China

Large language models (LLMs) can spell out tokens character by character with high accuracy, yet they struggle with more complex character-level tasks, such as identifying compositional subcomponents within tokens. In this work, we investigate how LLMs internally represent and utilize character-level information during the spelling-out process. Our analysis reveals that, although spelling out is a simple task for humans, it is not handled in a straightforward manner by LLMs. Specifically, we show that the embedding layer does not fully encode character-level information, particularly beyond the first character. As a result, LLMs rely on intermediate and higher Transformer layers to reconstruct character-level knowledge, where we observe a distinct “breakthrough” in their spelling behavior. We validate this mechanism through three complementary analyses: probing classifiers, identification of knowledge neurons, and inspection of attention weights.

EMNLP 2025

Spelling-out is not Straightforward: LLMs&#39; Capability of Tokenization from Token to Characters

interpretability

generation

## Welcome to EMNLP 2024! 
We are excited to welcome you to one of the most prominent conferences in the field of Natural Language Processing. This year, EMNLP 2024 is being held in a hybrid format,
offering both virtual and in-person participation in beautiful Miami. Due to a record-breaking number of submissions, we've expanded the total number of accepted papers to accommodate more cutting-edge research from around the globe.
### [Conference Handbook](https://drive.google.com/file/d/1WPROgxjLAC96AJL7Ugy0tEnYm7dkrbHt/view?usp=sharing)

You are required to register for this event. **Please register [here](https://2024.emnlp.org/registration/).** The EMNLP 2024 event page on Underline will be open to public one week prior to the event.

Please register!

EMNLP 2024

EMNLP 2024 will take place in Miami, Florida from Nov 12th to Nov 16th, 2024, at the Hyatt Regency Miami Hote and on Underline for remote participants.

Spelling-out is not Straightforward: LLMs' Capability of Tokenization from Token to Characters

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 advancement of large language models (LLMs), more concerns about their controllability have been raised in recent research. In this paper, we argue for the importance of Knowledge-Constrained Responsiveness (KCR), ensuring that LLMs comply with human-defined constraints. However, KCR is an implicit and unobservable capability of LLMs, functioning as a black box that currently cannot be quantitatively assessed. To address this, we first introduce the definition of "permitted boundary" and define the "boundary bias" to depict KCR. We propose six metrics to quantify the boundary bias of LLMs and subsequently assess the KCR. Furthermore, we establish a benchmark with two new datasets, KCR-SimpleQA and KCR-WebNLG, to evaluate the performance of LLMs. Our extensive experiments show that several tested LLMs still struggle to varying degrees when adhering to constraints, especially without the corresponding knowledge.

Permitted Knowledge Boundary: Evaluating the Knowledge-Constrained Responsiveness of Large Language Models

This study explores how bilingual language models develop complex internal representations. We employ sparse autoencoders to analyze internal representations of bilingual language models with a focus on the effects of training steps, layers, and model sizes. Our analysis shows that language models first learn languages separately, and then gradually form bilingual alignments, particularly in the middle layers. We also found that this bilingual tendency is stronger in larger models. Building on these findings, we demonstrate the critical role of bilingual representations in model performance by employing a novel method that integrates decomposed representations from a fully trained model into a mid-training model. Our results provide insights into how language models acquire bilingual capabilities.

How a Bilingual LM Becomes Bilingual: Tracing Internal Representations with Sparse Autoencoders

In-Context Learning (ICL) is an essential emergent ability of Large Language Models (LLMs), and recent studies introduce CoT to exemplars of ICL to enhance the reasoning capability, especially in mathematics tasks. However, given the continuous advancement of model capabilities, it remains unclear whether CoT exemplars still benefit recent, stronger models in such tasks. Through systematic experiments, we find that for recent strong models such as the Qwen2.5 series, adding traditional CoT exemplars does not improve reasoning performance compared to Zero-Shot CoT. Instead, their primary function is to align the output format with human expectations. We further investigate the effectiveness of enhanced CoT exemplars, constructed using answers from advanced models such as \texttt{Qwen2.5-Max} and \texttt{DeepSeek-R1}. Experimental results indicate that these enhanced exemplars still fail to improve the model's reasoning performance. Further analysis reveals that models tend to ignore the exemplars and focus primarily on the instructions, leading to no observable gain in reasoning ability. Overall, our findings highlight the limitations of the current ICL+CoT framework in mathematical reasoning, calling for a re-examination of the ICL paradigm and the definition of exemplars.

Revisiting Chain-of-Thought Prompting: Zero-shot Can Be Stronger than Few-shot

Safety alignment is critical in pre-trained large language models (LLMs) to generate responses aligned with human values and refuse harmful queries. Unlike LLM, the current safety alignment of VLMs is often achieved with post-hoc safety fine-tuning. However, these methods are less effective to white-box attacks. To address this, we propose textitAdversary-aware DPO (ADPO), a novel training framework that explicitly considers adversary. textitAdversary-aware DPO (ADPO) integrates adversarial training into DPO to enhance the safety alignment of VLMs under worst-case adversarial perturbations. textitADPO introduces two key components: (1) an adversarial-trained reference model that generates human-preferred responses under worst-case perturbations, and (2) an adversary-aware DPO loss that generates winner-loser pairs accounting for adversarial distortions. By combining these innovations, textitADPO ensures that VLMs remain robust and reliable even in the presence of sophisticated jailbreak attacks. Extensive experiments demonstrate that textitADPO outperforms baselines in terms of both safety alignment and general utility of VLMs.

Adversary-Aware DPO: Enhancing Safety Alignment in Vision Language Models via Adversarial Training

Link prediction in knowledge graphs (KGs) requires integrating structural information and semantic context to infer missing entities. While large language models (LLMs) offer strong generative reasoning capabilities, they often struggle with *structural sparsity*, *semantic ambiguity*, and limited exploitation of structural signals, especially under incomplete or zero-shot settings. To address these challenges, we propose **SLiNT** (**S**tructure-aware **L**anguage model with **I**njection and co**N**trastive **T**raining), a modular framework that injects KG-derived structural context into frozen LLMs for robust link prediction. Specifically, **Structure-Guided Neighborhood Enhancement (SGNE)** retrieves pseudo-neighbors to enrich sparse entities and mitigate missing context; **Dynamic Hard Contrastive Learning (DHCL)** introduces fine-grained supervision by interpolating hard positives and negatives to resolve entity-level ambiguity; and **Gradient-Decoupled Dual Injection (GDDI)** performs token-level structure-aware intervention without altering the LLM backbone. Experiments on WN18RR and FB15k-237 show that SLiNT outperforms both embedding-based and generation-based baselines, demonstrating the effectiveness of structure-aware representation learning for scalable knowledge graph completion.

SLiNT: Structure-aware Language Model with Injection and Contrastive Training for Knowledge Graph Completion

Understanding humor is a core aspect of social intelligence, yet it remains a significant challenge for Large Multimodal Models (LMMs). We introduce PixelHumor, a benchmark dataset of 2,800 annotated multi-panel comics designed to evaluate LMMs’ ability to interpret multimodal humor and recognize narrative sequences. Experiments with state-of-the-art LMMs reveal substantial gaps: for instance, top models achieve only 61% accuracy in panel sequencing, far below human performance. This underscores critical limitations in current models’ integration of visual and textual cues for coherent narrative and humor understanding. By providing a rigorous framework for evaluating multimodal contextual and narrative reasoning, PixelHumor aims to drive the development of LMMs that better engage in natural, socially aware interactions.

Humor in Pixels: Benchmarking Large Multimodal Models Understanding of Online Comics

Recent advances in Chain-of-Thought (CoT) prompting have substantially improved the reasoning capabilities of Large Language Models (LLMs). However, these methods often suffer from overthinking, leading to unnecessarily lengthy or redundant reasoning traces. Existing approaches attempt to mitigate this issue through curating multiple reasoning chains for training LLMs, but their effectiveness is often constrained by the quality of the generated data and prone to overfitting. To address the challenge, we propose Reasoning Compression Through Stepwise Trials (ReCUT), a novel method aimed at balancing the accuracy and length of reasoning trajectory. Specifically, ReCUT employs a stepwise exploration mechanism and a long-short switched sampling strategy, enabling LLMs to incrementally generate diverse reasoning paths. These paths are evaluated and used to construct preference pairs to train two specialized models (Gemini LLMs)—one optimized for reasoning accuracy, the other for shorter reasoning. A final integrated model is obtained by interpolating the parameters of these two models. Experimental results across multiple math reasoning datasets and backbone models demonstrate that ReCUT significantly reduces reasoning lengths by approximately 30-50%, while maintaining or improving reasoning accuracy compared to various baselines. All codes and data will be released via GitHub.

ReCUT: Balancing Reasoning Length and Accuracy in LLMs via Stepwise Trails and Preference Optimization

Text-Attributed Graphs (TAGs), which integrate text and graph structures, have recently gained traction, especially in web applications. However, as a graph structure, TAG representation learning (TAGRL) naturally inherits issues from Graph Neural Networks (GNNs), such as fairness. Moreover, previous TAGRL research has mainly focused on using LM-as-encoder to boost downstream task performance, with little consideration given to whether this process may raise additional concerns related to fairness and other safety-related issues. As the first work to explore fairness in TAGRL, this paper proposes the concept of evolving LM-as-encoder to LM-as-fair-encoder, developing a two-stage fairness-aware alignment process called FairTAG based on the observed issues. Specifically, we first mitigate the tendency of LMs to overfit to homophily during downstream tasks fine-tuning, followed by subgraph-level connection behavior preference optimization for selected anchor nodes. We provide theoretical support and demonstrate the feasibility of LM-as-fair-encoder through extensive experiments and ablation studies. We also show that FairTAG can be seamlessly integrated with fairness-enhancing strategies on the GNNs decoder side, thus innovatively constructing a plug-and-play learning framework.

Fair Text-Attributed Graph Representation Learning

Influence functions are important for quantifying the impact of individual training data points on a model's predictions. Although extensive research has been conducted on influence functions in traditional machine learning models, their application to large language models (LLMs) has been limited. In this work, we conduct a systematic study to address a key question: do influence functions work on LLMs? Specifically, we evaluate influence functions across multiple tasks and find that they consistently perform poorly in most settings. Our further investigation reveals that their poor performance can be attributed to: (1) inevitable approximation errors when estimating the iHVP component due to the scale of LLMs, (2) uncertain convergence during fine-tuning, and, more fundamentally, (3) the definition itself, as changes in model parameters do not necessarily correlate with changes in LLM behavior. Thus, our study suggests the need for alternative approaches for identifying influential samples.

Do Influence Functions Work on Large Language Models?

Large language models (LLMs) are increasingly integral as productivity assistants, but existing benchmarks fall short in rigorously evaluating their real-world instruction-following capabilities. Current benchmarks often (i) lack sufficient multilinguality, (ii) fail to capture the implicit constraints inherent in user requests, and (iii) overlook the complexities of multi-turn dialogue. To address these critical gaps and provide a more realistic assessment, we introduce ProductivityBench, a novel benchmark specifically designed for LLM-based productivity assistants. ProductivityBench distinguishes itself by featuring input prompts across 12 languages, incorporating intra-instance multilingual instructions, employing rigorous evaluation criteria to capture both explicit and implicit constraints, and including complex multi-turn dialogue scenarios with both accumulating constraints and context switches. Furthermore, to ensure the reliability evaluation, we refined constraints using an LLM validator. Extensive experiments demonstrate that ProductivityBench presents significantly greater challenges than existing benchmarks; for instance, a strong model like GPT-o1 achieved only a 69.07% overall pass rate. ProductivityBench offers a demanding and realistic assessment of LLM in practical productivity settings, highlighting their capabilities and limitations.

Downloads

Next from EMNLP 2025

Permitted Knowledge Boundary: Evaluating the Knowledge-Constrained Responsiveness of Large Language Models

Similar lecture

C-LLM: Learn to Check Chinese Spelling Errors Character by Character