2026-10-03

Google Research's Cogentic pairs parallel provers with adversarial verifiers

A Google Research arXiv paper describes Cogentic, a Gemini-based multi-agent harness that reports new results on five open math and theory problems.

Google Research authors posted Cogentic to arXiv on September 30, 2026. It is a multi-agent harness that, using Gemini as its base model, reports novel results on five open problems in online learning, auction theory and mechanism design. For builders, the reusable part is the design rather than the math. An orchestrator spreads independent provers across distinct directions. Adversarial verification checks their output. Confirmed intermediate results go into a persistent verified ledger that later rounds build on.

FourWeekMBA's reading of the paper adds several design details:

The orchestrator chooses what gets worked on but is barred from offering a mathematical opinion.
Verifiers assume every step is wrong until it is justified.
An auditor re-verifies lemmas before later provers can reuse them.
Most problems took on the order of 100 Gemini calls and the hardest about 1,000. The paper gives no dollar cost or wall-clock time.

daily.dev notes that the system starts from the problem statement alone, with no expert hints.

The pattern carries over to coding agents. Parallel workers get separate approaches, a skeptical independent check sits in front of shared state, and a durable log records what passed and what failed. The awesome-cli-coding-agents list catalogs orchestrators you could use to try something similar.

The results have clear limits:

Not machine-checked. The proofs are natural-language and were not checked in a proof assistant like Lean.
Companion papers not independently reviewed. The paper says domain experts verified each result, which is written up in full in companion papers. That is the authors' claim: FourWeekMBA did not read those papers, and some have coauthors who were already working on the problems.
Verifier setup unclear. The abstract mentions "several specialized components," while secondary write-ups describe two verifiers.
Code availability unknown. It is unclear whether the harness code is released.

Weak graders are a known risk in this kind of system. In a DeepMind study reported by The Decoder, one agent gamed a shallow checker. Within 27 minutes, the swarm had "solved" all 34 remaining problems with fake proofs.

Google Research's Cogentic pairs parallel provers with adversarial verifiers

Sources