China

This work presents a novel trie (prefix-tree)-based parallel decoding method that addresses the memory inefficiency of batch-based beam search. By sharing a single KV cache across beams with common prefixes, our approach dramatically reduces memory usage and enables efficient decoding. We evaluated our method across three attention architectures--Multi-Head Attention (Phi-3.5-mini-instruct), Grouped Query Attention (Llama-3.1-8B-Instruct), and Sliding Window Attention (Mistral-Small-24B-Instruct-2501)--using CNN/DailyMail for abstractive summarization and HumanEval for code generation. Our experiments demonstrate substantial memory savings (4–8×) and notable decoding speed improvements (up to 2.4×), without compromising generation quality. These results highlight the method&#39;s suitability for memory-constrained environments and large-scale deployments.

EMNLP 2025

Efficient Beam Search for Large Language Models Using Trie-Based Decoding

This work presents a novel trie (prefix-tree)-based parallel decoding method that addresses the memory inefficiency of batch-based beam search. By sharing a single KV cache across beams with common prefixes, our approach dramatically reduces memory usage and enables efficient decoding. We evaluated our method across three attention architectures--Multi-Head Attention (Phi-3.5-mini-instruct), Grouped Query Attention (Llama-3.1-8B-Instruct), and Sliding Window Attention (Mistral-Small-24B-Instruct-2501)--using CNN/DailyMail for abstractive summarization and HumanEval for code generation. Our experiments demonstrate substantial memory savings (4–8×) and notable decoding speed improvements (up to 2.4×), without compromising generation quality. These results highlight the method's suitability for memory-constrained environments and large-scale deployments.

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."<br>

*Message from the General Chair, Dirk Hovy*

[**Link to Conference Handbook**](https://drive.google.com/file/d/1johU5QqVVYO4RfH7QcIORr7qrVBdzdwC/view?usp=sharing)





<br>

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

Although large language models (LLM) have achieved remarkable performance, their enormous parameter counts hinder deployment on resource-constrained hardware. Low-rank compression can reduce both memory usage and computational demand, but applying a uniform compression ratio across all layers often leads to significant performance degradation, and previous methods perform poorly during decoding. To address these issues, we propose the Fine-grained Low-Rank Compressor (FLRC), which efficiently determines an optimal rank allocation for each layer, and incorporates progressive low-rank decoding to maintain text generation quality. Comprehensive experiments on diverse benchmarks demonstrate the superiority of FLRC, achieving up to a 17% improvement in ROUGE-L on summarization tasks compared to state-of-the-art low-rank compression methods, establishing a more robust and efficient framework to improve LLM inference.

FLRC: Fine-grained Low-Rank Compressor for Efficient LLM Inference

Most multilingual question-answering benchmarks, while covering a diverse pool of languages, do not factor in regional diversity in the information they capture and tend to be Western-centric. This introduces a significant gap in fairly evaluating multilingual models' comprehension of factual information from diverse geographical locations. To address this, we introduce XNationQA to investigate the cultural literacy of multilingual LLMs. XNationQA encompasses questions on the geography, culture, and history of nine countries, containing a total of 49,280 questions in seven languages. We benchmark eight standard multilingual LLMs on XNationQA. Our analyses uncover a considerable discrepancy in the model's accessibility to culturally specific facts across languages. Notably, we often find that a model demonstrates greater knowledge of cultural information in English than in the dominant language of the respective culture. The models show better performance in Western languages, though this does not translate to their being more literate for Western countries, which is counter-intuitive. Furthermore, we observe that models have a very limited ability to transfer knowledge across languages, particularly evident in open-source models.

 Do You Know About My Nation? Investigating Multilingual Language Models’ Cultural Literacy Through Factual Knowledge

Fine-tuning pre-trained models on targeted datasets enhances task-specific performance but often comes at the expense of generalization. Model merging techniques, which integrate multiple fine-tuned models into a single multi-task model through task arithmetic, offer a promising solution. However, task interference remains a fundamental challenge, leading to performance degradation and suboptimal merged models. Existing approaches largely overlook the fundamental roles of neurons, their connectivity, and activation, resulting in a merging process and a merged model that does not consider how neurons relay and process information. In this work, we present the first study that relies on neuronal mechanisms for model merging. We decompose task-specific representations into two complementary neuronal subspaces that regulate neuron sensitivity and input adaptability. Leveraging this decomposition, we introduce NeuroMerging, a novel merging framework developed to mitigate task interference within neuronal subspaces, enabling training-free model fusion across diverse tasks. Through extensive experiments, we demonstrate that NeuroMerging achieves superior performance compared to existing methods on multi-task benchmarks across both natural language and vision domains. Our findings highlight the importance of aligning neuronal mechanisms in model merging, offering new insights into mitigating task interference and improving knowledge fusion. Code will be released upon acceptance.

To See a World in a Spark of Neuron: Disentangling Multi-Task Interference for Training-Free Model Merging

Upgrading embedding models in production vector databases typically necessitates re-encoding the entire corpus and rebuilding the Approximate Nearest Neighbor (ANN) index, leading to significant operational disruption and computational cost. This paper presents Drift-Adapter, a lightweight, learnable transformation layer designed to bridge embedding spaces between model versions. By mapping new queries into the legacy embedding space, Drift-Adapter enables the continued use of the existing ANN index, effectively deferring full re-computation. We systematically evaluate three adapter parameterizations: Orthogonal Procrustes, Low-Rank Affine, and a compact Residual MLP, trained on a small sample of paired old/new embeddings. Experiments on MTEB text corpora and a CLIP image model upgrade (1M items) show that Drift-Adapter recovers 95–99% of the retrieval recall (Recall@10, MRR) of a full re-embedding, adding less than 10,μs query latency. Compared to operational strategies like full re-indexing or dual-index serving, Drift-Adapter dramatically reduces recompute costs (by over 100 times) and facilitates upgrades with near-zero operational interruption. We analyze robustness to varied model drift, training data size, scalability to billion-item systems, and the impact of design choices like diagonal scaling, demonstrating Drift-Adapter's viability as a pragmatic solution for agile model deployment.

Drift-Adapter: A Practical Approach to Near Zero-Downtime Embedding Model Upgrades in Vector Databases

Mixture-of-Experts (MoE) layers scale transformers by routing tokens to a sparse subset of feed-forward experts. Token-level routing, however, assigns an entire semantic spectrum to each expert, creating capacity bottlenecks, load-balancing pathologies, and limited specialisation. We introduce SliceMoE, an architecture that routes contiguous slices of a token's hidden vector. A d-dimensional embedding is partitioned into S slices, and for each slice, a lightweight shared router predicts the top-k experts. Experts operate on their assigned slices independently, and outputs are re-assembled, maintaining per-token FLOP efficiency. Because slices from different tokens interleave within an expert, utilisation is naturally smoother. We propose a slice-level capacity loss, cross-slice dropout, and efficient fused batched-GEMM kernels. Experiments on WikiText-103 language modelling, WMT En–De translation, and three text-classification datasets show SliceMoE attains up to 1.7x faster inference than dense baselines, 12–18% lower perplexity than parameter-matched token-MoE, and improved expert balance, with interpretable expertise over syntactic versus semantic sub-spaces.

SliceMoE: Routing Embedding Slices Instead of Tokens for Fine-Grained and Balanced Transformer Scaling

Large language models (LLMs) tend to follow maliciously crafted instructions to generate deceptive responses, posing safety challenges. How deceptive instructions alter the internal representations of LLM compared to truthful ones remains poorly understood beyond output analysis. To bridge this gap, we investigate when and how these representations ``flip'', such as from truthful to deceptive, under deceptive versus truthful/neutral instructions. Analyzing the internal representations of Llama-3.1-8B-Instruct and Gemma-2-9B-Instruct on a factual verification task, we find the model's instructed True/False output is predictable via linear probes across all conditions based on the internal representation. Further, we use Sparse Autoencoders (SAEs) to show that the Deceptive instructions induce significant representational shifts compared to Truthful/Neutral representations (which are similar), concentrated in early-to-mid layers and detectable even on complex datasets. We also identify specific SAE features highly sensitive to deceptive instruction and use targeted visualizations to confirm distinct truthful/deceptive representational subspaces.

When Truthful Representations Flip Under Deceptive Instructions?

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.

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

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.

Downloads

Next from EMNLP 2025

FLRC: Fine-grained Low-Rank Compressor for Efficient LLM Inference

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES