Austria

To enhance reasoning capabilities, research has explored incorporating special-purpose tokens into the training process to strengthen the learning mechanism of transformer-based LLMs. Building on prior research that showed inserting dummy tokens consecutively just before reasoning steps can enhance effectiveness, we introduce a novel approach termed Dynamic Inserting tokens Training (DIT). Our method identifies positions within sequences where model confidence is lowest identified by token log likelihood, strategically inserting [PAUSE] tokens to bolster the model&#39;s predictive capabilities for subsequent tokens. Experimental results across diverse datasets and models, from 2.7B model to 8B model, demonstrate that DIT consistently outperforms traditional fine-tuning and previous token insertion methods. With this simple yet impactful method, we achieve accuracy gains of up to 4.7%p on GSM8K, 3.23%p on AQUA-RAT, and pass@1 improvements of up to 3.4%p on MBPP. Our work shows a model-based, dynamic approach rather than a heuristic one, thereby broadening the scope of research in reasoning.

ACL 2025

Learning to Insert [PAUSE] Tokens for Better Reasoning

reasoning

### 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)

To enhance reasoning capabilities, research has explored incorporating special-purpose tokens into the training process to strengthen the learning mechanism of transformer-based LLMs. Building on prior research that showed inserting dummy tokens consecutively just before reasoning steps can enhance effectiveness, we introduce a novel approach termed Dynamic Inserting tokens Training (DIT). Our method identifies positions within sequences where model confidence is lowest identified by token log likelihood, strategically inserting [PAUSE] tokens to bolster the model's predictive capabilities for subsequent tokens. Experimental results across diverse datasets and models, from 2.7B model to 8B model, demonstrate that DIT consistently outperforms traditional fine-tuning and previous token insertion methods. With this simple yet impactful method, we achieve accuracy gains of up to 4.7%p on GSM8K, 3.23%p on AQUA-RAT, and pass@1 improvements of up to 3.4%p on MBPP. Our work shows a model-based, dynamic approach rather than a heuristic one, thereby broadening the scope of research in reasoning.

poster

Many applications of Large Language Models (LLMs) require them to either simulate people or offer personalized functionality, making the demographic representativeness of LLMs crucial for equitable utility. At the same time, we know little about the extent to which these models actually reflect the demographic attributes and behaviors of certain groups or populations in valid ways. Empirical research is divided, with conflicting findings on whether LLMs can accurately represent human (sub)populations. To shed light on this debate, we review 211 papers on the demographic representativeness of LLMs. We find that while 29% of the studies report positive conclusions, 30% of these do not evaluate LLMs across multiple demographic categories or within demographic subcategories. Another 35-47% fail to specify these subcategories altogether for gender and race. Out of the papers that do report subcategories, a majority do not include results on marginalized groups in their study. This suggests an inflated perception of LLM representativeness in the broader community. We recommend more precise evaluation methods and comprehensive documentation to ensure the responsible use of LLMs for social applications.

Do Large Language Models Represent the People? A Systematic Literature Review on the Demographic Representativeness of Large Language Models

Finetuning large language models with a variety of instruction-response pairs has enhanced their capability to understand and follow instructions. Current instruction tuning primarily relies on teacher models or human intervention to generate and refine the instructions and responses for training, which are costly, non-sustainable, and may lack diversity. In this paper, we introduce Mosaic Instruction Tuning (Mosaic-IT), a human/model-free compositional data synthesis method that can efficiently create rich and diverse augmentations from existing instruction tuning data to enhance the LLMs. Mosaic-IT randomly concatenates multiple instruction data into one and trains the model to produce the corresponding responses with predefined higher-level meta-instructions to strengthen its multi-step instruction-following and format-following skills. Our extensive evaluations demonstrate a superior performance and training efficiency of Mosaic-IT, which achieves consistent performance improvements over various benchmarks and a 
reduction in training costs compared with the original instruction tuning.

Mosaic-IT: Cost-Free Compositional Data Synthesis for Instruction Tuning

Natural language processing evaluation has made significant progress, largely driven by the proliferation of powerful large language mod-els (LLMs). New evaluation benchmarks are of increasing priority as the reasoning capabilities of LLMs are expanding at a rapid pace. In particular, while multi-document (MD) reasoning is an area of extreme relevance given LLM capabilities in handling longer-context inputs, few benchmarks exist to rigorously examine model behavior in this setting. Moreover, the multi-document setting is historically challenging for benchmark creation due to the expensive cost of annotating long inputs. In this work, we introduce MDBench, a new dataset for evaluating LLMs on the task of multi-document reasoning. Notably, MDBench is created through a novel synthetic generation process, allowing us to controllably and efficiently generate challenging document sets and the corresponding question-answer (QA) examples. Our novel technique operates on condensed structured seed knowledge, modifying it through LLM-assisted edits to induce MD-specific reasoning challenges. We then convert this structured knowledge into a natural text surface form, generating a document set and corresponding QA example. We analyze the behavior of popular LLMs and prompting techniques, finding that MDBench poses significant challenges for all methods, even with relatively short document sets. We also see our knowledge-guided generation technique (1) allows us to readily perform targeted analysis of MD-specific reasoning capabilities and (2) can be adapted quickly to account for new challenges and future modeling improvements.

MDBench: A Synthetic Multi-Document Reasoning Benchmark Generated with Knowledge Guidance

Speculative Decoding (SD) enforces strict distributional equivalence to the target model, limiting potential speed ups as distributions of near-equivalence achieve comparable outcomes in many cases. Furthermore, enforcing distributional equivalence means that users are unable to trade deviations from the target model distribution for further inference speed gains. To address these limitations, we introduce Fuzzy Speculative Decoding (FSD) - a decoding algorithm that generalizes SD by accepting candidate tokens purely based on the divergences between the target and draft model distributions. By allowing for controlled divergence from the target model, FSD enables users to flexibly trade generation quality for inference speed. Across several benchmarks, our method is able to achieve significant runtime improvements of over 5 tokens per second faster than SD at only an approximate 2% absolute reduction in benchmark accuracy. In many cases, FSD is even able to match SD benchmark accuracy at over 2 tokens per second faster, demonstrating that distributional equivalence is not necessary to maintain target model performance.

Fuzzy Speculative Decoding for a Tunable Accuracy-Runtime Tradeoff

Recent advancements in large language models (LLMs) have demonstrated remarkable capabilities in complex reasoning tasks, such as math problem-solving and code generation. However, multi-hop question answering (MHQA) over long contexts, which demands both robust knowledge-intensive reasoning and efficient processing of lengthy documents, remains a significant challenge. Existing approaches often struggle to balance these requirements, either neglecting explicit reasoning or incurring expensive computational costs due to full-attention mechanisms over long contexts. To address this, we propose **Search-in-Context (SIC)**, a novel framework that integrates Monte Carlo Tree Search (MCTS) with dynamic key-value (KV) retrieval to enable iterative, context-aware reasoning. SIC dynamically retrieves critical KV pairs (e.g., 4K tokens) at each step, prioritizing relevant evidence while mitigating the "lost in the middle" problem. Furthermore, the paper introduces a Process-Reward Model (PRM) trained on auto-labeled data to guide the MCTS process with stepwise rewards, promoting high-quality reasoning trajectories without manual annotation. Experiments on three long-context MHQA benchmarks (HotpotQA, 2WikiMultihopQA, MuSiQue) and a counterfactual multi-hop dataset demonstrate SIC’s superiority, achieving state-of-the-art performance while significantly reducing computational overhead.

Search-in-Context: Efficient Multi-Hop QA over Long Contexts via Monte Carlo Tree Search with Dynamic KV Retrieval

Knowledge-grounded question answering (QA) leverages knowledge graphs (KGs) to provide answers based on structured factual data. Existing approaches are mainly divided into Knowledge Graph Question Answering (KGQA) and Complex Query Answering (CQA). Both approaches have limitations: the first struggles when essential triplets related to the questions are missing in the given KG, while the second depends on structured first-order logic queries. To overcome these limitations, we propose a novel framework, termed Aqua-QA. Aqua-QA approximates query graphs from natural language questions, enabling reasoning over KGs. We evaluate Aqua-QA on challenging QA tasks where the entities or relations representing the input questions are excluded from the KGs, and complex logical reasoning is required to answer natural language questions. Experimental results on these datasets demonstrate that Aqua-QA outperforms existing methods, showcasing its effectiveness in handling complex reasoning tasks in knowledge-grounded QA settings.

Enhancing Complex Reasoning in Knowledge Graph Question Answering Through Query Graph Approximation

In this reproduction study, we revisit recent claims that self-attention implements kernel principal component analysis (KPCA) (Teo and Nguyen, 2024), positing that (i) value vectors $V$ capture the eigenvectors of the Gram matrix of the keys, and (ii) that self-attention projects queries onto the principal component axes of the key matrix $K$ in a feature space. Our analysis reveals three critical inconsistencies: (1) No alignment exists between learned self-attention value vectors and what is proposed in the KPCA perspective, with average similarity metrics (optimal cosine similarity $\leq 0.32$, linear CKA (Centered Kernel Alignment) $\leq 0.11$, kernel CKA $\leq 0.32$) indicating negligible correspondence; (2) Reported decreases in reconstruction loss $J_\text{proj}$, arguably justifying the claim that the self-attention
minimizes the projection error of KPCA, are misinterpreted, as the quantities involved differ by orders of magnitude ($\sim 10^3$); (3) Gram matrix eigenvalue statistics, introduced to justify that $V$ captures the eigenvector of the gram matrix, are irreproducible without undocumented implementation-specific adjustments. Across 10 transformer architectures, we conclude that the KPCA interpretation of self-attention lacks empirical support.

A Reproduction Study: The Kernel PCA Interpretation of Self-Attention Fails Under Scrutiny

Filtering and annotating textual data are routine tasks in many areas, like social media or news analytics. Automating these tasks allows to scale the analyses wrt. speed and breadth of content covered and decreases the manual effort required. Due to technical advancements in Natural Language Processing, specifically the success of large foundation models, a new tool for automating such annotation processes by using a text-to-text interface given written guidelines without providing training samples has become available. 

In this work, we assess these advancements \textit{in-the-wild} by empirically testing them in an annotation task on German Twitter data about social and political European crises. We compare the prompt-based results with our human annotation and preceding classification approaches, including Naive Bayes and a BERT-based fine-tuning/domain adaptation pipeline. Our results show that the prompt-based approach – despite being limited by local computation resources during the model selection – is comparable with the fine-tuned BERT but without any annotated training data. Our findings emphasize the ongoing paradigm shift in the NLP landscape, i.e., the unification of downstream tasks and elimination of the need for pre-labeled training data.

Zero-shot prompt-based classification: topic labeling in times of foundation models in German Tweets

Many pre-trained language models (PLMs) exhibit suboptimal performance on mid- and low-resource languages, largely due to limited exposure to these languages during pre-training. A common strategy to address this is to introduce new tokens specific to the target languages, initialize their embeddings, and apply continual pre-training on target-language data. Among such methods, OFA (Liu et al., 2024a) proposes a similarity-based subword embedding initialization heuristic that is both effective and efficient. However, OFA restricts target-language token embeddings to be convex combinations of a fixed number of source-language embeddings, which may limit expressiveness. To overcome this limitation, we propose HYPEROFA, a hypernetwork-based approach for more adaptive token embedding initialization. The hypernetwork is trained to map from an external multilingual word vector space to the PLM’s token embedding space using source-language tokens. Once trained, it can generate flexible embeddings for target-language tokens, serving as a good starting point for continual pretraining. Experiments demonstrate that HYPEROFA consistently outperforms random initialization baseline and matches or exceeds the performance of OFA in both continual pre-training convergence and downstream task performance. We make the code publicly available.

HYPEROFA: Expanding LLM Vocabulary to New Languages via Hypernetwork-Based Embedding Initialization

Deception is the intentional practice of twisting information. It is a nuanced societal practice deeply intertwined with human societal evolution, characterized by a multitude of facets. This research explores the problem of deception through the lens of psychology, employing a framework that categorizes deception into three forms: lies of omission, lies of commission, and lies of influence. The primary focus of this study is specifically on investigating only lies of omission. We propose a novel framework for deception detection leveraging NLP techniques. We curated an annotated dataset of 876,784 samples by amalgamating a popular large-scale fake news dataset and scraped news headlines from the Twitter handle of "Times of India", a well-known Indian news media house. Each sample has been labeled with four layers, namely: (i) the type of omission (speculation, bias, distortion, sounds factual, and opinion), (ii) colors of lies (black, white, grey, and red), and (iii) the intention of such lies (to influence, gain social prestige, etc) (iv) topic of lies (political, educational, religious, racial, and ethnicity). We present a novel multi-task learning [MTL] pipeline that leverages the dataless merging of fine-tuned language models to address the deception detection task mentioned earlier. Our proposed model achieved an impressive F1 score of 0.87, demonstrating strong performance across all layers including the type, color, intent, and topic aspects of deceptive content. Finally, our research aims to explore the relationship between the lies of omission and propaganda techniques. To accomplish this, we conducted an in-depth analysis, uncovering compelling findings. For instance, our analysis revealed a significant correlation between loaded language and opinion, shedding light on their interconnectedness. To encourage further research in this field, we will publicly release the SEPSIS dataset and models.

Downloads

Next from ACL 2025

Do Large Language Models Represent the People? A Systematic Literature Review on the Demographic Representativeness of Large Language Models

Similar lecture

Disentangling Reasoning Tokens and Boilerplate Tokens For Language Model Fine-tuning