François Fleuret asks: are formal proof checkers actually unused in RL training?
francoisfleuret · x · 2026-11-12
Noted AI researcher François Fleuret raised a technical question: while people repeatedly point out that a formal proof checker can verify a proof with a clean yes/no answer, he has never heard of one actually being used in the RL part of model training. Is that the case?
The question touches on a concrete gap in the RLVR space: proof correctness could in principle be rigorously rewarded via checkers like Lean/Coq, so if they're genuinely absent from training pipelines, the reason is worth exploring. The thread discusses an analogy between proof checking and clinical hypothesis verification.
More from Research
- Anthropic scaled 1,000 Claude agents to formalize Fermat's Last Theorem — SimplyAnnisa · 2026-11-12
- NeurIPS paper gives RNNs provable state-estimation guarantees via control-theory observers — chriswolfvision · 2026-11-12
- NeurIPS paper proves when RNN+MLP can provably recover true robot state — chriswolfvision · 2026-11-12
- Curated 9-paper reading list for LLM quantization beginners, from QLoRA to BitNet — East-Muffin-6472 · 2026-11-12
- A Quick Diagnostic Guide to 4 Common ML Training Errors and Their Causes — cneuralnetwork · 2026-11-12
- DLS runs 1.5x-2.4x faster than tangermeme on A100, verified within numerical precision — anshulkundaje · 2026-11-12