Google Paper: Data-Free On-Policy Self-Training Improves LLMs Across Text and Vision
google · hf · 2026-10-26
A Google/Ulm University paper introduces On-Policy Self-Training, a data-free self-improvement framework for LLMs.
Key points:
- Reformulates on-policy self-distillation as a self-training loop: guided only by coarse, high-level topic descriptions, the model synthesizes its own question-answer pairs and privileged reasoning traces — no human-curated prompts, ground-truth answers, or supervised signals needed.
- Self-distillation on fully synthetic privileged information improves the model across multiple domains while retaining capabilities on domains it wasn't trained on.
- Validated on both text and multimodal (VQA) tasks, achieving consistent self-improvement on standard language benchmarks.
Takeaway: foundation models harbor rich latent capabilities that can be amplified without external task supervision.
More from Research
- ETH Zurich paper: dynamic evaluation allocation matches 10x-budget model rankings — zouharvi · 2026-10-26
- DuoLM: Fast LLM-Based Personalized Recommendation via Compact Encoder-Ranker — GoogleDeepMind · 2026-10-26
- DISC Learns to Combine Time-Series Dataset Similarity Metrics for Transfer Prediction — Yen-Ku Liu · 2026-10-26
- Adobe's SliderEdit++ Improves Slider-Based Image Editing by 18.5% with New SlideBench — Arman Zarei · 2026-10-26
- DeepMind Builds Auditable AI Text Detectors with Sparse Autoencoders — GoogleDeepMind · 2026-10-26
- CoRE Framework Learns Rule Routing for Knowledge Retrieval Without Supervision — Zhisheng Qi · 2026-10-26