Meta AI's RL-XAR fixes AI slop via expert-aligned rubrics, plus 4 more papers explained
idanbeck · x · 2026-10-03
idanbeck has been making explainer videos for notable papers and compiled this week's batch with full links:
- Unslopping AI (Meta AI): introduces RL-XAR (Reinforcement Learning from eXpert-Aligned Rubrics), a training method targeting the lack of high-quality writing known as "AI slop." It collects top human-written samples, learns LLM-judged rubrics that score expert text above model generations, then runs RL against those rubrics, iterating until meta-optimization closes the gap. Tested on scientific paper sections, Pulitzer-novel continuations, and Wikipedia pages with large improvements over standard training. The core insight: maximum-likelihood pretraining reproduces context quality rather than exceeding it, and non-expert-annotated reward models cap out below expert writing.
- Other topics: Recursive Self-Improvement, dataset generation with zero seed data, PixelUMM, and Frontier Learning.
A handy weekly digest for anyone tracking frontier research without reading full papers.
More from Research
- Symbolic supervision gives video models equal reasoning at ~1/30 the compute, Rubik's Cube study finds — niloofar_mire · 2026-10-03
- Sony AI open-sources PAVAS, a physics-aware video-to-audio model accepted as CVPR 2026 Oral — mittu1204 · 2026-10-03
- Maryland RCT: GPT-4o tutor offering left students 4 points lower, 15% used it — lmoroney · 2026-10-03
- OpenAI paper: simulating deployment predicts LLM safety before release — JacquesThibs · 2026-10-03
- PyTorch Helion DSL integrated into vLLM beats CUTLASS and DeepGEMM with 10%+ throughput gains — PyTorch · 2026-10-03
- Neural memory as an unbounded 2D lattice: writes and reads through a tiny 3x3 bump — CFGeek · 2026-10-03