RLVR trains models to maximize task reward — long-term code maintainability is on you
bendee983 · x · 2026-11-22
A point worth internalizing for everyone coding with agents, from kunchenguid: RLVR post-training in current frontier LLMs is done at task level, so models simply won't care about long-term maintainability.
- Models write useless comments and test cases by default because these don't affect the current task's success — but make future tasks less efficient — so the behavior survived training and became part of the models.
- Rewarding long-term maintainability in post-training is likely very hard, since measuring it is inherently slow and expensive.
- bendee983 adds: models (especially RLVR-trained reasoning models) exist to maximize reward on the trained task; they have no long-term vision, don't care about developer experience, and don't know who will maintain the code in a year — that responsibility falls on the engineer.
- The fix isn't writing code yourself, but steering at the harness layer: right level of instruction detail, prompts, skills, lint rules, and looking a few steps ahead when no technical choice is perfect.
More from AGI Musings
- TIME now sells ads hidden from humans and shown only to AI agents — Specialist_Lab_4226 · 2026-11-22
- Kenneth Stanley: AI's Riemann proof vindicates 'Why Greatness Cannot Be Planned' — kenneth0stanley · 2026-11-22
- Kasparov on AI math: brute force beat humans at chess without thinking like us — stevenstrogatz · 2026-11-22
- Fields medallist Tim Gowers speaks at London March Against AI Extinction — birchlse · 2026-11-22
- Guardian investigates the uncharted trauma of breaking up with AI companions — TobyWalsh · 2026-11-22
- WaPo: 'Entitled and out of touch' AI CEOs heading for a fall as sentiment turns — GaryMarcus · 2026-11-22