Vibeleaderboard
← All Intel
Intel / article

Mathematical proof, powered by AI. And no hallucinations!

Source
amazon.science
Date

Context

An AI model can invent a mathematical proof that looks right, but a proof assistant will not accept one that is wrong, and Amazon has released tools that put the two in the same loop. A proof assistant is a program that checks a proof written in a precise formal language, step by step, and rejects any step that does not follow. In an Amazon Science post, Hanno Becker of Amazon's Automated Reasoning Group and Larry Paulson, a Cambridge emeritus professor of computational logic and Amazon Scholar, describe connecting AI agents to Isabelle, a proof assistant AWS already uses to verify its Nitro Isolation Engine. The submits its proof to the checker as it writes, so a made-up step is rejected and has to be repaired before the work goes on.

The tools are in AWS's open-source AutoCorrode verification suite. Most proof assistants are built around rich editors for people, so an agent needs a plainer channel, and these tools provide one through the Model Context Protocol, which the authors note all the leading AI agents support for connecting to tools. I/Q lets an agent take temporary control of a live Isabelle session on the user's computer. I/R lets an agent develop proofs in a background Isabelle process while the person keeps working, and I/Q can call I/R. Isabelle Assistant, built on I/Q, adds a chat window inside Isabelle where an agent can read and change the session.

The post's examples show what this looks like in practice. Anthropic's Claude formalized classic results, including the proof that the square root of two is irrational, and when asked only for "a nice simple example involving calculus", it proved that the integral of sine from 0 to π equals two. When the authors gave it a flawed informal argument about runs of consecutive non-prime numbers, Claude noticed the argument produced one number too few and corrected the theorem. These are textbook examples, not new mathematics; the authors say the hardest problems still need human thought and that handing an agent long, tiring work, such as cleaning up a messy proof, leaves people fresh for the key decisions. The checker stops hallucinated steps, and the authors also keep the proofs readable so a person can check them too, which they see as the way for an agent and a human to tackle hard problems together.

Key takeaways · AI-distilled
  • AWS released three Isabelle tools in its AutoCorrode verification suite: I/Q lets an agent take control of a live session, I/R lets an agent develop proofs in a separate subprocess while the user keeps working, and Isabelle Assistant wraps I/Q in a chat window.
  • The authors say interleaving proof checking with generation keeps the agent on track: the proof assistant checks each step as the agent writes, so the AI-generated proofs are error-free even when longer than a human would write them.
  • Given a flawed proof that N consecutive composite numbers start after N!, Claude noticed the argument only yields N-1 composites and corrected it to use (N+1)!, according to the AWS post.
  • The key novelty, per the authors, is composition: I/Q running as a general problem solver can call I/R to prove theorems, and I/R can also run standalone and autonomously for long stretches.
  • AWS already uses HOL Light to verify cryptographic code, including post-quantum cryptography, and Isabelle to verify the Nitro Isolation Engine. The authors stress that the proofs stay readable enough for people to check, not only the machine.
Terms in this piece · Glossary
  • AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
  • MCP — The Model Context Protocol — an open standard that lets any AI assistant plug into any tool or data source without custom integration code.
Read the source www.amazon.science
Recommended reads
Comments

Checking sign-in…

Loading comments…