arXiv subjects

Haozhe Hu

Publications and source records attributed to Haozhe Hu.

6 recordsLinked to original sources

arXiv2026

Beyond FLOPs: Benchmarking Real Inference Acceleration of LLM Pruning under a GEMM-Centric Taxonomy

Pruning has emerged as a dominant paradigm for accelerating large language model (LLM) inference, spanning a broad spectrum of methods that remove computation across tokens, layers, heads, dimensions, and attention patterns. Despite sharing the same objective, these pruning approaches induce fundamentally different execution behaviors, causing realized speedups to depend heavily on hardware and kernel implementations. Consequently, the practical acceleration benefits of different pruning families remain poorly understood. In this work, we introduce a GEMM-centric taxonomy that reorganizes existing pruning methods according to the logical \textbf{M}, \textbf{N}, and \textbf{K} dimensions of general matrix multiplication (GEMM). Leveraging this abstraction, we build a unified benchmarking framework that enables implementation-consistent comparison across the pruning design space and systematically characterizes the acceleration--quality Pareto frontier. Our results on Llama3.1-8B show that static depth pruning remains the strongest Pareto-optimal baseline and stays closest to its theoretical acceleration upper bound in memory-bounded scenarios. During prefill, the frontier transitions from static depth at low quality loss (0\%--4\%), to dynamic depth at moderate loss (5\%--16\%), and finally to static width pruning at higher loss levels (17\%--26\%). These findings establish the first unified view of the practical limits of pruning-based LLM acceleration and provide guidance for future pruning research. Code is available at https://github.com/EIT-NLP/LLM-Pruning/tree/main/PruningInferSim

cs.LG↗

arXiv2026

ReCache: Efficient KV Cache Reuse and Compression for Tool-Augmented LLM Agents

Agentic language models repeatedly encode tool and skill schemas that recur across requests in different combinations and orders, preventing standard prefix caching from reusing their key--value (KV) states. We introduce \textbf{ReCache}, a framework for independently caching resource representations while reducing their inference-time computational and memory overhead. Resource-wise attention removes cross-resource interactions and assigns resource-local positions, producing composition-invariant KV blocks. ReCache then restricts resource visibility to contribution-selected layer--KV-head-group routes and retains only invocation-critical fields through structural and semantic pruning. We evaluate ReCache on a benchmark assembled from seven public tool- and skill-use datasets, including resource-disjoint tests. Resource-wise attention matches dense invocation performance (82.3\% versus 82.4\% Inv-F1) while providing a 3.655$\times$ time-to-first-token speedup. The complete framework reduces allocated KV-tensor memory by 92.43\% and accelerates attention by 1.423$\times$. These results show that separating reusable schema encoding from selective resource access substantially reduces agentic inference costs with limited effectiveness loss. The code is available at https://github.com/EIT-NLP/ReCache.

cs.CL↗

arXiv2026

WIDE: Boosting Adaptive LLM Inference via Token-level Dynamic Width Pruning

Pruning is a promising approach for improving the efficiency of LLMs. Existing static structured pruning methods are hardware-friendly and can deliver practical throughput gains, but their input-agnostic computation allocation often causes substantial accuracy degradation under aggressive sparsity. Recent dynamic sparsity methods improve quality retention by adapting computation to individual inputs, yet they remain largely limited to coarse-grained structural decisions and their practical acceleration under real-world inference scenarios remains challenging. To address these challenges, we present WIDE, the first end-to-end differentiable token-level dynamic width pruning framework designed for both prefill and decode scenarios. WIDE enables fine-grained computation allocation by allowing each token to dynamically select attention-head groups and FFN-channel groups, extending dynamic pruning beyond layer-level decisions to neuron-block-level granularity. Through a two-stage training pipeline, WIDE learns effective token-wise sparse execution patterns and achieves substantially better quality retention than existing approaches. To make such fine-grained dynamic pruning practical, we further propose a pruning--kernel co-design framework that decomposes dynamic sparsity acceleration into mask reordering, hardware-agnostic block-level skipping, and hardware-dependent intra-block skipping, enabling efficient execution across different granularities. At 50% sparsity, WIDE provides 55.1% performance boost when compared to the state-of-the-art dynamic depth pruning under calibration-only settings. Under prefill and decoding inference workloads, WIDE achieves close-to-theoretical kernel-level speedups of up to 1.98x for prefill and 4.95x for decoding, as well as 1.68x and 1.55x end-to-end acceleration. Our code is available at https://github.com/EIT-NLP/LLM-Pruning/tree/main/WIDE.

cs.AI↗

arXiv2026

Beyond Single-Dimensional Compression: The Compound Sparsity Frontier of Large Language Models

Large language models (LLMs) are often compressed through static parameter pruning or dynamic token-level computation, yet aggressive sparsification can trigger rapid performance degradation beyond an essential sparsity boundary. This work asks \emph{whether combining these two mechanisms can delay such degradation by distributing the compression burden}. We study a minimalist compound sparsity framework that first applies low-rank approximation and channel pruning to obtain a statically compressed backbone, and then introduces lightweight routers for per-token dynamic layer skipping. This design enables independent control of parameter sparsity and token-level computation sparsity. Experiments across language understanding and modeling benchmarks show that compound sparsity consistently outperforms single-mechanism compression under the same total sparsity, delaying the decay point on understanding tasks and preserving stronger modeling performance. Further analysis reveals cross-dimensional interference between parameter pruning and token skipping, and shows that near-balanced allocation is most effective under a fixed sparsity budget. These results demonstrate that compound compression provides a practical way to improve LLM compression, while revealing a broader cross-dimensional sparsity boundary that ultimately limits further compression. Code will be available at https://github.com/EIT-NLP/LLM-Pruning.

cs.LG↗

arXiv2026

Chain Of Interaction Benchmark (COIN): When Reasoning meets Embodied Interaction

Generalist embodied agents must perform interactive, causally-dependent reasoning, continually interacting with the environment, acquiring information, and updating plans to solve long-horizon tasks before they could be adopted in real-life scenarios. For instance, retrieving an apple from a cabinet may require opening multiple doors and drawers before the apple becomes visible and reachable, demanding sequential interaction under partial observability. However, existing benchmarks fail to systematically evaluate this essential capability. We introduce COIN, a benchmark designed to assess interactive reasoning in realistic robotic manipulation through three key contributions. First, we construct COIN-50: 50 interactive tasks in daily scenarios, and create COIN-Primitive required by causally-dependent tasks, and COIN-Composition with mid-term complexity for skill learning and generalization evaluation. Second, we develop a low-cost mobile AR teleoperation system and collect the COIN-Primitive Dataset with 50 demonstrations per primitive task (1,000 in total). Third, we develop systematic evaluation metrics about execution stability and generalization robustness to evaluate CodeAsPolicy, VLA, and language-conditioned H-VLA approaches. Our comprehensive evaluation reveals critical limitations in current methods: models struggle with interactive reasoning tasks due to significant gaps between visual understanding and motor execution. We provide fine-grained analysis of these limitations.

cs.RO↗

arXiv2026

CS-MUNet: A Channel-Spatial Dual-Stream Mamba Network for Multi-Organ Segmentation

Recently Mamba-based methods have shown promise in abdominal organ segmentation. However, existing approaches neglect cross-channel anatomical semantic collaboration and lack explicit boundary-aware feature fusion mechanisms. To address these limitations, we propose CS-MUNet with two purpose-built modules. The Boundary-Aware State Mamba module employs a Bayesian-attention framework to generate pixel-level boundary posterior maps, injected directly into Mamba's core scan parameters to embed boundary awareness into the SSM state transition mechanism, while dual-branch weight allocation enables complementary modulation between global and local structural representations. The Channel Mamba State Aggregation module redefines the channel dimension as the SSM sequence dimension to explicitly model cross-channel anatomical semantic collaboration in a data-driven manner. Experiments on two public benchmarks demonstrate that CS-MUNet consistently outperforms state-of-the-art methods across multiple metrics, establishing a new SSM modeling paradigm that jointly addresses channel semantic collaboration and boundary-aware feature fusion for abdominal multi-organ segmentation.

cs.CV↗