← News

Agents & MCPSources: arXiv · 16 Nov 2026

Hybrid LLM and mechanical inference automates Move specification writing

A new arXiv paper pairs an LLM agent with weakest-precondition analysis to auto-generate formal specifications for Move smart contracts, cutting inference costs in the Move Prover.

In plain words

Formal verification proves code is correct, but writing the required specifications by hand is expensive, so the authors built a system that writes them automatically. An LLM agent, guided by a mechanical analysis technique, drafts contracts for Move smart contract functions. In their tests it finished most runs successfully and cost less when the mechanical analysis was switched on.

What was announced

The paper, an arXiv abstract by the authors of the study, argues that manual specification writing has long blocked wider adoption of formal verification, since proving high-level properties also requires contracts for many supporting functions that are mechanical and low-level. Their answer is a hybrid of mechanical and agentic specification inference for Move, a smart contract language, integrated into its verifier, the Move Prover. An LLM agent drives inference and verification through a tool that calls the prover and derives specifications from code using weakest-precondition (WP) analysis, a technique that works backward from a function's outcome to the conditions its inputs must satisfy.

The evaluation covers 26 targets, 24 of them drawn from public, previously unspecified Move functions in the Aptos framework, tested with a frontier model and a mid-tier model from the same family, named Sol 5.6 and Terra 5.6. The agent completed over 90% of its runs both with and without WP, which the authors say shows agentic inference is effective on its own. Adding WP lowered overall cost by 11-21%, and by 40-64% on targets without loops. The loop-free figures suggest the mechanical component helps most where control flow is simple, though the abstract does not report results for individual targets or models separately.

What it does not show

  • The evidence is the paper's abstract only; full methodology, per-target results and peer review status are not available in the supplied excerpt.
  • Reported completion and cost figures are author-reported aggregate results and have not been independently confirmed.
  • The abstract does not state whether the generated specifications were accepted into the Aptos codebase or how they perform on functions with loops beyond the aggregate cost range.

For security and IT teams

Scope
The work targets the Move smart contract language and its Move Prover verifier, with evaluation on functions from the Aptos framework.

Sources

Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 17 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.