name: aletheia-math-research
description: >
Agente de pesquisa matematica autonoma baseado em Aletheia (Feng et al., 2026).
Implementa loop Generator-Verifier-Reviser com verificacao Cora-Debate V1-V7
e desacoplamento thinking/output. Atinge nivel L2 (Publishable Research) em
dominios de teoria dos numeros, combinatoria, algebra e geometria.
Use quando precisar resolver problemas matematicos de pesquisa, gerar provas
autonomas, ou conduzir investigacao cientifica com verificacao iterativa.
spec: "SPEC-012"
version: "1.0"
category: research
tags: [aletheia, mathematics, research, generator-verifier, cora-debate, autonomous-science]
dependencies: [SPEC-001, CORA-Eval, cora-debate, reasoning-orchestrator-v11]
tdd_suite: "scripts/aletheia_engine.py"
ct_count: 10
reference: "Feng, T. et al. (2026). Towards Autonomous Mathematics Research. arXiv:2602.10177v3."
status: active
Aletheia Math Research Engine
Arquitetura
PROBLEM ──▶ [GENERATOR] ──▶ [VERIFIER (Cora V1-V7)] ──▶ pass? ──▶ SOLUTION
▲ │ │
│ ▼ fail │
└────── [REVISER] ◀────┘ │
(feedback loop) │
max_attempts = 10 (hyperparameter) │
3 Subagentes
| Subagente |
Funcao |
Inspiracao |
| Generator |
Produz solucao em linguagem natural com 16 tipos de raciocinio |
Feng et al. §2.2 |
| Verifier |
Verifica via Cora-Debate V1-V7 + deteccao de alucinacoes |
Feng et al. §2.2 |
| Reviser |
Corrige flaws identificados pelo Verifier |
Feng et al. §2.1 |
Benchmark (5 problemas)
| ID |
Dominio |
Dificuldade |
| IMO-2024-P1 |
Number Theory |
Olympiad |
| Erdos-1051 |
Combinatorics |
Research Open |
| FutureMath-Basic-1 |
Algebra |
PhD Exercise |
| Thue-Morse |
Combinatorics |
Olympiad |
| Goldbach-Variant |
Number Theory |
Research Open |
Uso
from aletheia_engine import AletheiaEngine, MathProblem
engine = AletheiaEngine(max_attempts=10, strictness=0.7)
problem = MathProblem(
id="My-Problem",
statement="Prove that...",
domain="number_theory",
difficulty="research_open"
)
session = engine.solve(problem)
print(session.status, session.final_solution)
Referencia Principal
Feng, T., Trinh, T.H., Bingham, G. et al. (2026).
Towards Autonomous Mathematics Research.
arXiv:2602.10177v3 [cs.LG].
Google DeepMind Superhuman Reasoning Team.
By