Skip to content
AI.info

The Pulse

Amazon releases tools to check AI-generated mathematical proofs

AWS has made three tools public for connecting AI agents to the Isabelle proof assistant. Amazon researchers describe how the tools can check formal proofs as an agent develops them.

Amazon releases tools to check AI-generated mathematical proofs

AWS has made three tools public for connecting AI agents to a proof assistant that can check their mathematical work as they develop it. In an Amazon Science post, researchers Hanno Becker and Larry Paulson describe the tools as a way to catch errors in AI-generated proofs without waiting until an entire proof is complete.

The tools connect an AI agent to the Isabelle proof assistant through the Model Context Protocol, or MCP. They are part of the AutoCorrode verification suite and give users different ways to work with an agent during a proof session.

Checking proofs as they develop

Formal proofs must be written in a precise logical language that a computer can check. Proof assistants provide an environment for writing and checking that material; the researchers say checking can be interleaved with generation so that an agent can revise its work as it goes.

The researchers describe other proof assistants, including Lean and HOL Light, alongside Isabelle. AWS uses HOL Light to formally verify high-performance cryptographic code and Isabelle to verify the Nitro Isolation Engine.

Some proof assistants offer a command-line interface through which a user enters commands and receives replies. Others have interfaces designed for human users; the post describes MCP as a connection between an agent and a tool such as a proof assistant.

Three ways to use Isabelle

The new tools are publicly available in AutoCorrode. I/Q lets an agent communicate with a live Isabelle session running on a user’s computers. The user can temporarily hand over control of the assistant to the AI to fill in formal material or assess the context and offer advice.

I/R takes a different approach: the agent can develop proofs or other formal material while the user continues working with the proof assistant. The agent still communicates directly with Isabelle to check whether its proofs work, but does not take control in the same way as I/Q.

Isabelle Assistant uses I/Q to provide a chat window through which an agent can observe and modify the user’s session, including developing or fixing proofs. The post describes it as less powerful than the full I/Q and I/R package, but available in a convenient window without additional setup.

Examples and a correction

The authors report using Isabelle Assistant with Claude Opus 4.6 to produce formal proofs for established mathematical examples. One example concerned the irrationality of the square root of two: the argument assumes it can be expressed as a fraction in lowest terms and reaches a contradiction.

The post calls the resulting formal proof idiomatic, though more detailed than necessary. Its authors note that the proof language is designed both to be checked by computer and to be read by people.

Another example asked for a proof that, for every integer N greater than zero, a block of N consecutive composite numbers exists. The researchers supplied Claude with a proposed proof that contained an error. According to the post, Claude identified that the argument provided only N minus one consecutive integers and corrected it.

The researchers also list a proof involving the definite integral of the sine function over the interval from zero to pi among their examples. They present the demonstrations as cases in which an AI agent assists mathematicians rather than replaces them.

Source

  1. Amazon Science

Related stories

Explore

More articles