Tag
The author agrees with poteto, criticizing adversarial code review between sub-agents as a waste of tokens, pointing out its three fatal flaws — meaningless nitpicking, over-engineering, and mismatched compute — and noting that top teams are shifting toward real testing, formal verification, and actual red-teaming to find issues.
A UIUC student is seeking compute sponsors for Loxeon, an Apache-2.0-to-be runtime where AI agent teams self-organize via formally specified primitives (identity, capability delegation, role rebinding, tree-based management, and epoch/ticket message validation), written in Rust and checked against a formal spec. The goal is to fairly benchmark whether multi-agent teams outperform single agents on real coding tasks (~$200 small / $1,500 full), with code, logs and results to be published regardless of outcome.
A researcher-in-the-loop pipeline combining LLM-guided program evolution, exact evaluation, and Lean formal proof checking discovers a new family of entanglement-assisted quantum error-correcting codes, closing 54 gaps in archived EAQECC tables with 115 verified records and 10 certified nonexistence results.
LeanGuard is a runtime that verifies every tool call from language-model agents against Lean4 policies with temporal logic semantics, cutting forbidden calls and prompt-injection success while enabling proof-guided repair of denied actions.
QwenBend is an experimental project where Qwen writes a Bend program, type-checks it against Bend2 rules, and executes it inside its own forward pass using a 53M-weight CPU (16 registers, 8k RAM, SwiGLU-based logic gates) with no external interpreter or compiler.
A new Meta and NYU paper shows that h1, a simple agent loop giving models fast Lean feedback and search tools, matches or beats hand-designed theorem-proving workflows: it outperformed Goedel-Architect on PutnamBench (598 vs 597 problems) at lower cost, and with Opus 4.8 solved 83/100 hard algebra problems versus 63 for Claude Code.
OpenFrontiers introduces a cross-disciplinary corpus of 12,151 open research problems plus a 200-problem retrospective benchmark of recently resolved mathematics problems with Lean 4 statements, evaluating seven LLM agents and finding that closed-book rediscovery of research-level results remains rare while live-web success largely reflects retrieval of published solutions.
Lasm compiles Lean 4 programs ahead of time to WebAssembly, enabling formally verified web apps that deploy to Node, serverless, and edge platforms from a single code base, with 117 proved theorems in its flagship Conduit example.
Proves the long-standing conjecture W(2;3,21) = 416 by Kullmann (2010) using a Lean 4 kernel-checked proof over 209 cases and 38.6M leaf formulas, produced by the neuro-symbolic Large Proof Model (LPM) with a 13.6 GB verifiable evidence package.
Podcast interview with Lean creator Leo de Moura on how AI is making formal verification practical, the need to verify AI-generated code and autonomous agents, and the vision of a verified software stack.
A wave of startups, including &Truth ($70M), Harmonic ($295M), and Theorem with Anthropic, are betting generative AI can make formal software verification practical, potentially transforming how reliable AI-generated code is validated.
AdaCore 展示了利用 AI（无需直接人工干预）在 SPARK 中生成经过形式化证明的 AES、SHA-256 和 HMAC-SHA-256 实现，正确性通过四种独立方式验证，呼应 Anthropic 关于 LLM 用于形式化软件验证的研究。
The paper presents EasyProver, a Lean proving system using recursive agents and asynchronous search that near-saturates PutnamBench and IMO benchmarks with an off-the-shelf frontier model, plus Aegis, a pipeline converting competitive programming into harder real-world Rust-to-Lean verification tasks for benchmarking and RL.
Anthropic 发布 formal-math GitHub 仓库，收录机器验证的 Lean 4 形式化证明，其中 Claude 生成的证明在 Lean 中形式化了一个计数 k-CNF 满足赋值的算法，从而推翻指数时间假设（ETH），并包含 Riemann zeta 零点与 3SUM/APSP 等结果的形式化。
Lanyon AI 从 11 月 1 日起向早期客户和测试者推出 Lanyon 的预发布版本，并在旧金山和芝加哥同时举办公开演示，致力于构建可形式化验证的科学模型。
该论文确定了在独立非同分布噪声和两两非负相关噪声下，选择最大噪声分数所产生的“赢家偏差”的尖锐常数界，并将全部定理在 Lean 4 中形式化验证，同时在 448 个模型选择流程上验证了结论的实际意义。
本文介绍 PaperFormalizer——一个基于 Lean 的作者参与式自动形式化流程，可从 LaTeX 论文出发构建语句依赖图、审计语句并将其形式化证明，同时在语义存在歧义时让作者做出决策，从而产出经 Lean 验证的论文版本。
A paper showing that a simple agentic harness (h1) with an efficient Lean 4 feedback interface achieves state-of-the-art formal theorem proving results on PutnamBench, FATE-X, and FormalConjectures, matching or beating specialized multi-step workflows and settling eight open problems. The authors argue that execution interfaces matter more than complex orchestration and warn against prescribing structured workflows before evaluating frontier models with efficient feedback.
A community research repository reports a sharper exponent bound for integer multiplication, achieving T(n)=O(n(log n)^(1-κ)) with κ=0.0023 under the OpenAI #109 framework, conditional on unverified hypotheses and not yet a complete Lean proof.