China

Rapid advances in Large Language Models (LLMs) have spurred demand for processing extended context sequences in contemporary applications. However, this progress faces two challenges: performance degradation due to sequence lengths out-of-distribution, and excessively long inference times caused by the quadratic computational complexity of attention. These issues limit LLMs in long-context scenarios. In this paper, we propose Dynamic Token-Level KV Cache Selection (*TokenSelect*), a training-free method for efficient and accurate long-context inference. *TokenSelect* builds upon the observation of non-contiguous attention sparsity, using QK dot products to measure per-head KV Cache criticality at token-level. By per-head soft voting mechanism, *TokenSelect* selectively involves a few critical KV cache tokens in attention calculation without sacrificing accuracy. To further accelerate *TokenSelect*, we design the Selection Cache based on observations of consecutive Query similarity and implemented the efficient Paged Dot Product Kernel, significantly reducing the selection overhead. A comprehensive evaluation of *TokenSelect* demonstrates up to 23.84times speedup in attention computation and up to 2.28times acceleration in end-to-end latency, while providing superior performance compared to state-of-the-art long-context inference methods.

EMNLP 2025

TokenSelect: Efficient Long-Context Inference and Length Extrapolation for LLMs via Dynamic Token-Level KV Cache Selection

long-context inference

efficient inference

large language models

### Welcome to The 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025)

Message from the General Chair: 
*It is my great pleasure and honor to welcome you to the 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025), held in beautiful Vienna, Austria, from July 27 to August 1, 2025. ACL2025continues our field’s tradition of excellence in scholarship, innovation, and inclusivity, and I am deeply grateful to the many volunteers who have worked tirelessly to bring this event to life.* 
[Read more](https://drive.google.com/file/d/1GI_hvOpjswAuYdUTromfeDiPpCcqidwg/view?usp=sharing)

To access this event page, you need to log in with the **email address you registered with**. Access credentials will be sent to your email from Underline - subject line "Welcome to ACL 2025". Please be sure to check your spam email folder if you do not see an email confirmation right away.

Please log in

To access this event page, you need to log in with the **email address you registered with**. 

ACL 2025

Welcome to The 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025)

poster

## Welcome!
"I am excited to welcome you to this year’s edition of the Conference on Empirical Methods in Natural Language Processing! Importantly, it marks the 30th edition of EMNLP. With over 8,000 submissions, more than 3,000 accepted papers, and thousands of attendees, we have come a long way from that first
workshop, which had 14 accepted papers. As the field looks ahead, Suzhou is the fitting location for celebrating this milestone: rooted in a long literary tradition, yet modern and forward-looking, and home to a large share of the NLP community." 

*Message from the General Chair, Dirk Hovy*

Celebrate 30 Years of EMNLP! 
EMNLP 2025 will be held in Suzhou, China from November 5th to November 9th, 2025.

Large language models (LLMs) are increasingly deployed with task-specific adapters catering to multiple downstream applications. In such a scenario, the additional compute associated with these apparently insignificant number of adapter parameters (typically less than 1% of the base model) turns out to be disproportionately significant during inference time (up to 2.5x times that of the base model). In this paper, we propose a new zero-latency fused low-rank adapter (zFLoRA) that introduces zero or negligible latency overhead on top of the base model. Experimental results on LLMs of size 1B, 3B and 7B show that zFLoRA compares favorably against the popular supervised fine-tuning benchmarks including low-rank adapters (LoRA) as well as full fine-tuning (FFT). Experiments are conducted on 18 different tasks across three different categories namely commonsense reasoning, math reasoning and summary-dialogue. Latency measurements made on NPU (Samsung Galaxy S25+) as well as GPU (NVIDIA H100) platforms show that the proposed zFLoRA adapters introduce zero to negligible latency overhead.

zFLoRA: Zero-Latency Fused Low-Rank Adapters

Efficient Key-Value (KV) cache management is essential for processing long text sequences in large language models (LLMs), where memory constraints often limit performance. Conventional KV eviction strategies, such as top-k selection based on attention scores, depend on static heuristics that fail to capture the evolving implicit dependencies among tokens during inference. To overcome this, we propose GraphKV, a graph-based framework that redefines token selection for KV cache compression. In GraphKV, textbftokens are modeled as textbfnodes with importance scores, and textbfedges represent their textbfsimilarity relationships. Through a decay-signal-propagation mechanism, token importance is dynamically updated by propagating information across the graph, enabling adaptive retention of the most contextually significant tokens. GraphKV can be seamlessly utilized in existing KV cache eviction methods such as SnapKV and PyramidKV in a plug-and-play manner. Codes are available in the supplementary materials and will be released on Github.

GraphKV: Breaking the Static Selection Paradigm with Graph-Based KV Cache Eviction

Data efficiency is crucial in domain-specific continual pre-training (CPT) of large language models (LLMs), especially under resource constraints. Aiming for "small data, big impact," this work addresses the limitations of existing domain-specific data selection strategies, which often rely on scarce labeled data or computationally expensive LLMs. We introduce CDF Sampling with Grammatical Complexity (CDF-GC), an annotation-independent, efficient and interpretable data selection framework for CPT. Our approach comprehensively evaluates grammatical complexity using lexical diversity and syntactic complexity, and employs a cumulative distribution function (CDF)-based sampling strategy to balance complexity and diversity. To validate the effectiveness of CDF-GC, we conducted experiments on a financial dataset. The results demonstrate that CDF-GC significantly outperforms baselines, achieving 2.0% improvement in financial QA at the same selection ratio and even surpassing full-data training by 1.7% using only 20% of the data.

Data-Efficient Selection via Grammatical Complexity in Continual Pre-training of Domain-Specific LLMs

This paper explores the lemmatization of multi-word expressions (MWEs) and proper names in Polish – tasks complicated by linguistic irregularities and historical factors. Instead of using rule-based methods, we apply a machine learning approach with fine-tuned plT5 and mT5 models. We trained and validated the models on enhanced gold-standard data from the 2019 PolEval task and evaluated the impact of additional fine-tuning on a silver-standard dataset derived from Wikipedia. Two setups were tested: one without context, and one using left-side context of the target MWE. Our best model achieved 86.23% AccCS, 89.43% AccCI, and a combined score of 88.79%, setting a new state-of-the-art for Polish MWE and named entity lemmatization, as confirmed by the PolEval maintainers. We also evaluated optimization and quantization techniques to reduce model size and inference time with minimal quality loss.

Lemmatization of Polish Multi-word Expressions

A recent study showed that large language models (LLMs) can reconstruct surprisingly long texts — up to thousands of tokens — via autoregressive generation from just one specially trained input embedding. In this work, we explore whether such reconstruction is possible without autoregression. We show that frozen LLMs can generate hundreds of accurate tokens in just one forward pass, when provided with only two learned embeddings. This reveals a surprising and underexplored capability of LLMs — multi-token generation without iterative decoding. We investigate the behaviour of these embeddings and provide insight into the type of information they encode. We also empirically show that although these representations are not unique for a given text, they form connected and local regions in embedding space — a property that suggests the potential of learning a dedicated encoder into that space.

Exploring the Hidden Capacity of LLMs for One-Step Text Generation

Transformer models face scalability challenges in causal language modeling (CLM) due to inefficient memory allocation for growing key-value (KV) caches, which strains compute and storage resources. Existing methods like Grouped Query Attention (GQA) and token-level KV optimization improve efficiency but rely on rigid resource allocation, often discarding "low-priority" tokens or statically grouping them, failing to address the dynamic spectrum of token importance. We propose mixSGA, a novel mixture-of-expert (MoE) approach that dynamically optimizes token-wise computation and memory allocation. Unlike prior approaches, mixSGA retains all tokens while adaptively routing them to specialized experts with varying KV group sizes, balancing granularity and efficiency. Our key novelties include: (1) a token-wise expert-choice routing mechanism guided by learned importance scores, enabling proportional resource allocation without token discard; (2) weight-sharing across grouped attention projections to minimize parameter overhead; and (3) an auxiliary loss to ensure one-hot routing decisions for training-inference consistency in CLMs. Extensive evaluations across Llama3, TinyLlama, OPT, and Gemma2 model families show mixSGA's superiority over static baselines. On instruction-following and continued pretraining tasks, mixSGA achieves higher ROUGE-L and lower perplexity under the same KV budgets.

Mixture of Weight-shared Heterogeneous Group Attention Experts for Dynamic Token-wise KV Optimization

This paper presents R-BPE, a lightweight framework for adapting existing Byte-Pair Encoding (BPE) tokenizers to better support a specified target language. It reuses tokens from user-excluded languages and creates ID-based maps to resolve the new tokens of the chosen language. We evaluate R-BPE on Arabic as a target language. R-BPE reduced subword fertility by an average of 24.4% across the LLaMA 3.1 8B, Command R 35B, and Qwen 3 8B models. Applied to LLaMA 3.1 8B in continued pretraining mode, R-BPE yields a 7.33% reduction in training time. On the ArabicMMLU benchmark, the resulting model improved by 5.09 points on five in-domain topics and matched the original model's overall performance. It also preserved performance on EnglishMMLU. R-BPE effectively leverages existing models' tokenizers, embedding layers, and performance to better support target languages without incurring model size changes. We release an R-BPE implementation that is compatible with HuggingFace interfaces and thereby readily applicable to a wide range of existing models at https://acr.ps/1L9GPmL.

R-BPE: Improving BPE-Tokenizers with Token Reuse

Large Language Models (LLMs) are increasingly costly to fine-tune due to their size, with embedding layers alone accounting for up to 20% of model parameters. While Parameter-Efficient Fine-Tuning (PEFT) methods exist, they largely overlook the embedding layer. In this paper, we introduce TinyTE, a novel PEFT approach that steers model behavior via minimal translational transformations in the embedding space. TinyTE modifies input embeddings without altering hidden layers, achieving achieving competitive performance while requiring approximately 0.0001% of the parameters needed for full fine-tuning. Experiments across architectures provide a new lens for understanding the relationship between input representations and model behavior---revealing them to be more flexible at their foundation than previously thought.

Language Models Can be Efficiently Steered via Minimal Embedding Layer Transformations

Scaling data and model size has driven recent advances in language modeling, but this strategy falters under scenarios with strict data constraints, as in the BabyLM Challenge. However, insights from Chinchilla highlights that smaller models trained on more data outperform larger counterparts trained inadequately, emphasizing the need for compact architectures. Furthermore, while embedding weight tying is a common parameter-saving technique, we find it significantly diminishes linguistic competence in compact models. In response, we explore alternative architectural strategies that preserve the parameter efficiency of tied models without sacrificing the representational benefits of untied embeddings. Consequently, we introduce SLlama a Llama3 architecture variant which incorporates targeted modifications - Repeated Reduced Hidden Size and Projection (RRHP), Permutated Weight Attention (PWA), Shared Projection Multi-Layer Perceptron (SPMLP), and Layer Weight Sharing - to compress transformer components. Without relying on distillation, SLlama achieves a 31.72% improvement in linguistic knowledge acquisition over the BabyLlama baseline, with a comparable GLUE score and significantly lower parameter count. These results demonstrate that well-designed, compact models can rival larger ones under strict data constraints.

SLlama: Parameter-Efficient Language Model Architecture for Enhanced Linguistic Competence Under Strict Data Constraints

Recent advances in large language models (LLMs) have enabled strong reasoning capabilities through Chain-of-Thought (CoT) prompting, which elicits step-by-step problem solving, but often at the cost of excessive verbosity in intermediate outputs, leading to increased computational overhead. We propose Sketch-of-Thought (SoT), a prompting framework that integrates cognitively inspired reasoning paradigms with linguistic constraints to reduce token usage while preserving reasoning accuracy. SoT is designed as a flexible, modular approach and is instantiated with three paradigms—Conceptual Chaining, Chunked Symbolism, and Expert Lexicons—each tailored to distinct reasoning tasks and selected dynamically at test-time by a lightweight routing model. Across 15 reasoning datasets spanning multiple domains, languages, and modalities, SoT achieves token reductions of up to 78% with minimal accuracy loss. In tasks such as mathematical and multi-hop reasoning, it even improves accuracy while shortening outputs.

Downloads

Next from EMNLP 2025

zFLoRA: Zero-Latency Fused Low-Rank Adapters

Similar lecture

RefreshKV: Updating Small KV Cache During Long-form Generation