RLVR Can't Prove What Its Verifier Can't: The ZFC Ceiling Argument
Liu_eroteme · x · 2026-09-29
A debate on the ceiling of RLVR (RL with verifiable rewards): the poster argues both pretraining and posttraining are limited — posttraining can only approach the best policies an external verifier can verify, so models will end up "equal to or slightly better than the best humans at any task" rather than superhuman.
In follow-up he sharpens the point: if the verifier is itself a subset of ZFC set theory, RLVR has no reward signal for theorems independent of ZFC — no possible policy gradient can distinguish them from "impossible." The core claim: RLVR is capped by the expressive power of its verifiers, so superhuman mathematical reasoning can't come from RLVR alone.
Related event: RLVR Is Capped by Verifiers, Limiting Path to Superhuman AI(3 posts)→
More from AGI Musings
- Findings of ACL already separates papers from talks, so 'AI-authored' rules are moot, scholar argues — ipeirotis · 2026-09-29
- Humanities will rule the future of higher education, scholar argues in AI era — begusgasper · 2026-09-29
- Education writer suggests keeping young kids away from AI for now — benjaminjriley · 2026-09-29
- US survey finds public AI concerns span many categories, likely worse after recent news — SpencrGreenberg · 2026-09-29
- Instruction tuning shows intelligence can exist without purpose or free will — sytelus · 2026-09-29
- Media takes AI x-risk seriously; next frontier question: AI consciousness — cccalum · 2026-09-29