RL framework teaches LLMs to self-learn from textbooks, lifting proof accuracy 51.5%→79.5%
_akhaliq · x · 2026-11-13
A new RL framework lets LLMs self-learn from structured textbooks: with fine-grained rewards, the policy autonomously retrieves, studies, and applies knowledge from long-form references (Study→Reason→Exercise), then improves via an RL loop with a judge scoring proofs.
Benchmark results shown in the paper: Self-Learning RL lifts ID-Book from 39.0 to 65.5, reaches 85.2 on OOD-Book and 86.7 on OOD-Subject, far above both the base model and No-Book RL, with overall proof accuracy rising from 51.5% to 79.5%.
Related event: BookNavigator: RL Framework Lets LLMs Self-Learn from Textbooks(2 posts)→
More from Research
- New research tackles the agent experience problem: knowledge trapped in weights — khademinori · 2026-11-13
- Computational chemistry platform Rowan launches gamified Intro to Optimization games — CatAstro_Piyush · 2026-11-13
- One Chart Explains the Linear Algebra Behind Modern AI, from Vectors to SVD — HankYeomans · 2026-11-13
- Codex agent shrinks a 27M-param transformer to 11.46M with equal loss via weight sharing — yoshiyama_akira · 2026-11-13
- Google open-sources Harness²: Agent = Model + Harness, recursive agent harnessing framework — 2C_ornot2C · 2026-11-13
- Question's Gambit: first-step reranking lifts agentic search accuracy to 90.5% — CShorten30 · 2026-11-13