Position-Aware Depth Decay Decoding (D3): Boosting Large Language Model Inference Efficiency (2025.findings-acl)
Copied to clipboard
| Challenge: | Recent dynamic computation methods show that not all components are required for inference, enabling a training-free pipeline. |
| Approach: | They propose a token-position aware layer skipping framework to save 1.5x times operations efficiently while maintaining performance. |
| Outcome: | The proposed algorithm achieves 1.5x speedup on large language models with no retraining and with comparable performance on the GSM8K and BBH benchmarks. |
Similar Papers
Balcony: A Lightweight Approach to Dynamic Inference of Generative Language Models (2025.emnlp-main)
Copied to clipboard
Benyamin Jamialahmadi, Parsa Kavehzadeh, Mehdi Rezagholizadeh, Parsa Farinneya, Hossein Rajabzadeh, Aref Jafari, Boxing Chen, Marzieh S. Tahaei
| Challenge: | Existing methods for dynamic inference are limited by hardware inefficiencies or performance degradation. |
| Approach: | They propose a framework for depth-based dynamic inference that freezes the pre-trained model and inserts additional transformer layers at selected exit points. |
| Outcome: | The proposed framework outperforms state-of-the-art methods such as Flextron and Layerskip on multiple models at various scales, as well as other leading compression techniques across a variety of benchmarks. |
Turning Trash into Treasure: Accelerating Inference of Large Language Models with Token Recycling (2025.acl-long)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) generate only one token at each decoding step, leading to high latency. |
| Approach: | They propose a speculative decoding paradigm that stores tokens in an adjacency matrix and employs a breadth-first-search algorithm to construct a draft tree. |
| Outcome: | The proposed method outperforms existing train-free methods by 30% and even a training method by 25%. |
Hardware-Aware Parallel Prompt Decoding for Memory-Efficient Acceleration of LLM Inference (2025.findings-emnlp)
Copied to clipboard
Hao Mark Chen, Wayne Luk, Yiu Ka Fai Cedric, Rui Li, Konstantin Mishchenko, Stylianos Venieris, Hongxiang Fan
| Challenge: | Auto-regressive decoding of Large Language Models results in significant overheads in hardware performance . a novel parallel prompt decoding approach is proposed to overcome these limitations . |
| Approach: | They propose a parallel prompt decoding that uses a single model for speculation and verification. |
| Outcome: | The proposed approach speeds up auto-regressive decoding of large language models 2.49 times . it can be used on mobileLlama to Vicuna-13B on a wide range of benchmarks . |
Exploring the Hidden Capacity of LLMs for One-Step Text Generation (2025.emnlp-main)
Copied to clipboard
| Challenge: | Large language models (LLMs) can reconstruct surprisingly long texts via autoregressive generation from just one trained input embedding. |
| Approach: | They show that large language models can reconstruct surprisingly long texts via autoregressive generation from just one trained input embedding. |
| Outcome: | The proposed model can generate hundreds of accurate tokens in one token-parallel forward pass, when provided with only two learned embeddings. |
Draft
& Verify: Lossless Large Language Model Acceleration via Self-Speculative Decoding (2024.acl-long)
Copied to clipboard
| Challenge: | Existing methods for accelerating Large Language Models have been criticized for their inference costs and inefficient decoding. |
| Approach: | They propose a self-speculative decoding approach for accelerating Large Language Models without an auxiliary model. |
| Outcome: | The proposed method achieves a speedup of up to 1.99 with no additional neural network training and no extra memory footprint. |
River-LLM: Large Language Model Seamless Exit Based on KV Share (2026.acl-long)
Copied to clipboard
| Challenge: | Existing methods to reduce latency and speed up early exits are costly and impose significant cost and energy consumption. |
| Approach: | They propose a lightweight KV-Shared Exit River framework that allows the backbone’s missing KV cache to be naturally generated and preserved during the exit process. |
| Outcome: | The proposed framework achieves 1.71 to 2.16 speedup while maintaining high generation quality. |
A Drop-In Solution for On-the-Fly Adaptation of Speculative Decoding in Large Language Models (2025.acl-long)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) are highly memory-intensive when performing real-time inference. |
| Approach: | They propose a technique that allows for speculative decoding to be run on the fly to maximize the efficiency of LLM inferences. |
| Outcome: | The proposed solution can lead to 3.55-16.48% speed improvement over the standard speculative decoding, and 1.2-3.4 over the default LLMs. |
LM-Infinite: Zero-Shot Extreme Length Generalization for Large Language Models (2024.naacl-long)
Copied to clipboard
| Challenge: | Currently, large language models (LLMs) train on short text segments due to the computational overhead quadratic in the input lengths of their Transformer architectures. |
| Approach: | They propose a method that allows LLMs pre-trained with 2K or 4K-long segments to generalize to up to 200M length inputs while retaining perplexity. |
| Outcome: | The proposed method achieves 2.7 decoding speed up and 7.5 memory saving over the original model. |
UniSpec: Training-Free Speculative Decoding for Robust LLM Acceleration Across Languages and Hardware (2026.acl-long)
Copied to clipboard
| Challenge: | Existing methods for speculative decoding ignore device-specific verification costs and lack of mechanisms to assess draft token quality. |
| Approach: | They propose a training-free, lossless speculative decoding framework that enables robust, plug-and-play LLM acceleration across diverse hardware configurations and languages. |
| Outcome: | The proposed framework outperforms existing training-free methods while maintaining identical output quality across different hardware environments. |
LayerSkip: Enabling Early Exit Inference and Self-Speculative Decoding (2024.acl-long)
Copied to clipboard
Mostafa Elhoushi, Akshat Shrivastava, Diana Liskovich, Basil Hosmer, Bram Wasti, Liangzhen Lai, Anas Mahmoud, Bilge Acun, Saurabh Agarwal, Ahmed Roman, Ahmed Aly, Beidi Chen, Carole-Jean Wu
| Challenge: | Large Language Models (LLMs) have been deployed to many applications, yet their high compute and memory requirements lead to high financial and energy costs when deployed to GPU servers. |
| Approach: | They propose an end-to-end solution to speed-up inference of large language models . they apply layer dropout, and show that it increases the accuracy of early exit at earlier layers without adding any auxiliary layers or modules to the model. |
| Outcome: | The proposed method shows speedups of up to 2.16x on summarization for CNN/DM documents, 1.82x on coding, and 2.0x on TOPv2 semantic parsing task. |