TACL 2025 论文列表

Detecting content generated by large language models (LLMs) is crucial for preventing misuse and building trustworthy AI systems. Although existing detection methods perform well, their robustness in out-of-distribution (OOD) scenarios is still lacking. In this paper, we hypothesize that, compared to features used by existing detection methods, the internal representations of LLMs contain more comprehensive and raw features that can more effectively capture and distinguish the statistical pattern differences between LLM-generated texts (LGT) and human-written texts (HWT). We validated this hypothesis across different LLMs and observed significant differences in neural activation patterns when processing these two types of texts. Based on this, we propose RepreGuard, an efficient statistics-based detection method. Specifically, we first employ a surrogate model to collect representation of LGT and HWT, and extract the distinct activation feature that can better identify LGT. We can classify the text by calculating the projection score of the text representations along this feature direction and comparing with a precomputed threshold. Experimental results show that RepreGuard outperforms all baselines with average 94.92% AUROC on both in-distribution and OOD scenarios, while also demonstrating robust resilience to various text sizes and mainstream attacks.1

Human evaluation is the gold standard for evaluating text generation models. However, it is expensive. In order to fit budgetary constraints, a random subset of the test data is often chosen in practice for human evaluation. However, randomly selected data may not accurately represent test performance, making this approach economically inefficient for model comparison. Thus, in this work, we develop and analyze a suite of selectors to get the most informative datapoints for human evaluation, taking the evaluation costs into account. We show that selectors based on variance in automated metric scores, diversity in model outputs, or Item Response Theory outperform random selection. We further develop an approach to distill these selectors to the scenario where the model outputs are not yet available. In particular, we introduce source-based estimators, which predict item usefulness for human evaluation just based on the source texts. We demonstrate the efficacy of our selectors in two common NLG tasks, machine translation and summarization, and show that only ∼70% of the test data is needed to produce the same evaluation result as the entire data.

Fine-tuning large language models (LLMs) on downstream tasks requires substantial computational resources. Selective-PEFT, a class of parameter-efficient fine-tuning (PEFT) methodologies, aims to mitigate these computational challenges by selectively fine-tuning only a small fraction of the model parameters. Although parameter-efficient, these techniques often fail to match the performance of fully fine-tuned models, primarily due to inherent biases introduced during parameter selection. Traditional selective-PEFT techniques use a fixed set of parameters selected using different importance heuristics, failing to capture parameter importance dynamically and often leading to suboptimal performance. We introduce ID3, a novel selective-PEFT method that calculates parameter importance continually, and dynamically unmasks parameters by balancing exploration and exploitation in parameter selection. Our empirical study on 16 tasks spanning natural language understanding, mathematical reasoning, and summarization demonstrates the effectiveness of our method compared to fixed-masking selective-PEFT techniques. We analytically show that ID3 reduces the number of gradient updates by a factor of two, enhancing computational efficiency. Since ID3 is robust to random initialization of neurons and operates directly on the optimization process, it is highly flexible and can be integrated with existing additive and reparameterization-based PEFT techniques such as Adapters and LoRA, respectively.1

Recent cognitive modeling studies have reported that larger language models (LMs) exhibit a poorer fit to human reading behavior (Oh and Schuler, 2023b; Shain et al., 2024; Kuribayashi et al., 2024), leading to claims of their cognitive implausibility. In this paper, we revisit this argument through the lens of mechanistic interpretability and argue that prior conclusions were skewed by an exclusive focus on the final layers of LMs. Our analysis reveals that next-word probabilities derived from internal layers of larger LMs align with human sentence processing data as well as, or better than, those from smaller LMs. This alignment holds consistently across behavioral (self-paced reading times, gaze durations, MAZE task processing times) and neurophysiological (N400 brain potentials) measures, challenging earlier mixed results and suggesting that the cognitive plausibility of larger LMs has been underestimated. Furthermore, we first identify an intriguing relationship between LM layers and human measures: Earlier layers correspond more closely with fast gaze durations, while later layers better align with relatively slower signals such as N400 potentials and MAZE processing times. Our work opens new avenues for interdisciplinary research at the intersection of mechanistic interpretability and cognitive modeling.1

Personalized dialogue generation aims to leverage persona profiles and dialogue history to generate persona-relevant and consistent responses. Mainstream models typically rely on token-level language model training with persona dialogue data, such as Next Token Prediction, to implicitly achieve personalization, meaning that these methods tend to neglect the given personas and generate generic responses. To address this issue, we propose a novel Persona-Aware Alignment Framework (PAL), which directly treats persona alignment as the training objective of dialogue generation. Specifically, PAL employs a two-stage training method including Persona-Aware Learning and Persona Alignment, equipped with an easy-to-use inference strategy Select then Generate, to improve persona sensitivity and generate more persona-relevant responses at the semantics level. Through extensive experiments, we demonstrate that our framework outperforms many state-of-the-art personalized dialogue methods and large language models.

Building high-quality datasets for specialized tasks is a time-consuming and resource-intensive process that often requires specialized domain knowledge. We propose Corpus Retrieval and Augmentation for Fine-Tuning (CRAFT), a method for generating synthetic datasets, given a small number of user-written few-shots that demonstrate the task to be performed. Given these examples, CRAFT uses large-scale public web-crawled corpora and similarity-based document retrieval to find other relevant human-written documents. Lastly, instruction-tuned large language models (LLMs) augment the retrieved documents into custom-formatted task samples, which then can be used for fine- tuning. We demonstrate that CRAFT can efficiently generate large-scale task-specific training datasets for four diverse tasks: biology, medicine, and commonsense question-answering (QA), as well as summarization. Our experiments show that CRAFT-based models outperform or match general LLMs on QA tasks, while exceeding models trained on human-curated summarization data by 46 preference points. CRAFT outperforms other synthetic dataset generation methods such as Self- and Evol-Instruct, and remains robust even when the quality of the initial few-shots varies.

Evaluating social biases in language models (LMs) is crucial for ensuring fairness and minimizing the reinforcement of harmful stereotypes in AI systems. Existing benchmarks, such as the Bias Benchmark for Question Answering (BBQ), primarily focus on Western contexts, limiting their applicability to the Indian context. To address this gap, we introduce BharatBBQ,1 a culturally adapted benchmark designed to assess biases in Hindi, English, Marathi, Bengali, Tamil, Telugu, Odia, and Assamese. BharatBBQ covers 13 social categories, including 3 intersectional groups, reflecting prevalent biases in the Indian sociocultural landscape. Our dataset contains 49,108 examples in one language that are expanded using translation and verification to 392,864 examples in eight different languages. We evaluate five multilingual LM families across zero- and few-shot settings, analyzing their bias and stereotypical bias scores. Our findings highlight persistent biases across languages and social categories and often amplified biases in Indian languages compared to English, demonstrating the necessity of linguistically and culturally grounded benchmarks for bias evaluation.

A broad range of NLP tasks involve selecting relevant text spans from given source texts. Despite this shared objective, such content selection tasks have traditionally been studied in isolation, each with its own modeling approaches, datasets, and evaluation metrics. In this work, we propose instruction-guided content selection (IGCS) as a beneficial unified framework for such settings, where the task definition and any instance-specific request are encapsulated as instructions to a language model. To promote this framework, we introduce IGCS-Bench, the first unified benchmark covering diverse content selection tasks. Further, we create a large generic synthetic dataset that can be leveraged for diverse content selection tasks, and show that transfer learning with these datasets often boosts performance, whether dedicated training for the targeted task is available or not. Finally, we address generic inference time issues that arise in LLM-based modeling of content selection, assess a generic evaluation metric, and overall propose the utility of our resources and methods for future content selection models.1

The task of image captioning has recently been gaining popularity, and with it the complex task of evaluating the quality of image captioning models. In this work, we present the first survey and taxonomy of over 70 different image captioning metrics and their usage in hundreds of papers, specifically designed to help users select the most suitable metric for their needs. We find that despite the diversity of proposed metrics, the vast majority of studies rely on only five popular metrics, which we show to be weakly correlated with human ratings. We hypothesize that combining a diverse set of metrics can enhance correlation with human ratings. As an initial step, we demonstrate that a linear regression-based ensemble method, which we call EnsembEval, trained on one human ratings dataset, achieves improved correlation across five additional datasets, showing there is a lot of room for improvement by leveraging a diverse set of metrics.1

Speaker attribution from speech transcripts is the task of identifying a speaker from the transcript of their speech based on patterns in their language use. This task is especially useful when the audio is unavailable (e.g., deleted) or unreliable (e.g., anonymized speech). Prior work in this area has primarily focused on the feasibility of attributing speakers using transcripts produced by human annotators. However, in real-world settings, one often only has more errorful transcripts produced by automatic speech recognition (ASR) systems. In this paper, we conduct what is, to our knowledge, the first comprehensive study of the impact of automatic transcription on speaker attribution performance. In particular, we study the extent to which speaker attribution performance degrades in the face of transcription errors, as well as how properties of the ASR system impact attribution. We find that attribution is surprisingly resilient to word-level transcription errors and that the objective of recovering the true transcript is minimally correlated with attribution performance. Overall, our findings suggest that speaker attribution on more errorful transcripts produced by ASR is as good, if not better, than attribution based on human-transcribed data, possibly because ASR transcription errors can capture speaker-specific features revealing of speaker identity.

Most existing work on event extraction has focused on sentence-level texts and presumes the identification of a trigger-span—a word or phrase in the input that evokes the occurrence of an event of interest. Event arguments are then extracted with respect to the trigger. Indeed, triggers are treated as integral to, and trigger detection as an essential component of, event extraction. In this paper, we provide the first investigation of the role of triggers for the more difficult and much less studied task of document-level event extraction. We analyze their usefulness in multiple end-to-end and pipelined transformer-based event extraction models for three document-level event extraction datasets, measuring performance using triggers of varying quality (human-annotated, LLM-generated, keyword-based, and random). We find that whether or not systems benefit from explicitly extracting triggers depends both on dataset characteristics (i.e., the typical number of events per document) and task-specific information available during extraction (i.e., natural language event schemas). Perhaps surprisingly, we also observe that the mere existence of triggers in the input, even random ones, is important for prompt-based in-context learning approaches to the task.

Interpretation of topics is crucial for their downstream applications. State-of-the-art evaluation measures of topic quality such as coherence and word intrusion do not measure how much a topic facilitates the exploration of a corpus. To design evaluation measures grounded on a task, and a population of users, we do user studies to understand how users interpret topics. We propose constructs of topic quality and ask users to assess them in the context of a topic and provide rationale behind evaluations. We use reflexive thematic analysis to identify themes of topic interpretations from rationales. Users interpret topics based on availability and representativeness heuristics rather than probability. We propose a theory of topic interpretation based on the anchoring-and-adjustment heuristic: users anchor on salient words and make semantic adjustments to arrive at an interpretation. Topic interpretation can be viewed as making a judgment under uncertainty by an ecologically rational user, and hence cognitive biases aware user models and evaluation frameworks are needed.

The development and evaluation of Large Language Models (LLMs) has primarily focused on their task-solving capabilities, with recent models even surpassing human performance in some areas. However, this focus often neglects whether machine-generated language matches the human level of diversity, in terms of vocabulary choice, syntactic construction, and expression of meaning, raising questions about whether the fundamentals of language generation have been fully addressed. This paper emphasizes the importance of examining the preservation of human linguistic richness by language models, given the concerning surge in online content produced or aided by LLMs. We adapt a comprehensive framework for evaluating LLMs from various linguistic diversity perspectives including lexical, syntactic, and semantic dimensions. Using this framework, we benchmark several state-of-the-art LLMs across all diversity dimensions, and conduct an in-depth analysis for syntactic diversity. Finally, we analyze how the design, development, and deployment choices of LLMs impact the linguistic diversity of their outputs, focusing on the creative task of story generation.

Powerful large language models (LLMs) are increasingly expected to be deployed with lower computational costs, enabling their capabilities on resource-constrained devices. Post-training quantization (PTQ) has emerged as a star approach to achieve this ambition, with best methods compressing weights to less than 2 bit on average. In this paper, we propose Channel-Relaxed Vector Quantization (CRVQ), a novel technique that significantly improves the performance of PTQ baselines at the cost of only minimal additional bits. This state-of-the-art extreme compression method achieves its results through two key innovations: (1) carefully selecting and reordering a very small subset of critical weight channels, and (2) leveraging extended codebooks to relax the constraint of critical channels. With our method, we demonstrate a 38.9% improvement over the current strongest sub-2-bit PTQ baseline, enabling nearer lossless 1-bit compression. Furthermore, our approach offers flexible customization of quantization bit-width and performance, providing a wider range of deployment options for diverse hardware platforms. Code and checkpoints are available at https://github.com/xuyuzhuang11/CRVQ.

Fine-tuning Large Language Models (LLMs) with Low-Rank Adaptation (LoRA) enhances adaptability while reducing computational costs. However, fine-tuning can compromise safety alignment, even with benign data, increasing susceptibility to harmful outputs. Existing safety alignment methods struggle to capture complex parameter shifts, leading to suboptimal safety-utility trade-offs. To address this issue, we propose Safe Pruning LoRA (SPLoRA), a novel pruning-based approach that selectively removes LoRA layers that weaken safety alignment, improving safety while preserving performance. At its core, we introduce Empirical-DIEM (E-DIEM), a dimension-insensitive similarity metric that effectively detects safety misalignment in LoRA-adapted models. We conduct extensive experiments on LLMs fine-tuned with mixed of benign and malicious data, and purely benign datasets, evaluating SPLoRA across utility, safety, and reliability metrics. Results demonstrate that SPLoRA outperforms state-of-the-art safety alignment techniques, significantly reducing safety risks while maintaining or improving model performance and reliability. Additionally, SPLoRA reduces inference overhead, making it a scalable and efficient solution for deploying safer and more reliable LLMs. The code is available at https://github.com/AoShuang92/SPLoRA.

Although large language models excel across many tasks, they can memorize training data and thereby expose private or copyrighted text. Most defenses target the pre-training stage, leaving memorization during fine-tuning–especially for domain adaptation and instruction tuning–poorly understood. We fine-tune Pythia, Llama3, and Mistral models spanning 1.4B–70B parameters on common evaluation datasets and track verbatim memorization throughout training. We find that memorization increases dramatically in the first few epochs, often significantly before either validation perplexity or evaluation performance is optimized. We use a simple but effective n-gram memorization score which reliably precedes verbatim memorization; using it as an early-stopping criterion mitigates memorization with minimal performance loss. Further, we introduce an n-gram–aware loss regularizer and show that it reduces memorization across all model families tested by up to 40% while minimizing evaluation performance trade-offs when compared to an existing memorization mitigation strategy. These results yield practical, scalable insights into memorization dynamics during language model fine-tuning.

Interpretability is a key challenge in fostering trust for Large Language Models (LLMs), which stems from the complexity of extracting reasoning from a model’s parameters. We present the Frame Representation Hypothesis, a theoretically robust framework grounded in the Linear Representation Hypothesis (LRH) to interpret and control LLMs by modeling multi-token words. Prior research explored LRH to connect LLM representations with linguistic concepts, but was limited to single token analysis. As most words are composed of several tokens, we extend LRH to multi-token words, thereby enabling usage on any textual data with thousands of concepts. To this end, we propose that words can be interpreted as frames, ordered sequences of vectors that better capture token-word relationships. Then, concepts can be represented as the average of word frames sharing a common concept. We showcase these tools through Top-k Concept-Guided Decoding, which can intuitively steer text generation using concepts of choice. We verify said ideas on Llama 3, Gemma 2, Phi 3, and Qwen-2-VL families, demonstrating gender and language biases, exposing harmful content, but also potential to remediate them, leading to safer and more transparent LLMs. Code is available at this https url.

Word-level quality estimation (QE) methods aim to detect erroneous spans in machine translations, which can direct and facilitate human post-editing. While the accuracy of word-level QE systems has been assessed extensively, their usability and downstream influence on the speed, quality, and editing choices of human post-editing remain understudied. In this study, we investigate the impact of word-level QE on machine translation (MT) post-editing in a realistic setting involving 42 professional post-editors across two translation directions. We compare four error-span highlight modalities, including supervised and uncertainty-based word-level QE methods, for identifying potential errors in the outputs of a state-of-the-art neural MT model. Post-editing effort and productivity are estimated from behavioral logs, while quality improvements are assessed by word- and segment-level human annotation. We find that domain, language and editors’ speed are critical factors in determining highlights’ effectiveness, with modest differences between human-made and automated QE highlights underlining a gap between accuracy and usability in professional workflows.

Pre-trained language models (PLMs) have driven substantial progress in natural language processing but remain vulnerable to adversarial attacks, raising concerns about their robustness in real-world applications. Previous studies have sought to mitigate the impact of adversarial attacks by introducing adversarial perturbations into the training process, either implicitly or explicitly. While both strategies enhance robustness, they often incur high computational costs. In this work, we propose a simple yet effective add-on module that enhances the adversarial robustness of PLMs by removing instance-level principal components, without relying on conventional adversarial defenses or perturbing the original training data. Our approach transforms the embedding space to approximate Gaussian properties, thereby reducing its susceptibility to adversarial perturbations while preserving semantic relationships. This transformation aligns embedding distributions in a way that minimizes the impact of adversarial noise on decision boundaries, enhancing robustness without requiring adversarial examples or costly training-time augmentation. Evaluations on eight benchmark datasets show that our approach improves adversarial robustness while maintaining comparable before-attack accuracy to baselines, achieving a balanced trade-off between robustness and generalization.