formal verification
30 articles about formal verification in AI news
FAME Framework Delivers Scalable, Formal Explanations for Complex Neural Networks
Researchers have introduced FAME (Formal Abstract Minimal Explanations), a new method that provides mathematically rigorous explanations for neural network decisions. The approach scales to large models while reducing explanation size through novel perturbation domains and LiRPA-based bounds, outperforming previous verification methods.
GPT-5.4 Pro Reportedly Solves Open Problem in FrontierMath, With Human Verification
Researchers Kevin Barreto and Liam Price used GPT-5.4 Pro to produce a construction for an open problem in FrontierMath, which mathematician Will Brian confirmed. A formal write-up is planned for publication.
Learning to Disprove: LLMs Fine-Tuned for Formal Counterexample Generation in Lean 4
Researchers propose a method to train LLMs for formal counterexample generation, a neglected skill in mathematical AI. Their symbolic mutation strategy and multi-reward framework improve performance on three new benchmarks.
Terence Tao Demonstrates AI's Growing Role in Formal Mathematics with Claude and Lean
Fields Medalist Terence Tao has released a video showing how Claude Code can be used to formalize mathematical proofs in Lean, highlighting AI's expanding capabilities in high-level mathematics.
How Spec-Driven Development Cuts Claude Code Review Time by 80%
A developer's experiment shows that writing formal, testable specifications in plain English before coding reduces Claude Code hallucinations and eliminates manual verification of every generated line.
Claude Fable 5 Solves String Theory Problem Stalled for Six Months
Claude Fable 5 solved a string theory problem stalled for six months. Professor Yuji Tachikawa says the model made a non-trivial observation and used SymPy for verification.
JetSpec hits 1,000 t/s on Qwen-8B with speculative decoding
JetSpec achieves 1,000 t/s on Qwen-8B with a B200 GPU, claiming superiority over prior speculative decoding methods, but lacks independent verification.
New Protocol Enables Self-Improving AI Agents with Auditable Lineage
Researchers have proposed a formal protocol for creating self-improving AI agent systems. The framework enables agents to autonomously evaluate and implement upgrades while maintaining auditable lineage and safe rollback options.
Rank, Don't Generate: A New Benchmark for Factual, Ranked Explanations in Recommendation Systems
A new research paper formalizes explainable recommendation as a statement-level ranking problem, not a generation task. It introduces the StaR benchmark, built from Amazon reviews, showing that simple popularity baselines can outperform state-of-the-art models in personalized explanation ranking.
FAOS Neurosymbolic Architecture Boosts Enterprise Agent Accuracy by 46% via Ontology-Constrained Reasoning
Researchers introduced a neurosymbolic architecture that constrains LLM-based agents with formal ontologies, improving metric accuracy by 46% and regulatory compliance by 31.8% in controlled experiments. The system, deployed in production, serves 21 industries with over 650 agents.
OpenAI Internal Model Reportedly Solves Three New Erdős Problems, Marking AI Advance in Pure Mathematics
An internal AI model at OpenAI has reportedly solved three previously unsolved mathematical problems from the Erdős collection. This development signals a potential leap in AI's capacity for abstract reasoning and formal theorem proving.
Fanvue Emerges as Primary Platform for AI-Generated Influencers, Explicitly Allowing Synthetic Creator Accounts
Fanvue, a subscription content platform, has positioned itself as the primary destination for AI-generated influencer accounts, explicitly permitting creators to monetize synthetic personas. This formalizes a niche market for AI-driven adult and influencer content.
Stepwise Neuro-Symbolic Framework Proves 77.6% of seL4 Theorems, Surpassing LLM-Only Approaches
Researchers introduced Stepwise, a neuro-symbolic framework that automates proof search for systems verification. It combines fine-tuned LLMs with Isabelle REPL tools to prove 77.6% of seL4 theorems, significantly outperforming previous methods.
Google DeepMind Proposes 'Intelligent AI Delegation' Framework for Dynamic Task Handoffs with Verifiable Trust
Google DeepMind researchers propose a formal framework for delegating tasks to AI agents, treating delegation as a structured process with dynamic trust models, verifiable proofs, and failure management. The system is designed to prevent over- or under-delegation and enable AI-to-AI task handoffs with clear accountability.
Mathematics Enters New Era as AI Generates Novel Proofs, Says Fields Medalist Terence Tao
Fields Medalist Terence Tao reveals AI is now producing unique mathematical proofs, though verification remains a bottleneck. He argues that to fully leverage AI, mathematicians must design problems that are easily checkable by both humans and machines.
Bridging Human Language and Machine Logic: New AI Framework Achieves Near-Perfect Translation Accuracy
Researchers have developed NL2LOGIC, an AI framework that translates natural language into formal logic with 99% syntactic accuracy. By using abstract syntax trees as an intermediate representation, the system dramatically improves semantic correctness and downstream reasoning performance.
27B Agent Beats Claude Opus 4.8, GPT-5.5 on Research Replication
A 27B agent named Replica reportedly beat Claude Opus 4.8 and GPT-5.5 on held-out research replication, per @omarsar0. No methodology or scores disclosed, so the claim is unverified but suggests efficiency can rival scale.
Anthropic's unreleased model pushes Riemann bound, tests 650 ideas
Anthropic's unreleased model raised the lower bound for the Riemann hypothesis, testing 650 ideas with 60 subagents, confirmed by mathematicians and Lean. This signals AI's growing role in mathematical discovery.
Epoch AI Opens FrontierMath's Unsolved Problems to Public Scrutiny After 2 Years
Epoch AI opened FrontierMath's unsolved problems to public scrutiny after two years, aiming to verify AI claims. The benchmark includes 1,000+ original math problems, with transparency seen as a step against benchmark gaming.
AI Disproves 87-Year-Old Conjecture, Finds Counterexample Humans Missed
AI disproves 87-year-old math conjecture, finding a counterexample humans missed, per @rohanpaul_ai.
NVIDIA Releases FP4 Quantized Kimi-K2.7-Code with 1T Parameters
NVIDIA released FP4 quantized Kimi-K2.7-Code on Hugging Face, a 1T-parameter model for Blackwell GPUs with claimed accuracy retention.
Hugging Face Papers: 35B Agent Matches Trillion-Parameter Performance
Hugging Face Daily Papers featured eight AI papers, including Orca (world model), Dockerless (62% SWE-bench), and a 35B agent matching trillion-parameter performance.
Mistral's Leanstral 1.5 hits 100% on miniF2F, finds 5 real bugs
Mistral's Leanstral 1.5 scores 100% on miniF2F, solves 587 Putnam problems, and finds 5 real bugs in open-source code.
How to Use MCP Servers for Financial Data
MCP servers turn financial data sources into auditable, replaceable protocol endpoints. For Claude Code users building agentic BFSI systems, this means 90% fewer custom integrations and regulator-ready logging.
Norway Bans AI Tools for Under-13s, Pointing to Record-Low PISA Scores Since 2015
Norway will prohibit generative AI tools in grades 1-7 from late August 2026, citing falling PISA scores since 2015. Secondary students may use AI only under supervision. The policy extends an earlier smartphone ban that demonstrably improved grades and reduced bullying, and is backed by planned leg
MiniMax M3 Exceeds Human Gold-Medal on Math Benchmarks via MaxProof
MiniMax's M3 exceeded human gold-medal on math benchmarks via MaxProof, but no scores or details were disclosed.
GitHub Spec Kit: Open-Source Tool to Fix Vibe Coding’s Core Flaw
GitHub released Spec Kit, an open-source toolkit that enforces specification-first workflows for AI coding, addressing vibe coding's tendency to generate code before requirements are clear.
Ontology-Grounded AI Agent Testing Hits 48.3% Regulatory Coverage vs.
Ontology-grounded AI agent testing achieves 48.3% regulatory coverage vs. 33.1% baseline in 1800-scenario pilot. Coverage advantage over RAG not robust after Bonferroni correction.
Google LEAP Scaffold Lifts Lean-IMO-Bench One-Shot Solve Rate from <10% to 70%
Google's LEAP scaffold lifts Lean-IMO-Bench one-shot solve rate from <10% to 70%, solving all 12 Putnam 2025 problems.
Anthropic Unveils TAI Research Agenda Targeting AI Economics, Threats, R&D
Anthropic's TAI will study four areas: economic diffusion, threats, wild AI, and AI-driven R&D. No budget disclosed.