Daily AI Paper

Daily AI Paper

2026-09-04

Archived
2026-09-04

Rethinking On-Policy Distillation of Large Language Models II: One Training Example

Zixuan Fu, Bingxiang He, Yuxin Zuo, Haohuan Huang, Jinqian Zhang, Ruhang Xiao, et al.

arXiv:2609.04172v1

Today’s most broadly interesting paper is about how little data large language models may actually need to learn a task through on-policy distillation. The authors ask a simple but surprising question: if a student model is trained from its own rollouts with teacher supervision, do you really need a full dataset, or could one example be enough? They find that one-shot distillation keeps improving for hundreds of steps and can recover most of the gain from full-data training. The key idea is state coverage: a single query can already drive the student through a large fraction of the same internal states that many examples would, so the training signal is much broader than it looks. More diverse queries help, but the bigger bottleneck is not data quantity so much as how slowly the student absorbs the supervision. That matters because it suggests we may be overpaying for data in some post-training pipelines, and that future progress could come from making optimization more step-efficient rather than simply scaling up the dataset.

Previous daily papers

2026-09-06SWE-Gate: Passing Functional Tests Is Not Enough for Software Engineering Agents2026-09-05From Deceptive Outputs to Deceptive Mechanisms: A Causal Framework for Language-Model Deception Research2026-09-05Last Translation Benchmark2026-09-03A Common Measure of Communication for Speech Brain-Computer Interfaces2026-09-03Discriminative World Models for Web Agents2026-09-03Post-Training Language Models for Gold-Medal Performance in Coding Competitions2026-09-02Mechanism Design for Alignment and Control2026-09-02Beyond Scores: Understanding LLM-as-a-Judge Mechanisms in Summarization Evaluation2026-09-02The Rise of Verbal Reinforcement Learning2026-09-01Stress-Testing Efficient Responsible-AI Evaluation: When Compute Savings Change Benchmark Conclusions2026-09-01Aspire: Can Models Self-Evolve from Vague Goals?2026-09-01PaperGym: Rubric-Centered Evolution for Research-Plan Generation2026-08-31DARTS: Decoder-Aware Representation Tuning via Surgery for Model Merging2026-08-31Video Generative Models as Geometry Learner2026-08-31When Robots Mishear Us: Mapping the Safety Risks of Voice-Controlled Embodied AI2026-08-30CritICL: Inference-Time Weak-to-Strong Generalization from Small Language Model Failure Modes2026-08-30Boosting LLM Exploration via Weak-Model Guidance in RLVR2026-08-30CLAP: Cross-Embodiment Video World Models are Zero-Shot Physical Simulators2026-08-30TTPO: Test-Time Policy Optimization2026-08-29TTPO: Test-Time Policy Optimization2026-08-28TTPO: Test-Time Policy Optimization2026-08-27Planetary Prediction Engine: Autonomous Geospatial Prediction via Intelligent Data Selection and Foundation Model Embeddings2026-08-26BrowserForge: Scaling Web Episode via Parallel Browser Sandboxes2026-08-25ConvergeFlow: Language Flow with Provable Convergence to Token Embeddings2026-08-24VIALS: A Benchmark for Visual Interpretation of Artifacts in the Life Sciences2026-08-23Inducing Task Models from Computer-Use Traces2026-08-22Inducing Task Models from Computer-Use Traces2026-08-21Inducing Task Models from Computer-Use Traces2026-08-20SPADE: Self-Play in Adaptive Synthetic Executable Environments2026-08-19From Corpora to Co-Evolving Capabilities: Capability-Centric Data Design for Generalist Image Generation2026-08-18Towards Computational Provenance: Carrying Causal-State Evidence in Generated Text2026-08-17Universal Thermodynamic Interatomic Potentials for Crystalline Materials2026-08-16OmniScientist: An Omni-Modal Omni-Discipline AI Scientist