Visual Autoregressive (VAR) models adopt a next-scale prediction paradigm, offering high-quality content generation with substantially fewer decoding steps. However, existing VAR models suffer from significant attention complexity and severe memory overhead due to the accumulation of key-value (KV) caches across scales. In this paper, we tackle this challenge by introducing KV cache compression into the next-scale generation paradigm. We begin with a crucial observation: attention heads in VAR models can be divided into two functionally distinct categories: Contextual Heads focus on maintaining semantic consistency, while Structural Heads are responsible for preserving spatial coherence. This structural divergence causes existing one-size-fits-all compression methods to perform poorly on VAR models. To address this, we propose HACK, a training-free Head-Aware KV cache Compression frameworK. HACK utilizes an offline classification scheme to separate head types, enabling it to apply pattern-specific compression strategies with asymmetric cache budgets for each category. By doing so, HACK effectively constrains the average KV cache length within a fixed budget B, reducing the theoretical attention complexity from O(n4) to O(Bn2). Extensive experiments on multiple VAR models across text-to-image and class-conditional tasks validate the effectiveness and generalizability of HACK. It achieves up to 70% KV cache compression without degrading output quality, resulting in memory savings and faster in- ference. For example, HACK provides a 1.75× memory reduction and a 1.57× speedup on Infinity-8B.

Mathematical reasoning remains an ongoing challenge for AI models, especially for geometry problems, which require both linguistic and visual signals. As the vision encoders of most MLLMs are trained on natural scenes, they often struggle to understand geometric diagrams, performing no better in geometry problem-solving than LLMs that only process text. This limitation is further amplified by the lack of effective methods for representing geometric relationships. To address these issues, we introduce the Diagram Formalization Enhanced Geometry Problem Solver (DFE-GPS), a new framework that integrates visual features, geometric formal language, and natural language representations. Specifically, we propose a novel synthetic data approach and construct a large-scale geometric dataset, SynthGeo228K, annotated with formal and natural language captions, designed to enhance the vision encoder to understand geometric structures better. Our framework improves MLLMs’ ability to process geometric diagrams and extends their application to open-ended tasks on the formalgeo7k dataset.

Large language models (LLMs) excel at processing long sequences, boosting demand for key-value (KV) caching. While recent efforts to evict KV cache have alleviated the inference burden, they often fail to allocate resources rationally across layers with different attention patterns. In this paper, we introduce Cascading and Adaptive KV cache Eviction (CAKE), a novel approach that frames KV cache eviction as a ``cake-slicing problem.'' CAKE assesses layer-specific preferences by considering attention dynamics in both spatial and temporal dimensions, allocates rational cache size for layers accordingly, and manages memory constraints in a cascading manner. This approach enables a global view of cache allocation, adaptively distributing resources across diverse attention mechanisms while maintaining memory budgets. CAKE also employs a new eviction indicator that considers the shifting importance of tokens over time, addressing limitations in existing methods that overlook temporal dynamics. Comprehensive experiments on LongBench and NeedleBench show that CAKE maintains model performance with only 3.2\% of the KV cache and consistently outperforms current baselines across various models and memory constraints, particularly in low-memory settings. Additionally, CAKE achieves over 10$\times$ speedup in decoding latency compared to full cache when processing contexts of 128K tokens with FlashAttention-2. Our code is available at https://github.com/antgroup/cakekv.