Austria

Recently, scaling test-time compute on Large Language Models (LLM) has garnered wide attention. However, there has been limited investigation of how various reasoning prompting strategies perform as scaling. In this paper, we focus on a standard and realistic scaling setting: majority voting. We systematically conduct experiments on 6 LLMs $\times$ 8 prompting strategies $\times$ 6 benchmarks. 
Experiment results consistently show that as the sampling time and computational overhead increase, complicated prompting strategies with superior initial performance gradually fall behind simple Chain-of-Thought.
We analyze this phenomenon and provide theoretical proofs. 
Additionally, we propose a probabilistic method to efficiently predict scaling performance and identify the best prompting strategy under large sampling times, eliminating the need for resource-intensive inference processes in practical applications.
Furthermore, we introduce two ways derived from our theoretical analysis to significantly improve the scaling performance. We hope that our research can promote to re-examine the role of complicated prompting, unleash the potential of simple prompting strategies, and provide new insights for enhancing test-time scaling performance. Code is available at https://github.com/MraDonkey/rethinking_prompting.

ACL 2025

Rethinking the Role of Prompting Strategies in LLM Test-Time Scaling: A Perspective of Probability Theory

findings and theory

majoritiy voting test-time scaling

large language models

## 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*

EMNLP 2025

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

poster

### 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)

This paper explores whether enhancing temporal reasoning capabilities in Large Language Models (LLMs) can improve the quality of timeline summarisation, the task of summarising long texts containing sequences of events, such as social media threads. We first introduce NarrativeReason, a novel dataset focused on temporal relationships among sequential events within narratives, distinguishing it from existing temporal reasoning datasets that primarily address pair-wise event relationships. Our approach then combines temporal reasoning with timeline summarisation through a knowledge distillation framework, where we first fine-tune a teacher model on temporal reasoning tasks and then distill this knowledge into a student model while simultaneously training it for the task of timeline summarisation. Experimental results demonstrate that our model achieves superior performance on out-of-domain mental health-related timeline summarisation tasks, which involve long social media threads with repetitions of events and a mix of emotions, highlighting the importance and generalisability of leveraging temporal reasoning to improve timeline summarisation.

Temporal reasoning for timeline summarisation in social media

Large language models (LLMs) face inherent performance bottlenecks under parameter constraints, particularly in processing critical tokens that demand complex reasoning. Empirical analysis reveals challenging tokens induce abrupt gradient spikes across layers, exposing architectural stress points in standard Transformers. Building on this insight, we propose Inner Thinking Transformer (ITT), which reimagines layer computations as implicit thinking steps. ITT dynamically allocates computation through Adaptive Token Routing, iteratively refines representations via Residual Thinking Connections, and distinguishes reasoning phases using Thinking Step Encoding. ITT enables deeper processing of critical tokens without parameter expansion. Evaluations across 162M-466M parameter models show ITT achieves 96.5% performance of a 466M Transformer using only 162M parameters, reduces training data by 43.2%, and outperforms Transformer/Loop variants in 11 benchmarks. By enabling elastic computation allocation during inference, ITT balances performance and efficiency through architecture-aware optimization of implicit thinking pathways.

Inner Thinking Transformer: Leveraging Dynamic Depth Scaling to Foster Adaptive Internal Thinking

Instruction-following large language models (LLMs), such as ChatGPT, have become widely popular among everyday users. However, these models inadvertently disclose private, sensitive information to their users, underscoring the need for machine unlearning techniques to remove selective information from the models. While prior work has focused on forgetting small, random subsets of training data at the instance-level, we argue that real-world scenarios often require the removal of an entire user data, which may require a more careful maneuver. In this study, we explore entity-level unlearning, which aims to erase all knowledge related to a target entity while preserving the remaining model capabilities. To address this, we introduce Opt-Out, an optimal transport-based unlearning method that utilizes the Wasserstein distance from the model's initial parameters to achieve more effective and fine-grained unlearning. We also present the first Entity-Level Unlearning Dataset (ELUDe) designed to evaluate entity-level unlearning. Our empirical results demonstrate that Opt-Out surpasses existing methods, establishing a new standard for secure and adaptable LLMs that can accommodate user data removal requests without the need for full retraining.

Opt-Out: Investigating Entity-Level Unlearning for Large Language Models via Optimal Transport

One key characteristic of the Chinese spelling check (CSC) task is that incorrect characters are usually similar to the correct ones in either phonetics or glyph. To accommodate this, previous works usually leverage confusion sets, which suffer from two problems, i.e., difficulty in determining which character pairs to include and lack of probabilities to distinguish items in the set. In this paper, we propose a light-weight plug-and-play DISC (i.e., decoding intervention with similarity of characters) module for CSC models. DISC measures phonetic and glyph similarities between characters and incorporates this similarity information only during the inference phase. This method can be easily integrated into various existing CSC models, such as ReaLiSe, SCOPE, and ReLM, without additional training costs. Experiments on three CSC benchmarks demonstrate that our proposed method significantly improves model performance, approaching and even surpassing the current state-of-the-art models.

DISC: Plug-and-Play Decoding Intervention with Similarity of Characters for Chinese Spelling Check

Large Language Models (LLMs) have achieved impressive accomplishments in recent years. However, the increasing memory consumption of KV cache has possessed a significant challenge to the inference system. Eviction methods have revealed the inherent redundancy within the KV cache, demonstrating its potential for reduction, particularly in deeper layers. However, KV cache reduction for shallower layers has been found to be insufficient. Based on our observation that, the KV cache exhibits a high degree of similarity. Based on this observation, we proposed a novel KV cache reduction method, SpindleKV, which balances both shallow and deep layers. For deep layers, we employ an attention weight based eviction method, while for shallow layers, we apply a codebook based replacement approach which is learnt by similarity and merging policy. Moreover, SpindleKV addressed the Grouped-Query Attention (GQA) dilemma faced by other attention based eviction methods. Experiments on two common benchmarks with three different LLMs shown that SpindleKV obtained better KV cache reduction effect compared to baseline methods, while preserving similar or even better model performance.

SpindleKV: A Novel KV Cache Reduction Method Balancing Both Shallow and Deep Layers

How can we quantize large language models while preserving accuracy? Quantization is essential for deploying large language models (LLMs) efficiently. Binary-coding quantization (BCQ) and uniform quantization (UQ) are promising quantization schemes that have strong expressiveness and optimizability, respectively. However, neither scheme leverages both advantages. In this paper, we propose UniQuan$_F$ (Unified Quantization with Flexible Mapping), an accurate quantization method for LLMs. UniQuan$_F$ harnesses both strong expressiveness and optimizability by unifying the flexible mapping technique in UQ and BCQ’s non-uniform quantization levels. We propose unified initialization, and local and periodic mapping techniques to optimize the parameters in UniQuan$_F$ precisely. After optimization, our unification theorem removes computational and memory overhead, allowing us to utilize the superior accuracy of UniQuan$_F$ without extra deployment costs induced by the unification. Experimental results demonstrate that UniQuan$_F$ outperforms existing UQ and BCQ methods, achieving up to 4.60% higher accuracy on GSM8K benchmark.

Unifying Uniform and Binary-coding Quantization for Accurate Compression of Large Language Models

With the growing demand to fit fine-grained user intents, faceted query-by-example (QBE), which retrieves similar documents conditioned on specific facets, has gained recent attention. However, prior approaches mainly depend on document-level comparisons using basic indicators like citations due to the lack of facet-level relevance datasets; yet, this limits their use to citation-based domains and fails to capture the intricacies of facet constraints. In this paper, we propose a multi-facet blending (FaBle) augmentation method, which exploits modularity by decomposing and recomposing to explicitly synthesize facet-specific training sets. We automatically decompose documents into facet units and generate (ir)relevant pairs by leveraging LLMs' intrinsic distinguishing capabilities; then, dynamically recomposing the units leads to facet-wise relevance-informed document pairs. Our modularization eliminates the need for pre-defined facet knowledge or labels. Further, to prove the FaBle's efficacy in a new domain beyond citation-based scientific paper retrieval, we release a benchmark dataset for educational exam item QBE. FaBle augmentation on 1K documents remarkably assists training in obtaining facet conditional embeddings.

Multi-Facet Blending for Faceted Query-by-Example Retrieval

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.

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

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.

Downloads

Next from ACL 2025

Temporal reasoning for timeline summarisation in social media

Similar lecture

Step-level Verifier-guided Hybrid Test-Time Scaling for Large Language Models