Papers with LLM inference
KV Pareto: Systems-Level Optimization of KV Cache and Model Compression for Long Context Inference (2026.eacl-industry)
Copied to clipboard
| Challenge: | Long-context Large Language Models (LLMs) face significant memory bottlenecks due to the linear growth of key-value (KV) cache with sequence length. |
| Approach: | They propose a framework that maps the trade-off frontier between total memory consumption and task accuracy across three complementary optimization techniques. |
| Outcome: | The proposed model-specific configurations achieve 68-78% total memory reduction with minimal (1-3%) accuracy degradation on long-context tasks. |
LinguaLinked: Distributed Large Language Model Inference on Mobile Devices (2024.acl-demos)
Copied to clipboard
| Challenge: | Recent research shows that large language models demonstrate enhanced capabilities in various language tasks. |
| Approach: | They introduce a system for decentralized, distributed LLM inference on mobile devices . they use optimized model assignment technique to segment LLMs and linear optimization to align segments with each device . |
| Outcome: | The proposed system performs well on high-end to low-end Android devices. |
ScaleLLM: A Resource-Frugal LLM Serving Framework by Optimizing End-to-End Efficiency (2024.emnlp-industry)
Copied to clipboard
Yuhang Yao, Han Jin, Alay Shah, Shanshan Han, Zijian Hu, Dimitris Stripelis, Yide Ran, Zhaozhuo Xu, Salman Avestimehr, Chaoyang He
| Challenge: | Large language models (LLMs) are widely used in commercial applications . low latency is crucial due to system latency, query concurrency, and computational resources constraints. |
| Approach: | They propose a system that can be resource-efficiently served by addressing bottlenecks beyond LLM inference . they propose 4.3 speed up over vLLM and 1.5 higher throughput . |
| Outcome: | The proposed system outperforms state-of-the-arts with 1.5 higher throughput . it achieves 4.3 speed up with 64 concurrent requests on Mixtral 8x7B . |
HierGR: Hierarchical Semantic Representation Enhancement for Generative Retrieval in Food Delivery Search (2025.acl-industry)
Copied to clipboard
Fuwei Zhang, Xiaoyu Liu, Xinyu Jia, Yingfei Zhang, Zenghua Xia, Fei Jiang, Fuzhen Zhuang, Wei Lin, Zhao Zhang
| Challenge: | Generative retrieval (GR) is an emerging search paradigm for food delivery search. |
| Approach: | They propose a method that harnesses the advanced query understanding capabilities of large language models to enhance the retrieval of results for complex and long-tail queries in food delivery search scenarios. |
| Outcome: | The proposed method increases the number of online orders by 0.68% for complex search intents. |
KNN-SSD: Enabling Dynamic Self-Speculative Decoding via Nearest Neighbor Layer Set Optimization (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large language models (LLMs) have proven highly capable in handling downstream tasks, but the token-by-token generation in autoregressive decoding results in quadratic computational complexity. |
| Approach: | They propose a method that proposes skipping certain layers to construct a draft model, which eliminates the need for additional parameters or training. |
| Outcome: | The proposed method achieves 1.31.6 speedup in LLM inference while being sensitive to domain shifts. |
A Systematic Study of Cross-Layer KV Sharing for Efficient LLM Inference (2025.naacl-short)
Copied to clipboard
| Challenge: | Recent studies have shown that sharing key-value (KV) cache across layers is effective in efficient inference of large language models. |
| Approach: | They propose a unified framework that covers several recent methods and their novel variants to investigate cross-layer KV sharing. |
| Outcome: | The proposed framework achieves higher throughput and better performance when reducing the size of the key-value cache by 2 while maintaining competitive performance. |
Adaptive Layer Selection for Layer-Wise Token Pruning in LLM Inference (2026.findings-acl)
Copied to clipboard
| Challenge: | Large language models (LLMs) have demonstrated remarkable capabilities in processing long contexts. |
| Approach: | They propose a training-free method that adaptively chooses the selection layer for KV cache reduction . they exploit the variance of token ranks ordered by attention score to optimize decoding . |
| Outcome: | The proposed method outperforms state-of-the-art token pruning methods on InfiniteBench, RULER, and NIAH benchmarks. |
Attention Tracker: Detecting Prompt Injection Attacks in LLMs (2025.findings-naacl)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) are vulnerable to prompt injection attacks, where malicious inputs manipulate the model into ignoring original instructions and executing designated actions. |
| Approach: | They propose a training-free method that tracks attention patterns on instruction to detect prompt injection attacks without additional inference. |
| Outcome: | The proposed method shows an improvement of up to 10.0% over existing methods and performs well even on small LLMs. |
TABED: Test-Time Adaptive Ensemble Drafting for Robust Speculative Decoding in LVLMs (2026.findings-eacl)
Copied to clipboard
Minjae Lee, Wonjun Kang, Byeongkeun Ahn, Christian Classen, Kevin Galim, Seunghyuk Oh, Minghao Yan, Hyung Il Koo, Kangwook Lee
| Challenge: | Large Vision Language Models (LVLMs) are advanced models that process multiple modalities, such as images, audio, and video, alongside text. |
| Approach: | They propose to use a method to generate and verify draft tokens in parallel . they compare existing methods with small draft models and observe performance fluctuations . |
| Outcome: | The proposed method achieves an average walltime speedup of 1.74 over autoregressive decoding and a 5% improvement over single drafting methods. |
Chain-in-Tree: Back to Sequential Reasoning in LLM Tree Search (2026.findings-acl)
Copied to clipboard
| Challenge: | Large language models excel at tasks such as mathematical and commonsense reasoning, but their performance improves further when additional test-time compute is allocated. |
| Approach: | They propose a plug-in framework that decides when to branch during search instead of expanding at every step. |
| Outcome: | The proposed framework reduces token generation, model calls, and runtime by 75-85% on GSM8K and Math500, with negligible or no accuracy loss. |
OCP: Outlier-Centric Probing for Dynamic Structured Pruning of LLMs (2026.acl-long)
Copied to clipboard
| Challenge: | Existing structured pruning methods fail to identify outlier-triggering tokens and uniform layer-wise sparsity misaligns with heterogeneous outlier distributions. |
| Approach: | They propose a framework that prioritizes capturing outlier-triggering tokens rather than reconstructing full hidden distributions. |
| Outcome: | Experiments on LLaMA2, LLama3 and OPT show that the proposed framework outperforms state-of-the-art methods and achieves 25% perplexity reduction at 1.6 speedup. |
Self-attention-based Graph-of-Thought for Math Problem Solving (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing methods for generating reasoning paths in a chain structure are inefficient and non-human-like. |
| Approach: | They propose a decoding method for a chain-based LLM that constructs a thought graph simultaneously as an LLM inference and generates reasoning steps with a graph-structured self-attention mechanism. |
| Outcome: | The proposed method improves reasoning accuracy without huge computational over-expensive LLMs and avoids performance degradation issues when the LLM is too small to comprehend complex prompts. |
PROMTEC: Fast LLM Inference Decoding using Prompt Multi-Lookup with Template Database and Common Sequences (2025.findings-acl)
Copied to clipboard
| Challenge: | Prompt Multi-Lookup, Template Datastore, and Common Sequences methods are used to accelerate the inference of large language models (LLMs). |
| Approach: | They propose a multi-faceted approach to accelerate the inference of large language models by leveraging Prompt Multi-Lookup, Template Datastore, and Common Sequences methods. |
| Outcome: | The proposed method achieves a 3.91 speedup on the miniF2F benchmark and 4.23 on the HumanEval benchmark. |
EOP-LLM: Energy Oriented Pruning for Large Language Models (2026.findings-acl)
Copied to clipboard
| Challenge: | Inference energy consumption has grown rapidly in large language models (LLMs) but existing methods focus on reducing FLOPs or latency rather than modeling or enforcing end-to-end inference energy constraints. |
| Approach: | They propose an energy-oriented dynamic pruning framework that enables LLM inference under explicit per-sequence energy budgets. |
| Outcome: | EOP-LLM outperforms state-of-the-art dynamic pruning baselines while adhering to per-sequence energy constraints. |
An Empirical Study on Cross-lingual Vocabulary Adaptation for Efficient Language Model Inference (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Cross-lingual vocabulary adaptation (CVA) methods have been proposed for adapting models to a target language . but effectiveness of these methods on increasing inference efficiency of generative large language models has not been explored. |
| Approach: | They propose to use cross-lingual vocabulary adaptation methods to adapt models to a target language to improve downstream performance. |
| Outcome: | The proposed methods significantly speed up models in four languages and four natural language understanding tasks. |
Unlocking Efficiency in Large Language Model Inference: A Comprehensive Survey of Speculative Decoding (2024.findings-acl)
Copied to clipboard
Heming Xia, Zhe Yang, Qingxiu Dong, Peiyi Wang, Yongqi Li, Tao Ge, Tianyu Liu, Wenjie Li, Zhifang Sui
| Challenge: | Large Language Models (LLMs) have a high inference latency stemming from autoregressive decoding. |
| Approach: | They propose a novel decoding paradigm that drafts multiple tokens and verifies them in parallel . they aim to provide a catalyst for further research on Speculative Decoding . |
| Outcome: | The proposed method drafts multiple tokens and verifies them in parallel . it can be used to accelerate inference in large language models. |
Towards Sustainable NLP: Insights from Benchmarking Inference Energy in Large Language Models (2025.naacl-long)
Copied to clipboard
| Challenge: | Large language models (LLMs) are recognized for their exceptional generative capabilities and versatility across various tasks. |
| Approach: | They conduct a comprehensive benchmarking of LLM inference energy across a wide range of NLP tasks to determine the impact of different models, tasks, prompts, and system-related factors on inference. |
| Outcome: | The proposed model energy benchmarks show that quantization and optimal batch sizes can significantly reduce energy usage. |
SPECTRA: Faster Large Language Model Inference with Optimized Internal and External Speculation (2025.acl-long)
Copied to clipboard
| Challenge: | Existing approaches to inference with Large Language Models (LLMs) are expensive and time-consuming. |
| Approach: | They propose a framework for accelerating large language model inference without additional training or modification to the original LLM. |
| Outcome: | The proposed framework outperforms state-of-the-art methods and achieves 4.08x speedups across benchmarks and LLM architectures. |
Fast Matrix Multiplications for Lookup Table-Quantized LLMs (2024.findings-emnlp)
Copied to clipboard
| Challenge: | generative (tokenby-token) inference is memory-bound and requires a large amount of memory to perform. |
| Approach: | They propose a lookup table engine for weight-quantized large language models that uses offline restructuring of the quantized weight matrix to minimize bit manipulations associated with unpacking. |
| Outcome: | The proposed kernel can be 2-4x faster than existing GEMM kernels while achieving performance gains of 1.5 to 2 times. |
GLiM: Integrating Graph Transformer and LLM for Document-Level Biomedical Relation Extraction with Incomplete Labeling (2025.findings-acl)
Copied to clipboard
Hao Fang, Yuejie Zhang, Rui Feng, Yingwen Wang, Qing Wang, Wen He, Xiaobo Zhang, Tao Zhang, Shang Gao
| Challenge: | Document-level relation extraction (DocRE) solves problems of document quality . number of entities and entity-pair relations increases, causing incomplete annotations . |
| Approach: | a framework that reduces the problem space using a graph-enhanced Transformer-based model is proposed . GLiM leverages large language models for reasoning to reduce the problem-space . |
| Outcome: | GLiM boosts average recall and F1 scores on biomedical datasets . compared with existing models, GLim outperforms existing models on biomedicine benchmarks compared to existing models . |
LAVa: Layer-wise KV Cache Eviction with Dynamic Budget Allocation (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Existing methods for cache compression are heuristic and lack dynamic budget allocation . cnn's john mccartney and johnny mccain present a new approach for cache eviction and dynamic budgets . |
| Approach: | They propose a unified framework for cache compression that minimizes information loss in transformer residual streams. |
| Outcome: | The proposed method consistently maintains top performance across task types. |
FLRC: Fine-grained Low-Rank Compressor for Efficient LLM Inference (2025.emnlp-main)
Copied to clipboard
| Challenge: | Low-rank compression can reduce memory usage and computational demand, but results are poor during decoding. |
| Approach: | They propose a fine-grained low-rank compression algorithm that determines optimal rank allocation for each layer and incorporates progressive low-ranked decoding to maintain text generation quality. |
| Outcome: | The proposed approach outperforms state-of-the-art methods on summarization tasks and on understanding tasks. |
MiniKV: Pushing the Limits of 2-Bit KV Cache via Compression and System Co-Design for Efficient Long Context Inference (2025.findings-acl)
Copied to clipboard
| Challenge: | State-of-the-art 2-bit KV cache quantization methods achieve excellent results in accelerating LLM inference while retaining accuracy on long context tasks. |
| Approach: | They propose a method based on 2-bit KV cache quantization with adaptive KV policies that retain LLM accuracy with only a subset of KV states. |
| Outcome: | The proposed method outperforms state-of-the-art methods on a wide range of long context tasks while retaining accuracy. |
Speculative Streaming: Efficient and Scalable Speculative Decoding with Multi-Stream Attention (2025.emnlp-main)
Copied to clipboard
Nikhil Bhendawade, Irina Belousova, Qichen Fu, Henry Mason, Antonie Lin, Mohammad Rastegari, Mahyar Najibi
| Challenge: | Speculative decoding is a prominent technique for accelerating LLM inference by leveraging an auxiliary draft model, but its effectiveness is limited by the autoregressive nature of draft generation. |
| Approach: | They propose a method that integrates speculative draft generation directly within the target model using multi-stream attention. |
| Outcome: | The proposed method improves acceptance but also latency and speculation latency, limiting overall speedup. |
CHESS: Optimizing LLM Inference via Channel-Wise Thresholding and Selective Sparsification (2024.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods for activation sparsification do not capture the relationship between activation and model performance. |
| Approach: | They propose a general activation sparsification approach using channel-wise thresholding and selective sparsifying to capture the relationship between activation and model performance. |
| Outcome: | The proposed approach reduces the number of activated neurons during inference by 1.27x over eight downstream tasks while activating fewer parameters than existing methods. |
HATA: Trainable and Hardware-Efficient Hash-Aware Top-k Attention for Scalable Large Model Inference (2025.findings-acl)
Copied to clipboard
Ping Gong, Jiawei Yi, Shengnan Wang, Juncheng Zhang, Zewen Jin, Ouxiang Zhou, Ruibo Liu, Guanbin Xu, Youhui Bai, Bowen Ye, Kun Yuan, Tong Yang, Gong Zhang, Renhai Chen, Feng Wu, Cheng Li
| Challenge: | Existing top-k attention methods struggle to strike a balance between efficiency and accuracy. |
| Approach: | They propose a top-k attention approach that integrates low-overhead techniques into the Top-k Attention process to achieve 7.2 speedup compared to vanilla full attention. |
| Outcome: | The proposed approach achieves 7.2 speedup compared to current top-k attention methods while maintaining model accuracy. |
SwiftKV: Fast Prefill-Optimized Inference with Knowledge-Preserving Model Transformation (2025.emnlp-main)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) are an integral enabler of enterprise applications such as summarization, retrieval augmented generation, and agentic workflows. |
| Approach: | They propose a model transformation and distillation procedure that prefills later layers’ KV cache using an earlier layer’s output, allowing prompt tokens to skip those later layers. |
| Outcome: | The proposed procedure can reduce prefill computation by 25-50% across several LLM families while incurring minimum quality degradation. |
Injecting Context via Situation Working Memory for Logical Reasoning with LLMs (2026.acl-long)
Copied to clipboard
| Challenge: | Recent advances in large language models have improved logical reasoning by injecting formal logic or explicit structured representations. |
| Approach: | a cognitively inspired method is proposed to help LLMs construct a mental representation of events . SituW builds a situation representation by decomposing text along these five dimensions . it also guides LLM inference with this evolving state . |
| Outcome: | a cognitively inspired method improves accuracy and predictability in large language models . SituW builds a mental representation by decomposing text along these dimensions . |
HeteroCache: A Dynamic Retrieval Approach to Heterogeneous KV Cache Compression for Long-Context LLM Inference (2026.acl-long)
Copied to clipboard
| Challenge: | Existing static compression methods suffer from coarse-grained caching and high I/O overhead. |
| Approach: | They propose a training-free dynamic compression framework that uses a sparse attention mechanism to categorize attention heads based on stability and similarity. |
| Outcome: | The proposed framework achieves state-of-the-art performance on long-context benchmarks and accelerates decoding by up to 3 compared to the original model with a 224K context. |