1. The evidence establishes feasibility more strongly than substantial practical value. Seven prescribed cases, deterministic maximum decoding in the integrated run, and finite tool/role/credential checks leave natural agent workloads and richer argument authorization largely unexplored.

  2. The comparative evidence does not establish a broad advance over the closest system alternatives. The 354.8× binding ratio is against a custom per-weight LowMC design, expressly not the complete Mystique implementation.

Questions for authors
  1. What incremental cost does private transcript handling and authorization add to identical bound infer- ence, and which workloads obtain a meaningful benefit from it?
  2. How does the binding bridge compare with a practical alternative under matched security, registration, and reuse assumptions?
Reviewer_2bvHOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents an interactive zero-knowledge prototype connecting approved private model weights to authorized tool calls. A streamed Reed–Solomon/Merkle consistency check links registered parameters to the authenticated values consumed by Q12 GPT-2 inference. The resulting hidden generation is checked through canonical decoding, private parsing, and credential-based policy evaluation. Accepted calls are bound to one-time capabilities enforced by a trusted gateway; tool execution itself remains trusted. The evaluation combines controlled implementation comparisons with seven bounded agent tasks, supporting the feasibility of this weight-to-action composition under the stated assumptions.

Critical strengths
  1. The paper addresses a meaningful gap between correct inference, approved model identity, and permission to act. Its strongest contribution is the concrete authenticated chain connecting registered parameters to the weights actually consumed by inference, then carrying verified hidden tokens through parsing, policy checks, and capability issuance. This provides useful systems knowledge beyond an isolated inference proof.

  2. The evaluation supports the paper’s principal implementation claims within clearly stated boundaries. Matched public-token inference ablations show substantial benefits from the fused Q12 path, while the model-binding experiment demonstrates a large improvement over the implemented per-weight LowMC comparator. The bounded agent cases additionally exercise permitted actions, explicit denial, and credential mismatch. The authors’ clear separation of verified authorization from trusted execution strengthens the interpretation of these results.

Critical weaknesses

My main reservation concerns the precision of the Q12 specification in Appendix A.6. The construction uses an authenticated nonzero check to exclude the integer representation zero while retaining p, although both have the same field residue. The text should explicitly identify the representation-level predicate and its correspondence to the integer reference. This matters to the claimed equivalence of the fused computation, but the current evidence establishes a specification ambiguity rather than an implementation defect. I consider it a localized and potentially straightforward issue to resolve, and it does not presently outweigh the paper’s contributions.

Questions for authors
  1. Could you specify the exact authenticated predicate in Appendix A.6 that excludes the all-zero limb representation while admitting the integer representation p? A constraint or pseudocode excerpt, together with endpoint checks showing acceptance of the intended representation and rejection of the alternative, would address my main remaining correctness concern.

  2. Could you point to the versioned integer-reference definitions and existing equivalence tests for the fused Q12 operators, particularly their rounding, clamping, and boundary behavior? This is a clarification request about the exact proved semantics; an expanded benchmark suite is not necessary.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.