While large language models (LLMs) demonstrate strong capabilities across diverse user queries, they still suffer from hallucinations, often arising from knowledge misalignment between pre-training and fine-tuning. To address this misalignment, we reliably estimate a fine-grained, instance-level knowledge score via multi-sampled inference. Using the knowledge score, we scale the learning signal according to the model’s existing knowledge, while encouraging explicit “I don’t know” responses for out-of-scope queries. Experimental results show that this approach allows the model to explicitly express uncertainty when it lacks knowledge, while maintaining accuracy on questions it can answer. Furthermore, we propose evaluation metrics for uncertainty, showing that accurate discrimination between known and unknown instances consistently improves performance.

Recent progress in Omni-models has yielded strong performance on benchmarks focused on static observations. However, their ability to maintain spatiotemporal continuity when predicting coherent future states from previous context remains underexplored. We address this by formulating Next Scene Prediction (NSP) as a diagnostic task for assessing multimodal near-future state prediction, requiring models to anticipate the immediate future from a given context. We introduce OmniNSP, a benchmark evaluating causal and cross-modal coherence across physical, logical, and social dynamics. Unlike prior tasks that often rely on textual or discriminative proxies, OmniNSP asks models to express predicted future states through text-mediated visual and speech synthesis. Our analysis shows that current Omni-models can produce plausible textual forecasts while still failing to preserve future-state consistency in generated outputs. A frozen-generator control substantially narrows cross-model visual-consistency gaps, attributing much of the original system-level difference to model-specific realization stacks rather than textual foresight alone. A modality-input ablation further shows that audio provides complementary predictive cues beyond vision, with the largest gains for logical transitions and contexts with informative speech. These results position OmniNSP as a diagnostic of both near-future prediction and the preservation of predicted state through multimodal realization.

Large Language Models (LLMs) are widely used to generate plausible text on online platforms, without revealing the generation process.As users increasingly encounter such black-box outputs, detecting hallucinations has become a critical challenge.To address this challenge, we focus on developing a hallucination detection framework for black-box generators.Motivated by the observation that hallucinations, once introduced, tend to persist, we sample future contexts.The sampled future contexts provide valuable clues for hallucination detection and can be effectively integrated with various sampling-based methods.We extensively demonstrate performance improvements across multiple methods using our proposed sampling approach.

Most testbeds for omni-modal models assess multimodal understanding via textual outputs, leaving it unclear whether these models can properly speak their answers. To study this, we introduce OmniACBench, a benchmark for evaluating context-grounded acoustic control in omni-modal models. Given a spoken instruction, a text script, and an image, a model must read the script aloud with an appropriate tone and manner. OmniACBench comprises 3,559 verified instances covering six acoustic features: speech rate, phonation, pronunciation, emotion, global accent, and timbre. Extensive experiments on eight models reveal their limitations in the proposed setting, despite their strong performance on prior textual-output evaluations. Our analyses show that the main bottleneck lies not in processing individual modalities, but in integrating multimodal context for faithful speech generation. Moreover, we identify three common failure modes—weak direct control, failed implicit inference, and failed multimodal grounding—providing insights for developing models that can verbalize responses effectively.

Reinforcement Learning with Verifiable Rewards (RLVR) can improve reasoning in large language models, but specialized non-English domains expose a data mismatch. Raw English domain text is often available, while target-language RLVR tasks with verifiable rewards are scarce. We introduce code-switched task synthesis for cross-lingual RLVR, a data-centric approach that derives target-language and code-switched RLVR task variants from otherwise unverifiable English domain text. Using fill-in-the-middle multiple-choice QA (FIM-MCQA) as a seed task format, we synthesize automatically verifiable tasks from domain passages and reshape them into cross-lingual variants. These variants preserve the same evidence-answer relation but change how English and target-language evidence are presented during RLVR. On Korean medical benchmarks and cybersecurity benchmarks across four target languages, code-switched training outperforms matched monolingual and bilingual-mixture baselines on average target-language performance while maintaining comparable English performance. Additional analyses provide evidence of better target-language input understanding and stronger cross-lingual alignment in both model outputs and internal representations. These results suggest that code-switched training helps target-language inputs access domain reasoning learned more strongly in English.

Although there has been a growing interest among industries in integrating generative LLMs into their services, limited experience and scarcity of resources act as a barrier in launching and servicing large-scale LLM-based services. In this paper, we share our experiences in developing and operating generative AI models within a national-scale search engine, with a specific focus on the sensitiveness of user queries. We propose a taxonomy for sensitive search queries, outline our approaches, and present a comprehensive analysis report on sensitive queries from actual users. We believe that our experiences in launching generative AI search systems can contribute to reducing the barrier in building generative LLM-based services.

The advancements in large language models (LLMs) have brought significant progress in NLP tasks. However, if a task cannot be fully described in prompts, the models could fail to carry out the task. In this paper, we propose a simple yet effective method to contextualize a task toward a LLM. The method utilizes (1) open-ended zero-shot inference from the entire dataset, (2) aggregate the inference results, and (3) finally incorporate the aggregated meta-information for the actual task. We show the effectiveness in text clustering tasks, empowering LLMs to perform text-to-text-based clustering and leading to improvements on several datasets. Furthermore, we explore the generated class labels for clustering, showing how the LLM understands the task through data.

Most prior safety research of large language models (LLMs) has focused on enhancing the alignment of LLMs to better suit the safety requirements of their use cases. However, internalizing such safeguard features into larger models brought challenges of higher training cost and unintended degradation of helpfulness. In this paper, we leverage a smaller LLM for both harmful query detection and safeguard response generation. We introduce our safety requirements and the taxonomy of harmfulness categories, and then propose a multi-task learning mechanism fusing the two tasks into a single model. We demonstrate the effectiveness of our approach, providing on par or surpassing harmful query detection and safeguard response performance compared to the publicly available LLMs.

Retrofitting techniques, which inject external resources into word representations, have compensated for the weakness of distributed representations in semantic and relational knowledge between words. However, the previous methods require additional external resources and strongly depend on the lexicon. To address the issues, we propose a simple extension of extrofitting, self-supervised extrofitting: extrofitting by its own word vector distribution. Our methods improve the vanilla embeddings on all of word similarity tasks without any external resources. Moreover, the method is also effective in various languages, which implies that our method will be useful in lexicon-scarce languages. As downstream tasks, we show its benefits in dialogue state tracking and text classification tasks, reporting better and generalized results compared to other word vector specialization methods.

We investigate the representation of pretrained language models and humans, using the idea of word definition modeling–how well a word is represented by its definition, and vice versa. Our analysis shows that a word representation in pretrained language models does not successfully map its human-written definition and its usage in example sentences. We then present a simple method DefBERT that integrates pretrained models with word semantics in dictionaries. We show its benefits on newly-proposed tasks of definition ranking and definition sense disambiguation. Furthermore, we present the results on standard word similarity tasks and short text classification tasks where models are required to encode semantics with only a few words. The results demonstrate the effectiveness of integrating word definitions and pretrained language models.

Despite recent explosion of interests in in-context learning, the underlying mechanism and the precise impact of the quality of demonstrations remain elusive.Intuitively, ground-truth labels should have as much impact in in-context learning (ICL) as supervised learning, but recent work reported that the input-label correspondence is significantly less important than previously thought.Intrigued by this counter-intuitive observation, we re-examine the importance of ground-truth labels in in-context learning.With the introduction of two novel metrics, namely Label-Correctness Sensitivity and Ground-truth Label Effect Ratio (GLER), we were able to conduct quantifiable analysis on the impact of ground-truth label demonstrations.Through extensive analyses, we find that the correct input-label mappings can have varying impacts on the downstream in-context learning performances, depending on the experimental configuration.Through additional studies, we identify key components, such as the verbosity of prompt templates and the language model size, as the controlling factor to achieve more noise-resilient ICL.

Natural language models often fall short when understanding and generating mathematical notation. What is not clear is whether these shortcomings are due to fundamental limitations of the models, or the absence of appropriate tasks. In this paper, we explore the extent to which natural language models can learn semantics between mathematical notation and their surrounding text. We propose two notation prediction tasks, and train a model that selectively masks notation tokens and encodes left and/or right sentences as context. Compared to baseline models trained by masked language modeling, our method achieved significantly better performance at the two tasks, showing that this approach is a good first step towards modeling mathematical texts. However, the current models rarely predict unseen symbols correctly, and token-level predictions are more accurate than symbol-level predictions, indicating more work is needed to represent structural patterns. Based on the results, we suggest future works toward modeling mathematical texts.

We present a new form of ensemble method–Devil’s Advocate, which uses a deliberately dissenting model to force other submodels within the ensemble to better collaborate. Our method consists of two different training settings: one follows the conventional training process (Norm), and the other is trained by artificially generated labels (DevAdv). After training the models, Norm models are fine-tuned through an additional loss function, which uses the DevAdv model as a constraint. In making a final decision, the proposed ensemble model sums the scores of Norm models and then subtracts the score of the DevAdv model. The DevAdv model improves the overall performance of the other models within the ensemble. In addition to our ensemble framework being based on psychological background, it also shows comparable or improved performance on 5 text classification tasks when compared to conventional ensemble methods.

We propose a novel and simple method for semi-supervised text classification. The method stems from the hypothesis that a classifier with pretrained word embeddings always outperforms the same classifier with randomly initialized word embeddings, as empirically observed in NLP tasks. Our method first builds two sets of classifiers as a form of model ensemble, and then initializes their word embeddings differently: one using random, the other using pretrained word embeddings. We focus on different predictions between the two classifiers on unlabeled data while following the self-training framework. We also use early-stopping in meta-epoch to improve the performance of our method. Our method, Delta-training, outperforms the self-training and the co-training framework in 4 different text classification datasets, showing robustness against error accumulation.

We propose post-processing method for enriching not only word representation but also its vector space using semantic lexicons, which we call extrofitting. The method consists of 3 steps as follows: (i) Expanding 1 or more dimension(s) on all the word vectors, filling with their representative value. (ii) Transferring semantic knowledge by averaging each representative values of synonyms and filling them in the expanded dimension(s). These two steps make representations of the synonyms close together. (iii) Projecting the vector space using Linear Discriminant Analysis, which eliminates the expanded dimension(s) with semantic knowledge. When experimenting with GloVe, we find that our method outperforms Faruqui’s retrofitting on some of word similarity task. We also report further analysis on our method in respect to word vector dimensions, vocabulary size as well as other well-known pretrained word vectors (e.g., Word2Vec, Fasttext).