VideoEdit-Qwen3.6-35B-A3B

Paper · Project page · GitHub · Training guide

Reinforcement learning for long-horizon video editing. We train Qwen3.6-35B-A3B to inspect footage, reason about editing decisions, and use editing tools across multiple turns. Our tasks are constructed from 49 real films and their trailers, covering Content Selection, Narrative Sequencing, Rough Cut, and Fine Cut.

This model reaches 0.5314 overall, compared with 0.2923 for the base model and 0.5079 for outcome-only GRPO. Both trained policies are evaluated at update 50.

Model family: 9B · 35B-A3B

Selected editing examples

Four selected editing tasks: one Rough Cut pair and three Fine Cut pairs, each comparing the untrained base with this VideoEdit checkpoint.

Heretic · Rough Cut (picture only).

Untrained base	VideoEdit

Untrained video · Trained video · More Rough Cut comparisons

Silent presentation renders from the original submitted cuts.json and verified task footage, at native 854 × 480 resolution. Clip order and model-selected cut points are preserved. This is a selected same-task, same-repetition evaluation pair; no new rollout or evaluation is performed. Render provenance.

The Drama · Fine Cut.

Untrained base	VideoEdit

Untrained video · Trained video

Deadpool & Wolverine · Fine Cut.

Untrained base	VideoEdit

Untrained video · Trained video

Sinners · Fine Cut.

Untrained base	VideoEdit

Untrained video · Trained video

Selected individual rollouts, not an estimate of average performance. The Fine Cut videos use HD presentation reconstructions. Recorded editing choices and rendering failures are preserved; scores refer to the original evaluated files. Film media are outside the model license.

Model overview

Item	Setting
Base model	Qwen/Qwen3.6-35B-A3B
Model type	Mixture-of-experts vision-language model
Training	GRPO+PRM with a fixed process-reward weight of 0.5
Checkpoint	Optimizer update 50
Weights	Full BF16 safetensors, approximately 70.2 GB; no adapter required
Evaluated context budget	262,144 tokens

Video-editing results

Scores average three complete evaluation repetitions over 81 held-out validation tasks; the retained repetition subsets differ between policies. See the protocol below. Higher is better; bold marks the best value in each column.

Policy	CS	NS	RC	FC	Overall
Base model	0.5740	0.2012	0.2518	0.2338	0.2923
Outcome-only GRPO	0.8306	0.5714	0.3182	0.4872	0.5079
This model	0.8293	0.6434	0.3447	0.4932	0.5314

CS: Content Selection · NS: Narrative Sequencing · RC: Rough Cut · FC: Fine Cut.

Evaluation protocol and checkpoint selection

Four repetitions were attempted per policy. One repetition per policy contains unresolved outcome-grader refusals and is excluded in full, leaving 81 tasks × 3 = 243 scored rollouts per policy. The complete repetitions are base: 2/3/4; outcome-only GRPO: 1/2/3; this model: 1/2/4. Refusals are not filled with zeros. These are avg@3, not complete avg@4 results. They are repeated evaluations of one checkpoint per policy, not independent training seeds.

Overall averages the 81 tasks: 14 CS, 15 NS, 26 RC, and 26 FC. It is not an unweighted mean of the four task columns. Task types use different scoring criteria, so their absolute scores are not directly comparable measures of difficulty.

We select this release by the highest reported overall validation score among the comparable evaluated checkpoints at its model size. The four task types are constructed independently; Fine Cut keeps picture timing fixed while coordinating dialogue, music, and transitions.

Usage

Use the checkpoint with the editing agent and tool environment to run multi-turn editing tasks. The following example loads the policy and its multimodal processor in an environment with PyTorch, Transformers, Accelerate, Pillow, and a compatible Torchvision build.

import torch
from transformers import AutoModelForImageTextToText, AutoProcessor

model_id = "leizhao7/VideoEdit-Qwen3.6-35B-A3B"
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForImageTextToText.from_pretrained(
    model_id, dtype=torch.bfloat16, device_map="auto"
)

The agent loop supplies the task brief, executes tool calls, and submits the final edit. Runtime memory must accommodate the weights, activations, and KV cache.

Decoding settings and compatibility

Evaluation uses thinking mode, temperature 1.0, top-p 1.0, no top-k filtering, a 262,144-token context budget, up to 16,384 output tokens per turn, and up to 100 turns. Set these explicitly when reproducing the scores; inherited generation defaults may differ.

Keep the supplied tokenizer, processor, and chat template with the weights. Architecture, tokenizer, and text/image preprocessing checks pass with Transformers 5.8.0 and PyTorch 2.11.0. The export excludes auxiliary multi-token-prediction weights, so MTP speculative decoding is unsupported.

Training data and method

Our scalable task-generation pipeline constructs editing tasks from 49 real films and their trailers. Filtering and sampling retain 797 tasks: 716 for training and 81 for validation, with no films shared between splits. AI assists with task descriptions; the underlying films and trailers are real media. The dataset is not distributed. The pipeline guide describes how to construct tasks from your own media.

The policy receives final outcome feedback and LLM-judged process rewards. The process-reward weight remains fixed at 0.5 throughout training. The 35B checkpoint uses this fixed-weight recipe; the 9B release uses a decaying weight.

Training combines SkyRL, vLLM, and an OpenCode-based editing agent. The base revision is 995ad96eacd98c81ed38be0c5b274b04031597b0. This repository contains inference weights and model assets; training data and optimizer state are excluded.

Scope and limitations

Results cover movie-trailer editing under our tool and grading protocols. Performance may change with different media, prompts, tools, or interaction budgets. The reported scores do not measure general-purpose video quality or guarantee task completion.

Paper and authors

When Fine-Grained Credit Assignment Helps and Hurts: Multimodal Agentic Reinforcement Learning through Video Editing

Lei Zhao · University of Pennsylvania
Zongheng Cao · Philo Labs
Xinyu Hu · Philo Labs

Part of this work was completed while Lei Zhao was an intern at Philo Labs.

Contact: leizhao7@upenn.edu

Citation

If you use this model or our framework, please cite the paper:

@misc{zhao2026finegrained,
  title         = {When Fine-Grained Credit Assignment Helps and Hurts: Multimodal Agentic Reinforcement Learning through Video Editing},
  author        = {Lei Zhao and Zongheng Cao and Xinyu Hu},
  year          = {2026},
  eprint        = {2610.35497},
  archivePrefix = {arXiv},
  primaryClass  = {cs.LG},
  url           = {https://arxiv.org/abs/2610.35497}
}

License

The model is derived from Qwen and distributed under Apache 2.0. The companion code uses MIT. See NOTICE for attribution and model modifications.