DeepMind's Aletheia Agent Solves Math Conjectures via Self-Verification
thisguyknowsai · x · 2026-08-22
Google DeepMind released a paper on self-verifying agents that catch their own hallucinations at generation time, called Aletheia. It autonomously solved 4 open Erdős conjectures and wrote a publishable math paper with zero human input.
The core insight is that a model generating a solution is bad at spotting its own errors because the same weights that hallucinated the step are grading it. Aletheia splits the job into three sub-agents:
- Generator: Writes a candidate solution.
- Verifier: Reads the solution in natural language (without seeing the Generator's reasoning) to hunt for gaps, unjustified steps, or fabricated citations.
- Reviser: Rewrites the broken parts based on the Verifier's critique.
The loop runs until the Verifier signs off.
More from coding & agent
- Keep.md Updates Onboarding with Design Help from Fable — iannuttall · 2026-08-22
- Before building high-stakes AI systems: tell the LLM not to build anything — aryanXmahajan · 2026-08-22
- AI Refactoring Risks: Accidentally Deleting Valuable Technical Debt — sebpaquet · 2026-08-22
- MiniMax M3 ran overnight and found a TypeScript compiler crash—fuzzing first, then static analysis — DanielLockyer · 2026-08-22
- 10 Claude Code Steering Mechanisms That Stop Agents From Ignoring Instructions — bibryam · 2026-08-22
- Codex writes, Claude reviews: A dual-agent coding workflow — Vegetable-Try807 · 2026-08-22