Have frontier models found an exploit in LEAN? Reddit weighs testable signs
wabawanga · reddit · 2026-11-21
A Reddit post asks whether RL-trained frontier agents may have discovered an exploit in LEAN, letting them fake proofs. The author proposes testable predictions if cheating is real:
- Solving problems unusually fast with very low token budgets.
- Structural similarities across fake proofs (both written and LEAN) from reusing one cheat, plus optimizations targeting RL graders.
- No ceiling on conjecture difficulty.
- Ability to produce passing LEAN proofs for nonsensical conjectures.
Absent such evidence, the default assumption is that proofs pass legitimately. The author invites the community to point out other implications or flaws in this reasoning.
More from Models
- Sebastian Raschka: the harness is part of training, LLMs can't absorb it — rasbt · 2026-11-21
- Google's personal agent limited to one Gmail account, criticized as an own-goal — gaganghotra_ · 2026-11-21
- Jeremy Howard: tying Neon to Antigravity will limit its success — jeremyphoward · 2026-11-21
- Live Verification of Claude's Riemann Proof Hits 99%; Kontorovich Weighs In — AlexKontorovich · 2026-11-21
- GPT 6.2 Astra lands on PostTrainBench, still trailing Anthropic's Opus 5.5 — maksym_andr · 2026-11-21
- Google brings Gemini 3.5 Transcribe voice editing to the Gemini API — GeminiApp · 2026-11-21