Code simulations teach video generators realistic physics with little data

The authors show that executable simulation code, used as privileged information during self-distillation, improves physical plausibility of generated videos more than supervised fine-tuning or reinforcement learning.

Chinese Tech

Harold Haodong Chen · Rui Peng · Lingteng Qiu · Kejie Qiu · Chao Xu · Zilong Dong

HKUST · Alibaba Group

Research Digest·11 November 2026·3 min read

In one line

Executable code as privileged context in on-policy self-distillation dramatically improves physical plausibility in video generation.

Before you read this

Three ideas this paper builds on

on-policy distillationA training method where a student model learns from a teacher using its own current policy's outputs, often improving reasoning and instruction following.128 papers →privileged informationInformation available only during training that improves performance but is not accessible at inference.80 papers →code as actionA paradigm where actions are expressed as executable code, such as method calls or program synthesis, rather than direct manipulation.29 papers →

Read it as

AI video generators often produce motion that defies physics. Standard training methods struggle: finished videos mismatch the model's outputs, while sparse rewards poorly guide complex interactions.

PhysOPSD tackles this by having the model act as its own teacher. It generates a simple, editable computer simulation code of a physical scene. This code acts as a perfect rulebook, cleanly separated from visual noise.

Models trained with this code privilege produced significantly more physically plausible motion, outperforming standard baselines with only 3K samples. An executable rulebook works better than static examples for teaching physics.

RSS

The Daily Edition

The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.

Concepts

on-policy distillation privileged information code as action agent-driven simulation counterfactual reasoning world model for agents