LLM judge quality hinges on pretraining scale; full online RL alone won't crack continual learning
herbiebradley · x · 2026-08-29
Herbie Bradley responds to two technical questions. On evaluating longer tasks with LLM judges, the difficulty isn't length but whether the judge is good enough to capture subtle qualitative judgements — which depends heavily on pretraining scale and grinding out data coverage. On continual learning, it depends on the in-weight method, but challenges like RL sample inefficiency make him skeptical that full online RL alone would solve it.
Related event: The S-Curve Debate: Can AI Progress Keep Stacking New Paradigms(16 posts)→
More from Research
- DeepPhylo infers phylogenetic trees without unrealistic evolutionary assumptions — anshulkundaje · 2026-10-15
- Wander preprint explores graph foundation models via shared objective and random-walk interface — mmbronstein · 2026-10-15
- CHI Reviewer: AI Slop Is Real, but Unusually Rigorous Papers Still Getting Through — IanArawjo · 2026-10-15
- Into the Unknown: CMU benchmark finds 114 of 135 AI-written research papers fail integrity checks — arankomatsuzaki · 2026-10-15
- 8 non-negotiable principles for building AI that actually learns from experience — tokenbender · 2026-10-15
- Benchling launches Science Foundries: multiplayer agent swarms for end-to-end science workflows — nlarusstone · 2026-10-15