ACL 2025 Papers

As the scale of large language models (LLMs) grows and natural language tasks become increasingly diverse, Parameter-Efficient Fine-Tuning (PEFT) has become the standard paradigm for fine-tuning LLMs. Among PEFT methods, LoRA is widely adopted for not introducing additional inference overhead. However, existing LoRA’s shared parameter space paradigm introduces parameter interference, leading to a gap in generalization performance for specific tasks compared to full fine-tuning. To address this issue, we propose a parameter-separated low-rank adapter, called Subspace Low-Rank Adaptation (SuLoRA). The core idea of SuLoRA is to account for task differences by decomposing LoRA’s parameter matrix into multiple independent subspaces and assigning them differentially to distinct tasks. This prevents interference across tasks and enhances the effectiveness of low-rank adaptation. Additionally, SuLoRA achieves higher rank expansion by freezing the A matrix, further improving generalization capability. We conduct extensive experiments on various NLP tasks, demonstrating that SuLoRA significantly outperforms LoRA in trainable parameter efficiency and overall model performance. Furthermore, we validate SuLoRA’s effectiveness in domain generalization and multi-modal tasks, showcasing its strong generalization ability.

Model pruning in transformer-based language models, traditionally seen as a means of computational savings, can enhance the model’s reasoning capabilities. In this work, we uncover the surprising phenomenon that the selective pruning of certain attention heads leads to improvements in reasoning performance, particularly on challenging tasks. Motivated by this observation, we propose SPRINT, a novel contrastive learning framework that dynamically selects the optimal head and layer to prune during inference. By aligning question embeddings with head embeddings, our approach identifies those pruned-head configurations that result in more accurate reasoning. Extensive experiments on the MATH dataset demonstrate that our method significantly outperforms traditional best-of-N and random head selection strategies on the MATH500 and GSM8K datasets.

Analysing the differences in how events are represented across texts, or verifying whether the language model generations hallucinate, requires the ability to systematically compare their content. To support such comparison, structured representation that captures fine-grained information plays a vital role.In particular, identifying distinct atomic facts and the discourse relations connecting them enables deeper semantic comparison. Our proposed approach combines structured discourse information extraction with a classifier, FDSpotter, for factual consistency verification. We show that adversarial discourse relations pose challenges for language models, but fine-tuning on our annotated data, DiscInfer, achieves competitive performance. Our proposed approach advances factual consistency verification by grounding in linguistic structure and decomposing it into interpretable components. We demonstrate the effectiveness of our method on the evaluation of two tasks: data-to-text generation and text summarisation. Our code and dataset will be publicly available on GitHub.

This paper introduces a pioneering methodology, termed StructTuning, to efficiently transform foundation Large Language Models (LLMs) into domain specialists. It significantly reduces the training corpus needs to a mere 5% while achieving an impressive 100% of traditional knowledge injection performance. Motivated by structured human education, we propose a novel two-stage strategy for knowledge injection and alignment: Structure-aware Continual Pre-Training (SCPT) and Structure-aware Supervised Fine-Tuning (SSFT). In the SCPT phase, we automatically extract the domain knowledge taxonomy and reorganize the training corpora, enabling LLMs to effectively link textual segments to targeted knowledge points within the taxonomy. In the SSFT phase, we explicitly prompt models to elucidate the underlying knowledge structure in their outputs, leveraging the structured domain insight to address practical problems. Our ultimate method was extensively evaluated across model architectures and scales on LongBench and MMedBench datasets, demonstrating superior performance against other knowledge injection methods. We also explored our method’s scalability across different training corpus sizes, laying the foundation to enhance domain-specific LLMs with better data utilization.

The rapid proliferation of fake news across multiple domains poses significant threats to society. Existing multi-domain detection models typically capture domain-shared semantic features to achieve generalized detection. However, they often fail to generalize well due to poor adaptability, which limits their ability to provide complementary features for detection, especially in data-constrained conditions. To address these challenges, we investigate the propagation-adaptive multi-domain fake news detection paradigm. We propose a novel framework, Structure-adaptive Adversarial Contrastive Learning (StruACL), to adaptively enable structure knowledge transfer between multiple domains. Specifically, we first contrast representations between content-only and propagation-rich data to preserve structural patterns in the shared representation space. Additionally, we design a propagation-guided adversarial training strategy to enhance the diversity of representations. Under the StruACL objective, we leverage a unified Transformer-based and graph-based model to jointly learn transferable semantic and structural features for detection across multiple domains. Experiments on seven fake news datasets demonstrate that StruACL-TGN achieves better multi-domain detection performance on general and data-constrained scenarios, showing the effectiveness and better generalization of StruACL.

Recently, large language models (LLMs) have shown significant progress, approaching human perception levels. In this work, we demonstrate that despite these advances, LLMs still struggle to reason using molecular structural information. This gap is critical because many molecular properties, including functional groups, depend heavily on such structural details. To address this limitation, we propose an approach that sketches molecular structures for reasoning. Specifically, we introduce Molecular Structural Reasoning (MSR) framework to enhance the understanding of LLMs by explicitly incorporating the key structural features. We present two frameworks for scenarios where the target molecule is known or unknown. We verify that our MSR improves molecular understanding through extensive experiments.

Although Transformers-based architectures excel at processing textual information, their naive adaptation for tabular data often involves flattening the table structure. This simplification can lead to the loss of essential inter-dependencies between rows, columns, and cells, while also posing scalability challenges for large tables. To address these issues, prior works have explored special tokens, structured embeddings, and sparse attention patterns. In this paper, we conduct a comprehensive analysis of tabular encoding techniques used in QA, which highlights the crucial role of attention sparsity in preserving structural information of tables. We also introduce a set of novel sparse attention mask designs for tabular data, that not only enhance computational efficiency but also preserve structural integrity, leading to better overall performance.

Multi-turn instruction following capability constitutes a core competency of large language models (LLMs) in real-world applications. Existing evaluation benchmarks predominantly focus on fine-grained constraint satisfaction and domain-specific capability assessment, yet overlook the crucial structural dependencies between dialogue turns that distinguish multi-turn from single-turn interactions. These structural dependencies not only reflect user intent but also establish an essential second dimension for the instruction following evaluation beyond constraint satisfaction. To address this gap, we propose StructFlowBench, a multi-turn instruction following benchmark with structural flow modeling. The benchmark defines an innovative structural flow framework with six fundamental inter-turn relationships. These relationships introduce novel structural constraints for model evaluation and also serve as generation parameters for creating customized dialogue flows tailored to specific scenarios. Adopting established LLM-based automatic evaluation methodologies, we conduct systematic evaluations of 13 leading open-source and closed-source LLMs. Experimental results reveal significant deficiencies in current models’ comprehension of multi-turn dialogue structures. The code is available at https://github.com/MLGroupJLU/StructFlowBench.

Large language models (LLMs) have made significant strides in natural language processing by leveraging their ability to comprehend and reason with factual knowledge. However, a significant amount of factual knowledge is stored in structured data, which has unique characteristics not typically encountered in the unstructured texts used for pretraining LLMs. To evaluate the capability of LLMs in handling facts structurally stored, we introduce a benchmark called StructFact, which includes meticulously annotated factual questions, spanning five tasks that reflect the intrinsic properties of structured data. This benchmark aims to delineate the strengths and limitations of LLMs in reasoning with structured data for knowledge-intensive tasks in practical applications. Extensive experiments conducted on 10 common LLMs have yielded several insights, one notable finding being that these models struggle significantly with the heterogeneity of structured data during reasoning.

The effective utilization of structured data, integral to corporate data strategies, has been challenged by the rise of large language models (LLMs) capable of processing unstructured information. This shift prompts the question: can LLMs interpret structured data directly in its unstructured form? We propose an automatic evaluation data generation method for assessing LLMs’ reasoning capabilities on structure-rich text to explore this. Our approach supports 8 structured languages and 29 tasks, generating data with adjustable complexity through controllable nesting and structural width. We introduce StrucText-Eval, a benchmark containing 5,800 pre-generated and annotated samples designed to evaluate how well LLMs understand and reason through structured text. StrucText-Eval is divided into two suites: a regular Test suite (3,712 samples) and a Test-Hard suite (2,088 samples), the latter emphasizing the gap between human and model performance on more complex tasks. Experimental results show that while open-source LLMs achieve a maximum accuracy of 74.9% on the standard dataset, their performance drops significantly to 45.8% on the harder dataset. In contrast, human participants reach an accuracy of 92.6% on StrucText-Eval-Hard, highlighting LLMs’ current limitations in handling intricate structural information.

Recent advancements in Generative AI and Large Language Models (LLMs) have enabled the creation of highly realistic synthetic content, raising concerns about the potential for malicious use, such as misinformation and manipulation. Moreover, detecting Machine-Generated Text (MGT) remains challenging due to the lack of robust benchmarks that assess generalization to real-world scenarios. In this work, we evaluate the resilience of state-of-the-art MGT detectors (e.g., Mage, Radar, LLM-DetectAIve) to linguistically informed adversarial attacks. We develop a pipeline that fine-tunes language models using Direct Preference Optimization (DPO) to shift the MGT style toward human-written text (HWT), obtaining generations more challenging to detect by current models. Additionally, we analyze the linguistic shifts induced by the alignment and how detectors rely on “linguistic shortcuts” to detect texts. Our results show that detectors can be easily fooled with relatively few examples, resulting in a significant drop in detecting performances. This highlights the importance of improving detection methods and making them robust to unseen in-domain texts. We release code, models, and data to support future research on more robust MGT detection benchmarks.

In Retrieval-Augmented Generation (RAG) and agent-based frameworks, the “Chain of Models” approach is widely used, where multiple specialized models work sequentially on distinct sub-tasks. This approach is effective but increases resource demands as each model must be deployed separately. Recent advancements attempt to address this by applying prompt tuning, which allows a shared base model to adapt to multiple tasks with minimal parameter changes. However, a key challenge remains: intermediate outputs, passed between models as plain text, require recomputation of hidden states (i.e., Key and Value (KV) states in Transformers) during inference. In this paper, we introduce FTHSS, a novel prompt-tuning method that enables models to share KV hidden states, eliminating redundant forward passes and reducing KV cache storage. By modifying input and attention masks during training, FTHSS allows models to effectively utilize KV hidden states from prior models in both single- and multi-round scenarios. Empirical results on four tasks show that FTHSS matches the performance of traditional model chains while improving inference efficiency.

Macroeconomic fluctuations and the narratives that shape them form a mutually reinforcing cycle: public discourse can spur behavioural changes leading to economic shifts, which then result in changes in the stories that propagate. We show that shifts in semantic embedding space can be causally linked to real-world market shocks or deviations from the expected market behaviour (sec:market_shocks). Furthermore, we show how partisanship can influence the predictive power of text for market fluctuations and shape reactions to those same shocks. We also provide some evidence that text-based signals are particularly salient during rare events such as COVID-19, highlighting the value of language data as an exogenous variable in economic forecasting. Our findings underscore the bidirectional relationship between news outlets and market shocks, offering a novel empirical approach to studying their effect on each other.

The widespread success of LLMs on NLP benchmarks has been accompanied by concerns that LLMs function primarily as stochastic parrots that reproduce texts similar to what they saw during pre-training, often erroneously. But what is the nature of their errors, and do these errors exhibit any regularities? In this work, we examine irrelevant context hallucinations, in which models integrate misleading contextual cues into their predictions. Through behavioral analysis, we show that these errors result from a structured yet flawed mechanism that we term _class-based (mis)generalization_, in which models combine abstract class cues with features extracted from the query or context to derive answers. Furthermore, mechanistic interpretability experiments on Llama-3, Mistral, and Pythia across 39 factual recall relation types reveal that this behavior is reflected in the model’s internal computations: (i) abstract class representations are constructed in lower layers before being refined into specific answers in higher layers, (ii) feature selection is governed by two competing circuits — one prioritizing direct query-based reasoning, the other incorporating contextual cues — whose relative influences determine the final output. Our findings provide a more nuanced perspective on the stochastic parrot argument: through form-based training, LLMs can exhibit generalization leveraging abstractions, albeit in unreliable ways based on contextual cues — what we term _stochastic chameleons_.

The rapid evolution of large language models (LLMs) has revolutionized natural language processing (NLP) tasks such as text generation, translation, and comprehension. However, the increasing computational demands and inference costs of these models present significant challenges. This study investigates the dynamic and efficient utilization of pre-trained weights from open-sourced LLMs of varying parameter sizes to achieve an optimal balance between computational efficiency and task performance. Drawing inspiration from the dual-process theory of human cognition, we introduce StitchLLM: a dynamic model routing framework that employs a powerful bottom model to process all queries, and uses a lightweight routing mechanism to allocate computational resources appropriately. Our novel framework optimizes efficiency and maintains performance, leveraging a trainable stitching layer for seamless integration of decoder layers across different LLMs. Experimental results demonstrate that StitchLLM improves system throughput while minimizing performance degradation, offering a flexible solution for deploying LLMs in resource-constrained settings.

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.

While personalized recommendations are often desired by users, it can be difficult in practice to distinguish cases of bias from cases of personalization: we find that models generate racially stereotypical recommendations regardless of whether the user revealed their identity intentionally through explicit indications or unintentionally through implicit cues. We demonstrate that when people use large language models (LLMs) to generate recommendations, the LLMs produce responses that reflect both what the user wants and who the user is. We argue that chatbots ought to transparently indicate when recommendations are influenced by a user’s revealed identity characteristics, but observe that they currently fail to do so. Our experiments show that even though a user’s revealed identity significantly influences model recommendations (p < 0.001), model responses obfuscate this fact in response to user queries. This bias and lack of transparency occurs consistently across multiple popular consumer LLMs and for four American racial groups.

Bias and stereotypes in language models can cause harm, especially in sensitive areas like content moderation and decision-making. This paper addresses bias and stereotype detection by exploring how jointly learning these tasks enhances model performance. We introduce StereoBias, a unique dataset labeled for bias and stereotype detection across five categories: religion, gender, socio-economic status, race, profession, and others, enabling a deeper study of their relationship. Our experiments compare encoder-only models and fine-tuned decoder-only models using QLoRA. While encoder-only models perform well, decoder-only models also show competitive results. Crucially, joint training on bias and stereotype detection significantly improves bias detection compared to training them separately. Additional experiments with sentiment analysis confirm that the improvements stem from the connection between bias and stereotypes, not multi-task learning alone. These findings highlight the value of leveraging stereotype information to build fairer and more effective AI systems.

Large language models (LLMs) have made remarkable strides in complex reasoning tasks, but their safety and robustness in reasoning processes remain unexplored, particularly in third-party platforms that facilitate user interactions via APIs. Existing attacks on LLM reasoning are constrained by specific settings or lack of imperceptibility, limiting their feasibility and generalizability. To address these challenges, we propose the Stepwise rEasoning Error Disruption (SEED) attack, which subtly injects errors into prior reasoning steps to mislead the model into producing incorrect subsequent reasoning and final answers. Unlike previous methods, SEED is compatible with zero-shot and few-shot settings, maintains the natural reasoning flow, and ensures covert execution without modifying the instruction. Extensive experiments on four datasets across four different models demonstrate SEED’s effectiveness, revealing the vulnerabilities of LLMs to disruptions in reasoning processes. These findings underscore the need for greater attention to the robustness of LLM reasoning to ensure safety in practical applications. Our code is available at: https://github.com/Applied-Machine-Learning-Lab/SEED-Attack