# Not All Visual Tokens Are Equally Safe to Remove:Consequence-Sensitive Visual Token Compression

arXiv: https://arxiv.org/abs/2608.09176 · HTML: https://mlsanity.com/paper/2608.09176/

## Key facts

- Cited by 2 papers in the map (data through 14 Oct 2026).
- Placed in Visual token pruning in multimodal models, one of 12 directions in Efficient LLM inference and serving.
- Builds on 10 of its 17 references that are in the map.
- First posted on arXiv 10 Aug 2026.

## Placed in

- Efficient LLM inference and serving / [Visual token pruning in multimodal models](https://mlsanity.com/inference/visual-token-pruning-in-multimodal-models/)

## Builds on (most cited first, 10 of 10)

- [LLaVA-OneVision: Easy Visual Task Transfer](https://mlsanity.com/paper/2408.03326/) (2024)
- [Making the V in VQA Matter: Elevating the Role of Image Understanding in Visual Question Answering](https://mlsanity.com/paper/1612.00837/) (2016)
- [Towards VQA Models That Can Read](https://mlsanity.com/paper/1904.08920/) (2019)
- [ChartQA: A Benchmark for Question Answering about Charts with Visual and Logical Reasoning](https://mlsanity.com/paper/2203.10244/) (2022)
- [DocVQA: A Dataset for VQA on Document Images](https://mlsanity.com/paper/2007.00398/) (2020)
- [PyramidDrop: Accelerating Your Large Vision-Language Models via Pyramid Visual Redundancy Reduction](https://mlsanity.com/paper/2410.17247/) (2024)
- [Prune Redundancy, Preserve Essence: Vision Token Compression in VLMs via Synergistic Importance-Diversity](https://mlsanity.com/paper/2603.09480/) (2026)
- [AdaptVision: Efficient Vision-Language Models via Adaptive Visual Acquisition](https://mlsanity.com/paper/2512.03794/) (2025)
- [Not All Errors Are Equal: Consequence-Aware Reasoning Compute Allocation](https://mlsanity.com/paper/2606.04402/) (2026)
- [EERO: Early Exit with Reject Option for Efficient Classification with limited budget](https://mlsanity.com/paper/2402.03779/) (2024)

## Built on by (most cited first, 2 of 2)

- [Preserving Mathematical Reasoning in Compressed Diffusion Language Models via Trajectory-Aware Low-Rank Approximation](https://mlsanity.com/paper/2610.03326/) (2026)
- [Decide Before You Look: Learning Which Retrieved Memories Deserve Pixels](https://mlsanity.com/paper/2610.07984/) (2026)

## Abstract (arXiv)

Visual token compression for vision--language models (VLMs) has largely relied on criteria such as attention, redundancy, and uncertainty to maximize average accuracy under a fixed compute budget, implicitly assuming that all errors carry equal cost. However, the consequence of an incorrect prediction on downstream tasks is rarely symmetric: misreading an invoice amount can be far more costly than misclassifying a background color. Motivated by this, we introduce consequence-sensitive visual token compression, which allocates visual computation across requests according to their potential error costs. Our method follows a calibrate-then-allocate procedure, estimating consequence-specific error-budget curves offline and applying the calibrated token budgets online using consequence signals available from question or task information. On a controlled within-task benchmark, high- and low-consequence questions are drawn from the same document images, so content alone cannot reveal which questions are costly to get wrong. In this setting, our method reduces high-stakes errors from 0.300 to 0.133 under the same total token budget, whereas a content-driven allocator performs no better than uniform allocation. Measuring how error rates change with token budget across different cost ratios, we derive an allocation frontier: uniform allocation is optimal when errors are equally costly, and token transfer toward high-consequence questions becomes increasingly beneficial as the cost gap grows. This allocation principle generalizes well across three dense vision-language benchmarks, two budget realization mechanisms (token deletion and resolution reallocation), two VLM architectures, and multiple token selection strategies. On a realistic mixed workload, consequence-sensitive allocation reduces cost-weighted error by 38% while achieving approximately 21% lower latency than full-resolution inference.