Austria

Despite the widespread use of Transformer-based text embedding models in NLP tasks, surprising ``sticky tokens&#39;&#39; can undermine the reliability of embeddings. These tokens, when repeatedly inserted into sentences, pull sentence similarity toward a certain value, disrupting the normal distribution of embedding distances and degrading downstream performance. In this paper, we systematically investigate such anomalous tokens, formally defining them and introducing an efficient detection method, Sticky Token Detector (STD), based on sentence and token filtering. Applying STD to 40 checkpoints across 14 model families, we discover a total of 868 sticky tokens. Our analysis reveals that these tokens often originate from special or unused entries in the vocabulary, as well as fragmented subwords from multilingual corpora. Notably, their presence does not strictly correlate with model size or vocabulary size. We further evaluate how sticky tokens affect downstream tasks like clustering and retrieval, observing significant performance drops of up to 50\%. Through attention-layer analysis, we show that sticky tokens disproportionately dominate the model’s internal representations, raising concerns about tokenization robustness. Our findings show the need for better tokenization strategies and model design to mitigate the impact of sticky tokens in future text embedding applications.

ACL 2025

Sticking to the Mean: Detecting Sticky Tokens in Text Embedding Models

anomalous token

sticky token

phrase/sentence embedding

semantic textual similarity

### 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**. 

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

Despite the widespread use of Transformer-based text embedding models in NLP tasks, surprising ``sticky tokens'' can undermine the reliability of embeddings. These tokens, when repeatedly inserted into sentences, pull sentence similarity toward a certain value, disrupting the normal distribution of embedding distances and degrading downstream performance. In this paper, we systematically investigate such anomalous tokens, formally defining them and introducing an efficient detection method, Sticky Token Detector (STD), based on sentence and token filtering. Applying STD to 40 checkpoints across 14 model families, we discover a total of 868 sticky tokens. Our analysis reveals that these tokens often originate from special or unused entries in the vocabulary, as well as fragmented subwords from multilingual corpora. Notably, their presence does not strictly correlate with model size or vocabulary size. We further evaluate how sticky tokens affect downstream tasks like clustering and retrieval, observing significant performance drops of up to 50\%. Through attention-layer analysis, we show that sticky tokens disproportionately dominate the model’s internal representations, raising concerns about tokenization robustness. Our findings show the need for better tokenization strategies and model design to mitigate the impact of sticky tokens in future text embedding applications.

poster

Although large language models (LLMs) excel in knowledge recall and reasoning, their static nature leads to outdated information as the real world evolves or when adapting to domain-specific knowledge, highlighting the need for effective knowledge injection. However, current research on knowledge injection remains superficial, mainly focusing on knowledge memorization and retrieval. This paper proposes a four-tier knowledge injection framework that systematically defines the levels of knowledge injection: memorization, retrieval, reasoning, and association. Based on this framework, we introduce DeepKnowledge, a synthetic experimental testbed designed for fine-grained evaluation of the depth of knowledge injection across three knowledge types (novel, incremental, and updated). We then explore various knowledge injection scenarios and evaluate the depth of knowledge injection for each scenario on the benchmark. Experimental results reveal key factors to reach each level of knowledge injection for LLMs and establish a mapping between the levels of knowledge injection and the corresponding suitable injection methods, aiming to provide a comprehensive approach for efficient knowledge injection across various levels. The code is available at [https://github.com/icip-cas/Knowledge-Learning-Toolkits](https://github.com/icip-cas/Knowledge-Learning-Toolkits).

Memorizing is Not Enough: Deep Knowledge Injection Through Reasoning

In dialogue state tracking (DST), in-context learning comprises a retriever that selects labeled dialogues as in-context examples and a DST model that uses these examples to infer the dialogue state of the query dialogue. Existing methods for constructing training data for retrievers suffer from three key limitations: (1) the synergistic effect of examples is not considered, (2) the linguistic characteristics of the query are not sufficiently factored in, and (3) scoring is not directly optimized for DST performance. Consequently, the retriever can fail to retrieve examples that would substantially improve DST performance. To address these issues, we present CombiSearch—a method that scores effective in-context examples based on their combinatorial impact on DST performance. 
Our evaluation on MultiWOZ shows that retrievers trained with CombiSearch surpass state-of-the-art models, achieving a 20× gain in data efficiency and generalizing well to the SGD dataset. 
Moreover, CombiSearch attains a 12\% absolute improvement in the upper bound DST performance over traditional approaches when no retrieval errors are assumed. 
This significantly increases the headroom for practical DST performance while demonstrating that existing methods rely on suboptimal data for retriever training.

Improving Dialogue State Tracking through Combinatorial Search for In-Context Examples

Efficient processing of long contexts in large language models (LLMs) is essential for real-world applications like retrieval-augmented generation and in-context learning, especially in resource-constrained environments such as edge computing. This paper explores the embedding-based context compression to reduce inference costs while preserving the downstream LLM configurations. We propose a decoupled compressor-LLM framework, pretrained on text reconstruction and completion tasks, designed to effectively preserve essential contextual information within condensed embedding representations. Our extensive experiments investigate pretraining, model configurations, compression rates, efficiency across tasks, and adaptability to various LLMs. Results demonstrate that our approach outperforms competitive baselines in three domains and across eight datasets while being adaptable to different downstream LLMs. We find that thorough pretraining and carefully selected compression rates, such as 4x and 16x, enable a lightweight compressor to achieve a good balance between accuracy and speed. These findings underscore the potential of embedding-based compression to enhance LLM efficiency and motivate further research in this area.

Pretraining Context Compressor for Large Language Models with Embedding-Based Memory

Emotional support dialogue systems aim to reduce help-seekers' distress and help them overcome challenges. While human values—core beliefs that shape an individual's priorities—are increasingly emphasized in contemporary psychological therapy for their role in fostering internal transformation and long-term emotional well-being, their integration into emotional support systems remains underexplored. To bridge this gap, we present a value-driven method for training emotional support dialogue systems designed to reinforce positive values in seekers. Notably, our model identifies which values to reinforce at each turn and how to do so, by leveraging online support conversations from Reddit. We evaluate the method across support skills, seekers' emotional intensity, and value reinforcement. Our method consistently outperforms various baselines, effectively exploring and eliciting values from seekers. Additionally, leveraging crowd knowledge from Reddit significantly enhances its effectiveness. Therapists highlighted its ability to validate seekers' challenges and emphasize positive aspects of their situations—both crucial elements of value reinforcement. Our work, being the first to integrate value reinforcement into emotional support systems, demonstrates its promise and establishes a foundation for future research.

Dialogue Systems for Emotional Support via Value Reinforcement

Retrieval-Augmented Generation (RAG) has been proven to be an effective approach to address the hallucination problem in large language models (LLMs). In current RAG systems, LLMs typically need to synthesize knowledge provided by two main external sources (user prompts and an external database) to generate a final answer. When the knowledge provided by the user conflicts with that retrieved from the database, a critical question arises: Does the LLM favor one knowledge source over the other when generating the answer? In this paper, we are the first to unveil a new phenomenon, Authority Bias, where the LLMs tend to favor the knowledge provided by the user even when it deviates from the facts; this new phenomenon is rigorously evidenced via our novel and comprehensive characterization of Authority Bias in six widely used LLMs and across diverse task scenarios. We propose a novel dataset specifically designed for detecting Authority Bias, called the Authority Bias Detection Dataset (ABDD), and introduce new, detailed metrics to measure Authority Bias. To mitigate Authority bias, we finally propose the Conflict Detection Enhanced Query (CDEQ) framework. We identify the sentences and atomic information that generate conflicts, perform a credibility assessment on the conflicting paragraphs, and ultimately enhance the query to detect perturbed text, thereby reducing Authority bias. Comparative experiments with widely used mitigation methods demonstrate that CDEQ exhibits both effectiveness and advancement, significantly enhancing the robustness of RAG systems.

LLMs Trust Humans More, That’s a Problem! Unveiling and Mitigating the Authority Bias in Retrieval-Augmented Generation

While Large Language Models (LLMs) demonstrate remarkable capabilities, their ability to autonomously execute complex real-world tasks remains limited. Accordingly, tool learning has emerged to enable LLMs to effectively leverage external tools to extend their capabilities. Current tool-learning paradigms like CoT/ReAct employ sequential tool invocation but suffer from constrained perception and inadequate task planning. Alternative approaches using search-based decision trees incur substantial computational overhead. To address these limitations, we propose DTA-Llama (Divide-Then-Aggregate Llama), a novel parallel tool invocation framework featuring: (1) A Directed Acyclic Graph (DAG) structure that transformed from traditional tree-based tool search paths, enabling parallel execution and contributing high-quality training data; (2) A process-thread-inspired inference mechanism that iteratively decomposes tasks into parallel tool-using subtasks while aggregating results for subsequent decisions. Experimental results show that our approach substantially enhances task performance while reducing token consumption and inference time. Llama2-7B, using our method, is comparable to the official parallel function calling method of GPT-3.5. The relevant code, dataset, and model weights are available at https://corn0205.github.io/.

Divide-Then-Aggregate: An Efficient Tool Learning Method via Parallel Tool Invocation

While many advanced LLMs are designed to handle long sequence data, we can still observe notable quality degradation even within the sequence limit. In this work, we introduce a novel approach called Scaling to Emphasize Attention for Long-context retrieval (SEAL), which enhances the retrieval performance of large language models (LLMs) over long contexts. We observe that specific attention heads are closely tied to long-context retrieval, showing positive or negative correlation with retrieval scores, and adjusting the strength of these heads boosts the quality of LLMs in long context by a large margin. Built on this insight, we propose a learning-based mechanism that leverages generated data to emphasize these heads. By applying SEAL, we achieve significant improvements in long-context retrieval performance across various tasks and models. Additionally, when combined with existing training-free context extension techniques, SEAL extends the contextual limits of LLMs while maintaining highly reliable outputs.

SEAL: Scaling to Emphasize Attention for Long-Context Retrieval

As Large Language Models (LLMs) advance, aligning them with human values is critical for their responsible development. Value principles serve as the foundation for clarifying alignment goals.
Multiple sets of value principles have been proposed, such as HHH (helpful, honest, harmless) and instructions for data synthesis in reinforcement learning from AI feedback (RLAIF). However, most of them are heuristically crafted, without consideration of three primary challenges in practical LLM alignment: 1) Comprehensiveness to deal with diverse and even unforeseen scenarios in which LLMs could be applied; 2) Precision to provide LLMs with clear and actionable guidance in specific scenarios; and 3) Compatability to avoid internal contracts between principles.
In this paper, we formalize quantitative metrics to evaluate value principles along the three desirable properties. Building on these metrics, we propose the Hierarchical Value Principle framework (HiVaP), which constructs a hierarchical principle set and retrieves principles tailored to each scenario in a cascading way, addressing above challenges.
Experimental results validate that the three metrics capture the effectiveness of value principles for LLM alignment, and our HiVaP framework that enhances these metrics leads to superior alignment. 

Warning: This paper contains several toxic and offensive statements.

Towards Better Value Principles for Large Language Model Alignment: A Systematic Evaluation and Enhancement

Text-guided image editing, fueled by recent advancements in generative AI, is becoming increasingly widespread. This trend highlights the need for a comprehensive framework to verify text-guided edits and assess their quality. To address this need, we introduce EditInspector, a novel benchmark for evaluation of text-guided image edits, based on human annotations collected using an extensive template for edit verification. We leverage EditInspector to evaluate the performance of state-of-the-art (SoTA) vision and language models in assessing edits across various dimensions, including accuracy, artifact detection, visual quality, seamless integration with the image scene, adherence to common sense, and the ability to describe edit-induced changes. Our findings indicate that current models struggle to evaluate edits comprehensively and frequently hallucinate when describing the changes. To address these challenges, we propose two novel methods that outperform SoTA models in both artifact detection and difference caption generation.

EditInspector: A Benchmark for Evaluation of Text-Guided Image Edits

We introduce a novel framework for analyzing sorting algorithms in pairwise ranking prompting (PRP), re-centering the cost model around LLM inferences rather than traditional pairwise comparisons. While classical metrics based on comparison counts have traditionally been used to gauge efficiency, our analysis reveals that expensive LLM inferences overturn these predictions; accordingly, our framework encourages strategies such as batching and caching to mitigate inference costs. We show that algorithms optimal in the classical setting can lose efficiency when LLM inferences dominate the cost under certain optimizations.

Downloads

Next from ACL 2025

Memorizing is Not Enough: Deep Knowledge Injection Through Reasoning

Similar lecture

A Text is Worth Several Tokens: Text Embedding from LLMs Secretly Aligns Well with The Key Tokens

Stay up to date with the latest Underline news!

PRODUCT

COMPANY

RESOURCES