ACL 2025 Papers

NLP research frequently grapples with multiple sources of variability—spanning runs, datasets, annotators, and more—yet conventional analysis methods often neglect these hierarchical structures, threatening the reproducibility of findings. To address this gap, we contribute a case study illustrating how linear mixed-effects models (LMMs) can rigorously capture systematic language-dependent differences (i.e., population-level effects) in a population of monolingual and multilingual language models. In the context of a bilingual hate speech detection task, we demonstrate that LMMs can uncover significant population-level effects—even under low-resource (small-N) experimental designs—while mitigating confounds and random noise. By setting out a transparent blueprint for repeated-measures experimentation, we encourage the NLP community to embrace variability as a feature, rather than a nuisance, in order to advance more robust, reproducible, and ultimately trustworthy results.

Data watermarking in language models injects traceable signals, such as specific token sequences or stylistic patterns, into copyrighted text, allowing copyright holders to track and verify training data ownership. Previous data watermarking techniques primarily focus on effective memorization after pretraining, while overlooking challenges that arise in other stages of the LLM pipeline, such as the risk of watermark filtering during data preprocessing, or potential forgetting through post-training, or verification difficulties due to API-only access. We propose a novel data watermarking approach that injects coherent and plausible yet fictitious knowledge into training data using generated passages describing a fictitious entity and its associated attributes. Our watermarks are designed to be memorized by the LLM through seamlessly integrating in its training data, making them harder to detect lexically during preprocessing. We demonstrate that our watermarks can be effectively memorized by LLMs, and that increasing our watermarks’ density, length, and diversity of attributes strengthens their memorization. We further show that our watermarks remain robust throughout LLM development, maintaining their effectiveness after continual pretraining and supervised finetuning. Finally, we show that our data watermarks can be evaluated even under API-only access via question answering.

Mitigating positional bias of language models (LMs) for listwise inputs is a well-known and important problem (e.g., lost-in-the-middle). While zero-shot order-invariant LMs have been proposed to solve this issue, their success on practical listwise problems has been limited. In this work, as a first contribution, we identify and overcome two limitations to make zero-shot invariant LMs more practical: (1) training and inference distribution mismatch arising from modifying positional ID assignments to enforce invariance, and (2) failure to adapt to mixture of order-invariant and sensitive inputs in practical listwise problems. Then, to overcome these issues we propose (1) RoToR, a zero-shot invariant LM for genuinely order-invariant inputs with minimal modifications of positional IDs, and (2) Selective Routing, an adaptive framework that handles both order-invariant and order-sensitive inputs in listwise tasks. On the Lost in the middle (LitM), Knowledge Graph QA (KGQA), and MMLU benchmarks, we show that RoToR with Selective Routing can effectively handle practical listwise input tasks in a zero-shot manner (https://github.com/soyoung97/RoToR)

We propose Row-Column Fine-Tuning(RoCoFT), a parameter-efficient fine-tuning method for large language models based on updating only a few rows and columns of the weight matrices in transformers. Through extensive experiments with medium-sized LMs like RoBERTa and DeBERTa, and larger LMs like Bloom-7B, Llama2-7B, and Llama2-13B, we show that our method gives comparable or better accuracies than state-of-the-art Parameter-Efficient Finetuning methods while also being more memory and computation-efficient. We also study the reason behind the effectiveness of our method with tools from neural tangent kernel theory. We empirically demonstrate that our kernel, constructed using a restricted set of row and column parameters, is numerically close to the full-parameter kernel and gives comparable classification performance. Ablation studies are conducted to investigate the impact of different algorithmic choices, including the robustness of RoCoFT to any selection of rows and columns, as well as the optimal rank for the effective implementation of our method.

One of the most widely used tasks for evaluating Large Language Models (LLMs) is Multiple-Choice Question Answering (MCQA). While open-ended question answering tasks are more challenging to evaluate, MCQA tasks are, in principle, easier to assess, as the model’s answer is thought to be simple to extract and is compared directly to a set of predefined choices. However, recent studies have started to question the reliability of MCQA evaluation, showing that multiple factors can significantly impact the reported performance of LLMs, especially when the model generates free-form text before selecting one of the answer choices. In this work, we shed light on the inconsistencies of MCQA evaluation strategies, which can lead to inaccurate and misleading model comparisons. We systematically analyze whether existing answer extraction methods are aligned with human judgment, and how they are influenced by answer constraints in the prompt across different domains. Our experiments demonstrate that traditional evaluation strategies often underestimate LLM capabilities, while LLM-based answer extractors are prone to systematic errors. Moreover, we reveal a fundamental trade-off between including format constraints in the prompt to simplify answer extraction and allowing models to generate free-form text to improve reasoning. Our findings call for standardized evaluation methodologies and highlight the need for more reliable and consistent MCQA evaluation practices.

Recent advancements in large language models (LLMs) have highlighted their potential across a variety of tasks, but their performance still heavily relies on the design of effective prompts. Existing methods for automatic prompt optimization face two challenges: lack of diversity, limiting the exploration of valuable and innovative directions and semantic drift, where optimizations for one task can degrade performance in others. To address these issues, we propose Residual Optimization Tree (RiOT), a novel framework for automatic prompt optimization. RiOT iteratively refines prompts through text gradients, generating multiple semantically diverse candidates at each step, and selects the best prompt using perplexity. Additionally, RiOT incorporates the text residual connection to mitigate semantic drift by selectively retaining beneficial content across optimization iterations. A tree structure efficiently manages the optimization process, ensuring scalability and flexibility. Extensive experiments across five benchmarks — covering commonsense, mathematical, logical, temporal, and semantic reasoning — demonstrate that RiOT outperforms both previous prompt optimization methods and manual prompting. Code will be released.

Zero-Shot Voice Conversion (VC) aims to transform the source speaker’s timbre into an arbitrary unseen one while retaining speech content. Most prior work focuses on preserving the source’s prosody, while fine-grained timbre information may leak through prosody, and transferring target prosody to synthesized speech is rarely studied. In light of this, we propose R-VC, a rhythm-controllable and efficient zero-shot voice conversion model. R-VC employs data perturbation techniques and discretize source speech into Hubert content tokens, eliminating much content-irrelevant information. By leveraging a Mask Generative Transformer for in-context duration modeling, our model adapts the linguistic content duration to the desired target speaking style, facilitating the transfer of the target speaker’s rhythm. Furthermore, R-VC introduces a powerful Diffusion Transformer (DiT) with shortcut flow matching during training, conditioning the network not only on the current noise level but also on the desired step size, enabling high timbre similarity and quality speech generation in fewer sampling steps, even in just two, thus minimizing latency. Experimental results show that R-VC achieves comparable speaker similarity to state-of-the-art VC methods with a smaller dataset, and surpasses them in terms of speech naturalness, intelligibility and style transfer performance.

Sarcasm is a complex form of sentiment expression widely used in human daily life. Previous work primarily defines sarcasm as a form of verbal irony, which covers only a subset of real-world sarcastic expressions. However, sarcasm serves multifaceted functions and manifests itself through various rhetorical devices, such as echoic mention, rhetorical question and hyperbole. To fully capture its complexity, this paper investigates fine-grained sarcasm classification through the lens of rhetorical devices, and introduces RedSD, a RhEtorical Device-Aware Sarcasm Dataset with counterfactually augmented data.To construct the dataset, we extract sarcastic dialogues from situation comedies (i.e., sitcoms), and summarize nine rhetorical devices commonly employed in sarcasm. We then propose a rhetorical device-aware counterfactual data generation pipeline facilitated by both Large Language Models (LLMs) and human revision. Additionally, we propose duplex counterfactual augmentation that generates counterfactuals for both sarcastic and non-sarcastic dialogues, to further enhance the scale and diversity of the dataset.Experimental results on the dataset demonstrate that fine-tuned models exhibit a more balanced performance compared to zero-shot models, including GPT-3.5 and LLaMA 3.1, underscoring the importance of integrating various rhetorical devices in sarcasm detection. Our dataset is avaliable at https://github.com/qqHong73/RedSD.

As Large Language Models (LLMs) are widely applied in various domains, the safety of LLMs is increasingly attracting attention to avoid their powerful capabilities being misused. Existing jailbreak methods create a forced instruction-following scenario, or search adversarial prompts with prefix or suffix tokens to achieve a specific representation manually or automatically. However, they suffer from low efficiency and explicit jailbreak patterns, far from the real deployment of mass attacks to LLMs. In this paper, we point out that simply rewriting the original instruction can achieve a jailbreak, and we find that this rewriting approach is learnable and transferable. We propose the **R**ewrite to **J**ailbreak (R2J) approach, a transferable black-box jailbreak method to attack LLMs by iteratively exploring the weakness of the LLMs and automatically improving the attacking strategy. The jailbreak is more efficient and hard to identify since no additional features are introduced. Extensive experiments and analysis demonstrate the effectiveness of R2J, and we find that the jailbreak is also transferable to multiple datasets and various types of models with only a few queries. We hope our work motivates further investigation of LLM safety. The code can be found at [https://github.com/ythuang02/R2J/.](https://github.com/ythuang02/R2J/)

Existing alignment methods share a common topology of information flow, where reward information is collected from humans, modeled with preference learning, and used to tune language models. However, this shared topology has not been systematically characterized, nor have its alternatives been thoroughly explored, leaving the problems of low data efficiency and unreliable generalization unaddressed. As a solution, we introduce a theory of **reward generalization** in reinforcement learning from human feedback (RLHF), focusing on the **topology of information flow** at both macro and micro levels. At the macro level, we portray the RLHF information flow as an autoencoding process over behavior distributions, formalizing the RLHF objective of distributional consistency between human preference and model behavior. At the micro level, we present *induced Bayesian networks* to model the impact of dataset topologies on reward generalization. Combining analysis on both levels, we propose **reward modeling from tree-structured preference information**. It is shown to reduce reward uncertainty by up to 𝛩(log n/loglog n) times compared to baselines, where n is the dataset size. Validation on three NLP tasks shows that it achieves an average win rate of 65% against baselines, thus improving reward generalization *for free* via topology design, while *reducing* the amount of data requiring annotation.

Historical documents represent an invaluable cultural heritage, yet have undergone significant degradation over time through tears, water erosion, and oxidation. Existing Historical Document Restoration (HDR) methods primarily focus on single modality or limited-size restoration, failing to meet practical needs. To fill this gap, we present a full-page HDR dataset (FPHDR) and a novel automated HDR solution (AutoHDR). Specifically, FPHDR comprises 1,633 real and 6,543 synthetic images with character-level and line-level locations, as well as character annotations in different damage grades. AutoHDR mimics historians’ restoration workflows through a three-stage approach: OCR-assisted damage localization, vision-language context text prediction, and patch autoregressive appearance restoration. The modular architecture of AutoHDR enables seamless human-machine collaboration, allowing for flexible intervention and optimization at each restoration stage. Experiments demonstrate AutoHDR’s remarkable performance in HDR. When processing severely damaged documents, our system improves OCR accuracy from 46.83% to 84.05%, with further enhancement to 94.25% through human-machine collaboration. We believe this work represents a significant advancement in automated historical document restoration and contributes substantially to cultural heritage preservation. The model and dataset are available at https://github.com/SCUT-DLVCLab/AutoHDR.

The advent of test-time scaling in large language models (LLMs), exemplified by OpenAI’s o1 series, has advanced reasoning capabilities by scaling computational resource allocation during inference. While successors like QwQ, Deepseek-R1 (R1) and LIMO replicate these advancements, whether these models truly possess test-time scaling capabilities remains underexplored. This study found that longer CoTs of these o1-like models do not consistently enhance accuracy; in fact, correct solutions are often shorter than incorrect ones for the same questions. Further investigation shows this phenomenon is closely related to models’ self-revision capabilities - longer CoTs contain more self-revisions, which often lead to performance degradation. We then compare sequential and parallel scaling strategies on QwQ, R1 and LIMO, finding that parallel scaling achieves better coverage and scalability. Based on these insights, we propose “Shortest Majority Vote”, a method that combines parallel scaling strategies with CoT length characteristics, significantly improving models’ test-time scalability compared to conventional majority voting approaches.

As large language models advance toward superhuman performance, ensuring their alignment with human values and abilities grows increasingly complex. Weak-to-strong generalization offers a promising approach by leveraging predictions from weaker models to guide stronger systems, but its effectiveness could be constrained by the inherent noise and inaccuracies in these weak predictions. To address this, we propose a theoretically grounded approach that replaces forward KL divergence—whose mass-covering behavior risks overfitting to imperfect weak signals—with reverse KL divergence. Reverse KL divergence’s zero-forcing effect prioritizes high-confidence predictions, effectively mitigating the influence of unreliable weak supervision. Theoretically, we extend existing bounds and derive tighter lower bounds for both forward and reverse KL divergence. Notably, when a sufficiently pre-trained strong model is fine-tuned on the last linear layer, reverse KL guarantees that it outperforms its weak supervisor by the magnitude of their disagreement. Empirically, we demonstrate that reverse KL and reverse cross-entropy not only enable strong models to outperform those trained with forward KL and standard cross-entropy across most settings, but also exhibit greater robustness to noisy labels.

Uncertainty Quantification (UQ) in Language Models (LMs) is key to improving their safety and reliability. Evaluations often use metrics like AUROC to assess how well UQ methods (e.g., negative sequence probabilities) correlate with task correctness functions (e.g., ROUGE-L). We show that mutual biases-when both UQ methods and correctness functions are biased by the same factors-systematically distort evaluation. First, we formally prove that any mutual bias non-randomly skews AUROC rankings, compromising benchmark integrity. Second, we confirm this happens empirically by testing 7 widely used correctness functions, from lexical-based and embedding-based metrics to LM-as-a-judge approaches, across 4 datasets × 4 models × 8 UQ methods. Our analysis showsthat length biases in correctness functions distort UQ assessments by interacting with length biases in UQ methods. We identify LM-as-a-judge methods as the least length-biased, offering a promising path for a fairer UQ evaluation.

Self-consistency improves reasoning by aggregating diverse stochastic samples, yet the dynamics behind its efficacy remain underexplored. We reframe self-consistency as a dynamic distributional alignment problem, revealing that decoding temperature not only governs sampling randomness but also actively shapes the latent answer distribution. Given that high temperatures require prohibitively large sample sizes to stabilize, while low temperatures risk amplifying biases, we propose a confidence-driven mechanism that dynamically calibrates temperature: sharpening the sampling distribution under uncertainty to align with high-probability modes, and promoting exploration when confidence is high. Experiments on mathematical reasoning tasks show this approach outperforms fixed-diversity baselines under limited samples, improving both average and best-case performance across varying initial temperatures without additional data or modules. This establishes self-consistency as a synchronization challenge between sampling dynamics and evolving answer distributions.

Traditional scaling laws in natural language processing suggest that increasing model size and training data enhances performance. However, recent studies reveal deviations, particularly in large language models, where performance improvements decelerate—a phenomenon known as sub-scaling. This paper revisits these scaling laws by examining the impact of data quality and training strategies on model performance. Through extensive empirical analysis of over 400 models, we identify high data density and non-optimal resource allocation as key factors contributing to sub-scaling. High data density leads to diminishing returns due to redundant information, while optimal resource allocation is crucial for sustained performance improvements. We propose a sub-optimal scaling law that better predicts performance in sub-scaling regimes, highlighting the importance of data quality and diversity.

Low-Rank Adaptation (LoRA) has emerged as a prominent technique for fine-tuning large foundation models. Despite its successes, the substantial parameter redundancy, which limits the capacity and efficiency of LoRA, has been recognized as a bottleneck. In this work, we systematically investigate the impact of redundancy in fine-tuning LoRA and reveal that reducing density redundancy does not degrade expressiveness. Based on this insight, we introduce Spectral-encoding Low-Rank Adaptation (SeLoRA), which harnesses the robust expressiveness of spectral bases to re-parameterize LoRA from a sparse spectral subspace. Designed with simplicity, SeLoRA enables seamless integration with various LoRA variants for performance boosting, serving as a scalable plug-and-play framework. Extensive experiments substantiate that SeLoRA achieves greater efficiency with fewer parameters, delivering superior performance enhancements over strong baselines on various downstream tasks, including commonsense reasoning, math reasoning, and code generation.

Large Language Models (LLMs) have shown remarkable performance across diverse tasks without domain-specific training, fueling interest in their potential for time-series forecasting. While LLMs have shown potential in zero-shot forecasting through prompting alone, recent studies suggest that LLMs lack inherent effectiveness in forecasting. Given these conflicting findings, a rigorous validation is essential for drawing reliable conclusions. In this paper, we evaluate the effectiveness of LLMs as zero-shot forecasters compared to state-of-the-art domain-specific models. Our experiments show that LLM-based zero-shot forecasters often struggle to achieve high accuracy due to their sensitivity to noise, underperforming even simple domain-specific models. We have explored solutions to reduce LLMs’ sensitivity to noise in the zero-shot setting, but improving their robustness remains a significant challenge. Our findings suggest that rather than emphasizing zero-shot forecasting, a more promising direction would be to focus on fine-tuning LLMs to better process numerical sequences. Our experimental code is available at https://github.com/junwoopark92/revisiting-LLMs-zeroshot-forecaster.

In-Context Learning (ICL) is a technique by which language models make predictions based on examples provided in their input context. Previously, their context window size imposed a limit on the number of examples that can be shown, making example selection techniques crucial for identifying the maximally effective set of examples. However, the recent advent of Long Context Language Models (LCLMs) has significantly increased the number of examples that can be included in context, raising an important question of whether ICL performance in a many-shot regime is still sensitive to the method of sample selection. To answer this, we revisit these approaches in the context of LCLMs through extensive experiments on 18 datasets spanning 4 tasks. Surprisingly, we observe that sophisticated example selection techniques do not yield significant improvements over a simple random sample selection method. Instead, we discover that the advent of LCLMs has fundamentally shifted the challenge of ICL from that of selecting the most effective examples to that of collecting sufficient examples to fill the context window. Specifically, in certain datasets, including all available examples does not fully utilize the context window; however, by augmenting the examples in context with a simple data augmentation approach, we substantially improve ICL performance by 5%.