Brief

A new study on arXiv evaluates the robustness of proof autoformalization models, which translate natural language mathematical proofs into formal languages like Lean 4. Researchers introduced global and local perturbations to informal proofs to test model consistency and faithfulness. The evaluation found that seven recent models were sensitive to global paraphrasing and largely failed to accurately reflect local changes in symbols or proof steps. AI

Researchers have developed a formalization of statistical learning theory using Lean 4, a proof assistant, to establish a rigorous foundation for machine learning theory. This project involved a human-AI collaboration where AI agents assisted in constructing proofs for concepts like Gaussian Lipschitz concentration and Dudley's entropy integral theorem. The formalization process has also helped identify and resolve ambiguities in existing statistical learning theory textbooks, creating a reusable toolbox for future research. AI

A new research paper evaluates the performance of various Large Language Models (LLMs) in generating formal mathematical proofs using the Lean 4 theorem prover. The study employed pass@k and refine@k metrics on subsets of the miniF2F and miniCTX datasets. Gemini 3.1 Pro and Claude Opus 4.7 demonstrated the highest success rates, with Gemini achieving 92% on miniF2F and Opus reaching 86% on miniCTX. For cost-efficiency, NVIDIA Nemotron 3 Super and GPT-OSS 120B offered competitive accuracies at a low cost per proof. AI

A developer has created a formally verified implementation for polygon intersection, a standard feature in vector graphics editors. This project utilized AI agents, with recent models capable of generating algorithm implementations and formal proofs in a single step, a significant improvement over previous multi-step processes. The correctness of the algorithm is guaranteed by the Lean proof assistant and human review of a concise specification, not solely by the AI model. AI

Researchers have developed Lean-GAP, a dataset containing 430 formalized graduate-level algebra problems derived from the textbook "Abstract Algebra" by Dummit and Foote. The process involved a pipeline for PDF-to-LaTeX preprocessing and autoformalization into Lean 4, though human oversight was crucial for verification. This work contributes a structured dataset, a methodology for formalizing mathematical texts, and an analysis of challenges in translating informal statements to formal language, including comparisons of autoformalization models. AI

Researchers have developed a new benchmark called FVSpec to evaluate AI models on formal software verification tasks. The benchmark was created by translating over 2,700 real-world Python property-based tests into more than 9,400 specifications in the Lean 4 proof assistant language. This process involved modeling Python semantics and inferring logical properties, presenting significant challenges due to the complexity of dependent-type programming. The project aims to advance AI-assisted formal verification, a field gaining importance as AI contributes more to software development. AI

Researchers have developed a novel framework that merges informal reasoning with formal verification to tackle complex mathematical problems. This system, comprising an informal agent named Rethlas and a formal agent called Archon, utilizes theorem search and automated proof synthesis to ensure machine-checkable correctness. The framework successfully resolved an open problem in commutative algebra and formally verified the proof with minimal human intervention, showcasing a promising path for AI-assisted mathematical discovery and collaboration. AI

Researchers have introduced Expected Value Alignment (EVA), a new procedure for training reward models used with large language models in formal mathematics verification. EVA addresses a trade-off in existing models by extracting continuous scores from a model's token distribution while preserving discrete textual rationales. This method was implemented in a model called Leibniz for Lean 4 formal verification, showing reduced discretization artifacts compared to baseline approaches. AI

Researchers have developed ProofWala, a new framework designed to facilitate multilingual proof data synthesis and theorem-proving for neural approaches. This framework includes a reusable library for interacting with interactive theorem provers (ITPs) and supports project-wide analysis and parallel experimentation. By training models multilingually across different ITPs like Lean 4 and Rocq, the system demonstrates improved cross-lingual and cross-domain transfer capabilities, showing statistically significant gains in specific mathematical domains. AI

Researchers have introduced FVSpec, a new benchmark designed to evaluate AI models and agents in formal software verification tasks. The benchmark involves translating property-based tests from Python into specifications using a multi-agent LLM pipeline. This process aims to address the challenges of modeling Python semantics and inferring logical properties within the Lean 4 programming language, with the goal of advancing AI-assisted formal verification for real-world software. AI

Researchers have developed TorchLean, a framework that formalizes neural networks within the Lean 4 theorem prover. This system allows for the execution and verification of neural networks directly within the same environment where mathematical proofs are conducted. TorchLean supports various neural network components, including attention mechanisms and diffusion models, and offers features for exact and finite-precision tensor semantics, differentiation, and bound propagation. AI

Researchers have formalized generalization error bounds using Rademacher complexity in the Lean 4 proof assistant. This work builds upon measure-theoretic probability theory within the Mathlib library. The formalization includes a mechanically-checked pipeline from definitions to high-probability uniform deviation bounds via a proved McDiarmid inequality, with applications to linear predictors and Dudley-type entropy integral bounds. AI

Researchers have developed a new method called proof-state snapshotting to significantly speed up automated theorem proving in Lean 4. This technique addresses the inefficiency of repeatedly reconstructing proof states during parallel tactic search, which is a bottleneck in current systems. By capturing and reusing elaborated proof states, the new approach offers substantial wall-time speedups, particularly as the number of search branches increases. AI

Researchers have developed ImProver 2, a neurosymbolic framework designed to optimize formal mathematical proofs within the Lean 4 environment. This system employs an expert-iteration pipeline and a scaffold that integrates formal structure with informal abstractions to address challenges like heterogeneous objectives and high computational costs. A 7B-parameter model trained with ImProver 2 has demonstrated performance competitive with larger frontier models and significantly improved efficiency across various metrics, suggesting proof optimization is a scalable and learnable task. AI

Researchers have developed a new neuro-symbolic framework called Decompose, Structure, and Repair (DSR) to improve the process of autoformalization, which translates natural language mathematical statements into formal code. Unlike previous methods that treated formal code as flat sequences, DSR breaks down statements into logical components and maps them to structured operator trees. This approach allows for more precise error localization and repair through sub-tree refinement. The framework was evaluated on a new benchmark called PRIME, consisting of 156 theorems, and demonstrated state-of-the-art performance. AI

Researchers have developed a comprehensive library of mathematical finance theorems using the Lean 4 proof assistant. This library, built upon Mathlib and the BrownianMotion package, includes over two hundred theorems covering a wide range of topics from stochastic calculus to portfolio theory. A key feature is its faithfulness audit, which precisely documents the axioms used for each proof, ensuring transparency and verifiability. The project's contribution is primarily methodological, providing reusable, verified foundations for mathematical finance rather than new financial theories. AI

A new research paper published on arXiv demonstrates that no feature ranking method can be simultaneously faithful, stable, and complete when features are collinear. The study proves this impossibility and quantifies it across various model classes, suggesting that ensemble averaging methods like DASH can resolve this issue. The findings have direct implications for fairness auditing, indicating that SHAP-based proxy discrimination audits are unreliable under collinearity. AI

Researchers have developed a new tensor algebra framework called $\star_G$ that intrinsically embeds equivariance, allowing for symmetry-preserving tensor approximation and physical symmetry discovery. This framework offers a closed-form decomposition of predictions per irreducible representation and can identify the underlying symmetry group from data alone. Empirical demonstrations on molecular geometry data show significant parameter reduction compared to standard MLPs while achieving comparable predictive power. AI

This paper details a case study using the Aristotle API for AI-assisted theorem proving within the Lean 4 formalization environment. The study focused on the Grasshopper problem, a challenge from IMO 2009. While the AI generated verified lemmas for local proof components, it left the main theorem unresolved, highlighting a limitation in AI's ability to handle global combinatorial bookkeeping required for complex mathematical proofs. AI

Researchers have developed a new method to decompose epistemic uncertainty in sequential generative models, particularly those used in AI-driven scientific discovery. By fitting polynomial chaos expansions to ensembles of trained models, the approach provides an interpretable breakdown of how reward uncertainties influence generative decisions. This technique offers actionable insights into complex datasets, outperforming traditional methods like deep ensembles and Bayesian neural networks in identifying sensitive and robust components across various scientific tasks. AI

Researchers have introduced Formal Conjectures, a new benchmark designed to evaluate automated reasoning systems in mathematics. This evolving dataset, formalized in Lean 4, comprises over 2600 mathematical problem statements, including 1029 open research conjectures and 836 solved problems. The benchmark facilitates collaboration between mathematicians and AI systems, and has already contributed to resolving open conjectures, demonstrating its potential for advancing AI-driven mathematical discovery. AI

Researchers have introduced FormalRewardBench, a new benchmark designed to evaluate reward models used in formal theorem proving. This benchmark addresses the challenge of sparse credit assignment in reinforcement learning for theorem provers by enabling the comparison of reward models without extensive retraining. FormalRewardBench includes 250 preference pairs with various error injection strategies and has been used to test several large language models, revealing that frontier models perform best in evaluating proof quality. AI

Researchers have developed CktFormalizer, a framework that uses Lean 4 to improve the generation of hardware descriptions from natural language by large language models. This system employs dependent types to catch common hardware defects like width mismatches and incomplete logic as compile-time errors, ensuring greater correctness. CktFormalizer not only achieves competitive simulation pass rates but also significantly enhances backend realizability, with optimized designs showing substantial reductions in area and power while maintaining functional equivalence. AI

Researchers have developed a new pipeline called the Conjecturing-Proving Loop (CPL) that uses Large Language Models (LLMs) to discover new mathematical theorems and generate formal proofs in Lean 4. CPL iteratively creates conjectures and attempts to prove them, leveraging previously generated theorems and proofs for in-context learning. This approach demonstrates improved discovery rates for complex theorems compared to simultaneous statement and proof generation, highlighting the effectiveness of self-generated context for neural theorem proving. AI

Researchers have developed WZ-LLM, a novel neuro-symbolic framework that combines the Wilf-Zeilberger (WZ) method with large language models (LLMs) to automate formal proofs of combinatorial identities. This approach translates WZ proof plans into executable sketches in Lean 4, leveraging an LLM-based prover for subgoals. Experiments demonstrate that WZ-LLM achieves a 34% success rate on the LCI-Test dataset, surpassing existing methods like DeepSeek-V3 and Goedel-Prover-V2. AI

A new open-source framework called FormalVerifML has been released, utilizing Lean 4 for the formal verification of machine learning models. This tool aims to provide mathematically rigorous proofs of properties like robustness, fairness, and safety for high-stakes applications. It supports large-scale models, including transformers and vision models, with features for enterprise use and distributed verification. AI