Professor estimates viral post-training algorithms work out of the box only ~5% of the time
Kangwook_Lee · x · 2026-09-16
Kangwook Lee (UW-Madison) quips that the probability a viral new post-training algorithm works out of the box for his LLM training is roughly 0.05. He notes that nearly a decade after reproducibility issues in RL were widely called out, little has changed, linking to the broader discussion.
More from Research
- CARLA veteran Ros shares synthetic data workflows to accelerate AV development — abursuc · 2026-09-16
- Grade AI like coworkers: open-source FrontierAgent framework ships with CLI TUI and fully local execution — aakashgupta · 2026-09-16
- CoLLAs 2026 talk: memorization may be unavoidable — curation, unlearning, pruning as strategies — gkdziugaite · 2026-09-16
- ETH Zürich robotic hand walks on its own fingers, no legs or wheels needed — lukas_m_ziegler · 2026-09-16
- Genome Biology opens collection on tumor microenvironment, welcomes AI and multi-omic methods — arjunrajlab · 2026-09-16
- Google Engineers' WikiSkill Turns Agent Execution History Into Validated, Reusable Skills — blaizedsouza · 2026-09-16