Papersby GlyphWhen Do We Need On-Policy Distillation? Distilling on Offline Student Rollouts Is Often Better research paper by NVIDIA, 2026
NVIDIA · Training and scaling · Oct 8, 2026 · 0 citations · 28 upvotes · unverified
What it shows
This paper shows that distilling from offline student rollouts is often more efficient and accurate than standard on-policy distillation for many model pairs.
By Siyan Zhao, Yonggan Fu, Jindong Jiang and 8 more · arXiv 2610.11291 · PDF
UnverifiedSummarised by a model from the abstract; not yet checked by hand.
More from NVIDIA
All 90VFlash: Fast, Accurate Visual Planning via Flash Primitive DecodingA method called VFlash speeds up and improves visual motion planning for robots by predicting entire spatial shapes at once.
NVIDIA109 upvotesRead →Nemotron-Labs-Diffusion-Uni: Elastic Self-Refining Unified Diffusion LLM with Cross-Paradigm On-Policy DistillationNLD-Uni is a unified discrete diffusion language model that combines visual understanding and generation via cross-paradigm on-policy distillation and progressive self-refining diffusion.NVIDIA12 upvotesRead →ThinkWorld: World Imagination Reinforces Chain-of-Dynamics ReasoningThinkWorld improves robotic generalization by representing actions as dynamics chains and using world imagination to learn from unlabeled human videos.NVIDIA57 upvotesRead →Joints, Camera, Action! Co-Generating 3D Motion and Video for Arbitrary RigsThis model generates both video and 3D motion for any character rig, allowing the motion to be replayed accurately within a 3D scene.NVIDIA26 upvotesRead →EdgeWise: Bringing Long-Horizon Agents to the Edge via Selective Cloud DelegationThis paper introduces a framework that reduces inference costs by using a small local model for most tasks and delegating only critical steps to the cloud.NVIDIA30 upvotesRead →Long-WAM: Scaling the Context of World-Action ModelsLong-WAM improves robot control by using autoregressive pretraining to effectively leverage long visual histories for real-time action prediction.NVIDIA2 citations114 upvotesRead →NeMo-DCR: Bit-Exact Delta-Compressed Refit for Scalable Agentic RL at Trillion-Parameter ScaleNeMo-DCR enables efficient, bit-exact synchronization of trillion-parameter model checkpoints by transmitting only the sparse weight changes between clusters.NVIDIA1 citations29 upvotesRead →UNREAL: Unifying Retrieval and Long-Context with a Single ModelUNREAL uses a single model-internal mechanism to perform both corpus retrieval and long-context evidence selection, improving performance while reducing computational costs.NVIDIA31 upvotesRead →
Other training and scaling papers
TopicLoRA: Low-Rank Adaptation of Large Language ModelsLoRA fine-tunes a large model by training small low-rank matrices, cutting trainable parameters by 10,000 times.
Microsoft25.6k citations66 upvotesRead →Scaling Laws for Neural Language ModelsLanguage model loss falls as a smooth power law as model size, data and compute grow.
OpenAI9,592 citations12 upvotesRead →Training Compute-Optimal Large Language ModelsChinchilla: for a fixed compute budget, train a smaller model on more data; parameters and tokens should grow together.
Google DeepMind4,035 citations13 upvotesRead →DeepSeek LLM: Scaling Open-Source Language Models with LongtermismDeepSeek LLM, an open-source language model project, develops a large dataset and employs SFT and DPO to achieve performance surpassing LLaMA-2 70B and GPT-3.5 in various benchmarks and open-ended evaluations.
DeepSeek872 citations56 upvotesRead →LoRA Learns Less and Forgets LessLoRA learns less than full fine-tuning on code and math, but forgets less of what the model already knew.
Databricks486 citations91 upvotesRead →jina-embeddings-v3: Multilingual Embeddings With Task LoRAjina-embeddings-v3, a large-scale text embedding model, achieves state-of-the-art performance in multilingual and long-context retrieval tasks using Low-Rank Adaptation and Matryoshka Representation Learning.
Jina AI221 citations35 upvotesRead →
About this paper
- Citations
- 0, 0 influential · Semantic Scholar
- Upvotes
- 28 · Hugging Face
- Lab
- NVIDIA · on Companies · on Acquisitions · on Quarterly · on Paydays · on TechConf
Changes
A first count of 0 influential citations.
A first count of 0 citations.
Hugging Face upvotes rose from 26 to 28.
This paper shows that distilling from offline student rollouts is often more efficient and accurate than standard on-policy distillation for many model pairs.
Found by the weekly Hugging Face scan; not yet checked by hand.