Austria

Generation with source attribution is important for enhancing the verifiability of retrieval-augmented generation (RAG) systems. However, existing approaches in RAG primarily link generated content to document-level references, making it challenging for users to locate evidence among multiple content-rich retrieved documents. To address this challenge, we propose Retrieval-Augmented Generation with Visual Source Attribution (VISA), a novel approach that combines answer generation with visual source attribution. Leveraging large vision-language models (VLMs), VISA identifies the evidence and highlights the exact regions that support the generated answers with bounding boxes in the retrieved document screenshots. To evaluate its effectiveness, we curated two datasets: Wiki-VISA, based on crawled Wikipedia webpage screenshots, and Paper-VISA, derived from PubLayNet and tailored to the medical domain. Experimental results demonstrate the effectiveness of VISA for visual source attribution on documents&#39; original look, as well as highlighting the challenges for improvement.

ACL 2025

VISA: Retrieval Augmented Generation with Visual Source Attribution

multimodal applications; retrieval-augmented generation; vision-language model

### 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)

Generation with source attribution is important for enhancing the verifiability of retrieval-augmented generation (RAG) systems. However, existing approaches in RAG primarily link generated content to document-level references, making it challenging for users to locate evidence among multiple content-rich retrieved documents. To address this challenge, we propose Retrieval-Augmented Generation with Visual Source Attribution (VISA), a novel approach that combines answer generation with visual source attribution. Leveraging large vision-language models (VLMs), VISA identifies the evidence and highlights the exact regions that support the generated answers with bounding boxes in the retrieved document screenshots. To evaluate its effectiveness, we curated two datasets: Wiki-VISA, based on crawled Wikipedia webpage screenshots, and Paper-VISA, derived from PubLayNet and tailored to the medical domain. Experimental results demonstrate the effectiveness of VISA for visual source attribution on documents' original look, as well as highlighting the challenges for improvement.

poster

Transformer-based models have achieved state-of-the-art performance in document classification but struggle with long-text processing due to the quadratic computational complexity in the self-attention module. Existing solutions, such as sparse attention, hierarchical models, and key sentence extraction, partially address the issue but still fall short when the input sequence is exceptionally lengthy. To address this challenge, we propose **IRIS** (**I**nterpretable **R**etrieval-Augmented Classification for long **I**nterspersed Document **S**equences), a novel, lightweight framework that utilizes retrieval to efficiently classify long documents while enhancing interpretability. IRIS segments documents into chunks, stores their embeddings in a vector database, and retrieves those most relevant to a given task using learnable query vectors. A linear attention mechanism then aggregates the retrieved embeddings for classification, allowing the model to process arbitrarily long documents without increasing computational cost and remaining trainable on a single GPU. Our experiments across six datasets show that IRIS achieves comparable performance to baseline models on standard benchmarks, and excels in three clinical note disease risk prediction tasks where documents are extremely long and key information is sparse. Furthermore, IRIS provides global interpretability by revealing a clear summary of key risk factors identified by the model. These findings highlight the potential of IRIS as an efficient and interpretable solution for long-document classification, particularly in healthcare applications where both performance and explainability are crucial.

IRIS: Interpretable Retrieval-Augmented Classification for Long Interspersed Document Sequences

Accurately predicting human scanpaths during
reading is vital for diverse fields and downstream tasks, from educational technologies to
automatic question answering. To date, however, progress in this direction remains limited by scarce gaze data. We overcome the
issue with ScanEZ, a self-supervised framework grounded in cognitive models of reading.
ScanEZ jointly models the spatial and temporal dimensions of scanpaths by leveraging synthetic data and a 3-D gaze objective inspired by
masked language modeling. With this framework, we provide evidence that two key factors
in scanpath prediction during reading are: the
use of masked modeling of both spatial and
temporal patterns of eye movements, and cognitive model simulations as an inductive bias
to kick-start training. Our approach achieves
state-of-the-art results on established datasets
(e.g., up to 31.4% negative log-likelihood improvement on CELER L1), and proves portable
across different experimental conditions.

ScanEZ: Integrating Cognitive Models with Self-Supervised Learning for Spatiotemporal Scanpath Prediction

In spoken communication, information is transmitted not only via words, but also through a rich array of non-verbal signals, including prosody---the non-segmental auditory features of speech. Do these different communication channels carry distinct information? Prior work has shown that the information carried by prosodic features is substantially redundant with that carried by the surrounding words. Here, we systematically examine the time scale of this relationship, studying how it varies with the length of past and future contexts. We find that a word's prosodic features require an extended past context (3-8 words across different features) to be reliably predicted. Given that long-scale contextual information decays in memory, prosody may facilitate communication by adding information that is locally unique. We also find that a word’s prosodic features show some redundancy with future words, but only with a short scale of 1-2 words, consistent with reports of incremental short-term planning in language production. Thus, prosody may facilitate communication by helping listeners predict upcoming material. In tandem, our results highlight potentially distinct roles that prosody plays in facilitating integration of words into past contexts and in helping predict upcoming words.

The time scale of redundancy between prosody and linguistic context

Reading is a process that unfolds across space and time. Standard modeling approaches, however, overlook much of the spatio-temporal dynamics involved in reading by relying on aggregated reading measurements---typically only focusing on fixation durations---and employing models with strong simplifying assumptions. In this paper, we propose a generative model that captures not only how long fixations last, but also where they land and when they occur. To this end, we model reading scanpaths via two conditionally independent distributions: one for fixation location and timing, and another for fixation duration.
The location (and timing) of fixation shifts, so-called saccades, are modeled using a spatio-temporal Hawkes process, which captures how each fixation excites the probability of a new fixation occurring near it in time and space. Empirically, our Hawkes process model exhibits higher likelihood on held-out reading data than baselines. The duration time of fixation events is modeled as a function of fixation-specific features convolved across time, thus capturing non-stationary delayed effects. We find that convolution-based approaches demonstrate weak predictive power when modeling disaggregated fixation durations. Similarly, our analysis of surprisal theory on disaggregated data reveals limited effectiveness in predicting both where fixations occur and how long they last.

Fine-Grained Spatio-Temporal Modeling of Reading Behavior

Large language models (LLMs) excel at few-shot in-context learning (ICL) without requiring parameter updates. 
However, as ICL demonstrations increase from a few to many, performance tends to plateau and eventually decline. 
We identify two primary causes for this trend: the suboptimal negative log-likelihood (NLL) optimization objective and the incremental data noise. 
To address these issues, we introduce \textit{DrICL}, a novel optimization method that enhances model performance through \textit{Differentiated} and \textit{Reweighting} objectives. 
Globally, DrICL utilizes differentiated learning to optimize the NLL objective, ensuring that many-shot performance surpasses zero-shot levels. 
Locally, it dynamically adjusts the weighting of many-shot demonstrations by leveraging cumulative advantages inspired by reinforcement learning, thereby mitigating the impact of noisy data.
Recognizing the lack of multi-task datasets with diverse many-shot distributions, we develop the \textit{Many-Shot ICL Benchmark} (ICL-50)-a large-scale benchmark of 50 tasks that cover shot numbers from 1 to 350 within sequences of up to 8,000 tokens-for both fine-tuning and evaluation purposes.
Experimental results demonstrate that LLMs enhanced with DrICL achieve significant improvements in many-shot setups across various tasks, including both in-domain and out-of-domain scenarios.
We release the code and dataset hoping to facilitate further research in many-shot ICL\footnote{https://github.com/xiaoqzhwhu/DrICL}.

More is not always better? Enhancing Many-Shot In-Context Learning with Differentiated and Reweighting Objectives

Large language models (LLMs) have showcased remarkable capabilities as autonomous agents when augmented with external tools. Equipped with fixed tool sets, LLMs struggle with addressing diverse user inquiries in open-world tasks. To evaluate and boost the performance of LLMs in dealing with complex demands in the real-world, we propose open-world function calling, where LLMs need to retrieve suitable tools from a pre-defined external tool library and use retrieved tools to resolve the user's problem. We introduce Meta-Tool, a versatile and plug-and-play tool retrieval system as the access of LLMs to external tool library. Drawing inspiration from the myriad of enhanced approaches associated with Retrieval-Augmented Generation (RAG), Meta-Tool employs a \textit{hypothesize-retrieve-invoke} framework. We further propose Meta-Bench, a comprehensive benchmark for evaluating LLMs in open-world function calling and associated tasks. Meta-Bench encompasses $2,800$ dialogues and $7,361$ tools, spanning ten distinct scenarios to provide robust and diverse test categories. In conjunction, we present MT-LLaMA, a finetuned version of LLaMA-3.1, which exhibits remarkable performance improvements. Our empirical experiments reveal that Meta-Tool significantly enhances the ability of advanced LLMs to retrieve and leverage the most suitable tools compared to previous tool retrieval methods. Moreover, our fine-tuning enables even smaller-sized LLMs to achieve comparable even exceeding results to GPT-4o. Both the benchmark and the model are made publicly available at https://github.com/qinshengqian/Meta-Tool to foster further research and development in the field.

Meta-Tool: Unleash Open-World Function Calling Capabilities of General-Purpose Large Language Models

Clinical coding, assigning standardized codes to medical notes, is critical for epidemiological research, hospital planning, and reimbursement. Neural coding models generally process entire discharge summaries, which are often lengthy and contain information that is not relevant to coding. We propose an approach that combines Named Entity Recognition (NER) and Assertion Classification (AC) to filter for clinically important content before supervised code prediction. On MIMIC-IV, a standard evaluation dataset, our approach achieves near-equivalent performance to a state-of-the-art full-text baseline while using only 22% of the content and reducing training time by over half. Additionally, mapping model attention to complete entity spans yields coherent, clinically meaningful explanations, capturing coding-relevant modifiers such as acuity and laterality. We release a newly annotated NER+AC dataset for MIMIC-IV, designed specifically for ICD coding. Our entity-centric approach lays a foundation for more transparent and cost-effective assisted coding.

Less is More: Explainable and Efficient ICD Code Prediction with Clinical Entities

Clinical diagnosis education requires students to master both systematic reasoning processes and comprehensive medical knowledge. While recent advances in Large Language Models (LLMs) have enabled various medical educational applications, these systems often provide direct answers that could reduce students' cognitive engagement and lead to fragmented learning. Motivated by these challenges, we propose DDxTutor, a framework that follows differential diagnosis principles to decompose clinical reasoning into teachable components. It consists of a structured reasoning module that analyzes clinical clues and synthesizes diagnostic conclusions, and an interactive dialogue framework that guides students through this process. To enable such tutoring, we construct DDxReasoning, a dataset of 933 clinical cases with fine-grained diagnostic steps verified by doctors. Our experiments demonstrate that fine-tuned LLMs achieve strong performance in generating structured teaching references and conducting interactive diagnostic tutoring dialogues. Human evaluation by medical educators and students validates the framework's potential and effectiveness for clinical diagnosis education. Our project is available at https://github.com/med-air/DDxTutor.

DDxTutor: Clinical Reasoning Tutoring System with Differential Diagnosis-Based Structured Reasoning

Empowering LLMs with the ability to precisely understand long contexts is crucial for many downstream applications. However, handling long contexts with conventional transformer architecture requires substantial training and inference resources. Existing context condensing methods cannot accurately understand the full context, as there is a considerable amount of information loss in the condensing process. To address these issues, we present **FocusLLM**, a framework designed to extend the fixed context length of any decoder-only LLM, allowing the model to focus on relevant information from very long sequences. FocusLLM first divides long text input into chunks based on the model's original context length. It then employs the **_dynamic condensing_** process to distill crucial information from each chunk. Ultimately, through the novel **_parallel decoding_** mechanism, FocusLLM can integrate the extracted information into its local context. FocusLLM stands out for great training efficiency and versatility: trained with an 8K input length and with much less training cost than previous methods, FocusLLM exhibits superior performance across downstream tasks and maintains strong language modeling ability when handling extensive long texts, even up to 400K tokens. Our code is available at https://github.com/leezythu/FocusLLM.

FocusLLM: Precise Understanding of Long Context by Dynamic Condensing

Recent advances in large language models (LLMs) have demonstrated remarkable capabilities on widely benchmarked high-resource languages. However, linguistic nuances of under-resourced languages remain unexplored. We introduce Batayan, a holistic Filipino benchmark that systematically evaluates LLMs across three key natural language processing (NLP) competencies: understanding, reasoning, and generation. Batayan consolidates eight tasks, three of which have not existed prior for Filipino corpora, covering both Tagalog and code-switched Taglish utterances. Our rigorous, native-speaker-driven adaptation and validation processes ensures fluency and authenticity to the complex morphological and syntactic structures of Filipino, alleviating the pervasive translationese bias in existing Filipino corpora. We report empirical results on a variety of open-source and commercial LLMs, highlighting significant performance gaps that signal the under-representation of Filipino in pre-training corpora, the unique hurdles in modeling Filipino's rich morphology and construction, and the importance of explicit Filipino language support. Moreover, we discuss the practical challenges encountered in dataset construction and propose principled solutions for building culturally and linguistically-faithful resources in under-represented languages. We also provide a public evaluation suite as a clear foundation for iterative, community-driven progress in Filipino NLP.

Downloads

Next from ACL 2025

IRIS: Interpretable Retrieval-Augmented Classification for Long Interspersed Document Sequences

Similar lecture

Evaluation of Attribution Bias in Generator-Informed Retrieval-Augmented Large Language Models