Can language models (LMs) self-refine their own responses? This question is increasingly relevant as more than 10% of real-world user interactions involve refinement requests (see Appendix G). Yet prior studies have largely tested LMs on verifiable tasks such as competition math or symbolic reasoning with simplified scaffolds, whereas users often pose open-ended queries and provide varying degrees of feedback about what went wrong. The recent advent of reasoning models that exhibit self-reflection patterns in their chain-of-thought further motivates this question. To address it, we introduce RefineBench, a benchmark of 1,002 challenging problems across 11 domains paired with a checklist-based evaluation framework. We evaluate two refinement modes: (1) guided refinement, where an LM is provided natural language feedback, and (2) self-refinement, where LMs attempt to improve without guidance. In the self-refinement setting, even frontier LMs such as Gemini 2.5 Pro and GPT-4.1 achieve modest baseline scores of 31.1 and 23.4, respectively, and most models fail to consistently improve across iterations (e.g., Gemini-2.5-Pro gains only +1.8%, while DeepSeek-R1 declines by –0.2%). By contrast, in guided refinement, both proprietary LMs and large open-weight LMs (>70B) can leverage targeted feedback to refine responses to near-perfect levels within five turns. These findings suggest that frontier LMs require breakthroughs to self-refine effectively when their initial responses are incorrect, and that RefineBench provides a valuable testbed for tracking progress.

Large Language Models (LLMs) excel at generating human-like text but are prone to hallucinations-fabricated outputs that deviate from reality. Addressing this is critical for ensuring trustworthy AI, especially in fields like medicine, law, and education. This paper introduces an entropy-based metric for detecting sentence-level hallucinations by assessing token prediction uncertainty. High entropy correlates with hallucinatory outputs, and we validate this metric’s effectiveness across different LLMs. Applications include hallucination flagging and improving user trust through uncertainty-driven feedback. This approach advances the reliability and transparency of LLMs, contributing to safer AI systems.

Recently, many vision-and-language models have demonstrated remarkable performance in various multimodal tasks using a zero-shot approach. However, they lack spatial reasoning abilities, particularly in understanding visuo-spatial information from another’s perspective, a skill known as visual perspective taking. In this paper, we investigate whether vision-and-language models can develop visual perspective-taking abilities in a zero-shot manner. Through a simple experiment, we demonstrate that these models still struggle to see the world from another’s viewpoint, suggesting that vision-and-language models are hindered by an egocentric bias.

The rapid evolution of generative AI necessitates robust safety evaluations. However, current safety datasets are predominantly English-centric, failing to capture specific risks in non-English, socio-cultural contexts such as Korean, and are often limited to the text modality. To address this gap, we introduce AssurAI, a new quality-controlled Korean multimodal dataset for evaluating the safety of generative AI. First, we define a taxonomy of 35 distinct AI risk factors, adapted from established frameworks by a multidisciplinary expert group to cover both universal harms and relevance to the Korean socio-cultural context. Second, leveraging this taxonomy, we construct and release AssurAI, a large-scale Korean multimodal dataset comprising 11,480 instances across text, image, video, and audio. Third, we apply the rigorous quality control process used to ensure data integrity, featuring a two-phase construction (i.e., expert-led seeding and crowdsourced scaling), triple independent annotation, and an iterative expert red-teaming loop. Our pilot study validates AssurAI's effectiveness in assessing the safety of recent LLMs. We release AssurAI to the public to facilitate the development of safer and more reliable generative AI systems for the Korean community.

While Large Language Models have gained attention, many service developers still rely on embedding-based models due to practical constraints. In such cases, the quality of fine-tuning data directly impacts performance, and English datasets are often used as seed data for training non-English models. In this study, we propose LANGALIGN, which enhances target language processing by aligning English embedding vectors with those of the target language at the interface between the language model and the task header. Experiments on Korean, Japanese, and Chinese demonstrate that LANGALIGN significantly improves performance across all three languages. Additionally, we show that LANGALIGN can be applied in reverse to convert target language data into a format that an English-based model can process.

Vision-and-Language Models (VLMs) have shown impressive capabilities on single-turn benchmarks, yet real-world applications often demand more intricate multi-turn dialogues. Existing multi-turn datasets (e.g, MMDU, ConvBench) only partially capture the breadth and depth of conversational scenarios encountered by users. In this work, we introduce MultiVerse, a novel multi-turn conversation benchmark featuring 647 dialogues - each averaging four turns - derived from a diverse set of 12 popular VLM evaluation benchmarks. With 484 tasks and 484 interaction goals, MultiVerse covers a wide range of topics, from factual knowledge and perception to advanced reasoning tasks such as mathematics and coding. To facilitate robust assessment, we propose a checklist-based evaluation method that leverages GPT-4o as the automated evaluator, measuring performance across 37 key aspects, including perceptual accuracy, linguistic clarity, and factual correctness. We evaluate 18 VLMs on MultiVerse, revealing that even the strongest models (e.g., GPT-4o) achieve only a 50

While large language models (LLMs) demonstrate strong performance in mathematical reasoning, most research evaluates their accuracy only based on final answers, often with problem types remaining consistent. This study explores whether LLMs can maintain consistent reasoning across different problem types. To investigate this, we converted the math word problem dataset GSM8K into a multiple-choice question format and tested LLMs on three different question types. Surprisingly, results showed that LLMs generated different outputs for the same math word problem, depending on the question type. Through error analysis, we developed a straightforward prompting technique to address this discrepancy, showing that it effectively enhances GPT-3.5’s performance and consistency.

Can language models (LMs) self-refine their own responses? This question is increasingly relevant as a wide range of real-world user interactions involve refinement requests. However, prior studies have largely tested LMs' refinement abilities on verifiable tasks such as competition math or symbolic reasoning with simplified scaffolds, whereas users often pose open-ended queries and provide varying degrees of feedback on what they desire. The recent advent of reasoning models that exhibit self-reflection patterns in their chains-of-thought further motivates this question. To analyze this, we introduce RefineBench, a benchmark of 1,000 challenging problems across 11 domains paired with a checklist-based evaluation framework. We evaluate two refinement modes: (1) guided refinement, where an LM is provided natural language feedback, and (2) self-refinement, where LMs attempt to improve without guidance. In the self-refinement setting, even frontier LMs such as Gemini 2.5 Pro and GPT-5 achieve modest baseline scores of 31.3

Transfer learning leverages the abundance of English data to address the scarcity of resources in modeling non-English languages, such as Korean. In this study, we explore the potential of Phrase Aligned Data (PAD) from standardized Statistical Machine Translation (SMT) to enhance the efficiency of transfer learning. Through extensive experiments, we demonstrate that PAD synergizes effectively with the syntactic characteristics of the Korean language, mitigating the weaknesses of SMT and significantly improving model performance. Moreover, we reveal that PAD complements traditional data construction methods and enhances their effectiveness when combined. This innovative approach not only boosts model performance but also suggests a cost-efficient solution for resource-scarce languages.

Speech Emotion Recognition (SER) involves analyzing human emotions conveyed through vocal expressions. Recently, self-supervised learning (SSL) has emerged as a powerful tool for SER by enabling the extraction of meaningful representations from large-scale unlabeled audio datasets, significantly reducing the need for manual labeling. However, most SSL-based SER methods employ Global Average Pooling (GAP) to derive a single audio representation, which unintentionally treats all segments, including both verbal and non-verbal, as equally important. This approach can result in the dilution of crucial emotional cues by irrelevant non-verbal segments, thereby limiting the effectiveness of the model. To address this challenge, we introduce Segmental Average Pooling (SAP), a novel pooling technique designed to selectively emphasize informative verbal segments while disregarding non-verbal parts. By combining both GAP and SAP for SSL features, our method is able to integrate the overall context from GAP with more focused segmental information from SAP, thereby capturing a richer representation of emotional cues. This dual-pooling strategy leads to significant improvements in SER performance, as demonstrated by our experiments, which achieved state-of-the-art results on the IEMOCAP dataset for English, and superior performance on the KEMDy19 dataset for Korean, in terms of both unweighted and weighted accuracies.

In real-world continual learning (CL) scenarios, tasks often exhibit intricate and unpredictable semantic shifts, posing challenges forfixedprompt management strategies which are tailored to only handle semantic shifts ofuniformdegree (i.e., uniformly mild or uniformly abrupt). To address this limitation, we propose anadaptiveprompting approach that effectively accommodates semantic shifts ofvaryingdegree where mild and abrupt shifts are mixed. AdaPromptCL employs the assign-and-refine semantic grouping mechanism that dynamically manages prompt groups in accordance with the semantic similarity between tasks, enhancing the quality of grouping through continuous refinement. Our experiment results demonstrate that AdaPromptCL outperforms existing prompting methods by up to 21.3%, especially in the benchmark datasets with diverse semantic shifts between tasks.

Mathematical question answering over long-form documents is challenging across domains like finance or Wikipedia due to the abundance of candidate arguments within evidence, which complicates recognizing proper arguments for mathematical reasoning and poses hard to learning. In this paper, we propose an approach for training a generator to improve argument recognition. Our method enhances the probabilities of proper arguments in a reasoning program generation so that the arguments comprising the ground truth have higher weights. The proposed approach consists of an argument aggregator to model the probabilities in each candidate generation and an argument set loss to compute the cross-entropy between that probability and the candidates’ existence in the ground truth in terms of the argument set. In our experiments, we show performance improvements of 3.62% and 3.98% in execution accuracy and program accuracy, respectively, over the existing FinQANet model based on a financial mathematical QA dataset. Also, we observed that the similarity of argument sets between the generated program and the ground truth improved by about 2.9%, indicating a mitigation of the misrecognition problem.

To increase social bonding with interlocutors, humans naturally acquire the ability to respond appropriately in a given situation by considering which conversational skill is most suitable for the response - a process we call skill-of-mind. For large language model (LLM)-based conversational agents, planning appropriate conversational skills, as humans do, is challenging due to the complexity of social dialogue, especially in interactive scenarios. To address this, we propose a skill-of-mind-annotated conversation dataset, named Multifaceted Skill-of-Mind, which includes multi-turn and multifaceted conversational skills across various interactive scenarios (e.g., long-term, counseling, task-oriented), grounded in diverse social contexts (e.g., demographics, persona, rules of thumb). This dataset consists of roughly 100K conversations. Using this dataset, we introduce a new family of skill-of-mind-infused LLMs, named Thanos, with model sizes of 1B, 3B, and 8B parameters. With extensive experiments, these models successfully demonstrate the skill-of-mind process and exhibit strong generalizability in inferring multifaceted skills across a variety of domains. Moreover, we show that Thanos significantly enhances the quality of responses generated by LLM-based conversational agents and promotes prosocial behavior in human evaluations.

This paper studies enhancing mathematical reasoning in Question Answering by improving the comprehension of arguments. In mathematical QA based on real-world documents, generating a reasoning program with correct arguments is challenging due to the many noisy inputs, unlike in general mathematical QA. In this study, we explore the potential for improving the performance of mathematical QA by enhancing the understanding and extraction of proper arguments to the Question. To this end, we propose an argument predictor to enhance the model's ability to comprehend and extract proper arguments for generating solution programs. Experiments conducted on the FinQA dataset demonstrate the improvement in the model's ability to identify relevant arguments, particularly in scenarios demanding multiple reasoning. The result shows improved program accuracy and execution accuracy, confirming the efficacy of our argument predictor.

This paper explores the \textit{image-sharing} capability of Large Language Models (LLMs), such as InstructGPT, ChatGPT, and GPT-4, in a zero-shot setting, without the help of visual foundation models. Inspired by the two-stage process of image-sharing in human dialogues, we propose a two-stage framework that allows LLMs to predict potential image-sharing turns and generate related image descriptions using our effective restriction-based prompt template. With extensive experiments, we unlock the \textit{image-sharing} capability of LLMs in zero-shot prompting, with GPT-4 achieving the best performance. Additionally, we uncover the emergent \textit{image-sharing} ability in zero-shot prompting, demonstrating the effectiveness of restriction-based prompts in both stages of our framework. Based on this framework, we augment the PhotoChat dataset with images generated by Stable Diffusion at predicted turns, namely PhotoChat++. To our knowledge, this is the first study to assess the \textit{image-sharing} ability of LLMs in a zero-shot setting without visual foundation models. The source code and the dataset will be released after publication.

This paper explores the image-sharing capability of Large Language Models (LLMs), such as GPT-4 and LLaMA 2, in a zero-shot setting. To facilitate a comprehensive evaluation of LLMs, we introduce the PhotoChat++ dataset, which includes enriched annotations (i.e., intent, triggering sentence, image description, and salient information). Furthermore, we present the gradient-free and extensible Decide, Describe, and Retrieve (DribeR) framework. With extensive experiments, we unlock the image-sharing capability of DribeR equipped with LLMs in zero-shot prompting, with ChatGPT achieving the best performance. Our findings also reveal the emergent image-sharing ability in LLMs under zero-shot conditions, validating the effectiveness of DribeR. We use this framework to demonstrate its practicality and effectiveness in two real-world scenarios: (1) human-bot interaction and (2) dataset augmentation. To the best of our knowledge, this is the first study to assess the image-sharing ability of various LLMs in a zero-shot setting. We make our source code and dataset publicly available at https://github.com/passing2961/DribeR.

Recently, large language and vision models (LLVMs) have received significant attention and development efforts due to their remarkable generalization performance across a wide range of tasks requiring perception and cognitive abilities. A key factor behind their success is their simple architecture, which consists of a vision encoder, a projector, and a large language model (LLM). Despite their achievements in advanced reasoning tasks, their performance on fundamental perception-related tasks (e.g., MMVP) remains surprisingly low. This discrepancy raises the question of how LLVMs truly perceive images and exploit the advantages of the vision encoder. To address this, we systematically investigate this question regarding several aspects: permutation invariance, robustness, math reasoning, alignment preserving and importance, by evaluating the most common LLVM's families (i.e., LLaVA) across 10 evaluation benchmarks. Our extensive experiments reveal several intriguing properties of current LLVMs: (1) they internally process the image in a global manner, even when the order of visual patch sequences is randomly permuted; (2) they are sometimes able to solve math problems without fully perceiving detailed numerical information; (3) the cross-modal alignment is overfitted to complex reasoning tasks, thereby, causing them to lose some of the original perceptual capabilities of their vision encoder; (4) the representation space in the lower layers (<25 enhancing visual understanding. Lastly, based on the above observations, we suggest potential future directions for building better LLVMs and constructing more challenging evaluation benchmarks.

Syntactic elements, such as word order and case markers, are fundamental in natural language processing. Recent studies show that syntactic information boosts language model performance and offers clues for people to understand their learning mechanisms. Unlike languages with a fixed word order such as English, Korean allows for varied word sequences, despite its canonical structure, due to case markers that indicate the functions of sentence components. This study explores whether Korean language models can accurately capture this flexibility. We note that incomplete word orders and omitted case markers frequently appear in ordinary Korean communication. To investigate this further, we introduce the Syntactically Incomplete Korean (SIKO) dataset. Through SIKO, we assessed Korean language models’ flexibility with incomplete syntax and confirmed the dataset’s training value. Results indicate these models reflect Korean’s inherent flexibility, accurately handling incomplete inputs. Moreover, fine-tuning with SIKO enhances the ability to handle common incomplete Korean syntactic forms. The dataset’s simple construction process, coupled with significant performance enhancements, solidifies its standing as an effective data augmentation technique. The SIKO will become accessible post-publication.

Large language models have exhibited remarkable proficiency in various domains and tasks including Numerical Reasoning task. However, datasets for Numerical Reasoning tasks targeted by LLMs often contain relatively few number of numerical values within the context. This raises doubts about whether LLMs can infer well and provide accurate answers even when presented with contexts including many numerical values. Indeed in a simple pilot study, we have found that LLM's ability to distinguish between essential and non-essential numerical values when solving numerical reasoning tasks is quite poor. To relieve this phenomenon, this paper proposes a framework to improve the robustness of LLMs in handling irrelevant numerical values for problem-solving by augmenting the original context of the problem while adding extra numerical values to the context. The experiment demonstrated that the model trained on the augmented dataset by our methodology increased robustness in Numerical Reasoning tasks with higher accuracies.

Humans share a wide variety of images related to their personal experiences within conversations via instant messaging tools. However, existing works focus on (1) image-sharing behavior in singular sessions, leading to limited long-term social interaction, and (2) a lack of personalized image-sharing behavior. In this work, we introduce Stark, a large-scale long-term multi-modal conversation dataset that covers a wide range of social personas in a multi-modality format, time intervals, and images. To construct Stark automatically, we propose a novel multi-modal contextualization framework, Mcu, that generates long-term multi-modal dialogue distilled from ChatGPT and our proposed Plan-and-Execute image aligner. Using our Stark, we train a multi-modal conversation model, Ultron 7B, which demonstrates impressive visual imagination ability. Furthermore, we demonstrate the effectiveness of our dataset in human evaluation. We make our source code and dataset publicly available.