China

We propose new static word embeddings optimised for sentence semantic representation. We first extract word embeddings from a pre-trained Sentence Transformer, and improve them with sentence-level principal component analysis, followed by either knowledge distillation or contrastive learning. During inference, we represent sentences by simply averaging word embeddings, which requires little computational cost. We evaluate models on both monolingual and cross-lingual tasks and show that our model substantially outperforms existing static models on sentence semantic tasks, and even rivals a basic Sentence Transformer model (SimCSE) on some data sets. Lastly, we perform a variety of analyses and show that our method successfully removes word embedding components that are irrelevant to sentence semantics, and adjusts the vector norms based on the influence of words on sentence semantics.

EMNLP 2025

Static Word Embeddings for Sentence Semantic Representation

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.

Multimodal Large Language Models (MLLMs) are increasingly used in Personalized Image Aesthetic Assessment (PIAA), offering a scalable alternative to expert evaluation. However, their outputs may reflect subtle biases shaped by demographic cues such as gender, age, or education. In this work, we introduce **AesBiasBench**, a benchmark designed to evaluate MLLMs along two complementary axes: (1) the presence of *stereotype bias*, measured by how aesthetic evaluations vary across demographic groups; and (2) the *alignment* between model outputs and real human aesthetic preferences. Our benchmark spans three subtasks, Aesthetic Perception, Assessment, and Empathy, and introduces structured metrics (IFD, NRD, AAS) to quantify both bias and alignment. We evaluate 19 MLLMs, including proprietary models (e.g., GPT-4o, Claude-3.5-Sonnet) and open-source models (e.g., InternVL-2.5, Qwen2.5-VL). Results show that smaller models exhibit stronger stereotype bias, while larger models better align with human preferences. Adding identity information often amplifies bias, particularly in emotional judgment. These findings highlight the need for identity-aware evaluation frameworks for subjective vision-language tasks.

AesBiasBench: Evaluating Bias and Alignment in Multimodal Language Models for Personalized Image Aesthetic Assessment

Large Language Models (LLMs) often don't perform as expected under Domain Shift or after Instruct-tuning. A reliable indicator of LLM performance in these settings could assist in decision-making. We present a method that uses the known performance in high-resource domains and fine-tuning settings to predict performance in low-resource domains or base models respectively. In our paper, we formulate the task of performance prediction, construct a dataset for and train regression models to predict the said change in performance. Our proposed methodology is lightweight and, in practice, can help researchers & practitioners decide if resources should be allocated for data labeling and LLM Instruct-tuning.

DA-Pred: Performance Prediction for Text Summarization under Domain-Shift and Instruct-Tuning

Argumentative essay generation (AEG) is a complex task that requires advanced semantic understanding, logical reasoning, and organized integration of perspectives. Despite showing a promising performance, current efforts often overlook the dynamical and hierarchical nature of structural argumentative planning, and struggle with flexible rhetorical expression, leading to limited argument divergence and rhetorical optimization. Inspired by human debate behavior and Bitzer's rhetorical situation theory, we propose a debate-driven rhetorical framework for argumentative writing. The uniqueness lies in three aspects: (1) dynamic assesses the divergence of viewpoints and progressively reveals the hierarchical outline of arguments based on a depth-then-breadth paradigm, improving the perspective divergence within argumentation; (2) simulates human debate through iterative defender-attacker interactions, improving the logical coherence of arguments; (3) incorporates Bitzer's rhetorical situation theory to flexibly select appropriate rhetorical techniques, enabling the rhetorical expression. Experiments on four benchmarks validate that our approach significantly improves logical depth, argumentative diversity, and rhetorical persuasiveness over existing state-of-the-art models.

Plan Dynamically, Express Rhetorically: A Debate-Driven Rhetorical Framework for Argumentative Writing

Sparse Autoencoders (SAEs) have been proposed as an unsupervised approach to learn a decomposition of a model’s latent space. This enables useful applications, such as fine-grained steering of model outputs without requiring labeled data. Current steering methods identify SAE features to target by analyzing the input tokens that activate them. However, recent work has highlighted that activations alone do not fully describe the effect of a feature on the model’s output. In this work we draw a distinction between two types of features: input features, which mainly capture patterns in the model’s input, and output features, those that have a human-understandable effect on the model’s output. We propose input and output scores to characterize and locate these types of features, and show that high values for both scores rarely co-occur in the same features. These findings have practical implications: After filtering out features with low output scores, steering with SAEs results in a 2–3x improvement, matching the performance of existing supervised methods.

SAEs Are Good for Steering – If You Select the Right Features

Large language models (LLMs) have achieved impressive performance across natural language processing (NLP) tasks. As real-world applications increasingly demand longer context windows, continued pretraining and supervised fine-tuning (SFT) on long-context data has become a common approach. While the effects of data length in continued pretraining have been extensively studied, their implications for SFT remain unclear. In this work, we systematically investigate how SFT data length influences LLM behavior on short-context tasks. Counterintuitively, we find that long-context SFT improves short-context performance, contrary to the commonly observed degradation from long-context pretraining. To uncover the underlying mechanisms of this phenomenon, we first decouple and analyze two key components, Multi-Head Attention (MHA) and Feed-Forward Network (FFN), and show that both independently benefit from long-context SFT. We further study their interaction and reveal a knowledge preference bias: long-context SFT promotes contextual knowledge, while short-context SFT favors parametric knowledge, making exclusive reliance on long-context SFT suboptimal. Finally, we demonstrate that hybrid training mitigates this bias, offering explainable guidance for fine-tuning LLMs.

When Long Helps Short: How Context Length in Supervised Fine-tuning Affects Behavior of Large Language Models

Bit-flip errors (BFEs) are hardware faults where individual bits in memory or processing units are unintentionally flipped. These errors pose a significant threat to neural network reliability because even small changes in model parameters can lead to large shifts in outputs. Large language models (LLMs) are particularly vulnerable on resource-constrained or outdated hardware. Such hardware often lacks error-correction mechanisms and faces aging issues, leading to instability under the vast parameter counts and heavy computational loads of LLMs. While the impact of BFEs on traditional networks like CNNs is relatively well-studied, their effect on the complex architecture of transformers remains largely unexplored. Firstly, this paper presents a comprehensive systematic analysis of BFE vulnerabilities in key LLM components, revealing distinct sensitivities across parameters, activations, and gradients during fine-tuning and inference. Secondly, based on our findings, we introduce a novel defense strategy FlipGuard: (i) exponent bit protection, and (ii) a self-correction based fine-tuning mechanism, to address BFE consequences. FlipGuard minimizes performance degradation while significantly enhancing robustness against BFEs. Experiments demonstrate a 9.27 reduction in accuracy drop under 1 BFEs on the SST-2 dataset using BERT, and a 36.35-point improvement in perplexity on the Wikitext-103 dataset using GPT-2, compared to unprotected models. These results show the potential of our approach in enabling reliable LLM deployment on diverse and less reliable hardware platforms.

Bit-Flip Error Resilience in LLMs: A Comprehensive Analysis and Defense Framework

Large Language Models (LLMs) are capable of generating persuasive Natural Language Explanations (NLEs) to justify their answers. However, the faithfulness of these explanations should not be readily trusted at face value. Recent studies have proposed various methods to measure the faithfulness of NLEs, typically by inserting perturbations at the explanation or feature level. We argue that these approaches are neither comprehensive nor correctly designed according to the established definition of faithfulness. Moreover, we highlight the risks of grounding faithfulness findings on out-of-distribution samples. In this work, we leverage a causal mediation technique called activation patching, to measure the faithfulness of an explanation towards supporting the explained answer. Our proposed metric, Causal Faithfulness quantifies the consistency of causal attributions between explanations and the corresponding model outputs as the indicator of faithfulness. We experimented across models varying from 2B to 27B parameters and found that models that underwent alignment-tuning tend to produce more faithful and plausible explanations. We find that Causal Faithfulness is a promising improvement over existing faithfulness tests by taking into account the model’s internal computations and avoiding out-of-distribution concerns that could otherwise undermine the validity of faithfulness assessments.

Towards Faithful Natural Language Explanations: A Study Using Activation Patching in Large Language Models

Sequence classification is essential in NLP for understanding and categorizing language patterns in tasks like sentiment analysis, intent detection, and topic classification. Transformer-based models, despite achieving state-of-the-art performance, have inherent limitations due to quadratic time and memory complexity, restricting their input length. Although extensive efforts have aimed at reducing computational demands, processing extensive contexts remains challenging. To overcome these limitations, we propose ResFormer, a novel neural network architecture designed to model varying context lengths efficiently through a cascaded methodology. ResFormer integrates an reservoir computing network featuring a non-linear readout to effectively capture long-term contextual dependencies in linear time. Concurrently, short-term dependencies within sentences are modeled using a conventional Transformer architecture with fixed-length inputs. Experiments demonstrate that ResFormer significantly outperforms baseline models of DeepSeek-Qwen and ModernBERT, delivering an accuracy improvement of up to +22.3% on the EmoryNLP dataset and consistent gains on MultiWOZ, MELD, and IEMOCAP. In addition, EchoFormer exhibits reduced memory consumption, underscoring its effectiveness and efficiency in modeling extensive contextual information.

ResFormer: All-Time Reservoir Memory for Long Sequence Classification

We investigate how large language models (LLMs) perform latent multi-hop reasoning in prompts like "Wolfgang Amadeus Mozart's mother's spouse is". To analyze this process, we introduce logit flow, an interpretability method that traces how logits propagate across layers and positions toward the final prediction. Using logit flow, we identify four distinct stages in single-hop knowledge prediction: (A) entity subject enrichment, (B) entity attribute extraction, (C) relation subject enrichment, and (D) relation attribute extraction. Extending this analysis to multi-hop reasoning, we find that failures often stem from the relation attribute extraction stage, where conflicting logits reduce prediction accuracy. To address this, we propose back attention, a novel mechanism that enables lower layers to leverage higher-layer hidden states from different positions during attention computation. With back attention, a 1-layer transformer achieves the performance of a 2-layer transformer. Applied to five LLMs, back attention improves accuracy on five reasoning datasets, demonstrating its effectiveness in enhancing latent multi-hop reasoning ability.

Back Attention: Understanding and Enhancing Multi-Hop Reasoning in Large Language Models

Large pre-trained language models have demonstrated impressive capabilities, but there is still much to learn about how they operate. In this study, we conduct an investigation of the autoregressive transformer’s ability to perform basic addition operations. Specifically, by using causal analysis we found that a few different attention heads in the middle layers control the addition carry, with each head processing carries of different lengths. Due to the lack of global focus on the sequence within these attention heads, the model struggles to handle long-sequence addition tasks. By performing inference intervention on mistral-7B, partial task performance can be restored, with the accuracy on 20-digit long-sequence additions from 2% to 38%. Through fine-tuning, a new mechanism branches out for handling complex cases, yet it still faces challenges with length generalization. Our research reveals how the models perform basic arithmetic task, and further provides insights into the debate on whether these models are merely statistical.

Downloads

Next from EMNLP 2025

AesBiasBench: Evaluating Bias and Alignment in Multimodal Language Models for Personalized Image Aesthetic Assessment

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES