Can open-weight models write challenging Verus proofs for TLA+ specs? We tested @NVIDIAAI's Nemotron 3.5 Super VL and the answer is...absolutely, yes! Our fine-tuning boosts performance from 43% to 73% on the hardest benchmark. Read our latest report on novel datasets, fine-tuning 3.5 Super VL, and bridging the gap to real-world software verification: reasonable.io/blog/post-trai…

Formal verification of real-world software and systems requires models and toolkits capable across a wide variety of programming languages and verification techniques. This focus on polyglot verification is one reason our translation and proof generation pipeline is so valuable. We also need proofs about a system's real-world behavior, which Verus (auto-active verifier built for Rust) enables. All credit to Kseniia Strelbytska, Anna Mészáros, Balazs and the team working on this (amongst many other frontier-pushing initiatives) - Our paper “TLA-Verus: Generating and Solving Temporal Proof Obligations with LLM Agents,” has been accepted to the VeriCodeGen Workshop at NeurIPS 2026! In brief: developers use TLA+ to design systems, describe their behaviour, and test them for flaws. But its main tool (TLC) only tests small versions of the system, and the design stays separate from the implementation code. We translate TLA+ designs into Verus, a verifier for Rust. There, LLM agents write proofs about the system’s behaviour for any run, at any size. Since the design now lives in Rust, it can later be linked to the actual code. Our team generated 3,000+ fully machine-checked proofs of system properties, alongside a dataset to evaluate how well current LLMs write temporal proofs in Verus. - Paper: openreview.net/forum?id=ncFI… - New to TLA+? Read Reasonable’s intro: reasonable.io/blog/tla-tutor… - Reasonable is hiring! apply.workable.com/reasonabl…

Kseniia and Balázs are part of this cohort - brilliant impact with us already, setting the bar incredibly high. You can read more about their work at Reasonable shortly

We are hiring unusual CS, maths, machine learning and formal methods superstars: 🟥 shaping our research, engineering, and product development as our next Member of Technical Staff: apply.workable.com/reasonabl… 💙 driving growth-by-hiring as our Talent Partner: apply.workable.com/reasonabl… 🟡 building our operations in a tight, high-trust partnership as the EA to our founders: apply.workable.com/reasonabl… Play an early lead role in our growth - join to build and scale one of Europe’s most exciting AI startups! 🚀

Teach machines how to write provably correct code. On a boat. On the Danube. Join us: reasonable.io/about/

Smaller, faster, capable: writing machine-checked proofs with a 30B open-weight model We fine-tuned NVIDIA’s latest open model Nemotron 3.5 Lightning on 4B tokens of synthetic Verus data. The result: it beats a model ~50x its size on per-attempt pass rate while nearly matching it on pass@3, with faster token generation than any similarly sized open-weight model we tested. Explore how the models we tested fail at formal proofs, how they cheat, and the effects of fine-tuning: reasonable.io/blog/verificat…