In a six-task computer benchmark, a training method designed to track a model's reliance on images recorded the best reported balance between learning new multimodal tasks and retaining earlier performance. Its final average accuracy, known as AvgAcc, was 62.5%, the highest figure in the continual-learning comparison. The measure is meant to capture the trade-off between stability, keeping earlier abilities, and plasticity, learning the next task.
The study tackles multimodal unsupervised continual post-training: a model receives task streams one after another and must learn from image-question pairs without answer labels while retaining capabilities picked up earlier. The central question is whether those two goals can be met together. The supplied evaluation does not test people, animals or deployed systems.
How the approach uses visual information
VDA, short for Visual Dependence-Aware training, treats distortion in a model's visual-dependence pattern as a signal of forgetting. It combines Visually Constrained Optimal Transport, or VC-OT, to preserve old-task structure with Visually Modulated Adaptation, or VMA, to emphasize tokens grounded in the image when learning a new task. In the paper's token-level measure, visual dependence compares predictions made with the real image against predictions under a counterfactual visual condition. A positive score signals visual facilitation; a non-positive score suggests the prediction is being driven by language.
The reported evaluation covered TextVQA, SciVQA, StockQA, GQA, DriveLM and PMC-VQA. The primary setup used Qwen2.5-VL-7B, with VDA and the continual-learning baselines tuned with LoRA at rank 128. The comparison also included six unsupervised post-training methods and four multimodal-language-model continual-learning methods. In a separate hyperparameter analysis, the best reported performance occurred at rank 128, while a buffer of 1,000 produced an AvgAcc of 62.5.
What the benchmark found
The headline comparison favored VDA on all three summary measures. AvgLA, the average accuracy immediately after each task is learned, was 63.0%. AvgF, the average drop on earlier tasks, was 0.6, the lowest forgetting figure reported. Alongside the 62.5% AvgAcc, those scores gave VDA the best reported stability-plasticity result.
Against named comparators, VDA's 62.5% AvgAcc was 3.7 percentage points above ScPO, 2.4 points above SEEKR-MLLM and 8.6 points above the frozen Qwen2.5-VL-7B backbone. These are reported point estimates, and the study provides no uncertainty estimate for them. The figures therefore show how the methods ranked in this experiment, not a measured margin that can be assumed to hold in every model or task.
The component tests showed different scores for stripped-down versions of the design. In the reported VMA-only ablation, AvgLA changed from 61.7 to 63.8 and AvgAcc from 51.0 to 56.9, while AvgF was 8.2. The complete VDA configuration recorded AvgAcc of 62.5%, AvgLA of 63.0% and AvgF of 0.6 in the ablation analysis.
A preservation analysis focused on VC-OT. The configuration reported 14.27% drift in VD strength and 0.0385 drift in the VD distribution, alongside AvgAcc of 62.5 and AvgF of 0.6. The token-level L1 alternative had AvgAcc of 59.1 and AvgF of 0.4. Those measures describe how the model's visual-dependence pattern moved; they are not direct measures of reliability for users.
On TextVQA, another mechanism-level measure, RRAR, ended at 4.204 for VDA, 0.499 above baseline. Removing VMA or VC-OT in the reported comparison left RRAR at 3.986 and 3.914, respectively. The paper uses this result to discuss question-relevant visual attention, but RRAR is not a direct measure of human comprehension, hallucination rates or real-world decisions.
Why the result remains preliminary
Several boundaries are important. The study is a nonrandomized computational comparison centered on the reported primary backbone and task sequence; the main text also refers to additional random orders and model families in an unavailable appendix. The number of image-question pairs per task is not reported, and the supplied analysis reports no confidence intervals, p-values or repeated-run variability. Those omissions make it difficult to judge robustness beyond the stated benchmark.
Whether the approach's reported gains transfer across more MLLM families and task orders, or whether better VD and RRAR scores translate into fewer hallucinations and better multimodal decisions, remains open. The document is an arXiv preprint dated 26 August 2026, so the findings should be read as early computational evidence rather than established deployment performance.
Paper data and sources
Original title: A Visual Dependence-Aware Framework for Multimodal Unsupervised Continual Post-Training
Authors: Kaichen Li, Zhilin Zhu, Jianhao Huang et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text