Jul 2026· International Conference on Intelligent Computing· pp. 520-530· 0 citations· 30 references
Computer Science
TL;DR
An Adaptive Depth Sparse Framework (AdaDSF) that converts off-the-shelf pre-trained LLMs into depth-sparse models without full retraining and consistently yields smaller accuracy degradation than strong baselines including MoD, D-LLM, and DLO.
Abstract
Large language models (LLMs) achieve strong generation and reasoning performance, but the Transformer architecture incurs high inference cost. Existing acceleration methods often rely on task-specific fine-tuning or training from scratch, increasing adaptation cost and limiting cross-task usability. We present an Adaptive Depth Sparse Framework (AdaDSF) that converts off-the-shelf pre-trained LLMs into depth-sparse models without full retraining. Our key insight is that layers contribute unequally to representation transformation, characterized by the cosine similarity between layer input and output hidden states. Based on this, AdaDSF assigns layer-wise token retention ratios from similarity statistics, uses a lightweight router to select informative tokens at each layer, and introduces a feature-preserving alignment objective to match intermediate and final representations between sparse and dense models. On GPT-NeoX and Qwen2.5 over language modeling and commonsense reasoning, AdaDSF substantially reduces inference FLOPs while preserving performance close to dense counterparts. Under comparable sparsity, AdaDSF consistently yields smaller accuracy degradation than strong baselines including MoD, D-LLM, and DLO.
With the rapid development of large-scale pre-trained language models based on Transformer architectures, their high computational and memory costs have become a major obstacle to deployment, especially in resource-constrained environments. Traditional pruning methods typically depend on full gradient-based importance estimation, and they necessitate prior finetuning of the model to achieve satisfactory performance. This process often results in intolerable resource consumption. This paper proposes REP-LIE, a new approach to enable resource-efficient pruning during the process of finetuning. REP-LIE leverages the gradients of LoRA low-rank matrices to estimate the importance of weights without requiring full gradient computation. To address the inherent randomness in importance estimation, a stability score is introduced, serving as the basis for iterative pruning of unimportant model parameters. The pruned model is further finetuned through lightweight updates, eliminating the need for full-parameter optimization in the process of finetuning. Extensive experiments on both medium-scale encoder models and large-scale generative models (LLaMA-7B and Mistral-7B) demonstrate that REP-LIE still achieves competitive performance compared to existing approaches.
Peng Liu, Hui-Bing Zeng, Yi-Qun Zhang et al.· 0 citations
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.
Large language models (LLMs) remain expensive to fine-tune because full-parameter updates require substantial memory, compute, and per-task storage. We study whether saliency signals originally developed for pruning can be reused to choose where a model should adapt. We propose Super, a sparse parameter-efficient fine-tuning (PEFT) method that fixes a small trainable support using a Wanda-style activation-weighted magnitude score [Sun et al., 2023] computed from a calibration pass. We then introduce Supra, a hybrid adapter that combines this sparse update with LoRA while preserving a matched trainable-parameter budget through a simple budget-splitting rule. In single-seed Math17K arithmetic experiments on Llama-3.2-1B and Meta-Llama-3-8B, the best Super/Supra variants achieve the highest average accuracy among the tested schedule-selected adapter configurations. We also include a PaFi-style magnitude-only support as a closest training-free sparse baseline and find that low-score supports under both magnitude and Wanda-style orderings can be effective. These results suggest that simple pruning-inspired orderings can provide useful fixed sparse supports for PEFT, especially when combined with low-rank adapters.
Ivan D. Ilin, Philip Zmushko, Peter Richtárik· 0 citations
Large Language Models (LLMs) trained using Chain-of-Thought (CoT) supervision have achieved state-of-the-art performance on complex reasoning tasks. However, the generation of long reasoning chains introduces substantial computational overhead during inference, limiting their deployment in low-latency and resource-constrained environments. This paper proposes AdaptiReason, a novel framework that dynamically compresses intermediate reasoning steps based on task difficulty and model confidence without requiring retraining of the underlying base model. AdaptiReason employs a lightweight difficulty estimator to determine the appropriate reasoning depth for each input, followed by a learned token-pruning policy that eliminates redundant or low-information reasoning steps. Experimental evaluation on the MATH, GSM8K, and ARC-Challenge benchmarks demonstrates that AdaptiReason reduces the average number of generated tokens by 3.7× while preserving 98.2% of the baseline reasoning accuracy. Furthermore, the proposed framework is model-agnostic and can be seamlessly integrated with instruction-tuned LLMs without requiring access to model parameters, relying solely on output logits for adaptive reasoning compression. The results demonstrate that AdaptiReason significantly improves inference efficiency while maintaining high reasoning performance, making it suitable for real-time and resource-constrained LLM applications.
V. A, Mithaguru, Amrita Kundu et al.· 2026 4th International Confe...· 0 citations
Although Low-Rank Adaptation (LoRA) rev-olutionized parameter-efficient fine-tuning, it often incurs an inference overhead due to the extra computation required by adapter layers. While most literature focuses on maximizing accuracy or minimizing parameter counts, this paper prioritizes single-request inference performance in the unmerged adapter setting, where adapters must remain decoupled from the base model at runtime. By analyzing LoRA adapters on GPUs, we identify segmented function calls as the primary source of this latency. To address this, we propose G rouped A daptive W eight S haring (GAWS), a novel adapter design based on structured Kronecker product decomposition . Experiments on T5-3B, GPT-2 Large, LLaMA3.2-3B, and RoBERTa-Large show that GAWS reduces latency to about 40% of the gap between the unmerged LoRA and the base model, while maintaining parameter efficiency and comparable accuracy. This positions GAWS as a Pareto-efficient solution for deploying adapted LLMs in latency-sensitive settings, balancing the low latency of compressed adapters with the accuracy of LoRA. The source code is available at: https://github.com/SamsungLabs/GAWS .
Eman Alsuradi, Junhyung Lee, Kyeng-Hun Lee et al.· Annual Meeting of the Associ...· 0 citations
Large language models (LLMs) have been applied to sequential recommendation by incorporating collaborative signals through input conditioning or model adaptation. However, existing approaches often require LLM fine-tuning, additional architectural modules, representation distillation, or item-level conditioning over long interaction histories, increasing computational and deployment costs. We propose REPREC, a lightweight framework that conditions a frozen LLM using compact user-level representations. REPREC maps a fixed-size embedding from a frozen sequential encoder into a small set of learned soft tokens through an MLP injector, training only the injector while leaving both pretrained backbones unchanged. Our extensive experiments demonstrate that REPREC consistently improves recommendation performance across different sequential encoders, LLM backbones, and user activity levels. Its compact conditioning mechanism also makes REPREC computationally efficient during both training and inference. Moreover, training with short histories while evaluating with longer contexts retains 94--99\% of full-history performance while achieving an average $1.50\times$ per-epoch training speedup. The code is available at: https://github.com/phdbotcode/REPREC
Harshini Kavuru, Dwipam Katariya, Giri Iyengar et al.· 0 citations