Harnessing Code Agents for Automatic Software Verification

2026-07-07 · Source: Takara TLDR - Daily AI Papers · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Expert, quick

Summary

Aria, an LLM code agent utilizing models like Claude Code, significantly advances automatic software verification by achieving full proof coverage without expert intervention. Unlike prior approaches that impose fixed proof strategies, Aria allows the LLM to choose its own method, guided by a verification harness. This harness enforces soundness, completeness, and termination. Evaluation shows Aria proves all 4,257 lemmas of Iris's core modules and 217 Rust standard library lemmas. It also proves all 318 reglang lemmas, vastly outperforming previous LLM provers. Furthermore, Aria demonstrated generality by proving 72 lemmas on iris-lean, an unfinished Lean 4 port, confirming its applicability beyond Coq.

Key takeaway

For research scientists and software engineers focused on ensuring software correctness in critical systems, this work demonstrates that LLM code agents can fully automate formal verification. You should consider integrating such agents, like those powered by Claude Opus 4.7, into your development and testing pipelines. This approach drastically reduces the need for specialized Coq or Lean expertise, enabling scalable and complete proof generation for complex concurrent and memory-manipulating programs, thereby enhancing overall system reliability.

Key insights

LLM code agents, when unconstrained by fixed strategies and guided by verification harnesses, can fully automate formal software proof generation.

Principles

• Formal verification's scalability is limited by expert effort.
• Fixed proof strategies constrain LLM effectiveness in verification.
• Verification harnesses ensure proof soundness, completeness, and termination.

Method

A general LLM code agent receives a lemma, generates proofs, and operates under a verification harness that provides feedback and hard constraints to ensure the proof is sound, complete, and terminating, closing it via the prover's kernel.

In practice

• Apply to Iris separation logic for concurrent programs.
• Utilize Claude Opus 4.7 for proof generation.
• Extend verification to Lean 4-based projects.

Topics

• Formal Verification
• LLM Code Agents
• Software Correctness
• Coq
• Iris Separation Logic
• Claude Opus

Best for: AI Scientist, Research Scientist, Software Engineer

Related on AIssential

• VeruSAGE: A Study of Agent-Based Verification for Rust Systems — LLMs, when paired with specialized agent systems, can effectively generate correctness proofs for Rust system software.
• Autoformalization of Agent Instructions into Policy-as-Code — An LLM-driven autoformalization pipeline translates natural language policies into formally verified Cedar policies for enhanced agent safety.
• Components of A Coding Agent — Agentic harnesses significantly enhance LLM coding capabilities by integrating tools, context, and memory management.
• Exploring LLM-based Verilog Code Generation with Data-Efficient Fine-Tuning and Testbench Automation — Multi-agent LLMs can automate testbench generation, creating high-quality data for efficient Verilog code generation.
• LEAP: Supercharging LLMs for Formal Mathematics with Agentic Frameworks — Agentic frameworks like LEAP enable LLMs to achieve state-of-the-art formal theorem proving by bridging informal reasoning with verifiable proofs.
• Beyond Correctness: Enhancing Architectural Reasoning in Code LLMs via Scalable Labeling with Agentic Judgment — An agentic LLM pipeline scalably labels architectural quality, significantly improving code LLM reasoning and patch conformance.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Takara TLDR - Daily AI Papers.