An arXiv preprint reports that a video-language model kept almost all of its average performance on four long-video benchmarks while retaining only 25% of its visual tokens. On Qwen3-VL-8B, the proposed dynamic system, VTCDy, reached 97.8% average retention at the 0.25 setting. In this measure, 100% represents the average score of the same model without compression, so a figure just below 100% indicates a small reported difference in the tested average. The result comes from multiple-choice VideoQA tasks on MLVU, LongVideoBench, LVBench and Video-MME, spanning short clips to videos lasting up to two hours.
The stated aim is to apply video-coding principles to compress the tokens used by existing multimodal large language models, or MLLMs. The baseline method, called VTC, moves the logic of reference prediction and residual coding into visual feature space. It builds semantic I-frames and P-frames, then selects P-frame tokens by residual energy, a way of measuring how much information remains after reference prediction. VTC is described as training-free and requiring no fine-tuning.
VTCDy adds three pieces to VTC: Dynamic Resolution Input, Dynamic Token Allocation and Spatial Coverage Top-K. The first changes the resolution supplied to the model, while the other two govern how many tokens are assigned and which spatial tokens are kept. The study evaluates the methods on three video MLLMs: Qwen3-VL-8B, LLaVA-OneVision and LLaVA-OneVision-2.
Results across models
At 50% retention, the reported average retention was 100.1% for Qwen3-VL, 101.6% for LLaVA-OneVision and 100.3% for LLaVA-OneVision-2. At 25%, the two LLaVA backbones were reported at 102.1% and 97.8% on average. Because the metric is relative, values above 100% are reported as slightly higher aggregate scores than the uncompressed comparison.
On Qwen3-VL, the paper reports 1.4 percentage points higher average performance than FlashVID. In a matched comparison at 25% retention, VTCDy's reported retention was 100.5% on LongVideoBench and 98.2% on MLVU. Those numbers were 1.9 and 2.4 points above the second-best retention, respectively.
The paper also reports a model-side latency comparison under a matched budget equivalent to 64 full-resolution frames. At 25% retention, the figures were 4,109 milliseconds and 1,644 milliseconds on one RTX 3090, a reported 2.50 times difference. The latency result is therefore tied to that specific setup, not a general hardware-independent measure.
The reported module tests followed the same incremental logic. On MLVU, the average values moved from 67.3 for VTC to 68.2 after DyRSO, 68.9 after DyTA and 69.2 after SC-TopK. On LVBench, the reported score changed from 40.1 to 41.2.
What the tests leave open
The evidence is limited to three MLLM backbones, four long-video benchmarks and the reported retention settings. Latency was measured on one RTX 3090. The supplied analysis reports no inferential statistical tests or uncertainty intervals, leaving open how stable the differences would be across repeated runs, additional model families, datasets, sampling schemes and hardware.
The results show what VTCDy did in the tested configurations; they do not establish that it will outperform VTC or other baselines in general, or that lower latency will translate into lower energy or cost. Those broader questions remain unresolved by the reported benchmark comparisons.
The manuscript is an arXiv preprint, listed as arXiv:2608.28008v1 and dated 28 Aug 2026. The paper says its code is available through the supplied GitHub repository.
Paper data and sources
Original title: Visual Token Coding for Video Multimodal Large Language Models
Authors: Chenxin Fang, Tao Chen, JunChao You et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-28
DOI: Not available
Original paper · Full text