Papers with accuracy loss
LLMC: Benchmarking Large Language Model Quantization with a Versatile Compression Toolkit (2024.emnlp-industry)
Copied to clipboard
Ruihao Gong, Yang Yong, Shiqiao Gu, Yushi Huang, Chengtao Lv, Yunchen Zhang, Dacheng Tao, Xianglong Liu
| Challenge: | Existing quantization techniques have been categorized as 'simple' and 'highly efficient' however, their configurations vary from each other and cannot be fairly compared . |
| Approach: | They propose a plug-and-play compression toolkit to explore the impact of quantization. |
| Outcome: | The proposed toolkit explores the impact of quantization on large language models. |
FastSeq: Make Sequence Generation Faster (2021.acl-demo)
Copied to clipboard
Yu Yan, Fei Hu, Jiusheng Chen, Nikhil Bhendawade, Ting Ye, Yeyun Gong, Nan Duan, Desheng Cui, Bingyu Chi, Ruofei Zhang
| Challenge: | Transformer-based models have made tremendous impact in natural language generation, but inference speed is still a bottleneck due to large model size and intensive computing involved in auto-regressive decoding process. |
| Approach: | They propose an attention cache optimization, an efficient algorithm for detecting repeated n-grams, and an asynchronous generation pipeline with parallel I/O to accelerate sequence generation without loss of accuracy. |
| Outcome: | The proposed framework can accelerate the sequence generation by 4x to 9x with a simple one-line code change for a set of widely used and diverse models. |
Zero-Shot Dynamic Quantization for Transformer Inference (2022.emnlp-industry)
Copied to clipboard
| Challenge: | Existing methods for quantizing models require calibration or modification of parameters . run-time inference of such large models is costly due to large computational requirements . |
| Approach: | They propose a run-time method for quantizing BERT-like models to 8-bit integers . they demonstrate that the method can be used on many NLP tasks without calibration steps . |
| Outcome: | The proposed method reduces the accuracy loss associated with quantizing BERT-like models to 8-bit integers. |
RoBSA: RoPE-based Blockwise Sparse Multi-head Latent Attention (2026.acl-long)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) have advanced in recent years, scaling up in both parameter count and context length. |
| Approach: | They propose a method to compute attention over a subset of context tokens and to implement token selection in a blockwise manner. |
| Outcome: | The proposed method reduces end-to-end inference latency by up to 2.55x with minimal accuracy loss compared to full attention in long-context scenarios for very large models. |
L4Q: Parameter Efficient Quantization-Aware Fine-Tuning on Large Language Models (2025.acl-long)
Copied to clipboard
| Challenge: | Quantization-aware PEFT methods have been developed to reduce memory and computational costs associated with large language models. |
| Approach: | They propose a method that integrates Quantization-Aware Training (QAT) with LoRA to reduce memory overhead and improve model accuracy. |
| Outcome: | The proposed method significantly reduces QAT’s memory overhead while preserving the advantage of QAT in producing fully quantized LLMs with high accuracy. |
IMPACT: Importance-Aware Activation Space Reconstruction (2026.acl-long)
Copied to clipboard
| Challenge: | Large language models (LLMs) achieve strong performance across domains but remain difficult to deploy in resource-constrained environments due to their massive size. |
| Approach: | They propose an importance-aware activation reconstruction framework that links compression to its effect on model performance. |
| Outcome: | Experiments show that IMPACT reduces model size by 55.4% while maintaining accuracy comparable to or better than state-of-the-art models. |
Can We Steer Reasoning Direction by Thinking Intervention? (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Large Reason Models suffer from overthinking and erroneous reasoning problems due to the lack of fine-grained control over their reasoning behaviors. |
| Approach: | They propose a paradigm to enable fine-grained control over LRMs’ reasoning behaviors by aligning reasoning trajectories with specific cognitive patterns. |
| Outcome: | The proposed paradigm achieves integration intervention throughout model reasoning processes. |
When to Continue Thinking: Adaptive Thinking Mode Switching for Efficient Reasoning (2025.findings-emnlp)
Copied to clipboard
Xiaoyun Zhang, Jingqing Ruan, Xing Ma, Yawen Zhu, Haodong Zhao, Hao Li, Jiansong Chen, Ke Zeng, Xunliang Cai
| Challenge: | Large reasoning models (LRMs) incur excessive computational overhead due to redundant reasoning, especially on simple tasks. |
| Approach: | They propose an Adaptive Self-Recovery Reasoning framework that suppresses unnecessary reasoning and enables implicit recovery. |
| Outcome: | The proposed framework suppresses unnecessary reasoning and enables implicit recovery. |
Towards Realistic Single-Task Continuous Learning Research for NER (2021.findings-emnlp)
Copied to clipboard
Justin Payan, Yuval Merhav, He Xie, Satyapriya Krishna, Anil Ramakrishna, Mukund Sridhar, Rahul Gupta
| Challenge: | Academic datasets are often static and contain data that is annotated all at once based on fixed annotation guidelines. |
| Approach: | They propose to build a single-task continuous learning dataset from an existing dataset and release it along with the code to the research community. |
| Outcome: | The proposed model is based on an existing dataset and released to the research community. |
Massive End-to-end Speech Recognition Models with Time Reduction (2024.naacl-long)
Copied to clipboard
Weiran Wang, Rohit Prabhavalkar, Haozhe Shan, Zhong Meng, Dongseong Hwang, Qiujia Li, Khe Chai Sim, Bo Li, James Qin, Xingyu Cai, Adam Stooke, Chengjian Zheng, Yanzhang He, Tara Sainath, Pedro Moreno Mengibar
| Challenge: | Using the neural architecture of Google’s universal speech model, we reduce the frame rate and speed up training and inference. |
| Approach: | They propose to use the neural architecture of Google’s universal speech model with additional funnel pooling layers to significantly reduce the frame rate and speed up training and inference. |
| Outcome: | The proposed methods work with both connectionist temporal classification (CTC) and RNN-Transducer (RNN-T) and over two domains. |
Offloaded Reasoning: Efficient Inference for Large Language Models via Modular Reasoning and Refinement (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Large language models (LLMs) demonstrate strong reasoning capabilities but are expensive to run at inference time, limiting their practical deployment. |
| Approach: | They propose Offloaded Reasoning, a modular strategy where a lightweight model generates intermediate reasoning traces that are then used by a larger model to produce the final answer. |
| Outcome: | The proposed approach achieves faster inferences than full large-model reasoning with minimal accuracy loss while recovering or exceeding full accuracy at substantially lower cost. |
Pathologies of Neural Models Make Interpretations Difficult (D18-1)
Copied to clipboard
| Challenge: | Existing methods for NLP use input reduction to determine a word's importance . human accuracy degrades when shown the reduced examples instead of the original . |
| Approach: | They propose a process that iteratively removes the least important word from an input . they show human models make the same predictions with high confidence . |
| Outcome: | The proposed methods expose pathological behaviors of neural models . human experiments show that reduced examples lack information to support the prediction of any label . |
BadFair: Backdoored Fairness Attacks with Group-conditioned Triggers (2024.findings-emnlp)
Copied to clipboard
| Challenge: | *BadFair* is a backdoored fairness attack methodology that discriminates and produces incorrect results for specific groups. |
| Approach: | They propose a backdoored fairness attack methodology that discriminates and produces incorrect results for specific groups. |
| Outcome: | The proposed model achieves 85% success rate in attacks aimed at target groups on average while only incurring a minimal accuracy loss. |
FlattenQuant: Breaking through the Inference Compute-bound for Large Language Models with Per-tensor Quantization (2024.lrec-main)
Copied to clipboard
| Challenge: | Large language models (LLMs) have demonstrated state-of-the-art accuracies across tasks, but their latency and GPU memory consumption limit their performance. |
| Approach: | They propose a method which flattens the tensor to achieve low bit per-tensori quantization with minimal accuracy loss. |
| Outcome: | The proposed method achieves low bit per-tensor quantization with minimal accuracy loss. |
ClusterAttn: KV Cache Compression under Intrinsic Attention Clustering (2025.acl-long)
Copied to clipboard
| Challenge: | Existing methods for sparse attention apply the same pattern across different attention heads and inputs, but fail to capture the intrinsic attention clustering in large language models. |
| Approach: | They propose a training-free sparse attention method that provides an efficient prompt cache compression scheme under intrinsic attention clustering for efficient LLM inference. |
| Outcome: | The proposed method reduces memory usage by 10%–65% and increases throughput by 2.6–4.8 times with no accuracy loss. |
PruMUX: Augmenting Data Multiplexing with Model Compression (2023.findings-acl)
Copied to clipboard
| Challenge: | Prior work has investigated methods like model pruning, knowledge distillation, and data multiplexing to increase model throughput without sacrificing accuracy. |
| Approach: | They propose to combine structured pruning and data multiplexing methods to increase model throughput without sacrificing accuracy. |
| Outcome: | The proposed method achieves 7.5-29.5X throughput improvement over a BERT-base model with accuracy threshold from 80% to 74%. |
FAQ: Mitigating Quantization Error via Regenerating Calibration Data with Family-Aware Quantization (2026.findings-acl)
Copied to clipboard
| Challenge: | representativeness and universality of calibration data remain a bottleneck in quantization accuracy. |
| Approach: | They propose a framework that leverages prior knowledge from LLMs to generate calibration samples . their framework reduces accuracy loss by up to 28.5% compared to baseline . |
| Outcome: | Experiments show that family-aware quantization reduces accuracy loss by up to 28.5% compared to baseline data. |
Squeezed Attention: Accelerating Long Context Length LLM Inference (2025.acl-long)
Copied to clipboard
Coleman Richard Charles Hooper, Sehoon Kim, Hiva Mohammadzadeh, Monishwaran Maheswaran, Sebastian Zhao, June Paik, Michael W. Mahoney, Kurt Keutzer, Amir Gholami
| Challenge: | Emerging Large Language Models require long input context to perform complex tasks. |
| Approach: | They propose an algorithm to reduce the complexity of attention with respect to the fixed context length. |
| Outcome: | The proposed method reduces the complexity of attention from linear to logarithmic with respect to the fixed context length. |
EdgeFormer: Latency-Aware Collaborative Multi-Head Attention of Transformer Inference in Edge Networks (2026.acl-long)
Copied to clipboard
| Challenge: | Existing methods for inference in centralized cloud pose privacy risks due to sensitive data. |
| Approach: | They propose a latency-aware framework for distributed Transformer inference in resource-constrained edge networks. |
| Outcome: | The proposed framework achieves 2.01 times inference acceleration over state-of-the-art baselines with leq1.06% accuracy loss, maintaining robustness under varying edge conditions. |