Shanghai AI Lab: RL self-learning lifts Intern-S2 proof accuracy from 51.5% to 79.5%
internlm · hf · 2026-11-12
Researchers from Shanghai AI Lab, SJTU and collaborators study LLM self-learning — autonomously retrieving, studying, and applying knowledge from long-form reference materials on the fly, beyond parametric knowledge or fixed RAG.
- Task: textbook-based math problem solving — the model must navigate textbook exposition, learn relevant principles, and construct proofs for exercises with reference solutions withheld.
- BookNavigator: an interface supporting this human-inspired study process, combined with RL using fine-grained rewards to incentivize self-learning.
- Results: Intern-S2-Preview-35B's overall proof accuracy rises from 51.5% to 79.5%, beating RL-trained textbook-free and fixed-retrieval baselines, with gains across in-domain and out-of-domain textbooks and subjects.
- Generalization: improvements of 1.7 and 2.3 points over the base model on BrowseComp-Plus and CL-bench, suggesting RL can foster a domain-agnostic self-learning strategy for complex long-context tasks.
More from Research
- StepFun's P²-MOPD beats multi-teacher distillation with half the training steps — teortaxesTex · 2026-11-12
- Stanford professor: virtual cell models don't need billion-parameter foundation models — tweaked scVI did fine — anshulkundaje · 2026-11-12
- 2048 TPUv4s and deep domain expertise: professor dissects how a gene-regulation model was trained — anshulkundaje · 2026-11-12
- Professor: gene-regulation model was built on open academic data from ENCODE, GTex and Cell x Gene — anshulkundaje · 2026-11-12
- Stanford professor hails new gene-regulation model as a real paradigm shift toward causal cis-trans models — anshulkundaje · 2026-11-12
- François Fleuret asks: are formal proof checkers actually unused in RL training? — francoisfleuret · 2026-11-12