论文检索

Dhruv Gupta, Gayathri Ganesh Lakshmy, Yiqing Xie

In this work, we conduct an in-depth analysis of code retrieval by systematically masking specific features while preserving code functionality. Our discoveries include: (1) although trained on code, current retrievers heavily rely on surface-level textual features (e.g., docstrings, identifier names), and (2) they exhibit a strong bias towards well-documented code, even if the documentation is irrelevant. Based on our discoveries, we propose SACL, a framework that enriches textual information and reduces bias by augmenting code or structural knowledge with semantic information. Extensive experiments show that SACL substantially improves code retrieval (e.g., by 12.8% / 9.4% / 7.0% Recall@1 on HumanEval / MBPP / SWE-Bench-Lite), which also leads to better code generation performance (e.g., by 4.88% Pass@1 on HumanEval).

Arti Rani, Shweta Singh, Nihar Ranjan Sahoo, Gaurav Kumar Nayak

Large Language Models (LLMs) have achieved remarkable success on question answering (QA) tasks, yet they often encode harmful biases that compromise fairness and trustworthiness. Most existing bias mitigation approaches are restricted to predefined categories, limiting their ability to address novel or context-specific emergent biases. To bridge this gap, we tackle the novel problem of open-set bias detection and mitigation in text-based QA. We introduce _OpenBiasBench_, a comprehensive benchmark designed to evaluate biases across a wide range of categories and subgroups, encompassing both known and previously unseen biases. Additionally, we propose _Open-DeBias_, a novel, data-efficient, and parameter-efficient debiasing method that leverages adapter modules to mitigate existing social and stereotypical biases while generalizing to unseen ones. Compared to the state-of-the-art BMBI method, Open-DeBias improves QA accuracy on BBQ dataset by nearly **48%** on ambiguous subsets and **6%** on disambiguated ones, using adapters fine-tuned on just a small fraction of the training data. Remarkably, the same adapters, in a zero-shot transfer to Korean BBQ, achieve **84% accuracy**, demonstrating robust language-agnostic generalization. Through extensive evaluation, we also validate the effectiveness of Open-DeBias across a broad range of NLP tasks, including StereoSet and CrowS-Pairs, highlighting its robustness, multilingual strength, and suitability for general-purpose, open-domain bias mitigation. The project page is available at: [https://sites.google.com/view/open-debias25](https://sites.google.com/view/open-debias25)

Patrícia Schmidtová, Ondrej Dusek, Saad Mahamood

We examine evaluation of faithfulness to input data in the context of hotel highlights—brief LLM-generated summaries that capture unique features of accommodations. Through human evaluation campaigns involving categorical error assessment and span-level annotation, we compare traditional metrics, trainable methods, and LLM-as-a-judge approaches. Our findings reveal that simpler metrics like word overlap correlate surprisingly well with human judgments (r=0.63), often outperforming more complex methods when applied to out-of-domain data. We further demonstrate that while LLMs can generate high-quality highlights, they prove unreliable for evaluation as they tend to severely under- or over-annotate. Our analysis of real-world business impacts shows incorrect and non-checkable information pose the greatest risks. We also highlight challenges in crowdsourced evaluations.

Yangyi Li, Mengdi Huai

Large language models (LLMs) have shown strong capabilities, enabling concise, context-aware answers in question answering (QA) tasks. The lack of transparency in complex LLMs has inspired extensive research aimed at developing methods to explain large language behaviors. Among existing explanation methods, natural language explanations stand out due to their ability to explain LLMs in a self-explanatory manner and enable the understanding of model behaviors even when the models are closed-source. However, despite these promising advancements, there is no existing work studying how to provide valid uncertainty guarantees for these generated natural language explanations. Such uncertainty quantification is critical in understanding the confidence behind these explanations. Notably, generating valid uncertainty estimates for natural language explanations is particularly challenging due to the auto-regressive generation process of LLMs and the presence of noise in medical inquiries. To bridge this gap, in this work, we first propose a novel uncertainty estimation framework for these generated natural language explanations, which provides valid uncertainty guarantees in a post-hoc and model-agnostic manner. Additionally, we also design a novel robust uncertainty estimation method that maintains valid uncertainty guarantees even under noise. Extensive experiments on QA tasks demonstrate the desired performance of our methods.

Rajarshi Haldar, Julia Hockenmaier

As Natural Language Generation (NLG) continues to be widely adopted, properly assessing it has become quite difficult. Lately, using large language models (LLMs) for evaluating these generations has gained traction, as they tend to align more closely with human preferences than conventional n-gram or embedding-based metrics. In our experiments, we show that LLM judges have low intra-rater reliability in their assigned scores across different runs. This variance makes their ratings inconsistent, almost arbitrary in the worst case, making it difficult to measure how good their judgments actually are. We quantify this inconsistency across different NLG tasks and benchmarks and see if judicious use of LLM judges can still be useful following proper guidelines.

Jessica Maghakian, Raunak Sinha, Max Schettewi, Gunkirat Kaur

Retrieval-Augmented Generation (RAG) is the current state-of-the-art method for mitigating the shortcomings of large language models (LLMs) by incorporating external knowledge sources to provide more relevant and accurate responses to user queries. However building performant RAG systems for real use-cases typically requires heavy investment from NLP experts, such as fine-tuning embedding models for specialized domains, experimenting with text chunking strategies and other niche hyperparameter tunings. We propose Embedding-Free RAG, a model-agnostic approach that enables the deployment of a one-size-fits-all RAG pipeline for user-provided grounding documents. Unlike traditional RAG, which relies on embedding models for information retrieval, Embedding-Free RAG leverages the generalized reasoning abilities of LLMs in a novel algorithmic framework during the retrieval stage. Extensive experiments demonstrate that Embedding-Free RAG outperforms existing state-of-the-art methods, achieving up to 4.6x higher F1 scores and up to 2x better question answering accuracy across a wide range of challenging domains.

Diyam Akra, Mohammed Khalilia, Mustafa Jarrar

Multidialectal Arabic POS tagging is challenging due to the morphological richness and high variability among dialects. While POS tagging for MSA has advanced thanks to the availability of annotated datasets, creating similar resources for dialects remains costly and labor-intensive. Increasing the size of annotated datasets does not necessarily result in better performance. Active learning offers a more efficient alternative by prioritizing annotating the most informative samples. This paper proposes an active learning approach for multidialectal Arabic POS tagging. Our experiments revealed that annotating approximately 15,000 tokens is sufficient for high performance. We further demonstrate that using a fine-tuned model from one dialect to guide the selection of initial samples from another dialect accelerates convergence—reducing the annotation requirement by about 2,000 tokens. In conclusion, we propose an active learning pipeline and demonstrate that, upon reaching its defined stopping point of 16,000 annotated tokens, it achieves an accuracy of 97.6% on the Emirati Corpus.

Ananya Malik, Nazanin Sabri, Melissa M. Karnaze, Mai ElSherief

Large Language Models’ (LLMs) ability to converse naturally is empowered by their ability to empathetically understand and respond to their users. However, emotional experiences are shaped by demographic and cultural contexts. This raises an important question: Can LLMs demonstrate equitable empathy across diverse user groups? We propose a framework to investigate how LLMs’ cognitive and affective empathy vary across user personas defined by intersecting demographic attributes. Our study introduces a novel intersectional analysis spanning 315 unique personas, constructed from combinations of age, culture, and gender, across four LLMs. Results show that attributes profoundly shape a model’s empathetic responses. Interestingly, we see that adding multiple attributes at once can attenuate and reverse expected empathy patterns. We show that they broadly reflect real-world empathetic trends, with notable misalignments for certain groups, such as those from Confucian culture. We complement our quantitative findings with qualitative insights to uncover model behaviour patterns across different demographic groups. Our findings highlight the importance of designing empathy-aware LLMs that account for demographic diversity to promote more inclusive and equitable model behaviour.

Ananya Malik, Kartik Sharma, Shaily Bhatt, Lynnette Hui Xian Ng

Large Language Models (LLMs) offer a lucrative promise for scalable content moderation, including hate speech detection. However, they are also known to be brittle and biased against marginalised communities and dialects. This requires their applications to high-stakes tasks like hate speech detection to be critically scrutinized. In this work, we investigate the robustness of hate speech classification using LLMs particularly when explicit and implicit markers of the speaker’s ethnicity are injected into the input. For explicit markers, we inject a phrase that mentions the speaker’s linguistic identity. For the implicit markers, we inject dialectal features. By analysing how frequently model outputs flip in the presence of these markers, we reveal varying degrees of brittleness across 3 LLMs and 1 LM and 5 linguistic identities. We find that the presence of implicit dialect markers in inputs causes model outputs to flip more than the presence of explicit markers. Further, the percentage of flips varies across ethnicities. Finally, we find that larger models are more robust. Our findings indicate the need for exercising caution in deploying LLMs for high-stakes tasks like hate speech detection.

Takumi Goto, Yusuke Sakai, Taro Watanabe

Reference-free evaluation metrics for grammatical error correction (GEC) have achieved high correlation with human judgments.However, these metrics are not designed to evaluate adversarial systems that aim to obtain unjustifiably high scores. The existence of such systems undermines the reliability of automatic evaluation, as it can mislead users in selecting appropriate GEC systems. In this study, we propose adversarial attack strategies for four reference-free metrics: SOME, Scribendi, IMPARA, and LLM-based metrics, and demonstrate that our adversarial systems outperform the current state-of-the-art. These findings highlight the need for more robust evaluation methods.

Ixak Sarasua, Ander Corral, Xabier Saralegi

This paper investigates how open-weight instruction-tuned large language models (LLMs) can be efficiently adapted to low-resource languages without requiring costly large-scale post-training. We introduce DIPLomA (Decoupled Instruction-Preserving Language Adaptation), a lightweight delta-based transfer strategy that provides a practical and effective solution for this scenario. DIPLomA decouples language adaptation from post-training alignment by first continually pretraining a foundational LLM on a modest amount of monolingual target-language data while anchoring on English replay, and then injecting instruction-following capabilities via delta-based weight merging from the instructed counterpart of the base LLM. We evaluate DIPLomA on Basque and validate its generality on Welsh and Swahili, demonstrating consistent and substantial gains in instruction-following, linguistic proficiency, and safety. Compared to strong baselines, our method achieves average relative improvements of 50 points in Basque, 63 in Welsh, and 51 in Swahili, while preserving the original model’s multilingual performance. These results highlight DIPLomA as an effective, resource-efficient strategy for bringing high-quality instruction alignment to underrepresented languages at scale.

Devvrat Joshi, Islem Rekik

Extracting entities and relations from scientific text is challenging due to long sentences with densely packed entities. Pipeline approaches address this by first extracting entities and then predicting relations between all possible entity pairs. Since the relation extraction phase operates over this exhaustive set, the inclusion of candidate pairs that may be semantically related but lack syntactic proximity introduces precision errors, ultimately reducing Rel+ F1 metric. We propose a simple yet effective syntactic filtering method based on dependency parsing to prune unlikely entity pairs before relation prediction. By leveraging syntactic proximity in the dependency parse tree, our approach retains structurally plausible pairs and reduces false positives in downstream relation classification. Our method is grounded in consistent statistical patterns observed across all evaluated datasets, reinforcing its generalizability and effectiveness. We integrate this filtering step into architectures such as PL-Marker and HGERE, and evaluate its impact across multiple datasets. Our method improves Rel+ F1 scores significantly by an absolute increase of 3.5–10.3% on SciERC, SciER, and ACE05 datasets. These results highlight the importance of syntactic cues for accurate relation extraction in complex domains like scientific literature.

Rahul Atul Bhope, Praveen Venkateswaran, K. R. Jayaram, Vatche Isahagian, Vinod Muthusamy, Nalini Venkatasubramanian

Developers using LLMs and LLM-based agents in their applications have provided plenty of anecdotal evidencethat in-context-learning (ICL) is fragile. In this paper, we show that in addition to the quantity and quality of examples, the order in which the in-context examples are listed in the prompt affects the output of the LLM and, consequently, their performance. While prior work has explored improving ICL through dataset-dependent techniques, we introduce , a purely inference-time, dataset-free optimization method that efficiently determines the best example order. OptiSeq leverages log probabilities of LLM-generated outputs to systematically prune the search space of possible orderings and recommend the best order(s) by distinguishing orderings that yield high levels of accuracy and those that underperform. Extensive empirical evaluation on multiple LLMs, datasets, and prompts demonstrates that OptiSeq improves accuracy by 5.5 - 10.5 percentage points across multiple tasks.

Jianfei Zhao, Feng Zhang, Xin Sun, Chong Feng

Due to the unidirectional masking mechanism, Decoder-Only models propagate information from left to right. LVLMs (Large Vision-Language Models) follow the same architecture, with visual information gradually integrated into semantic representations during forward propagation. Through systematic analysis, we observe that over 80% of the visual information is absorbed into the semantic representations. However, the model’s attention still predominantly focuses on the visual representations. This misalignment between the attention distribution and the actual information flow undermines the model’s visual understanding ability and contributes to hallucinations.To address this issue, we enhance the model’s visual understanding by leveraging the core information embedded in semantic representations. Specifically, we identify attention heads that focus on core semantic representations based on their attention distributions. Then, through a two-stage optimization paradigm, we propagate the advantages of these attention heads across the entire model, aligning the attention distribution with the actual information flow.We evaluate our method on three image captioning benchmarks using five different LVLMs,demonstrating its effectiveness in significantly reducing hallucinations. Further experiments reveal a trade-off between reduced hallucinations and richer details. Notably, our method allows for manual adjustment of the model’s conservativeness, enabling flexible control to meet diverse real-world requirements.

Pouya Pezeshkpour, Moin Aminnaseri, Estevam Hruschka

Vision-language models (VLMs) have demonstrated impressive performance by effectively integrating visual and textual information to solve complex tasks. However, it is not clear how these models reason over the visual and textual data together, nor how the flow of information between modalities is structured. In this paper, we examine how VLMs reason by analyzing their biases when confronted with scenarios that present conflicting image and text cues—a common occurrence in real-world applications. To uncover the extent and nature of these biases, we build upon existing benchmarks to create five datasets containing mismatched image-text pairs, covering topics in mathematics, science, and visual descriptions. Our analysis shows that VLMs favor text in simpler queries but shift toward images as query complexity increases. This bias correlates with model scale, with the difference between the percentage of image- and text-preferred responses ranging from +56.8% (image favored) to -85.1% (text favored), depending on the task and model. In addition, we explore three mitigation strategies: simple prompt modifications, modifications that explicitly instruct models on how to handle conflicting information (akin to chain-of-thought prompting), and a task decomposition strategy that analyzes each modality separately before combining their results. Our findings indicate that the effectiveness of these strategies in identifying and mitigating bias varies significantly and is closely linked to the model’s overall performance on the task and the specific modality in question. We released our dataset and code.

Ziqing Wang, Chengsheng Mao, Xiaole Wen, Yuan Luo, Kaize Ding

Medical Multimodal Large Language Models (Med-MLLMs) have shown great promise in medical visual question answering (Med-VQA). However, when deployed in low-resource settings where abundant labeled data are unavailable, existing Med-MLLMs commonly fail due to their medical reasoning capability bottlenecks: (i) the intrinsic reasoning bottleneck that ignores the details from the medical image; (ii) the extrinsic reasoning bottleneck that fails to incorporate specialized medical knowledge. To address those limitations, we propose AMANDA, a training-free agentic framework that performs medical knowledge augmentation via LLM agents. Specifically, our intrinsic medical knowledge augmentation focuses on coarse-to-fine question decomposition for comprehensive diagnosis, while extrinsic medical knowledge augmentation grounds the reasoning process via biomedical knowledge graph retrieval. Extensive experiments across eight Med-VQA benchmarks demonstrate substantial improvements in both zero-shot and few-shot Med-VQA settings. The code is available at https://github.com/REAL-Lab-NU/AMANDA.

Juan Pablo Munoz, Jinjie Yuan

Test-Time Compute (TTC) has emerged as a powerful paradigm for enhancing the performance of Large Language Models (LLMs) at inference, leveraging strategies such as Test-Time Training (TTT) and Retrieval-Augmented Generation (RAG). However, the optimal adaptation strategy varies across queries, and indiscriminate application of TTC strategy incurs substantial computational overhead. In this work, we introduce Reward-Guided Test-Time Compute (RTTC), a novel framework that adaptively selects the most effective TTC strategy for each query via a pretrained reward model, maximizing downstream accuracy across diverse domains and tasks. RTTC operates in a distributed server-client architecture, retrieving relevant samples from a remote knowledge base and applying RAG or lightweight fine-tuning on client devices only when necessary. To further mitigate redundant computation, we propose Query-State Caching, which enables the efficient reuse of historical query states at both retrieval and adaptation levels. Extensive experiments across multiple LLMs and benchmarks demonstrate that RTTC consistently achieves superior accuracy compared to vanilla RAG or TTT, validating the necessity of adaptive, reward-guided TTC selection and the potential of RTTC for scalable, high-performance language model adaptation.

Ting Cai, Stephen Sheen, AnHai Doan

Expanding the abbreviated column names of tables, such as “esal” to “employee salary”, is critical for many downstream NLP tasks for tabular data, such as NL2SQL, table QA, and keyword search. This problem arises in enterprises, domain sciences, government agencies, and more. In this paper, we make three contributions that significantly advance the state of the art. First, we show that the synthetic public data used by prior work has major limitations, and we introduce four new datasets in enterprise/science domains, with real-world abbreviations. Second, we show that accuracy measures used by prior work seriously undercount correct expansions, and we propose new synonym-aware measures that capture accuracy much more accurately. Finally, we develop Columbo, a powerful LLM-based solution that exploits context, rules, chain-of-thought reasoning, and token-level analysis. Extensive experiments show that Columbo significantly outperforms NameGuess, the current most advanced solution, by 4-29%, over five datasets. Columbo has been used in production on EDI, a major data lake for environmental sciences.

Leif Azzopardi, Yashar Moshfeghi

Political bias in Large Language Models (LLMs) presents a growing concern for the responsible deployment of AI systems. Traditional audits often attempt to locate a model’s political position as a point estimate, masking the broader set of ideological boundaries that shape what a model is willing or unwilling to say. In this paper, we draw upon the concept of the Overton Window as a framework for mapping these boundaries: the range of political views that a given LLM will espouse, remain neutral on, or refuse to endorse. To uncover these windows, we applied an auditing-based methodology, called PRISM, that probes LLMs through task-driven prompts designed to elicit political stances indirectly. Using the Political Compass Test, we evaluated twenty-eight LLMs from eight providers to reveal their distinct Overton Windows. While many models default to economically left and socially liberal positions, we show that their willingness to express or reject certain positions varies considerably, where DeepSeek models tend to be very restrictive in what they will discuss and Gemini models tend to be most expansive. Our findings demonstrate that Overton Windows offer a richer, more nuanced view of political bias in LLMs and provide a new lens for auditing their normative boundaries.

Parker Seegmiller, Kartik Mehta, Soumya Saha, Chenyang Tao, Shereen Oraby, Arpit Gupta, Tagyoung Chung, Mohit Bansal, Nanyun Peng

Recent works improving LLM math reasoning with synthetic data have used unique setups, making comparison of data synthesis strategies impractical. This leaves many unanswered questions about the roles of different factors in the synthetic data pipeline, such as the impact of filtering low-quality problems. To address this gap, we introduce FLAMES, a Framework for LLM Assessment of Math rEasoning Data Synthesis, and perform a systematic study of 10 existing data synthesis strategies and multiple other factors impacting the performance of synthetic math reasoning data. Our FLAMES experiments provide several valuable insights about the optimal balance of difficulty and diversity of synthetic data. First, data agents designed to increase problem complexity lead to best improvements on most math metrics. Second, with a fixed data generation budget, keeping higher problem coverage is more important than keeping only problems with reliable solutions. Third, GSM8K- and MATH-based synthetic data can lead to improvements on competition-level benchmarks, showcasing easy-to-hard generalization. Leveraging insights from our FLAMES experiments, we design two novel data synthesis strategies for improving out-of-domain generalization and robustness. Further, we develop the FLAMES dataset, an effective blend of our novel and existing data synthesis strategies, outperforming public datasets on OlympiadBench (+15.7), CollegeMath (+4.5), GSMPlus (+6.5), and MATH (+3.1). Fine-tuning Qwen2.5-Math-7B on the FLAMES dataset achieves 81.4% on MATH, surpassing larger Llama3 405B, GPT-4o and Claude 3.5 Sonnet.