[
    {
        "id": "osp-22085",
        "type": "article-journal",
        "title": "Beyond Selection: Token Parameterization for Extreme Visual Token Compression",
        "author": [
            {
                "family": "Zhong",
                "given": "Rui"
            },
            {
                "family": "Li",
                "given": "Yu"
            },
            {
                "family": "Yan",
                "given": "Zheyu"
            },
            {
                "family": "Zhuo",
                "given": "Cheng"
            }
        ],
        "URL": "https://omanscience.com/en/articles/beyond-selection-token-parameterization-for-extreme-visual-token-compression",
        "language": "en",
        "issued": {
            "date-parts": [
                [
                    2026
                ]
            ]
        },
        "abstract": "Visual-token compression is effective for improving the efficiency of vision-language models, but under extreme compression budgets, token pruning can break visual grounding while learned resamplers increase parameter count, attention cost, and training complexity. We revisit compression through a token parameterization lens, separating (i) basis transformation and structured truncation (retained subspace/compressibility) from (ii) coordinate organization (optimization and cross-modal alignment). This view yields two coupled objectives, compressibility and learnability, which we formalize as unified functionals. Guided by these objectives, we design Braco, a lightweight four-step coder that combines transform-basis truncation, input-independent basis-coordinate embeddings, budget-dependent orthogonal re-parameterization, and learned spatial residual tokens from lightweight pooling. Experiments show that Braco forms the favorable empirical accuracy-efficiency frontier under $23\\times$--$64\\times$ compression and remains competitive at $144\\times$, reaching 95.2% accuracy while reducing prefill FLOPs by 84.2%--86.7% relative to the uncompressed upper bound. Against prior methods, Braco matches or improves accuracy while achieving up to approximately 36% end-to-end speedup and using $16.6\\times$/$78.8\\times$ lower compressor latency/FLOPs."
    }
]