Papers with inference throughput

17 papers
ArcLight: A Lightweight LLM Inference Architecture for Many-Core CPUs (2026.acl-demo)

Copied to clipboard

Challenge: Existing frameworks for large language model (LLM) inference on CPUs overlook overhead of cross-NUMA memory access.
Approach: They propose a lightweight LLM inference architecture designed from the ground up for many-core CPUs.
Outcome: Experimental results show that ArcLight surpasses the performance ceiling of mainstream frameworks, achieving up to 46% higher inference throughput.
DVAGen: Dynamic Vocabulary Augmented Generation (2025.emnlp-demos)

Copied to clipboard

Challenge: Existing dynamic vocabulary approaches struggle to generalize to novel or out-of-vocabulary words, limiting their flexibility in handling diverse token combinations.
Approach: They propose an open-source framework for training, evaluation, and visualization of dynamic vocabulary-augmented language models.
Outcome: The proposed framework validates the effectiveness of dynamic vocabulary-augmented language models on modern LLMs and shows support for batch inference significantly improving inference throughput.
Small Agents, Big Gains: Journey-Aware and Critic-Guided Simulation for Long-Horizon Shopping Dialogues (2026.acl-industry)

Copied to clipboard

Challenge: e-commerce assistants must support inspiration, comparison, and tool-grounded fact-checking . lack of data-coverage and verification problem hampers efficient, deployable models . eaa: "training trajectories must cover diverse user workflows with high fidelity"
Approach: They propose a system that synthesizes diverse, faithful, and policy-aligned shopping trajectories . a small model can significantly outperform same-size baselines and surpass a large-model baseline .
Outcome: The proposed model outperforms existing models and surpasses large models with 8 higher inference throughput.
SWAN: An Efficient and Scalable Approach for Long-Context Language Modeling (2025.emnlp-main)

Copied to clipboard

Challenge: Existing decoder-only models struggle with context lengths beyond their training distribution.
Approach: They propose a causal Transformer architecture that generalizes robustly to sequence lengths longer than seen during training.
Outcome: The proposed decoder-only architecture can generalize robustly to longer contexts . it is more computationally efficient than the standard Transformer architecture, the authors say .
ECHO-LLaMA: Efficient Caching for High-Performance LLaMA Training (2025.emnlp-industry)

Copied to clipboard

Challenge: ECHO-LLaMA transforms LLa MA models into shared KV caching across certain layers, significantly reducing KV computational complexity while maintaining or improving language performance.
Approach: They propose an efficient LLaMA architecture that transforms LLama models into shared KV caching across certain layers, reducing computational complexity while maintaining or improving language performance.
Outcome: ECHO-LLaMA achieves up to 77% higher token-per-second throughput during training, up to 16% higher Model FLOPs Utilization (MFU) and up to 14% lower loss when trained on an equal number of tokens.
VPTQ: Extreme Low-bit Vector Post-Training Quantization for Large Language Models (2024.emnlp-main)

Copied to clipboard

Challenge: Recent research has focused on pushing weight-only quantization to extremely low-bit due to numerical representation limitations.
Approach: They propose a vector-based quantization approach that pushes LLMs to extremely low-bit . they propose scalar-based weight quantization that reduces memory requirements and optimizes storage costs .
Outcome: The proposed method reduces model quantization perplexity by 0.01-0.34 on LLaMA-2, 0.38-0.68 on mistral-7B, 4.41-7.34, on llaMA-3 on QA tasks on average.
Align Attention Heads Before Merging Them: An Effective Way for Converting MHA to GQA (2025.findings-emnlp)

Copied to clipboard

Challenge: Large language models (LLMs) have demonstrated exceptional performance across diverse natural language processing tasks.
Approach: They propose a method for converting multi-head attention into grouped-query attention with any compression ratio of KV heads.
Outcome: The proposed method can compress up to 87.5% KV heads of LLaMA2-7B model and 75% Kv heads of Sheared-LLa MA-1.3B with acceptable performance degradation.
Improving the Throughput of Diffusion-based Large Language Models via a Training-Free Confidence-Aware Calibration (2026.findings-acl)

Copied to clipboard

Challenge: CadLLM is a plug-and-play model-agnostic with KV caching based dLLMs.
Approach: They propose a lightweight adaptive method that can control the generation block size, step size, and threshold based on the average confidence score of unmasked tokens.
Outcome: The proposed method can increase throughput by up to 1.1-2.28x over the state-of-the-art model with competitive accuracy.
EEE-QA: Exploring Effective and Efficient Question-Answer Representations (2024.lrec-main)

Copied to clipboard

Challenge: Current approaches to question answering rely on pre-trained language models like RoBERTa.
Approach: They propose a pooling approach that embeds all answer candidates with the question . they also propose enabling cross-reference between answer choices .
Outcome: The proposed methods improve throughput and memory efficiency with little sacrifice in performance.
HeteroSpec: Leveraging Contextual Heterogeneity for Efficient Speculative Decoding (2026.acl-long)

Copied to clipboard

Challenge: Autoregressive decoding limits the inference throughput of Large Language Models due to its sequential dependency.
Approach: They propose a framework that allocates verification effort in proportion to candidate uncertainty.
Outcome: Speculative decoding achieves an average speedup over state-of-the-art methods . a small subset of high-confidence predictions accounts for most successful verifications .
Layer-Condensed KV Cache for Efficient Inference of Large Language Models (2024.acl-long)

Copied to clipboard

Challenge: Using a key-value cache, memory consumption is a bottleneck for high-throughput language models.
Approach: They propose a method that only computes and caches the KVs of a small number of layers, thus saving memory consumption and improving inference throughput.
Outcome: The proposed method achieves higher throughput and competitive performance than standard transformers and is orthogonal to existing transformer memory-saving techniques.
PuMer: Pruning and Merging Tokens for Efficient Vision Language Models (2023.acl-long)

Copied to clipboard

Challenge: Large-scale vision language models use Transformers to perform cross-modal interactions . state-of-the-art models are memory intensive and expensive due to quadratic complexity .
Approach: They propose a token reduction framework that uses text-informed Pruning and modality-aware Merging strategies to progressively reduce the tokens of input image and text.
Outcome: The proposed framework improves inference speed and memory footprint on four vision language tasks.
SpecCache: Speculative KV Cache Reuse for Efficient RAG Serving (2026.acl-long)

Copied to clipboard

Challenge: Retrieval-Augmented Generation (RAG) improves LLMs but faces high prefill latency during long contexts.
Approach: They propose a method that uses deep-layer hidden-state norms to guide token selection . they propose to use deep-layered hidden-status norms as a proxy to guide the token selection.
Outcome: The proposed SpecCache outperforms state-of-the-art (SOTA) benchmarks.
On Efficient Language and Vision Assistants for Visually-Situated Natural Language Understanding: What Matters in Reading and Reasoning (2024.emnlp-main)

Copied to clipboard

Challenge: Recent advances in language and vision assistants have showcased impressive capabilities but suffer from a lack of transparency, limiting broader research and reproducibility.
Approach: They propose to redefine the design of vision-language models by identifying key components and creating efficient models with constrained inference costs.
Outcome: The proposed models achieve significant improvements in inference throughput while maintaining high performance.
You Only Use Reactive Attention Slice When Retrieving From Long Context (2025.findings-emnlp)

Copied to clipboard

Challenge: Existing retrieval techniques for language models are limited due to their reliance on lexical similarity and are computationally expensive to train.
Approach: They propose a training-free and fine-tuning-free attention-based retrieval technique that uses a reaction score heuristic to quantify how an LM’s self-attention “reacts” to a user query.
Outcome: The proposed approach improves QA task accuracy by 15% and inference throughput by 31% compared to embedding-based retrieval.
LoRA on the Go: Instance-level Dynamic LoRA Selection and Merging (2026.acl-long)

Copied to clipboard

Challenge: Low-Rank Adaptation (LoRA) is a parameter-efficient approach for fine-tuning large language models.
Approach: They propose a low-rank Adaptation framework that automatically selects and merges LoRA adapters at the instance level without additional training.
Outcome: The proposed framework outperforms training-based baselines on some tasks upto a margin of 3.6% while remaining competitive on other tasks and maintaining inference throughput.
MUTANT: A Recipe for Multilingual Tokenizer Design (2026.acl-long)

Copied to clipboard

Challenge: Subword tokenization schemes such as Byte Pair Encoding (BPE) are widely adopted, but their effectiveness in multilingual settings remains understudied.
Approach: They propose a multilingual tokenizer that produces linguistically coherent tokens for multilingual LLMs.
Outcome: The proposed tokenizer improves fertility score by 39.5% over LLaMA4 and 18% over Sutra.

What is GenGO?

GenGO is an NLP powered publication search system. It currenctly indexes 30k+ papers from ACL Anthology, and implements multi-aspect summarization, semantic search, and more!

Information

About
Limitations