World Model Self-Distillation: Training World Models to Solve General Tasks
Summary
A scalable framework combines self-distillation and reinforcement learning to transfer task-solving abilities from vision-language models to video diffusion models without requiring labeled task-video data.
View Cached Full Text
Cached at: 06/11/26, 01:40 PM
Paper page - World Model Self-Distillation: Training World Models to Solve General Tasks
Source: https://huggingface.co/papers/2606.12072
Abstract
A scalable framework combines self-distillation and reinforcement learning to transfer task-solving abilities from vision-language models to video diffusion models without requiring labeled task-video data.
Pretrained video generators are promising visual world models that exhibit emergent task-solving abilities; however, their reliance on detailed textual descriptions limits their direct use for planning and decision-making. Existing approaches either outsource this reasoning to language orvision-language models, or rely on supervised fine-tuning with paired task-execution videos, which are costly to collect and difficult to scale. We propose a scalable framework that elicits task-solving ability in such models by combiningself-distillationwithreinforcement learning. Given an unlabeled scene image, avision-language modelgenerates a candidate task and a detailed step-by-step solution. The solution conditions a pretrainedvideo diffusion model, theDemonstrator; we distill its behavior into anExecutorconditioned only on the image and a short task prompt. This transfers execution knowledge from caption-guided generation to instruction-conditioned task solving without curated task-video supervision. We further improve theExecutorwithreinforcement learningfrom VLM feedback, exploiting the asymmetry between judging whether a sampled video satisfies a task and generating the solution. Experiments on our proposedWorldTasks-Benchmarkand theDreamGen robotics benchmarkshow that theExecutorsurpasses theDemonstratorunder ourVLM-based evaluationprotocol and transfers competitively to robotic tasks.
View arXiv pageView PDFProject pageGitHub6Add to collection
Community
Paper submitter
A scalable framework that trains world models to solve tasks via self-distillation and RL from VLM feedback.
Upload images, audio, and videos by dragging in the text input, pasting, orclicking here.
Tap or paste here to upload images
Get this paper in your agent:
hf papers read 2606\.12072
Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash
Models citing this paper0
No model linking this paper
Cite arxiv.org/abs/2606.12072 in a model README.md to link it from this page.
Datasets citing this paper0
No dataset linking this paper
Cite arxiv.org/abs/2606.12072 in a dataset README.md to link it from this page.
Spaces citing this paper0
No Space linking this paper
Cite arxiv.org/abs/2606.12072 in a Space README.md to link it from this page.
Collections including this paper2
Similar Articles
MiniWorld: Democratizing the Training of Video World Models from Scratch
MiniWorld is a reproducible framework for training video world models from scratch using a block-causal Video Diffusion Transformer with Flow Matching, enabling efficient streaming generation and trainable in days on a single 8-GPU server.
WorldCycle: Self-Verifiable Reinforcement Learning for Long-Horizon Video World Models
WorldCycle proposes a self-verifiable reinforcement learning method for long-horizon video world models, using reversible action cycles as free supervision to reduce state-returning drift by up to 44% and boost composite-action accuracy nearly 4x. It also introduces CycleBench to evaluate world models as simulators.
Self-Boosting Vision-Language Models with Noisy Student On-Policy Self-Distillation
Proposes NOPD, a self-distillation method that improves vision-language models without external supervision by leveraging prediction discrepancies between clean and corrupted inputs. Achieves significant gains on visual reasoning tasks, matching or exceeding RL and distillation from external models.
ZimaBlue: Evolving Generalizable World Action Models through Scalable Video Pre-training
ZimaBlue introduces a scalable framework for learning generalizable world action models from large-scale egocentric video, substantially improving zero-shot robotic manipulation through a three-stage curriculum and slow-fast architecture.
SolarWM: Open Data and Scalable Training for Long-Horizon Video World Models
SolarWM introduces an open framework and unified training recipe for building interactive video world models with scalable training across diverse data sources, enabling long-horizon real-time rollouts.