← 研踪 PaperTrailarXiv:2611.19745 · v2
TIVS: Mapping the Sensitivity Landscape of Visual Token Transformations in MLLMs
首次提交 2026-11-15 11:13
版本更新 2026-11-17 11:16
雷达观测 2026-11-18 19:43
摘要
Recent efforts to improve the efficiency of multimodal large language models (MLLMs) have explored visual token pruning, quantization, and feature compression. However, the properties that determine whether transformed visual tokens remain usable by a frozen decoder are not well characterized. We introduce \textbf{TIVS (Token Identity Versus Statistics)}, a training-free intervention framework that separately studies sequence permutation, token retention, numerical precision, and statistical replacement under controlled conditions. On 192 POPE questions, three token permutations change TinyLLaVA accuracy by only percentage points on average, while changing 4.69\% of individual predictions. Diagonal and rank-64 Gaussian replacements decrease accuracy by 25.0 and 28.1 points on the same subset. Further experiments with exact moment constraints, quantization error, token retention, and additional visual tasks reveal how strongly the effects depend on the original feature values and the evaluation setting. These findings provide an empirical basis for assessing visual token transformations in frozen MLLMs and designing more reliable multimodal interfaces.
作者备注:9 pages, 12 figures, 3 tables
观测到的版本
相邻记录可能跳过中间版本;首次发现 v2 及以上也会标为版本更新。
v2 · 2026-11-17 11:16 · 最新观测版本
TIVS: Mapping the Sensitivity Landscape of Visual Token Transformations in MLLMs
Recent efforts to improve the efficiency of multimodal large language models (MLLMs) have explored visual token pruning, quantization, and feature compression. However, the properties that determine whether transformed visual tokens remain usable by a frozen decoder are not well characterized. We introduce \textbf{TIVS (Token Identity Versus Statistics)}, a training-free intervention framework that separately studies sequence permutation, token retention, numerical precision, and statistical replacement under controlled conditions. On 192 POPE questions, three token permutations change TinyLLaVA accuracy by only percentage points on average, while changing 4.69\% of individual predictions. Diagonal and rank-64 Gaussian replacements decrease accuracy by 25.0 and 28.1 points on the same subset. Further experiments with exact moment constraints, quantization error, token retention, and additional visual tasks reveal how strongly the effects depend on the original feature values and the evaluation setting. These findings provide an empirical basis for assessing visual token transformations in frozen MLLMs and designing more reliable multimodal interfaces.
查看此版本解读 →查看此版本原文 ↗v1 · 2026-11-15 11:13
TIVS: Mapping the Sensitivity Landscape of Visual Token Transformations in MLLMs
Recent efforts to improve the efficiency of multimodal large language models (MLLMs) have explored visual token pruning, quantization, and feature compression. However, the properties that determine whether transformed visual tokens remain usable by a frozen decoder are not well characterized. We introduce \textbf{TIVS (Token Identity Versus Statistics)}, a training-free intervention framework that separately studies sequence permutation, token retention, numerical precision, and statistical replacement under controlled conditions. On 192 POPE questions, three token permutations change TinyLLaVA accuracy by only percentage points on average, while changing 4.69\% of individual predictions. Diagonal and rank-64 Gaussian replacements decrease accuracy by 25.0 and 28.1 points on the same subset. Further experiments with exact moment constraints, quantization error, token retention, and additional visual tasks reveal how strongly the effects depend on the original feature values and the evaluation setting. These findings provide an empirical basis for assessing visual token transformations in frozen MLLMs and designing more reliable multimodal interfaces.
查看此版本解读 →查看此版本原文 ↗