China

Conventional speculative decoding (SD) methods utilize a predefined length policy for proposing drafts, which implies the premise that the target model smoothly accepts the proposed draft tokens. However, reality deviates from this assumption: the oracle draft length varies significantly, and the fixed-length policy hardly satisfies such a requirement. Moreover, such discrepancy is further exacerbated in scenarios involving complex reasoning and long-form generation, particularly under test-time scaling for reasoning-specialized models. Through both theoretical and empirical estimation, we establish that the discrepancy between the draft and target models can be approximated by the draft model&#39;s prediction entropy: a high entropy indicates a low acceptance rate of draft tokens, and vice versa. Based on this insight, we propose SVIP: Self-Verification Length Policy for Long-Context Speculative Decoding, which is a training-free dynamic length policy for speculative decoding systems that adaptively determines the lengths of draft sequences by referring to the draft entropy. Experimental results on mainstream SD benchmarks as well as reasoning-heavy benchmarks demonstrate the superior performance of SVIP, achieving up to 17% speedup on MT-Bench at 8K context compared with fixed draft lengths, and 22% speedup for QwQ in long-form reasoning.

EMNLP 2025

Draft Model Knows When to Stop: Self-Verification Speculative Decoding for Long-Form Generation

long-context generation

speculative decoding

large language models

### Welcome to 2025 Annual Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics

Welcome to the 2025 meeting of the Nations of the Americas Chapter of the Association for Computational Linguistics! I am proud to help organize the first NAACL conference to carry the new name of our organization, one that emphasizes inclusion for all of the Americas. I am also pleased to welcome you to Albuquerque, New Mexico, a state whose unique blend of cultural influences will make for an excellent backdrop for NAACL 2025, especially with this year’s special theme on NLP in a Multicultural World. 
**[Continue reading...](https://drive.google.com/file/d/1jX-qGhqVSZZCIrAnJaz798pflu5Irrdn/view?usp=sharing)**

*- Colin Cherry, Google, NAACL 2025 General Chair* 

[![](https://assets.underline.io/markdown_image/1/image/b087f8a4dc5816d6e1a6514e59c59ac3.png)](https://drive.google.com/file/d/1T96GzPqObXrMln2BMByCSXTSizjTg69P/view?usp=sharing)

You need to log in with the email address you registered with. Access credentials have been sent to your email. 

Please be sure to check your spam and other email folders if you do not see an email confirmation right away.

Please log in to explore this event.

To access NAACL 2025 event page you are required to register. Please follow [**this link**](https://2025.naacl.org/registration/registration/) to register. Access will depend on your registration type.

Please register!

NAACL 2025

Welcome to 2025 Annual Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics

Conventional speculative decoding (SD) methods utilize a predefined length policy for proposing drafts, which implies the premise that the target model smoothly accepts the proposed draft tokens. However, reality deviates from this assumption: the oracle draft length varies significantly, and the fixed-length policy hardly satisfies such a requirement. Moreover, such discrepancy is further exacerbated in scenarios involving complex reasoning and long-form generation, particularly under test-time scaling for reasoning-specialized models. Through both theoretical and empirical estimation, we establish that the discrepancy between the draft and target models can be approximated by the draft model's prediction entropy: a high entropy indicates a low acceptance rate of draft tokens, and vice versa. Based on this insight, we propose SVIP: Self-Verification Length Policy for Long-Context Speculative Decoding, which is a training-free dynamic length policy for speculative decoding systems that adaptively determines the lengths of draft sequences by referring to the draft entropy. Experimental results on mainstream SD benchmarks as well as reasoning-heavy benchmarks demonstrate the superior performance of SVIP, achieving up to 17% speedup on MT-Bench at 8K context compared with fixed draft lengths, and 22% speedup for QwQ in long-form reasoning.

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.

With the continuous advancement in the performance of large language models (LLMs), their demand for computational resources and memory has significantly increased, which poses major challenges for efficient inference on consumer-grade devices and legacy servers. These devices typically feature relatively weaker GPUs and stronger CPUs. Although techniques such as parameter offloading and partial offloading can alleviate GPU memory pressure to some extent, their effectiveness is limited due to communication latency and suboptimal hardware resource utilization. To address this issue, we propose Dovetail—a lossless inference acceleration method that leverages the complementary characteristics of heterogeneous devices and the advantages of speculative decoding. Dovetail deploys a draft model on the GPU to perform preliminary predictions, while a target model running on the CPU validates these outputs. By reducing the granularity of data transfer, Dovetail significantly minimizes communication overhead. To further improve efficiency, we optimize the draft model specifically for heterogeneous hardware environments by reducing the number of draft tokens to lower parallel verification latency, increasing model depth to enhance predictive capabilities, and introducing a Dynamic Gating Fusion (DGF) mechanism to improve the integration of feature and embedding information. We conduct comprehensive evaluations of Dovetail across various consumer-grade GPUs, covering multiple tasks and mainstream models. Experimental results on 13B models demonstrate that Dovetail achieves inference speedups ranging from 1.79× to 10.1× across different devices, while maintaining consistency and stability in the distribution of generated texts.

Dovetail: A CPU/GPU Heterogeneous Speculative Decoding for LLM inference

Multi-lingual ability transfer has become increasingly important for the broad application of large language models (LLMs). Existing work highly relies on training with the multi-lingual ability-related data, which may not be available for low-resource languages. To solve it, we propose a **M**ulti-lingual **A**bilities **E**xtraction and **C**ombination approach, named as **MAEC**. Our key idea is to decompose and extract language-agnostic ability-related weights from LLMs, and combine them across different languages by simple addition and subtraction operations without training. Specifically, our MAEC consists of the extraction and combination stages. In the extraction stage, we firstly locate key neurons that are highly related to specific abilities, and then employ them to extract the transferable ability-related weights. In the combination stage, we further select the ability-related tensors that mitigate the linguistic effects, and design a combining strategy based on them and the language-specific weights, to build the multi-lingual ability-enhanced LLM. To assess the effectiveness of our approach, we conduct extensive experiments on LLaMA-3 8B on mathematical and scientific tasks in both high-resource and low-resource lingual scenarios. Experiment results have shown that MAEC can effectively and efficiently extract and combine the advanced abilities, achieving **comparable performance with PaLM**. We will publicly release our code and data.

Extracting and Combining Abilities For Building Multi-lingual Ability-enhanced Large Language Models

Large Language Models (LLMs) have demonstrated remarkable capabilities, yet their further evolution is often hampered by the scarcity of high-quality training data and the heavy reliance of traditional methods on expert-labeled data. This reliance sets a ceiling on LLM performance and is particularly challenging in low data resource scenarios where extensive supervision is unavailable. To address this issue, we propose a novel paradigm named LANCE (**LAN**guage models as **C**ontinuous self-**E**volving data engineers) that enables LLMs to train themselves by autonomously generating, cleaning, reviewing, and annotating data with preference information. Our approach demonstrates that LLMs can serve as continuous self-evolving data engineers, significantly reducing the time and cost of post-training data construction. Through iterative fine-tuning on Qwen2 series models, we validate the effectiveness of LANCE across various tasks, showing that it can maintain high-quality data generation and continuously improve model performance. Across multiple benchmark dimensions, LANCE results in an average score enhancement of **3.64** for Qwen2-7B and **1.75** for Qwen2-7B-Instruct. This autonomous data construction paradigm not only lessens reliance on human experts or external models but also ensures data aligns with human preferences, offering a scalable path for LLM self-improvement, especially in contexts with limited supervisory data. Code is available at: https://anonymous.4open.science/r/LANCE.

Language Models as Continuous Self-Evolving Data Engineers

Mixture-of-Experts (MoE) models offer immense capacity via sparsely gated expert subnetworks, yet adapting them to multiple domains without catastrophic forgetting remains an open challenge. Existing approaches either incur prohibitive computation, suffer cross-domain interference, or require separate runs per domain. We propose DES-MoE, a dynamic expert specialization framework for multi-domain adaptation of Mixture-of-Experts models. DES-MoE addresses catastrophic forgetting through three innovations: (1) an adaptive router balancing pre-trained knowledge retention and task-specific updates via distillation, (2) real-time expert-domain correlation mapping to isolate domain-specific gradients, and (3) a three-phase adaptive fine-tuning schedule that progressively freezes non-specialized parameters. Evaluated on six domains (math, code, law, etc.), DES-MoE matches single-domain ESFT performance while training one unified model, reduces forgetting by 89% compared to full fine-tuning as domains scale from 2 to 6, and achieves 68% faster convergence than conventional methods. Our work establishes dynamic expert isolation as a scalable paradigm for multi-task MoE adaptation.

Dynamic Expert Specialization: Towards Catastrophic Forgetting-Free Multi-Domain MoE Adaptation

Large language models have demonstrated exceptional capabilities across diverse tasks, but their fine-tuning demands significant memory, posing challenges for resource-constrained environments. Zeroth-order (ZO) optimization provides a memory-efficient alternative by eliminating the need for backpropagation. However, ZO optimization suffers from high gradient variance, and prior research has largely focused on single-task learning, leaving its application to multi-task learning unexplored. Multi-task learning is crucial for leveraging shared knowledge across tasks to improve generalization, yet it introduces unique challenges under ZO settings, such as amplified gradient variance and collinearity. In this paper, we present MaZO, the first framework specifically designed for multi-task LLM fine-tuning under ZO optimization. MaZO tackles these challenges at the parameter level through two key innovations: a weight importance metric to identify critical parameters and a multi-task weight update mask to selectively update these parameters, reducing the dimensionality of the parameter space and mitigating task conflicts. Experiments demonstrate that MaZO achieves state-of-the-art performance, surpassing even multi-task learning methods designed for first-order optimization.

MaZO: Masked Zeroth-Order Optimization for Multi-Task Fine-Tuning of Large Language Models

Recent work proposed state-space models (SSMs) as an efficient alternative to transformer-based LLMs. Can these models be pruned to further reduce their computation costs? We adapt several pruning methods to the SSM structure, and apply them to four SSM-based LLMs across multiple tasks. We find that such models are quite robust to some pruning methods (e.g., WANDA), while using other methods lead to fast performance degradation.

On Pruning State-Space LLMs

Low-rank adaptation (LoRA) efficiently adapts LLMs to downstream tasks by decomposing LLMs' weight update into trainable low-rank matrices for fine-tuning. However, the random low-rank matrices may introduce massive task-irrelevant information, while their recomposed form suffer from limited representation spaces under low-rank operations. Such dense and choked adaptation in LoRA impairs the adaptation performance of LLMs on downstream tasks. To address these challenges, this paper proposes OHoRA, an orthogonal high-rank adaptation for parameter-efficient fine-tuning on LLMs. According to the information bottleneck theory, OHoRA decomposes LLMs' pre-trained weight matrices into orthogonal basis vectors via QR decomposition and splits them into two low-redundancy high-rank components to suppress task-irrelevant information. It then performs dynamic rank-elevated recomposition through Kronecker product to generate expansive task-tailored representation spaces, enabling precise LLM adaptation and enhanced generalization. OHoRA effectively operationalizes the information bottleneck theory to decompose LLMs' weight matrices into low-redundancy high-rank components and recompose them in rank-elevated manner for more task-tailored representation spaces and precise LLM adaptation. Empirical evaluation shows OHoRA’s effectiveness by outperforming LoRA and its variants and achieving comparable performance to full fine-tuning with only 0.0371% trainable parameters.

An Orthogonal High-Rank Adaptation for Large Language Models

The high costs of customizing large language models (LLMs) fundamentally limit their adaptability to user-specific needs. Consequently, LLMs are increasingly offered as cloud-based services, a paradigm that introduces critical limitations: providers struggle to support personalized customization at scale, while users face privacy risks when exposing sensitive data. To address this dual challenge, we propose Customized Black-box Prompt Tuning (CBP-Tuning), a novel framework that facilitates efficient local customization while preserving bidirectional privacy. Specifically, we design a two-stage framework: (1) a prompt generator trained on the server-side to capture domain-specific and task-agnostic capabilities, and (2) user-side gradient-free optimization that tailors soft prompts for individual tasks. This approach eliminates the need for users to access model weights or upload private data, requiring only a single customized vector per task while achieving effective adaptation. Furthermore, the evaluation of CBP-Tuning in the commonsense reasoning, medical and financial domain settings demonstrates superior performance compared to baselines, showcasing its advantages in task-agnostic processing and privacy preservation.

CBP-Tuning: Efficient Local Customization for Black-box Large Language Models

We propose Paired by the Teacher (PbT), a two-stage teacher–student pipeline for synthesizing accurate input–output pairs without any human labeling or existing parallel data. In many low-resource natural language generation (NLG) scenarios, practitioners may have only raw outputs, like recaps, highlights, or questions, or only raw inputs, such as dialogues, articles, or paragraphs, but seldom both sides of the parallel data, unless we perform human labeling. This mismatch forces small models to learn from very few examples or rely on costly, broad-scope synthetic examples produced by large LLMs. In PbT, a teacher LLM first transforms each unpaired example into a concise intermediate representation (IR), and a student model learns to invert this transformation to reconstruct the original input from the IR. This enables us to pair each output with its generated input, creating high-quality paired data. We evaluate PbT on five benchmarks—dialogue summarization (SAMSum, DialogSum), document summarization (XSum, CNNDM), and question generation (SQuAD)—and an unpaired setting on SwitchBoard (paired with DialogSum summaries). An 8B student trained only on PbT data outperforms models trained on 70 B teacher-generated corpora and other unsupervised baselines, closing the gap to human-annotated pairs to within 2 ROUGE points. Human evaluation on SwitchBoard further confirms that only PbT meets target summary lengths with concise, faithful outputs, while all baselines remain overly verbose.

Paired by the Teacher: Turning Unpaired Data into High-Fidelity Pairs for Low-Resource Text Generation

Large Language Models (LLMs) have demonstrated a remarkable understanding of language nuances through instruction tuning, enabling them to effectively tackle various natural language processing tasks. Recent research has focused on the quality of instruction data rather than the quantity of instructions. However, existing high-quality instruction selection methods rely on external models or rules, overlooking the intrinsic association between pre-trained model and instruction data, making it difficult to select data that align with the preferences of pre-trained model. To address this challenge, we propose a strategy that utilizes noise injection to identify the quality of instruction data, without relying on external model. We also implement the strategy of combining inter-class diversity and intra-class diversity to improve model performance. The experimental results demonstrate that our method significantly outperforms the model trained on the entire dataset and established baselines. Our study provides a new perspective on noise injection in the field of instruction tuning, and also illustrates that the pre-trained model itself should be considered in defining high-quality. Additionally, we publish our selected high-quality instruction data.

Downloads

Next from EMNLP 2025

Dovetail: A CPU/GPU Heterogeneous Speculative Decoding for LLM inference

Similar lecture

Decoding Speculative Decoding