The document presents a paper and Lean formalization claiming a 2^{o(n)}-time algorithm for 3-SAT that refutes ETH, with asymptotic results for counting #3-SAT; it notes the result is theoretical and not practical, and describes the repository infrastructure, proofs, axioms used, and validation workflow.
Translated title. Original: LLM'lerle Yazılım Doğrulaması: Büyük Ölçekli ve Otomatik Testlerin Yeni Çağı
Formal software verification is increasingly enabled by large language models, enabling automatic, large-scale proof-based checking that can reveal and fix errors beyond traditional testing; specs often arise from test suites, and verification aims to prove programs meet defined correctness across all scenarios.
PaperFormalizer converts LaTeX papers into Lean-verified versions by building a statement graph, auditing for faithfulness and ambiguity, and proposing fixes the author can approve. It outputs an annotated paper, a change report, and Lean code, enabling reuse of verified results across researchers.
Tests for nondeterministic LLM apps should verify consistent outcomes across runs, not exact wording; include authorization, state correctness, and layered evaluation with checks and human review.
BiVerify combines backward checks of answers with forward checks of reasoning, adjusting its verification process when either check fails. Across five math benchmarks, it outperformed the study’s strongest public baseline and either verification direction used alone.
The paper assesses whether LLM-generated proofs remain valid after program refactoring. It introduces RefactorProof to generate numerous variants and compares prover-based proofs to reference proofs, revealing brittleness and differing survival rates, while noting that repair effort does not always align with proof stability.
Palomar records fixed-version mathematical claims, verifies proofs with Lean, and publishes the statement, libraries used, and review comments.
The text says a quantum code for [11,0]]_5 was already known and corresponds to an AME(11,4) state, which provides another approach to a solution.
The document discusses trends in formal verification for software and hardware, emphasizing full-stack verification and the trade-offs between using Lean, Dafny, or other verification-first languages. It covers Cedar's production vs. reference implementations, challenges of Lean in production, and the role of liquid types and SMT solvers in reducing verification effort, while noting that advances in proof generation and token costs influence where the field is headed.
DecompGym benchmarks how well LLM-powered agents can design and implement decomposition solvers, testing 81 instances across Benders and column generation with strict success criteria. Results show partial success, common failures, and the need for robust checks and targeted feedback to fix method-audit gaps.