# PACE: A Unified Condense-and-Extract Paradigm for Fast VLM Inference

arXiv: https://arxiv.org/abs/2608.27206 · HTML: https://mlsanity.com/paper/2608.27206/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Placed in Multimodal reasoning, one of 17 directions in LLM reasoning.
- Builds on 21 of its 29 references that are in the map.
- First posted on arXiv 27 Aug 2026.

## Placed in

- LLM reasoning / [Multimodal reasoning](https://mlsanity.com/lr/multimodal-reasoning/)
- Multimodal LLMs / vision-language models / [Visual token pruning and efficiency](https://mlsanity.com/mllm/visual-token-pruning-and-efficiency/)

## Builds on (most cited first, 21 of 21)

- [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://mlsanity.com/paper/2010.11929/) (2020)
- [FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness](https://mlsanity.com/paper/2205.14135/) (2022)
- [InstructBLIP: Towards General-purpose Vision-Language Models with Instruction Tuning](https://mlsanity.com/paper/2305.06500/) (2023)
- [FlashAttention-2: Faster Attention with Better Parallelism and Work Partitioning](https://mlsanity.com/paper/2307.08691/) (2023)
- [Towards VQA Models That Can Read](https://mlsanity.com/paper/1904.08920/) (2019)
- [InternVL3.5: Advancing Open-Source Multimodal Models in Versatility, Reasoning, and Efficiency](https://mlsanity.com/paper/2508.18265/) (2025)
- [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)
- [AWQ: Activation-aware Weight Quantization for LLM Compression and Acceleration](https://mlsanity.com/paper/2306.00978/) (2023)
- [Stop Looking for Important Tokens in Multimodal Language Models: Duplication Matters More](https://mlsanity.com/paper/2502.11494/) (2025)
- [Beyond Attention or Similarity: Maximizing Conditional Diversity for Token Pruning in MLLMs](https://mlsanity.com/paper/2506.10967/) (2025)
- [Don't Just Chase "Highlighted Tokens" in MLLMs: Revisiting Visual Holistic Context Retention](https://mlsanity.com/paper/2510.02912/) (2025)
- [Accelerating Streaming Video Large Language Models via Hierarchical Token Compression](https://mlsanity.com/paper/2512.00891/) (2025)
- [MMTok: Multimodal Coverage Maximization for Efficient Inference of VLMs](https://mlsanity.com/paper/2508.18264/) (2025)
- [SCOPE: Saliency-Coverage Oriented Token Pruning for Efficient Multimodel LLMs](https://mlsanity.com/paper/2510.24214/) (2025)
- [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)
- [Are We Using the Right Benchmark: An Evaluation Framework for Visual Token Compression Methods](https://mlsanity.com/paper/2510.07143/) (2025)
- [Accelerating Vision Transformers with Adaptive Patch Sizes](https://mlsanity.com/paper/2510.18091/) (2025)
- [ViCO: A Training Strategy towards Semantic Aware Dynamic High-Resolution](https://mlsanity.com/paper/2510.12793/) (2025)
- [DocPrune:Efficient Document Question Answering via Background, Question, and Comprehension-aware Token Pruning](https://mlsanity.com/paper/2604.22281/) (2026)

## Abstract (arXiv)

Vision-Language Models (VLMs) demonstrate exceptional visual reasoning capabilities, yet their inference costs escalate rapidly with the proliferation of visual tokens. Existing visual token pruning methods exhibit two fundamental limitations. First, most approaches operate exclusively post-vision encoder, leaving the substantial latency of the visual encoding phase unoptimized. Second, under strict token budgets, these methods often fail to jointly preserve holistic visual contexts and fine-grained details, leading to performance degradation. To address these bottlenecks, we propose PACE (Pixel-Adaptive Condense and Extract), a training-free inference framework that accelerates both the vision encoder and the Large Language Model (LLM) via a unified Condense-and-Extract paradigm. During the Condense stage, an Adaptive Pixel Compressor (APC) evaluates visual information density prior to encoding, adaptively downsampling redundant inputs, curtailing encoder computation while preserving global context and essential visual cues. In the Extract stage, a Dynamic Dual-Attention Extractor (DDAE) selectively retains visual tokens via a fusion of internal visual signals from the encoder and semantic signals from the LLM, safeguarding task-critical details. By integrating PACE into Qwen2.5-VL-7B, the model retains 93.8% of its original performance while utilizing only 10% of the visual tokens, yielding a 3.1x speedup in time to first token (TTFT). Our code is available at https://github.com/jjL357/PACE.