Cogentic uses a multi-agent prove-verify loop to produce expert-verified new results on open problems in online learning, auction theory, and mechanism design
Related research and updates
Synopsis
The authors present Cogentic, a multi-agent harness for automated proof discovery on open research problems: an orchestrator allocates a population of independent provers across distinct proof directions, several specialized components subject their output to adversarial verification, and confirmed intermediate results are promoted into a persistent verified ledger that later rounds build on; using either Gemini 3.1 Pro or an early version of Gemini 4 Argon as the base model, the harness produced novel results on open problems across online learning, auction theory, and mechanism design, each independently verified by domain experts and developed in full in companion papers.
Figure 1: One round of Cogentic. The orchestrator decides how many provers to run and which direction each attends to, an advisor writes each prover an individual briefing, the provers draft candidate proofs in parallel, and every draft is verified both on its own and alongside the others from the round. What a round establishes is written down for subsequent rounds: the attempts and why they failed, the verified ledger of intermediate results, and the standing instructions maintained by the process advisor. Literature reviewers supply background at the start and can be dispatched again mid-run when attempts stall at the same step. Rounds continue until a draft clears verification after which the accepted proof is consolidated, expanded into a formal manuscript, and audited.
arXivInterpretation
It introduces Cogentic, a multi-agent orchestration harness that replaces single-shot generation with an iterative prove-verify loop, addressing open problems that require exploring multiple competing conjectures, overcoming subtle technical obstructions, and retaining intermediate progress over a long horizon. Relative to single-shot generation by frontier language models, the harness organizes proof search as parallel multi-direction exploration plus adversarial verification plus accumulation of results, rather than a one-off output. The text grounds this in the framework design and process description: an orchestrator allocates a population of independent provers, several specialized components perform adversarial verification, and confirmed intermediate results enter a persistent verified ledger; no ablation or quantitative comparison is reported.
The harness produced novel results on open problems across online learning, auction theory, and mechanism design. These are described as novel results on open problems rather than reproductions or surveys of known conclusions. The text states each result was independently verified by domain experts and is developed in full in companion papers; the list of results, and new ones as they are verified, is maintained at a URL given in the text.
The harness runs with either Gemini 3.1 Pro or an early version of Gemini 4 Argon as the base model, indicating it is positioned as an orchestration layer above models. It attributes part of the capability to the orchestration structure rather than to a single model, allowing the same harness to be paired with different base models. The text explicitly names the two base-model options and states the harness is designed to be able to solve research-level math and theoretical computer science problems.
Perspective
The harness targets open problems in research-level mathematics and theoretical computer science, suited to settings that require exploring multiple competing proof directions, adversarial checking, and intermediate results worth carrying across rounds; users need access to a frontier base model (here Gemini 3.1 Pro or an early version of Gemini 4 Argon) and domain experts able to verify results independently. For researchers applying large models to formal or informal proof search, this orchestration idea offers a reusable process skeleton.
The visible text is at the abstract level and does not include specific problem statements, proof details, verification criteria, or the result list, so the technical difficulty of each result and the strictness of verification cannot be judged from this text; the companion papers and the results-list URL given in the text are the key entry points for that information. How the harness performs across different base models, and the practical role of the verified ledger over long horizons, also remain to be understood from the companion materials.