https://emsenn.net/tags/formalization/ Recent content in Formalization on emsenn.net Hugo en Sat, 07 Mar 2026 00:00:00 +0000 https://emsenn.net/library/domains/engineering/domains/tech/domains/computing/domains/artificial-intelligence/domains/agents/texts/ai-math-agents-survey-2026/ Sat, 07 Mar 2026 00:00:00 +0000 https://emsenn.net/library/domains/engineering/domains/tech/domains/computing/domains/artificial-intelligence/domains/agents/texts/ai-math-agents-survey-2026/ <h2 id="ai-agents-for-formal-mathematics-external-landscape-survey"><a href="#ai-agents-for-formal-mathematics-external-landscape-survey" class="heading-anchor" aria-label="Link to this section">¶</a>AI Agents for Formal Mathematics: External Landscape Survey </h2> <p>Survey of external tools, research, and practices for LLM agents doing formal mathematics with proof assistants. Focused on what exists outside this repository that could improve agent mathematical capability here.</p> <h2 id="1-llm-based-mathematical-reasoning-agents"><a href="#1-llm-based-mathematical-reasoning-agents" class="heading-anchor" aria-label="Link to this section">¶</a>1. LLM-Based Mathematical Reasoning Agents </h2> <h3 id="deepseek-prover-v2-deepseek-2025"><a href="#deepseek-prover-v2-deepseek-2025" class="heading-anchor" aria-label="Link to this section">¶</a>DeepSeek-Prover-V2 (DeepSeek, 2025) </h3> <p>Formal theorem proving in Lean 4. Key innovation: subgoal decomposition — the model writes an informal proof sketch first, decomposes into formal subgoals, then proves each subgoal. Trained with GRPO (not PPO) reinforcement learning. Achieved 88.9% on MiniF2F-test, 49 out of 658 problems on FormalMATH. Open-weight (7B and 671B MoE variants).</p>