VeriLoop E2 现已发布，以 27B 开源权重模型在代码、数学和物理任务中引入验证器治理推理。该系统在 SWE-bench Pro 得分 76.2%、Terminal-Bench 2.1 得分 88.8%、AIME 2026 得分 98.3%、GPQA Diamond 得分 93.9%、Apex 2025 得分 89.6%。
VeriLoop E2 is now available, bringing verifier-governed reasoning to code, mathematics, and physics in a 27B open-weight model.
🤖 https://modelscope.ai/models/tsinghua-sigs-robot-lab/VeriLoop-E2
🏆 The evaluated E2 system scores 76.2% on SWE-bench Pro, 88.8% on Terminal-Bench 2.1, 98.3% on AIME 2026, 93.9% on GPQA Diamond, and 89.6% on Apex 2025.
🔁 VeriLoop-Governed Recurrence proposes candidates, verifies them externally, and commits only improvements without regression.
🧪 Two evidence-backed scientific demonstrations cover a 67.35% Riemann ζ zero-proportion certificate and Asymptotic Graviton Tomography. Neither claims to solve the underlying open problem.
📚 Built on Qwen3.8-27B with a 262K native context window and OpenAI-compatible vLLM serving validated at 131K.
📜 Model weights and public inference utilities: Apache 2.0. The production VeriLoop Harness is not included, and the scientific artifacts carry separate terms.
来源：ModelScope · x.com