A coding agent that builds a scene before a video generator adds its appearance could help improve the physical plausibility of generated clips. In Video-Coder, an arXiv preprint published November 6, 2026, researchers report higher physics-focused benchmark scores when MiniMax-H3 receives a programmed 3D draft alongside a text prompt. Peer review is not verified, and the results are author-reported.
Initialized from Qwen3.5-9B, the agent writes Blender Python programs specifying objects, spatial layout, motion and camera behavior. An interactive environment executes the code and returns errors, content checks and rendered frames so the agent can revise it. Successful scenes become untextured gray draft videos that guide a fixed downstream generator. The environment also reruns submitted scripts from scratch to check reproducibility of the draft.
Training combines supervised fine-tuning with reinforcement learning. One reward checks whether the program executes and passes environment checks; another uses a model to judge the final video against the prompt and a reference. The execution reward’s weight declines during training. Those references are synthetic MiniMax-H3 videos, rather than recordings of real physical events. The paper gives conflicting dataset sizes: 6,000 supervised and 2,000 reinforcement-learning examples in the methods, versus 10,000 and 4,000 in the setup and conclusion.
On PhyGenBench, which assesses mechanics, optics, thermal and material consistency using the authors’ described GPT-4o judging protocol, MiniMax-H3’s average rose from 0.61 to 0.67. That is an absolute gain of 0.06 on a 0–1 scale, rather than the approximately seven points claimed in the abstract. With one additional generator, OmniWeaving, scores rose from 0.48 to 0.51; the accompanying claim of a 16.28-point gain is inconsistent with the table.
On the selected physical dimensions of VBench-2.0, MiniMax-H3’s average increased from 54.58% to 78.19%, a gain of 23.61 percentage points. This is not a full-suite VBench-2.0 score. Mechanics reached 88.67%, thermotics 70.88% and material 75.02%. The average exceeded the listed CogVideoX-1.5 result of 77.04%, although material performance remained below that model’s 83.19%.
Training comparisons support the combined approach within this pipeline: PhyGenBench scores were 0.63 after supervised training alone, 0.65 without declining execution-reward weighting, 0.64 without video rewards and 0.63 without execution rewards, versus 0.67 for the full method. These comparisons do not establish that every component is universally necessary.
The evidence has substantial limits. Prompt counts, repeated generations, independent replications, confidence intervals and significance tests are not reported. Benchmark overlap checks, separation of parameter selection from final evaluation and comparable baseline inference budgets are undocumented. Prompt filtering favors scenes describable through geometric shapes and programmed motion, excluding some appearance-dependent and facial-expression content. Synthetic references and model judges may introduce biases; plausibility scores do not establish physical correctness or general reliability. Training reportedly used eight H100 GPUs plus 16 for rollout video generation, while inference cost and latency are unquantified. The authors say they released data, models and code, but availability was not independently checked.
Source: https://arxiv.org/html/2611.09894v1
Cover illustration generated with AI; conceptual, not research evidence.