ACL 2025 Papers

Attributing answers to source documents is an approach used to enhance the verifiability of a model’s output in retrieval-augmented generation (RAG). Prior work has mainly focused on improving and evaluating the attribution quality of large language models (LLMs) in RAG, but this may come at the expense of inducing biases in the attribution of answers. We define and examine two aspects in the evaluation of LLMs in RAG pipelines, namely attribution sensitivity and bias with respect to authorship information. We explicitly inform an LLM about the authors of source documents, instruct it to attribute its answers, and analyze (i) how sensitive the LLM’s output is to the author of source documents, and (ii) whether the LLM exhibits a bias towards human-written or AI-generated source documents. We design an experimental setup in which we use counterfactual evaluation to study three LLMs in terms of their attribution sensitivity and bias in RAG pipelines. Our results show that adding authorship information to source documents can significantly change the attribution quality of LLMs by 3 to 18%. We show that LLMs can have an attribution bias towards explicit human authorship, which can serve as a competing hypothesis for findings of prior work that shows that LLM-generated content may be preferred over human-written contents. Our findings indicate that metadata of source documents can influence LLMs’ trust, and how they attribute their answers. Furthermore, our research highlights attribution bias and sensitivity as a novel aspect of the vulnerability of LLMs.

Recent advancements in visual generative models have enabled high-quality image and video generation, opening diverse applications. However, evaluating these models often demands sampling hundreds or thousands of images or videos, making the process computationally expensive, especially for diffusion-based models with inherently slow sampling. Moreover, existing evaluation methods rely on rigid pipelines that overlook specific user needs and provide numerical results without clear explanations. In contrast, humans can quickly form impressions of a model’s capabilities by observing only a few samples. To mimic this, we propose the Evaluation Agent framework, which employs human-like strategies for efficient, dynamic, multi-round evaluations using only a few samples per round, while offering detailed, user-tailored analyses. It offers four key advantages: 1) efficiency, 2) promptable evaluation tailored to diverse user needs, 3) explainability beyond single numerical scores, and 4) scalability across various models and tools. Experiments show that Evaluation Agent reduces evaluation time to 10% of traditional methods while delivering comparable results. The Evaluation Agent framework is fully open-sourced to advance research in visual generative models and their efficient evaluation.

In applications such as dialogue systems, personalized recommendations, and personal assistants, large language models (LLMs) need to retain and utilize historical information over the long term to provide more accurate and consistent responses. Although long-term memory capability is crucial, recent studies have not thoroughly investigated the memory performance of large language models in long-term tasks. To address this gap, we introduce the Long-term Chronological Conversations (LOCCO) dataset and conduct a quantitative evaluation of the long-term memory capabilities of large language models. Experimental results demonstrate that large language models can retain past interaction information to a certain extent, but their memory decays over time. While rehearsal strategies can enhance memory persistence, excessive rehearsal is not an effective memory strategy for large models, unlike in smaller models. Additionally, the models exhibit memory preferences across different categories of information. Our study not only provides a new framework and dataset for evaluating the long-term memory capabilities of large language models but also offers important references for future enhancements of their memory persistence.

In commonsense generation, given a set of input concepts, a model must generate a response that is not only commonsense bearing, but also capturing multiple diverse viewpoints. Numerous evaluation metrics based on form- and content-level overlap have been proposed in prior work for evaluating the diversity of a commonsense generation model. However, it remains unclear as to which metrics are best suited for evaluating the diversity in commonsense generation. To address this gap, we conduct a systematic meta-evaluation of diversity metrics for commonsense generation. We find that form-based diversity metrics tend to consistently overestimate the diversity in sentence sets, where even randomly generated sentences are assigned overly high diversity scores. We then use an Large Language Model (LLM) to create a novel dataset annotated for the diversity of sentences generated for a commonsense generation task, and use it to conduct a meta-evaluation of the existing diversity evaluation metrics. Our experimental results show that content-based diversity evaluation metrics consistently outperform the form-based counterparts, showing high correlations with the LLM-based ratings. We recommend that future work on commonsense generation should use content-based metrics for evaluating the diversity of their outputs.

To create culturally inclusive vision-language models (VLMs), developing a benchmark that tests their ability to address culturally relevant questions is essential. Existing approaches typically rely on human annotators, making the process labor-intensive and creating a cognitive burden in generating diverse questions. To address this, we propose a semi-automated framework for constructing cultural VLM benchmarks, specifically targeting multiple-choice QA. This framework combines human-VLM collaboration, where VLMs generate questions based on guidelines, a small set of annotated examples, and relevant knowledge, followed by a verification process by native speakers. We demonstrate the effectiveness of this framework through the creation of K-Viscuit, a dataset focused on Korean culture. Our experiments on this dataset reveal that open-source models lag behind proprietary ones in understanding Korean culture, highlighting key areas for improvement. We also present a series of further analyses, including human evaluation, augmenting VLMs with external knowledge, and the evaluation beyond multiple-choice QA. Our dataset is available at https://huggingface.co/datasets/ddehun/k-viscuit.

Typically, when evaluating Theory of Mind, we consider the beliefs of others to be binary: held or not held. But what if someone is unsure about their own beliefs? How can we quantify this uncertainty? We propose a new suite of tasks, challenging language models (LMs) to model the uncertainty of participants in a dialogue. We design these tasks around conversation forecasting, where the goal is to predict the probability of an unobserved conversation outcome. Uniquely, we view conversation agents themselves as forecasters, asking an LM to predict the uncertainty of an individual from their language use. We experiment with scaling methods, bagging, and demographic context for this regression task, conducting experiments on three dialogue corpora (social, negotiation, task-oriented) with eight LMs. While LMs can explain up to 7% variance in the uncertainty of others, we highlight the difficulty of the tasks and room for future work, especially in tasks that require explicit shifts in perspective.

Various metrics exist for evaluating sequence labeling problems (strict span matching, token oriented metrics, token concurrence in sequences, etc.), each of them focusing on certain aspects of the task. In this paper, we define a comprehensive set of formal properties that captures the strengths and weaknesses of the existing metric families and prove that none of them is able to satisfy all properties simultaneously. We argue that it is necessary to measure how much information (correct or noisy) each token in the sequence contributes depending on different aspects such as sequence length, number of tokens annotated by the system, token specificity, etc. On this basis, we introduce the Sequence Labelling Information Contrast Model (SL-ICM), a novel metric based on information theory for evaluating sequence labeling tasks. Our formal analysis and experimentation show that the proposed metric satisfies all properties simultaneously

The scaling of causal language models in size and training data enabled them to tackle increasingly complex tasks. Despite the development of sophisticated tests to reveal their new capabilities, the underlying basis of these complex skills remains unclear. We argue that complex skills might be explained using simpler ones, represented by linguistic concepts. As an initial step in exploring this hypothesis, we focus on the lexical-semantic concept of synonymy, laying the groundwork for research into its relationship with more complex skills. We develop a comprehensive test suite to assess various aspects of synonymy under different conditions, and evaluate causal open-source models ranging up to 10 billion parameters. We find that these models effectively recognize synonymy but struggle to generate synonyms when prompted with relevant context.

Personalized tool utilization is essential for aligning large language models (LLMs) with user preference in interaction scenarios with various tools. However, most of the current benchmarks primarily focus on either personalization of text generation or direct tool-utilizing, without considering both. In this work, we introduce a novel benchmark ETAPP for evaluating personalized tool invocation, establishing a sandbox environment, and a comprehensive dataset of 800 testing cases covering diverse user profiles. To improve the accuracy of our evaluation, we propose a key-point-based LLM evaluation method, mitigating biases in the LLM-as-a-judge system by manually annotating key points for each test case and providing them to LLM as the reference. Additionally, we evaluate the excellent LLMs and provide an in-depth analysis. Furthermore, we investigate the impact of different tool-invoking strategies on LLMs’ personalization performance and the effects of fine-tuning in our task. The effectiveness of our preference-setting and key-point-based evaluation method is also validated. Our findings offer insights into improving personalized LLM agents. Our code is available at https://github.com/hypasd-art/ETAPP.

Video captioning can be used to assess the video understanding capabilities of Multimodal Large Language Models (MLLMs).However, existing benchmarks and evaluation protocols suffer from crucial issues, such as inadequate or homogeneous creation of key points, exorbitant cost of data creation, and limited evaluation scopes. To address these issues, we propose an automatic framework, named AutoCaption, which leverages Monte Carlo Tree Search (MCTS) to construct numerous and diverse descriptive sentences (i.e., key points) that thoroughly represent video content in an iterative way. This iterative captioning strategy enables the continuous enhancement of video details such as actions, objects’ attributes, environment details, etc. We apply AutoCaption to curate MCTS-VCB, a fine-grained video caption benchmark covering video details, thereby enabling a comprehensive evaluation of MLLMs on the video captioning task. We evaluate more than 20 open- and closed-source MLLMs of varying sizes on MCTS-VCB. Results show that MCTS-VCB can effectively and comprehensively evaluate the video captioning capability, with Gemini-1.5-Pro achieving the highest F1 score of 71.2. Interestingly, we fine-tune InternVL2.5-8B with the AutoCaption-generated data, which helps the model achieve an overall improvement of 25.0% on MCTS-VCB and 16.3% on DREAM-1K, further demonstrating the effectiveness of AutoCaption. The code and data are available at https://github.com/tjunlp-lab/MCTS-VCB.

This paper explores the effectiveness of Multimodal Large Language models (MLLMs) as assistive technologies for visually impaired individuals. We conduct a user survey to identify adoption patterns and key challenges users face with such technologies. Despite a high adoption rate of these models, our findings highlight concerns related to contextual understanding, cultural sensitivity, and complex scene understanding, particularly for individuals who may rely solely on them for visual interpretation. Informed by these results, we collate five user-centred tasks with image and video inputs, including a novel task on Optical Braille Recognition. Our systematic evaluation of twelve MLLMs reveals that further advancements are necessary to overcome limitations related to cultural context, multilingual support, Braille reading comprehension, assistive object recognition, and hallucinations. This work provides critical insights into the future direction of multimodal AI for accessibility, underscoring the need for more inclusive, robust, and trustworthy visual assistance technologies.

We present a novel evaluation framework designed to assess the lexical proficiency and linguistic creativity of Transformer-based Language Models (LMs). We validate the framework by analyzing the performance of a set of LMs of different sizes, in both mono- and multilingual configuration, across tasks involving the generation, definition, and contextual usage of lexicalized words, neologisms, and nonce words. To support these evaluations, we developed a novel dataset of lexical entries for the Italian language, including curated definitions and usage examples sourced from various online platforms. The results highlight the robustness and effectiveness of our framework in evaluating multiple dimensions of LMs’ linguistic understanding and offer an insight, through the assessment of their linguistic creativity, on the lexical generalization abilities of LMs.

Large Language Models (LLMs) have shown promise in automating high-labor data tasks, but the adoption of LLMs in high-stake scenarios faces two key challenges: their tendency to answer despite uncertainty and their difficulty handling long input contexts robustly.We investigate commonly used off-the-shelf LLMs’ ability to identify low-confidence outputs for human review through “check set selection”–a process where LLMs prioritize information needing human judgment.Using a case study on social media monitoring for disaster risk management,we define the “check set” as a list of tweets escalated to the disaster manager when the LLM has the least confidence, enabling human oversight within budgeted effort.We test two strategies for LLM check set selection: *individual confidence elicitation* – LLMs assesses confidence for each tweet classification individually, requiring more prompts with shorter contexts, and *direct set confidence elicitation* – LLM evaluates confidence for a list of tweet classifications at once, using less prompts but longer contexts.Our results reveal that set selection via individual probabilities is more reliable but that direct set confidence merits further investigation.Direct set selection challenges include inconsistent outputs, incorrect check set size, and low inter-annotator agreement. Despite these challenges, our approach improves collaborative disaster tweet classification by outperforming random-sample check set selection, demonstrating the potential of human-LLM collaboration.

Given the increasing use of synthetic data in language model (LM) post-training, an LM’s ability to generate high-quality data has become nearly as crucial as its ability to solve problems directly. While prior works have focused on developing effective data generation methods, they lack systematic comparison of different LMs as data generators in a unified setting. To address this gap, we propose AgoraBench, a benchmark that provides standardized settings and metrics to evaluate LMs’ data generation abilities. Through synthesizing 1.26 million training instances using 6 LMs and training 99 student models, we uncover key insights about LMs’ data generation capabilities. First, we observe that LMs exhibit distinct strengths. For instance, GPT-4o excels at generating new problems, while Claude-3.5-Sonnet performs better at enhancing existing ones. Furthermore, our analysis reveals that an LM’s data generation ability doesn’t necessarily correlate with its problem-solving ability. Instead, multiple intrinsic features of data quality—including response quality, perplexity, and instruction difficulty—collectively serve as better indicators. Finally, we demonstrate that strategic choices in output format and cost-conscious model selection significantly impact data generation effectiveness. Our code, checkpoints, and data are all publicly available at https://github.com/neulab/data-agora.

Recent advancements in Large Language Models (LLMs) have showcased striking results on existing logical reasoning benchmarks, with some models even surpassing human performance. However, the true depth of their competencies and robustness in reasoning tasks remains an open question. To this end, in this paper, we focus on two popular reasoning tasks: arithmetic reasoning and code generation. Particularly, we introduce (i) a general ontology of perturbations for math and coding questions, (ii) a semi-automatic method to apply these perturbations, and (iii) two datasets, GSMore and HumanEval-Core, respectively, of perturbed math and coding problems to probe LLM capabilities in numeric reasoning and coding tasks.Through comprehensive evaluations of both closed-source and open-source LLMs, we show a significant performance drop across all the models against the perturbed questions, suggesting that the current LLMs lack robust problem solving skills and structured reasoning abilities in many areas, as defined by our ontology.

With the rapid development of large language models (LLMs), LLM-as-a-judge has emerged as a widely adopted approach for text quality evaluation, including hallucination evaluation. While previous studies have focused exclusively on single-context evaluation (e.g., discourse faithfulness or world factuality), real-world hallucinations typically involve mixed contexts, which remains inadequately evaluated. In this study, we use summarization as a representative task to comprehensively evaluate LLMs’ capability in detecting mixed-context hallucinations, specifically distinguishing between factual and non-factual hallucinations. Through extensive experiments across direct generation and retrieval-based models of varying scales, our main observations are: (1) LLMs’ intrinsic knowledge introduces inherent biases in hallucination evaluation; (2) These biases particularly impact the detection of factual hallucinations, yielding a significant performance bottleneck; and (3) the fundamental challenge lies in effective knowledge utilization, balancing between LLMs’ intrinsic knowledge and external context for accurate mixed-context hallucination evaluation.

Sentence simplification (SS) focuses on adapting sentences to enhance their readability and accessibility. While large language models (LLMs) match task-specific baselines in English SS, their performance in Portuguese remains underexplored. This paper presents a comprehensive performance comparison of 26 state-of-the-art LLMs in Portuguese SS, alongside two simplification models trained explicitly for this task and language. They are evaluated under a one-shot setting across scientific, news, and government datasets. We benchmark the models with our newly introduced Gov-Lang-BR corpus (1,703 complex-simple sentence pairs from Brazilian government agencies) and two established datasets: PorSimplesSent and Museum-PT. Our investigation takes advantage of both automatic metrics and large-scale linguistic analysis to examine the transformations achieved by the LLMs. Furthermore, a qualitative assessment of selected generated outputs provides deeper insights into simplification quality. Our findings reveal that while open-source LLMs have achieved impressive results, closed-source LLMs continue to outperform them in Portuguese SS.

This paper aims to identify directional causal relations between events, including the existence and direction of causality. Previous studies mainly adopt prompt learning paradigm to predict a causal answer word based on a Pre-trained Language Model (PLM) for causality existence identification. However, the indecision in selecting answer words from some synonyms and the confusion of indicating opposite causal directions with the same answer word raise more challenges in directional causality identification. Inspired by the strong capabilities of pre-trained Generative Language Models (GLMs) in generating responses or statements, we propose to instruct a GLM to generate causality statements and identify directional event causality by evaluating the generated statements. Specifically, we propose an Instructive Generation and Statement Evaluation method to identify both the existence and direction of causality. We first fine-tune a GLM to instructively generate causality statements based on event description inputs. Then, we evaluate the rationality of the generated statements to determine the existence and direction of event causalities. Experiments on the ESC and MAVEN datasets show that our method significantly outperforms state-of-the-art algorithms, even with fewer training data.

As large language models (LLMs) become an important way of information access, there have been increasing concerns that LLMs may intensify the spread of unethical content, including implicit bias that hurts certain populations without explicit harmful words. In this paper, we conduct a rigorous evaluation of LLMs’ implicit bias towards certain demographics by attacking them from a psychometric perspective to elicit agreements to biased viewpoints. Inspired by psychometric principles in cognitive and social psychology, we propose three attack approaches, i.e., Disguise, Deception, and Teaching. Incorporating the corresponding attack instructions, we built two benchmarks: (1) a bilingual dataset with biased statements covering four bias types (2.7K instances) for extensive comparative analysis, and (2) BUMBLE, a larger benchmark spanning nine common bias types (12.7K instances) for comprehensive evaluation. Extensive evaluation of popular commercial and open-source LLMs shows that our methods can elicit LLMs’ inner bias more effectively than competitive baselines. Our attack methodology and benchmarks offer an effective means of assessing the ethical risks of LLMs, driving progress toward greater accountability in their development.