Composed Image Retrieval (CIR) aims to search an image of interest using a combination of a reference image and modification text as the query. Despite recent advancements, this task remains challenging due to limited training data and laborious triplet annotation processes. To address this issue, this paper proposes to synthesize the training triplets to augment the training resource for the CIR problem. Specifically, we commence by training a modification text generator exploiting large-scale multimodal models and scale up the CIR learning throughout both the pretraining and fine-tuning stages. During pretraining, we leverage the trained generator to directly create Modification Text-oriented Synthetic Triplets(MTST) conditioned on pairs of images. For fine-tuning, we first synthesize reverse modification text to connect the target image back to the reference image. Subsequently, we devise a two-hop alignment strategy to incrementally close the semantic gap between the multimodal pair and the target image. We initially learn an implicit prototype utilizing both the original triplet and its reversed version in a cycle manner, followed by combining the implicit prototype feature with the modification text to facilitate accurate alignment with the target image. Extensive experiments validate the efficacy of the generated triplets and confirm that our proposed methodology attains competitive recall on both the CIRR and FashionIQ benchmarks.
Generative recommendation has emerged as a promising paradigm aiming at directly generating the identifiers of the target candidates. Most existing methods attempt to leverage prior knowledge embedded in Pre-trained Language Models (PLMs) to improve the recommendation performance. However, they often fail to accommodate the differences between the general linguistic knowledge of PLMs and the specific needs of recommendation systems. Moreover, they rarely consider the complementary knowledge between the multimodal information of items, which represents the multi-faceted preferences of users. To facilitate efficient recommendation knowledge transfer, we propose a novel approach called Multimodal Quantitative Language for Generative Recommendation (MQL4GRec). Our key idea is to transform items from different domains and modalities into a unified language, which can serve as a bridge for transferring recommendation knowledge. Specifically, we first introduce quantitative translators to convert the text and image content of items from various domains into a new and concise language, known as quantitative language, with all items sharing the same vocabulary. Then, we design a series of quantitative language generation tasks to enrich quantitative language with semantic information and prior knowledge. Finally, we achieve the transfer of recommendation knowledge from different domains and modalities to the recommendation task through pre-training and fine-tuning. We evaluate the effectiveness of MQL4GRec through extensive experiments and comparisons with existing methods, achieving improvements over the baseline by 11.18\%, 14.82\%, and 7.95\% on the NDCG metric across three different datasets, respectively.
Large language models (LLMs) have demonstrated great potential for automating the evaluation of natural language generation. Previous frameworks of LLM-as-a-judge fall short in two ways: they either use zero-shot setting without consulting any human input, which leads to low alignment, or fine-tune LLMs on labeled data, which requires a non-trivial number of samples. Moreover, previous methods often provide little reasoning behind automated evaluations. In this paper, we propose HypoEval, Hypothesis-guided Evaluation framework, which first uses a small corpus of human evaluations to generate more detailed rubrics for human judgments and then incorporates a checklist-like approach to combine LLM's assigned scores on each decomposed dimension to acquire overall scores. With only 30 human evaluations, HypoEval achieves state-of-the-art performance in alignment with both human rankings (Spearman correlation) and human scores (Pearson correlation), on average outperforming G-Eval by 11.86% and fine-tuned Llama-3.1-8B-Instruct with at least 3 times more human evaluations by 11.95%. Furthermore, we conduct systematic studies to assess the robustness of HypoEval, highlighting its effectiveness as a reliable and interpretable automated evaluation framework.
Large language models (LLMs), known for their comprehension capabilities and extensive knowledge, have been increasingly applied to recommendation systems (RS). Given the fundamental gap between the mechanism of LLMs and the requirement of RS, researchers have focused on fine-tuning LLMs with recommendation-specific data to enhance their performance. Language Modeling Loss (LML), originally designed for language generation tasks, is commonly adopted. However, we identify two critical limitations of LML: 1) it exhibits significant divergence from the recommendation objective; 2) it erroneously treats all fictitious item descriptions as negative samples, introducing misleading training signals. To address these limitations, we propose a novel Masked Softmax Loss (MSL) tailored for fine-tuning LLMs on recommendation. MSL improves LML by identifying and masking invalid tokens that could lead to fictitious item descriptions during loss computation. This strategy can effectively avoid the interference from erroneous negative signals and ensure well alignment with the recommendation objective supported by theoretical guarantees. During implementation, we identify a potential challenge related to gradient vanishing of MSL. To overcome this, we further introduce the temperature coefficient and propose an Adaptive Temperature Strategy (ATS) that adaptively adjusts the temperature without requiring extensive hyperparameter tuning. Extensive experiments conducted on four public datasets further validate the effectiveness of MSL, achieving an average improvement of 42.24% in NDCG@10. The code is available at https://github.com/WANGBohaO-jpg/MSL.
Generative models, such as GPT and BERT, have significantly improved performance in tasks like text generation and summarization. However, hallucinations "where models generate non-factual or misleading content" are especially problematic in smaller-scale architectures, limiting their real-world applicability.In this paper, we propose a unified Virtual Mixture-of-Experts (MoE) fusion strategy that enhances inference performance and mitigates hallucinations in a single Qwen 1.5 0.5B model without increasing the parameter count. Our method leverages multiple domain-specific expert prompts (with the number of experts being adjustable) to guide the model from different perspectives. We apply a statistical outlier truncation strategy based on the mean and standard deviation to filter out abnormally high probability predictions, and we inject noise into the embedding space to promote output diversity. To clearly assess the contribution of each module, we adopt a fixed voting mechanism rather than a dynamic gating network, thereby avoiding additional confounding factors. We provide detailed theoretical derivations from both statistical and ensemble learning perspectives to demonstrate how our method reduces output variance and suppresses hallucinations. Extensive ablation experiments on dialogue generation tasks show that our approach significantly improves inference accuracy and robustness in small models. Additionally, we discuss methods for evaluating the orthogonality of virtual experts and outline the potential for future work involving dynamic expert weight allocation using gating networks.
Generating long, coherent text remains a challenge for large language models (LLMs), as they lack hierarchical planning and structured organization in discourse generation. We introduce Structural Alignment, a novel method that aligns LLMs with human-like discourse structures to enhance long-form text generation. By integrating linguistically grounded discourse frameworks into reinforcement learning, our approach guides models to produce coherent and well-organized outputs. We employ a dense reward scheme within a Proximal Policy Optimization framework, assigning fine-grained, token-level rewards based on the discourse distinctiveness relative to human writing. Two complementary reward models are evaluated: the first improves readability by scoring surface-level textual features to provide explicit structuring, while the second reinforces deeper coherence and rhetorical sophistication by analyzing global discourse patterns through hierarchical discourse motifs, outperforming both standard and RLHF-enhanced models in tasks such as essay generation and long-document summarization. All training data and code will be publicly shared at https://github.com/minnesotanlp/struct_align.
Large Language Models have significantly advanced natural language processing, achieving remarkable performance in tasks such as language generation, translation, and reasoning. However, their substantial computational requirements restrict deployment to high-end systems, limiting accessibility on consumer-grade devices. This challenge is especially pronounced for under-resourced languages like those spoken in the Iberian Peninsula, where relatively limited linguistic resources and benchmarks hinder effective evaluation. This work presents a comprehensive evaluation of compact state-of-the-art LLMs across several essential NLP tasks tailored for Iberian languages. The results reveal that while some models consistently excel in certain tasks, significant performance gaps remain, particularly for languages such as Basque. These findings highlight the need for further research on balancing model compactness with robust multilingual performance
Formulating statements that support diverse or controversial stances on specific topics is vital for platforms that enable user expression, reshape political discourse, and drive social critique and information dissemination. With the rise of Large Language Models (LLMs), controllable text generation towards specific stances has become a promising research area with applications in shaping public opinion and commercial marketing. However, current datasets often focus solely on pure texts, lacking multimodal content and effective context, particularly in the context of stance detection. In this paper, we formally define and study the new problem of stance-driven controllable content generation for tweets with text and images, where given a multimodal post (text and image/video), a model generates a stance-controlled response. To this end, we create the Multimodal Stance Generation Dataset (StanceGen2024), the first resource explicitly designed for multimodal stance-controllable text generation in political discourse. It includes posts and user comments from the 2024 U.S. presidential election, featuring text, images, videos, and stance annotations to explore how multimodal political content shapes stance expression. Furthermore, we propose a Stance-Driven Multimodal Generation (SDMG) framework that integrates weighted fusion of multimodal features and stance guidance to improve semantic consistency and stance control. We release the dataset and code (https://anonymous.4open.science/r/StanceGen-BE9D) for public use and further research.
LLMs' remarkable ability to tackle a wide range of language tasks opened new opportunities for collaborative human-AI problem solving. LLMs can amplify human capabilities by applying their intuitions and reasoning strategies at scale. We explore whether human guides can be simulated, by generalizing from human demonstrations of guiding an AI system to solve complex language problems. We introduce CoLa, a novel self-guided learning paradigm for training automated $\textit{guides}$ and evaluate it on two QA datasets, a puzzle-solving task, and a constrained text generation task. Our empirical results show that CoLa consistently outperforms competitive approaches across all domains. Moreover, a small-sized trained guide outperforms a strong model like GPT-4 when acting as a guide. We compare the strategies employed by humans and automated guides by conducting a human study on a QA dataset. We show that automated guides outperform humans by adapting their strategies to reasoners' capabilities and conduct qualitative analyses highlighting distinct differences in guiding strategies.
Large vision-language models (LVMs) hold a great promise for automating medical report generation, potentially reducing the burden of manual reporting. State-of-the-art (SOTA) research fine-tunes general LVMs with medical data to align radiology images to corresponding medical reports. However, there are two key factors that limit these LVM's performance. Firstly, LVMs lack complex reasoning capability that leads to logical inconsistencies and potential diagnostic errors in generated reports. Secondly, LVMs lack reflection mechanism that leads to an inability to discover errors in the thinking process. To address these gaps, we propose LVMed-R2, a new fine-tuning strategy that introduces complex reasoning and reflection mechanisms for LVMs to enhance medical report generation. To the best of our knowledge, this is the first work to introduce complex reasoning to the medical report generation (MRG) task. Our proposed complex reasoning contains medical knowledge injection and perception-enhancing modules which improve the accuracy of LVMs diagnosis, coupled with a perception tree to provide guidance to limit the perception range. Further, the reflection mechanism forces self-verification for outputs to correct for potential errors. We experimented by fine-tuning LVMs with our proposed LVMed-R2 strategy, using IU-Xray and MIMIC-CXR datasets. Our results, measured on natural language generation (NLG) metrics and clinical efficacy (CE) metrics, demonstrate that LVMs fine-tuned with the proposed reflection mechanism possess the ability to correct outputs and complex reasoning effectively and improve LVMs performance for MRG.
Generative information extraction using large language models, particularly through few-shot learning, has become a popular method. Recent studies indicate that providing a detailed, human-readable guideline-similar to the annotation guidelines traditionally used for training human annotators can significantly improve performance. However, constructing these guidelines is both labor- and knowledge-intensive. Additionally, the definitions are often tailored to meet specific needs, making them highly task-specific and often non-reusable. Handling these subtle differences requires considerable effort and attention to detail. In this study, we propose a self-improving method that harvests the knowledge summarization and text generation capacity of LLMs to synthesize annotation guidelines while requiring virtually no human input. Our zero-shot experiments on the clinical named entity recognition benchmarks, 2012 i2b2 EVENT, 2012 i2b2 TIMEX, 2014 i2b2, and 2018 n2c2 showed 25.86%, 4.36%, 0.20%, and 7.75% improvements in strict F1 scores from the no-guideline baseline. The LLM-synthesized guidelines showed equivalent or better performance compared to human-written guidelines by 1.15% to 4.14% in most tasks. In conclusion, this study proposes a novel LLM self-improving method that requires minimal knowledge and human input and is applicable to multiple biomedical domains.
Large Language Models (LLMs) have demonstrated impressive performance across diverse domains, yet they still encounter challenges such as insufficient domain-specific knowledge, biases, and hallucinations. This underscores the need for robust evaluation methodologies to accurately assess LLM-based applications. Traditional evaluation methods, which rely on word overlap or text embeddings, are inadequate for capturing the nuanced semantic information necessary to evaluate dynamic, open-ended text generation. Recent research has explored leveraging LLMs to mimic human reasoning and decision-making processes for evaluation purposes known as LLM-as-a-judge framework. However, these existing frameworks have two significant limitations. First, they lack the flexibility to adapt to different text styles, including various answer and ground truth styles, thereby reducing their generalization performance. Second, the evaluation scores produced by these frameworks are often skewed and hard to interpret, showing a low correlation with human judgment. To address these challenges, we propose a novel dynamic multi-agent system that automatically designs personalized LLM judges for various natural language generation applications. This system iteratively refines evaluation prompts and balances the trade-off between the adaptive requirements of downstream tasks and the alignment with human perception. Our experimental results show that the proposed multi-agent LLM Judge framework not only enhances evaluation accuracy compared to existing methods but also produces evaluation scores that better align with human perception.
Large language models (LLMs) demonstrate increasing capabilities in creative text generation, yet systematic evaluations of their humor production remain underexplored. This study presents a comprehensive analysis of 13 state-of-the-art LLMs across five architectural families, evaluating their performance in generating technically relevant humor for software developers. Through a full factorial design testing 715 unique configurations of temperature settings and prompt variations, we assess model outputs using five weighted criteria: humor quality, domain relevance, concept originality, tone precision, and delivery efficiency. Our methodology employs rigorous statistical analysis including ANOVA, correlation studies, and quadratic regression to identify optimal configurations and architectural influences. Results reveal significant performance variations across models, with certain architectures achieving 21.8% superiority over baseline systems. Temperature sensitivity analysis demonstrates that 73% of models achieve peak performance at lower stochasticity settings (<= 0.5), though optimal ranges vary substantially by architecture. We identify distinct model clusters: compact high-performers maintaining efficiency-quality balance versus verbose specialists requiring longer outputs for marginal gains. Statistical validation confirms model architecture explains 38.7% of performance variance, with significant correlations between humor quality and concept originality. The study establishes practical guidelines for model selection and configuration, demonstrating how temperature adjustments and architectural considerations impact humor generation effectiveness. These findings advance understanding of LLM capabilities in creative technical writing and provide empirically validated configuration strategies for developers implementing humor-generation systems.
Despite a large deaf and dumb population of 1.7 million, Bangla Sign Language (BdSL) remains a understudied domain. Specifically, there are no works on Bangla text-to-gloss translation task. To address this gap, we begin by addressing the dataset problem. We take inspiration from grammatical rule based gloss generation used in Germany and American sign langauage (ASL) and adapt it for BdSL. We also leverage LLM to generate synthetic data and use back-translation, text generation for data augmentation. With dataset prepared, we started experimentation. We fine-tuned pretrained mBART-50 and mBERT-multiclass-uncased model on our dataset. We also trained GRU, RNN and a novel seq-to-seq model with multi-head attention. We observe significant high performance (ScareBLEU=79.53) with fine-tuning pretrained mBART-50 multilingual model from Facebook. We then explored why we observe such high performance with mBART. We soon notice an interesting property of mBART -- it was trained on shuffled and masked text data. And as we know, gloss form has shuffling property. So we hypothesize that mBART is inherently good at text-to-gloss tasks. To find support against this hypothesis, we trained mBART-50 on PHOENIX-14T benchmark and evaluated it with existing literature. Our mBART-50 finetune demonstrated State-of-the-Art performance on PHOENIX-14T benchmark, far outperforming existing models in all 6 metrics (ScareBLEU = 63.89, BLEU-1 = 55.14, BLEU-2 = 38.07, BLEU-3 = 27.13, BLEU-4 = 20.68, COMET = 0.624). Based on the results, this study proposes a new paradigm for text-to-gloss task using mBART models. Additionally, our results show that BdSL text-to-gloss task can greatly benefit from rule-based synthetic dataset.
Evaluating the quality of generated text automatically remains a significant challenge. Conventional reference-based metrics have been shown to exhibit relatively weak correlation with human evaluations. Recent research advocates the use of large language models (LLMs) as source-based metrics for natural language generation (NLG) assessment. While promising, LLM-based metrics, particularly those using smaller models, still fall short in aligning with human judgments. In this work, we introduce ContrastScore, a contrastive evaluation metric designed to enable higher-quality, less biased, and more efficient assessment of generated text. We evaluate ContrastScore on two NLG tasks: machine translation and summarization. Experimental results show that ContrastScore consistently achieves stronger correlation with human judgments than both single-model and ensemble-based baselines. Notably, ContrastScore based on Qwen 3B and 0.5B even outperforms Qwen 7B, despite having only half as many parameters, demonstrating its efficiency. Furthermore, it effectively mitigates common evaluation biases such as length and likelihood preferences, resulting in more robust automatic evaluation.
Academic writing requires both coherent text generation and precise citation of relevant literature. Although recent Retrieval-Augmented Generation (RAG) systems have significantly improved factual accuracy in general-purpose text generation, their ability to support professional academic writing remains limited. In this work, we introduce ScholarCopilot, a unified framework designed to enhance existing large language models for generating professional academic articles with accurate and contextually relevant citations. ScholarCopilot dynamically determines when to retrieve scholarly references by generating a retrieval token [RET], which is then used to query a citation database. The retrieved references are fed into the model to augment the generation process. We jointly optimize both the generation and citation tasks within a single framework to improve efficiency. Our model is built upon Qwen-2.5-7B and trained on 500K papers from arXiv. It achieves a top-1 retrieval accuracy of 40.1% on our evaluation dataset, outperforming baselines such as E5-Mistral-7B-Instruct (15.0%) and BM25 (9.8%). On a dataset of 1,000 academic writing samples, ScholarCopilot scores 16.2/25 in generation quality -- measured across relevance, coherence, academic rigor, completeness, and innovation -- significantly surpassing all existing models, including much larger ones like the Retrieval-Augmented Qwen2.5-72B-Instruct. Human studies further demonstrate that ScholarCopilot, despite being a 7B model, significantly outperforms ChatGPT, achieving 100% preference in citation quality and over 70% in overall usefulness.
Recent advances in large language models (LLMs) have introduced the novel paradigm of using LLMs as judges, where an LLM evaluates and scores the outputs of another LLM, which often correlates highly with human preferences. However, the use of LLM-as-a-judge has been primarily studied in English. In this paper, we evaluate this framework in Russian by introducing the Russian Error tyPes Annotation dataset (REPA), a dataset of 1k user queries and 2k LLM-generated responses. Human annotators labeled each response pair expressing their preferences across ten specific error types, as well as selecting an overall preference. We rank six generative LLMs across the error types using three rating systems based on human preferences. We also evaluate responses using eight LLM judges in zero-shot and few-shot settings. We describe the results of analyzing the judges and position and length biases. Our findings reveal a notable gap between LLM judge performance in Russian and English. However, rankings based on human and LLM preferences show partial alignment, suggesting that while current LLM judges struggle with fine-grained evaluation in Russian, there is potential for improvement.
Text generated by language models (LMs) can degrade into repetitive cycles, where identical word sequences are persistently repeated one after another. Prior research has typically treated repetition as a unitary phenomenon. However, repetitive sequences emerge under diverse tasks and contexts, raising the possibility that it may be driven by multiple underlying factors. Here, we experimentally explore the hypothesis that repetition in LMs can result from distinct mechanisms, reflecting different text generation strategies used by the model. We examine the internal working of LMs under two conditions that prompt repetition: one in which repeated sequences emerge naturally after human-written text, and another where repetition is explicitly induced through an in-context learning (ICL) setup. Our analysis reveals key differences between the two conditions: the model exhibits varying levels of confidence, relies on different attention heads, and shows distinct pattens of change in response to controlled perturbations. These findings suggest that distinct internal mechanisms can interact to drive repetition, with implications for its interpretation and mitigation strategies. More broadly, our results highlight that the same surface behavior in LMs may be sustained by different underlying processes, acting independently or in combination.
Current research on LoRA primarily focuses on minimizing the number of fine-tuned parameters or optimizing its architecture. However, the necessity of all fine-tuned LoRA layers during inference remains underexplored. In this paper, we investigate the contribution of each LoRA layer to the model's ability to predict the ground truth and hypothesize that lower-layer LoRA modules play a more critical role in model reasoning and understanding. To address this, we propose a simple yet effective method to enhance the performance of large language models (LLMs) fine-tuned with LoRA. Specifically, we identify a ``boundary layer'' that distinguishes essential LoRA layers by analyzing a small set of validation samples. During inference, we drop all LoRA layers beyond this boundary. We evaluate our approach on three strong baselines across four widely-used text generation datasets. Our results demonstrate consistent and significant improvements, underscoring the effectiveness of selectively retaining critical LoRA layers during inference.
This paper explores the task of Complex Visual Text Generation (CVTG), which centers on generating intricate textual content distributed across diverse regions within visual images. In CVTG, image generation models often rendering distorted and blurred visual text or missing some visual text. To tackle these challenges, we propose TextCrafter, a novel multi-visual text rendering method. TextCrafter employs a progressive strategy to decompose complex visual text into distinct components while ensuring robust alignment between textual content and its visual carrier. Additionally, it incorporates a token focus enhancement mechanism to amplify the prominence of visual text during the generation process. TextCrafter effectively addresses key challenges in CVTG tasks, such as text confusion, omissions, and blurriness. Moreover, we present a new benchmark dataset, CVTG-2K, tailored to rigorously evaluate the performance of generative models on CVTG tasks. Extensive experiments demonstrate that our method surpasses state-of-the-art approaches.