AACL 2025 Papers

Retrieval-augmented generation (RAG) still forwards raw passages to large-language models, so private facts slip through. Prior defenses are either (i) heavyweight—full DP training that is impractical for today’s 70B-parameter models—or (ii) over-zealous—blanket redaction of every named entity, which slashes answer quality. We introduce VAGUE-Gate, a lightweight, locally differentially-private gate deployable in front of any RAG system. A precision pass drops low-utility tokens under a user budget ε, then up to k(ε) high-temperature paraphrase passes further cloud residual cues; post-processing guarantees preserve the same ε-LDP bound.To measure both privacy and utility, we release BlendPriv (3k blended-sensitivity QA pairs) and two new metrics: a lexical Information-Leakage Score and an LLM-as-Judge score. Across eight pipelines and four SOTA LLMs, VAGUE-Gate at ε = 0.3 lowers lexical leakage by 70% and semantic leakage by 1.8 points (1–5 scale) while retaining 91% of Plain-RAG faithfulness with only a 240 ms latency overhead. All code, data, and prompts are publicly released:- Code: https://github.com/arshiahemmat/LDP_RAG - Dataset: https://huggingface.co/datasets/AliMnp/BlendPriv

Survey papers play a critical role in scientific communication by consolidating progress across a field. Recent advances in Large Language Models (LLMs) offer a promising solution by automating key steps in the survey-generation pipeline, such as retrieval, structuring, and summarization. However, existing LLM-based approaches often struggle with maintaining coherence across long, multi-section surveys and providing comprehensive citation coverage. To address these limitations, we introduce SurveyGen-I, an automatic survey generation framework that combines coarse-to-fine retrieval, adaptive planning, and memory-guided generation. SurveyGen-I performs survey-level retrieval to construct the initial outline and writing plan, then dynamically refines both during generation through a memory mechanism that stores previously written content and terminology, ensuring coherence across subsections. When the system detects insufficient context, it triggers fine-grained subsection-level retrieval. During generation, SurveyGen-I leverages this memory mechanism to maintain coherence across subsections. Experiments across six scientific domains demonstrate that SurveyGen-I consistently outperforms previous works in content quality, consistency, and citation coverage. The code is available at https://github.com/SurveyGens/SurveyGen-I.

Recent advances in multimodal question answering have primarily focused on combining heterogeneous modalities or fine-tuning multimodal large language models. While these approaches have shown strong performance, they often rely on a single, generalized reasoning strategy, overlooking the unique characteristics of each modality ultimately limiting both accuracy and interpretability. To address these limitations, we propose MAMMQA, a multi-agent QA framework for multimodal inputs spanning text, tables, and images. Our system includes two Visual Language Model (VLM) agents and one text-based Large Language Model (LLM) agent. The first VLM decomposes the user query into sub-questions and sequentially retrieves partial answers from each modality. The second VLM synthesizes and refines these results through cross-modal reasoning. Finally, the LLM integrates the insights into a cohesive answer. This modular design enhances interpretability by making the reasoning process transparent and allows each agent to operate within its domain of expertise. Experiments on diverse multimodal QA benchmarks demonstrate that our cooperative, multi-agent framework consistently outperforms existing baselines in both accuracy and robustness.

Mental health plays a crucial role in the overall well-being of an individual. In recent years, digital platforms have increasingly been used to expand mental health and emotional support. However, there are persistent challenges related to limited user accessibility, internet connectivity, and data privacy, which highlight the need for an offline, smartphone-based solutions. To address these challenges, we propose EmoSApp (Emotional Support App): an entirely offline, smartphone-based conversational app designed to provide mental health and emotional support. EmoSApp leverages a language model, specifically the LLaMA-3.2-1B-Instruct, which is fine-tuned and quantized on a custom-curated “Knowledge Dataset” comprising 14,582 mental health QA pairs along with multi-turn conversational data, enabling robust domain expertise and fully on-device inference on resource-constrained smartphones.Through qualitative evaluation with students and mental health professionals, we demonstrate that EmoSApp has the ability to respond coherently and empathetically, provide relevant suggestions to user’s mental health problems, and maintain interactive dialogue. Additionally, quantitative evaluations on nine commonsense and reasoning benchmarks, along with two mental health specific datasets, demonstrate EmoSApp’s effectiveness in low-resource settings. By prioritizing on-device deployment and specialized domain-specific adaptation, EmoSApp serves as a blueprint for future innovations in portable, secure, and highly tailored AI-driven mental health support.

Recent advances in speech‐enabled AI, including Google’s NotebookLM and OpenAI’s speech-to-speech API, are driving widespread interest in voice interfaces across sectors such as finance, health, agritech, legal services, and call‐centers in the global north and south. Despite this momentum, there exists no publicly available application-specific model evaluation that caters to Africa’s linguistic diversity. We present \textbf{AfriSpeech‑MultiBench}, the first domain‐specific evaluation suite for over 100 African English accents across 10+ countries and seven application domains: Finance, Legal, Medical, General dialogue, Call Center, Named Entities, and Hallucination Robustness. We benchmark a diverse range of open, closed, unimodal ASR and multimodal LLM-based speech recognition systems using both spontaneous and non-spontaneous speech conversations drawn from various open African accented English speech datasets. Our empirical analysis reveals systematic variation: open‐source ASR excels in spontaneous speech contexts but degrades on noisy, non‐native dialogue; multimodal LLMs are more accent‐robust yet struggle with domain‐specific named entities; proprietary models deliver high accuracy on clean speech but vary significantly by country and domain. Smaller models fine‐tuned on African English achieve competitive accuracy with lower latency, a practical advantage for deployment. By releasing this benchmark, we empower practitioners and researchers to select voice technologies suited to African use‐cases, fostering inclusive voice applications for undeserved communities.

Large Language Models (LLMs) have shown remarkable capabilities, but their development has primarily focused on English and other high-resource languages, leaving many languages underserved. We present our latest Hindi-English bi-lingual LLM with ~3% average improvement in benchmark scores over both languages, outperforming models twice its size. Using a curated dataset composed of English and Hindi instruction data of 485K samples, we instruction tuned models such as Qwen-2.5-14B-Instruct and Phi-4 to improve performance over both English and Hindi. Our experiments encompassing seven different LLMs of varying parameter sizes and over 140 training attempts with varying English-Hindi training data ratios demonstrated that it is possible to significantly improve multilingual performance without compromising native performance. Further, our approach avoids resource-intensive techniques like vocabulary expansion or architectural modifications, thus keeping the model size small. Our results indicate that modest fine-tuning with culturally and locally informed data can bridge performance gaps without incurring significant computational overhead. We release our training code, datasets, and models under mit and apache licenses to aid further research towards under-represented and low-resource languages.

Unlike document-level AI-generated text (AIGT) detection, sentence-level AIGT detection remains underexplored, despite its importance for addressing collaborative writing scenarios where humans modify AIGT suggestions on a sentence-by-sentence basis. Prior sentence-level detectors often neglect the valuable context surrounding the target sentence, which may contain crucial linguistic artifacts that indicate a potential change in authorship. We propose GL-CLiC, a novel technique that leverages both Global and Local signals of Coherence and Lexical Complexity, which we operationalize through discourse analysis and CEFR-based vocabulary sophistication. GL-CLiC models local coherence and lexical complexity by examining a sentence’s relationship with its neighbors or peers, complemented with its document-wide analysis. Our experimental results show that GL-CLiC achieves superior performance and better generalization across domains compared to existing methods.

Although Farsi is widely spoken, no comprehensive benchmark exists for assessing commonsense reasoning in language models. We therefore present FarSense, a 6‐task benchmark for Farsi covering True/False judgment, multiple-choice questions, Explanation, Cause‐Effect inference, Counterfactual reasoning, and Knowledge Completion. Starting from Farsi‐Wikipedia, we filtered noise and retained ~4,210 passages, rewrote them into realistic daily scenarios, and derived the above tasks from each scenario. Scenario and task generation quality was first judged via native‐speaker annotations on outputs from five major LLMs—GPT‐4o, Gemini-2.5-Flash, Mistral-Large, Qwen‐Plus, and DeepSeek‐Chat. Gemini-2.5-Flash demonstrated the highest performance, leading to its use in generating a large-scale dataset, subsequently finalized through meticulous two-step human validation. Using FarSense, we measured the commonsense ability of the same five flagship LLMs and also fine‐tuned six compact models (1B–24B parameters) before re‐evaluating them. To ensure broad applicability, task wording was designed to minimize dialectal, cultural, or religious bias. Experiments show that targeted fine‐tuning yields substantial gains, confirming FarSense as a reliable, openly licensed resource for advancing reproducible commonsense understanding research in Farsi NLP. We publicly release all code and data at https://github.com/KamyarZeinalipour/FarSense.

Although the use of persona prompting in large language models appears to trigger different styles of generated text, it is unclear whether these translate into measurable behavioral differences. Furthermore, little work has studied whether these differences, when they do exist, can affect decision-making in an adversarial strategic environment. We investigate the impact of persona prompting on strategic performance in PERIL, a world domination board game. Specifically, we compare the effectiveness of persona-derived heuristics to those chosen manually. Our findings reveal that personality traits intuitively associated with strategic thinking do appear to improve game performance, but only when an additional mediator is used to translate personas into heuristic values. We introduce this mediator as a structured translation process, inspired by exploratory factor analysis, that maps LLM-generated inventory responses into strategic heuristics. Results indicate our method enhances heuristic reliability and face validity when compared to directly inferred heuristics, allowing us to better study the effect of persona types on decision-making behaviors. These insights advance our understanding of how persona prompting influences LLM-based decision-making and propose a novel heuristic generation method that adds to the growing body of work applying psychometric principles to LLMs.

Large language models (LLMs) provide detailed and impressive responses to queries in English. However, are they really consistent at responding to the same query in other languages? The popular way of evaluating for multilingual performance of LLMs requires expensive-to-collect annotated datasets. Further, evaluating for tasks like open-ended generation, where multiple correct answers may exist, is nontrivial. Instead, we propose to evaluate the predictability of model response across different languages. In this work, we propose a framework to evaluate LLM’s cross-lingual consistency based on a simple Translate then Evaluate strategy. We instantiate this evaluation framework along two dimensions of consistency: information and empathy. Our results reveal pronounced inconsistencies in popular LLM responses across thirty languages, with severe performance deficits in certain language families and scripts, underscoring critical weaknesses in their multilingual capabilities. These findings necessitate cross-lingual evaluations that are consistent along multiple dimensions. We invite practitioners to use our framework for future multilingual LLM benchmarking.

Pretraining on massive corpora has given rise to large language models (LLMs) with multi-task capabilities. However, real-world applications often require more specialized training, as is the case of NL2Code. We approach this specialization through the lens of data selection, i.e., identifying a subset of a large corpus that aligns with a desired target distribution—a challenge that remains under-explored within NL2Code. Existing methods are typically designed for selecting instruction-tuning data, and might not easily scale to large-scale code repositories; while methods for NL2Code do exist, they primarily rely on coarse heuristics—–such as repo stars—–for filtering. To bridge this gap, we propose FINDR, an efficient data selection method that extends logistic regression with feature-wise importance reweighting—marking it, to our knowledge, the first fine-grained solution to NL2Code pretraining. Our method uses hashed n-grams and code-aware features to capture code-specific patterns, and then apply informative priors to reweight feature importance when computing influence scores. Extensive experiments on NL2Python and NL2SQL, with two model families, show that FINDR consistently outperforms strong baselines in both execution accuracy and token efficiency. Notably, pretraining on only 2% of FINDR-selected data boosts Gemma by over 29% in both domains, even surpassing CodeGemma (pretrained on 300x more examples) by 10% in Python.

Spatial Reasoning is an important component of human cognition and is an area in which the latest Vision-language models (VLMs) show signs of difficulty. The current analysis papers often use image captioning tasks and visual question answering. In this work, we propose using the Referring Expression Comprehension task instead as a platform for the evaluation of spatial reasoning by VLMs. This platform provides the opportunity for a deeper analysis of spatial comprehension and grounding abilities when there is 1) ambiguity in object detection, 2) complex spatial expressions with a longer sentence structure and multiple spatial relations, and 3) expressions with negation (‘not’). In our analysis, we use task-specific architectures as well as large VLMs and highlight their strengths and weaknesses in dealing with these specific situations. While all these models face challenges with the task at hand, the relative behaviors depend on the underlying models and the specific categories of spatial semantics (topological, directional, proximal, etc.). Our results highlight these challenges and behaviors and provide insight into research gaps and future directions.

Dense Retrieval (DR) models have proven to be effective for Document Retrieval and Information Grounding tasks. Usually, these models are trained and optimized for improving the relevance of top-ranked documents for a given query. Previous work has shown that popular DR models are sensitive to the query and document lexicon: small variations of it may lead to a significant difference in the set of retrieved documents. In this paper, we propose a variation of the Multi-Negative Ranking loss for training DR that improves the coherence of models in retrieving the same documents with respect to semantically similar queries. The loss penalizes discrepancies between the top-k ranked documents retrieved for diverse but semantically equivalent queries. We conducted extensive experiments on various datasets, MS-MARCO, Natural Questions, BEIR, and TREC DL 19/20. The results show that (i) models optimizes by our loss are subject to lower sensitivity, and, (ii) interestingly, higher accuracy.

Recent advances in large language models (LLMs) have significantly impacted data science workflows, giving rise to specialized data science agents designed to automate analytical tasks. Despite rapid adoption, systematic benchmarks evaluating the efficacy and limitations of these agents remain scarce. In this paper, we introduce a comprehensive benchmark specifically crafted to reflect real-world user interactions with data science agents by observing usage of our commercial applications. We evaluate three LLMs: Claude-4.0-Sonnet, Gemini-2.5-Flash, and OpenAI-o4-Mini across three approaches: zero-shot with context engineering, multi-step with context engineering, and with SmolAgent. Our benchmark assesses performance across a diverse set of eight data science task categories, additionally exploring the sensitivity of models to common prompting issues, such as data leakage and slightly ambiguous instructions. We further investigate the influence of temperature parameters on overall and task-specific outcomes for each model and approach. Our findings reveal distinct performance disparities among the evaluated models and methodologies, highlighting critical factors that affect practical deployment. The benchmark dataset and evaluation framework introduced herein aim to provide a foundation for future research of more robust and effective data science agents.

In recent times, there has been a shift towards adapting sequential recommendation to LLM paradigm to harness the capabilities of LLMs. These methods typically formulate recommendation data into natural language and train the model to forecast the subsequent item in an auto-regressive manner. Despite their notable success, the significant computational burden during inference poses a major challenge to their practical implementation. In this study, we aim to streamline current LLM-based recommendation models and introduce a straightforward yet highly effective model Lite-LLM4Rec. The primary objective of Lite-LLM4Rec is to ensure efficient inference for the sequential recommendation task. Lite-LLM4Rec circumvents the step-by-step beam search decoding by employing a direct item projection head to produce ranking scores in one step. This design arises from our empirical finding that beam search decoding is ultimately unnecessary for sequential recommendations. Additionally, Lite-LLM4Rec introduces a hierarchical LLM structure crafted to efficiently handle the extensive contextual information of items and redundant computation issue, thus diminishing computational overhead while enjoying the power of LLMs. Experiments on four publicly available datasets validate the efficacy of Lite-LLM4Rec in enhancing both performance and inference efficiency (notably 46.8% performance improvement and 99.48% efficiency improvement on ML-1m) compared to existing LLM-based methods. Our implementations are available at: https://github.com/HanbingWang2001/Lite-LLM4Rec-PyTorch.

Multi-agent systems with smaller language models (SLMs) present a viable alternative to single agent systems powered by large language models (LLMs) for addressing complex problems. In this work, we study how these alternatives compare in terms of both effectiveness and efficiency. To study this trade-off, we instantiate single and multi-agent systems for the complex problems in the AppWorld environment using different sized language models.We find that difficulties with long-trajectory learning in smaller language models (SLMs) limit their performance. Even when trained for specialized roles, SLMs fail to learn all subtasks effectively. To address this issue, we introduce a simple progressive sub-task training strategy, which introduces new sub-tasks progressively in each training epoch. We find that this novel strategy, analogous to instance level curriculum learning, consistently improves the effectiveness of multi-agents at all configurations. Our Pareto analysis shows that fine-tuned multi-agent systems yield better effectiveness-efficiency trade-offs. Additional ablations and analyses shows the importance of our progressive training strategy and its ability to reduce subtask error rates.

In recent years, machine translation has evolved with the integration of multimodal information. Infusion of multi-modality into translation tasks decreases ambiguation and enhances translation scores. Common modalities include images, speech, and videos, which provide additional context alongside the text to be translated. While multimodal translation with images has been extensively studied, video-guided machine translation (VMT) has gained increasing attention, particularly since Wang et al. 2019 first explored this task. In this paper, we provide a comprehensive overview of VMT, highlighting its unique challenges, methodologies, and recent advancements. Unlike previous surveys that primarily focus on image-guided multimodal machine translation, this work explores the distinct complexities and opportunities introduced by adding video as a modality to the translation task.

Test-time scaling has emerged as an effective approach for improving language model performance by utilizing additional compute at inference time. Recent studies have shown that overriding end-of-thinking tokens (e.g., replacing "/think>” with “Wait”) can extend reasoning steps and improve accuracy. In this work, we explore whether a dedicated continue-thinking token can be learned to trigger extended reasoning. We augment distilled versions of DeepSeek-R1 with a single learned "<|continue-thinking|>” token, training only its embedding via reinforcement learning while keeping the model weights frozen. Our experiments show that this learned token achieves improved accuracy on standard math benchmarks compared to both the baseline model and a test-time scaling approach that uses a fixed token (e.g., “Wait”) for budget forcing. In particular, we observe that in cases where the fixed-token approach enhances the base model’s accuracy, our method achieves a markedly greater improvement. For example, on the GSM8K benchmark, the fixed-token approach yields a 1.3% absolute improvement in accuracy, whereas our learned-token method achieves a 4.2% improvement over the base model that does not use budget forcing.

Many real-world eligibility problems, ranging from medical diagnosis to tax planning, can be mapped to decision problems expressed in natural language, wherein a model must make a binary choice based on the features of the user. Large-scale domains such as legal codes or frequently updated funding opportunities render human annotation (e.g., web forms or decision trees) impractical, suggesting a need for agents that can automatically assist in decision-making. Since relevant information is often only known to the user, it is important that these agents can ask the right questions. To evaluate this task, we propose BeNYfits, a new benchmark for determining user eligibility for multiple overlapping social benefits opportunities through interactive decision-making. Our experiments show that current language models struggle with frequent hallucinations, with GPT-4o scoring only 35.7 F1 using a ReAct-style chain-of-thought. We therefore introduce ProADA, a novel approach that uses program synthesis to assist in decision-making by mapping dialog planning to a code generation problem and using gaps in structured data to determine the best next action. Our agent, ProADA, improves the F1 score to 56.2 while using nearly the same number of dialog turns.