Video-HopChain-8B

Qwen3-VL-8B-Instruct trained with GRPO in two stages, first on a general video question answering mixture and then on Video-HopChain with Confidence-Gated Exploration (CGE). This is Video-HopChain-8B, the "+ standard RL + Video-HopChain + CGE" row of Table 1 in the paper, at global_step_10 of the second stage.

Paper: Video-HopChain: Multi-Hop Questions and Confidence-Gated Exploration for Video Reasoning Models (arXiv:2609.25773)
Dataset: ngqtrung/Video-HopChain
Project page: ngquangtrung57.github.io/video-hopchain-page
Code: github.com/ngquangtrung57/video-hopchain
Collection: Video-HopChain
Initialization: ngqtrung/Video-HopChain-8B-Standard-RL, the standard-RL stage of Table 1

Results

Accuracy in percent, evaluated with lmms-eval at 100 frames per video under one setting for every row. "In-domain" is the 1,000-question held-out Video-HopChain split.

Model	Video-MME	PerceptionComp	Video-MMMU	Video-Holmes	VCRBench	MMR-V	LongVideo-Reason	VRBench	mean	in-domain
Qwen3-VL-8B-Instruct	64.0	28.1	63.5	40.7	32.0	42.7	72.6	74.7	52.3	13.4
+ standard RL	65.8	34.3	63.0	47.4	35.9	43.4	76.1	77.6	55.4	13.4
+ standard RL + CGE	67.9	34.4	62.5	48.5	35.1	46.4	78.8	79.5	56.6	16.4
+ standard RL + Video-HopChain	68.6	36.2	64.7	48.6	42.1	45.9	77.8	79.4	57.9	21.2
this model	69.2	37.4	67.5	48.6	44.9	46.6	78.9	81.0	59.3	23.2

VCRBench is reported on its multiple-choice subset.

Usage

from transformers import AutoProcessor, Qwen3VLForConditionalGeneration

model = Qwen3VLForConditionalGeneration.from_pretrained(
    "ngqtrung/Video-HopChain-8B", dtype="auto", device_map="auto")
processor = AutoProcessor.from_pretrained("ngqtrung/Video-HopChain-8B")

The weights are BF16. The model was trained to reason inside <think>...</think> and to put the final answer in <answer>\boxed{...}</answer>, so use the same system prompt as training. It is in the dataset rows and in the repository.

Training

Confidence-Gated Exploration recovers zero-variance groups in GRPO (also called advantage collapse). With 8 rollouts per question, CGE draws the first 4 normally. If those 4 are all correct or all incorrect, the group carries no reward variance and no gradient, so CGE draws the last 4 with the policy's top token masked wherever its probability exceeds tau = 0.95, inside the reasoning span only. The masked positions are dropped from the loss while all 8 rollouts enter the group advantage, so the method spends no extra rollouts.


Base model	Qwen3-VL-8B-Instruct
Stage 1	GRPO on a 105,993-row general video QA mixture, 24 frames
Stage 2	GRPO + CGE on Video-HopChain, 22,550 rows, 140 frames
Rollouts per question	8: the first 4 as usual, the last 4 under the mask
Mask threshold	0.95, reasoning span only
Reward	exact match on the integer sum, plus a format term
Hardware	4 nodes of 8 H100 80GB, asynchronous verl trainer

Appendix C of the paper gives every hyperparameter.

Limitations

Evaluated at one model scale only. The dataset is synthetic and verified against captions rather than against the frames, so a caption error can reach a label. The paper reports both limitations, together with the exploration settings that we did not ablate.

Citation

@misc{nguyenquang2026videohopchain,
  title         = {Video-HopChain: Multi-Hop Questions and Confidence-Gated Exploration for Video Reasoning Models},
  author        = {Nguyen Quang, Trung and Dong, Yuhao and Sun, Shuo and Liu, Shuai and Tian, Shulin and Yap, Kim-Hui and Liu, Ziwei},
  year          = {2026},
  eprint        = {2609.25773},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2609.25773}
}