ViSS-8B

Supervising the Visible State: Separating Representation from View Selection in Streaming Video

Yihao Hu*, Pan Wang*, Yitong Lin*, Zhihao Wen†, Deyuan Liu, Hang Wang, Mingda Wan, Yiwen Zhao, Yutong Li, Yuchen Mao, Tao Lin†

Equal contribution. †Corresponding author.

Collection · Source code · Installation · Training · Evaluation

Overview

Visible-State Supervision (ViSS) trains a vision–language model on source-grounded questions about event identity, order, accumulation, relation changes, and whether the available evidence suffices to answer. Test-Time Causal View Search (TVS) keeps the model fixed and acquires complementary recent, event, and compact-history views.

The model is initialized from Qwen3-VL-8B-Instruct and trained for one epoch on 7,677 examples, with learning rate 2e-6 and global batch size 12. The vision tower is frozen; the language model and multimodal projector are trainable. Model-only ViSS and TVS use the same checkpoint.

Released files

• Four original BF16 safetensors shards (17.53 GB in total).
• Model configuration, tokenizer, chat template, and native processor.
• locked_processor/: the processor used for the recent-view evaluation.
• code/: training, inference, TVS composition, evaluation adapters, public configuration templates, documentation, and synthetic tests.

The release contains no training annotations, raw media, company experiment records, prediction files, evaluation summaries, optimizer states, or internal training manifests. Training data are not being released in this version.

Download

hf download mis2/ViSS-8B --local-dir checkpoints/ViSS-8B

Load the model

Use a Transformers version supporting Qwen3-VL; the retained model configuration specifies Transformers 4.57.3.

import torch
from transformers import AutoProcessor, Qwen3VLForConditionalGeneration

model_id = "mis2/ViSS-8B"
processor = AutoProcessor.from_pretrained(
    model_id, subfolder="locked_processor"
)
model = Qwen3VLForConditionalGeneration.from_pretrained(
    model_id,
    dtype=torch.bfloat16,
    device_map="auto",
)

The root processor is the native training processor. The locked_processor/ subdirectory preserves the recent-view evaluation processor. The supplied evaluator implements causal sampling and benchmark prompt conventions; loading the processor alone does not implement TVS.

Use the code

cd checkpoints/ViSS-8B/code
python -m pip install -e ".[inference,test]"
python -m pytest -q tests

See the evaluation guide to bind the public configuration template to your own processor, benchmark query exports, and model. Benchmark media and official scoring installations must be obtained separately. TVS additionally requires calibration inputs prepared independently of benchmark evaluation.

License

The source code follows the project's Apache-2.0 license. The model is derived from Qwen3-VL-8B-Instruct, which is also distributed under Apache-2.0. External benchmark and media terms remain with their respective providers.

Citation

@misc{hu2026viss,
  title={Supervising the Visible State: Separating Representation from View Selection in Streaming Video},
  author={Yihao Hu and Pan Wang and Yitong Lin and Zhihao Wen and Deyuan Liu and Hang Wang and Mingda Wan and Yiwen Zhao and Yutong Li and Yuchen Mao and Tao Lin},
  year={2026},
  note={Manuscript}
}