Kyomin Jung

Hyojeong Yu, Hyukhun Koh, Minsung Kim et al.

Formality transfer is commonly framed as a symmetric bidirectional task between informal and formal registers. We argue that this framing conceals a supervision design flaw in existing benchmarks such as GYAFC: binary human rewrites encode relative stylistic shifts rather than absolute human notions of formality. Consequently, models learn to generate pseudo-formal outputs that satisfy benchmark labels while failing to produce genuinely formal language. We quantify this misalignment by re-evaluating benchmark formal labels under a human-aligned definition of formality, revealing substantial discrepancies that propagate to consistent informal-to-formal failures across model families. To address this issue, we reconceptualize formality transfer as a graded dimension rather than a binary attribute. We introduce a three-level spectrum: informal, casual, and formal, where casual serves as an explicit intermediate state that clarifies supervision signals. Based on this framework, we introduce 3LF, a dataset providing parallel supervision across all three levels. Training on 3LF substantially reduces informal-to-formal failures and improves alignment with human perception. For example, GPT-4.1-nano improves from 0.06 to 0.88 F1 in the informal-to- formal direction despite 3LF being significantly smaller than GYAFC. We further demonstrate that these gains cannot be reproduced through in-context learning alone and provide qualitative analyses of ambiguity-driven errors and meaning distortions. Overall, our findings demonstrate how supervision design shapes stylistic alignment and highlight the importance of alignment-aware benchmark construction in controllable text generation.

Hwanhee Lee, Kang Min Yoo, Joonsuk Park et al.

Despite the recent advances in abstractive summarization systems, it is still difficult to determine whether a generated summary is factual consistent with the source text. To this end, the latest approach is to train a factual consistency classifier on factually consistent and inconsistent summaries. Luckily, the former is readily available as reference summaries in existing summarization datasets. However, generating the latter remains a challenge, as they need to be factually inconsistent, yet closely relevant to the source text to be effective. In this paper, we propose to generate factually inconsistent summaries using source texts and reference summaries with key information masked. Experiments on seven benchmark datasets demonstrate that factual consistency classifiers trained on summaries generated using our method generally outperform existing models and show a competitive correlation with human judgments. We also analyze the characteristics of the summaries generated using our method. We will release the pre-trained model and the code at https://github.com/hwanheelee1993/MFMA.

Minbeom Kim, Jahyun Koo, Hwanhee Lee et al.

As large language models become increasingly integrated into daily life, detecting implicit toxicity across diverse contexts is crucial. To this end, we introduce LifeTox, a dataset designed for identifying implicit toxicity within a broad range of advice-seeking scenarios. Unlike existing safety datasets, LifeTox comprises diverse contexts derived from personal experiences through open-ended questions. Experiments demonstrate that RoBERTa fine-tuned on LifeTox matches or surpasses the zero-shot performance of large language models in toxicity classification tasks. These results underscore the efficacy of LifeTox in addressing the complex challenges inherent in implicit toxicity. We open-sourced the dataset\footnote{\url{https://huggingface.co/datasets/mbkim/LifeTox}} and the LifeTox moderator family; 350M, 7B, and 13B.

Minbeom Kim, Hwanhee Lee, Kang Min Yoo et al.

Steering language generation towards objectives or away from undesired content has been a long-standing goal in utilizing language models (LM). Recent work has demonstrated reinforcement learning and weighted decoding as effective approaches to achieve a higher level of language control and quality with pros and cons. In this work, we propose a novel critic decoding method for controlled language generation (CriticControl) that combines the strengths of reinforcement learning and weighted decoding. Specifically, we adopt the actor-critic framework to train an LM-steering critic from non-differentiable reward models. And similar to weighted decoding, our method freezes the language model and manipulates the output token distribution using called critic, improving training efficiency and stability. Evaluation of our method on three controlled generation tasks, namely topic control, sentiment control, and detoxification, shows that our approach generates more coherent and well-controlled texts than previous methods. In addition, CriticControl demonstrates superior generalization ability in zero-shot settings. Human evaluation studies also corroborate our findings.

Hwanhee Lee, Cheoneum Park, Seunghyun Yoon et al.

Despite the recent advancements in abstractive summarization systems leveraged from large-scale datasets and pre-trained language models, the factual correctness of the summary is still insufficient. One line of trials to mitigate this problem is to include a post-editing process that can detect and correct factual errors in the summary. In building such a post-editing system, it is strongly required that 1) the process has a high success rate and interpretability and 2) has a fast running time. Previous approaches focus on regeneration of the summary using the autoregressive models, which lack interpretability and require high computing resources. In this paper, we propose an efficient factual error correction system RFEC based on entities retrieval post-editing process. RFEC first retrieves the evidence sentences from the original document by comparing the sentences with the target summary. This approach greatly reduces the length of text for a system to analyze. Next, RFEC detects the entity-level errors in the summaries by considering the evidence sentences and substitutes the wrong entities with the accurate entities from the evidence sentences. Experimental results show that our proposed error correction system shows more competitive performance than baseline methods in correcting the factual errors with a much faster speed.

Nakyeong Yang, Minsung Kim, Seunghyun Yoon et al.

With the explosion of multimedia content, video moment retrieval (VMR), which aims to detect a video moment that matches a given text query from a video, has been studied intensively as a critical problem. However, the existing VMR framework evaluates video moment retrieval performance, assuming that a video is given, which may not reveal whether the models exhibit overconfidence in the falsely given video. In this paper, we propose the MVMR (Massive Videos Moment Retrieval for Faithfulness Evaluation) task that aims to retrieve video moments within a massive video set, including multiple distractors, to evaluate the faithfulness of VMR models. For this task, we suggest an automated massive video pool construction framework to categorize negative (distractors) and positive (false-negative) video sets using textual and visual semantic distance verification methods. We extend existing VMR datasets using these methods and newly construct three practical MVMR datasets. To solve the task, we further propose a strong informative sample-weighted learning method, CroCs, which employs two contrastive learning mechanisms: (1) weakly-supervised potential negative learning and (2) cross-directional hard-negative learning. Experimental results on the MVMR datasets reveal that existing VMR models are easily distracted by the misinformation (distractors), whereas our model shows significantly robust performance, demonstrating that CroCs is essential to distinguishing positive moments against distractors. Our code and datasets are publicly available: https://github.com/yny0506/Massive-Videos-Moment-Retrieval.

Nakyeong Yang, Yunah Jang, Hwanhee Lee et al.

Multi-task language models show outstanding performance for various natural language understanding tasks with only a single model. However, these language models utilize an unnecessarily large number of model parameters, even when used only for a specific task. This paper proposes a novel training-free compression method for multi-task language models using a pruning method. Specifically, we use an attribution method to determine which neurons are essential for performing a specific task. We task-specifically prune unimportant neurons and leave only task-specific parameters. Furthermore, we extend our method to be applicable in low-resource and unsupervised settings. Since our compression method is training-free, it uses few computing resources and does not destroy the pre-trained knowledge of language models. Experimental results on the six widely-used datasets show that our proposed pruning method significantly outperforms baseline pruning methods. In addition, we demonstrate that our method preserves performance even in an unseen domain setting.

Yongil Kim, Yerin Hwang, Hyeongu Yun et al.

Vulnerability to lexical perturbation is a critical weakness of automatic evaluation metrics for image captioning. This paper proposes Perturbation Robust Multi-Lingual CLIPScore(PR-MCS), which exhibits robustness to such perturbations, as a novel reference-free image captioning metric applicable to multiple languages. To achieve perturbation robustness, we fine-tune the text encoder of CLIP with our language-agnostic method to distinguish the perturbed text from the original text. To verify the robustness of PR-MCS, we introduce a new fine-grained evaluation dataset consisting of detailed captions, critical objects, and the relationships between the objects for 3, 000 images in five languages. In our experiments, PR-MCS significantly outperforms baseline metrics in capturing lexical noise of all various perturbation types in all five languages, proving that PR-MCS is highly robust to lexical perturbations.

Hyukhun Koh, Haesung Pyun, Nakyeong Yang et al.

In Task Oriented Dialogue (TOD) system, detecting and inducing new intents are two main challenges to apply the system in the real world. In this paper, we suggest the semantic multi-view model to resolve these two challenges: (1) SBERT for General Embedding (GE), (2) Multi Domain Batch (MDB) for dialogue domain knowledge, and (3) Proxy Gradient Transfer (PGT) for cluster-specialized semantic. MDB feeds diverse dialogue datasets to the model at once to tackle the multi-domain problem by learning the multiple domain knowledge. We introduce a novel method PGT, which employs the Siamese network to fine-tune the model with a clustering method directly.Our model can learn how to cluster dialogue utterances by using PGT. Experimental results demonstrate that our multi-view model with MDB and PGT significantly improves the Open Intent Induction performance compared to baseline systems.

Yerin Hwang, Yongil Kim, Hyunkyung Bae et al.

To address the data scarcity issue in Conversational question answering (ConvQA), a dialog inpainting method, which utilizes documents to generate ConvQA datasets, has been proposed. However, the original dialog inpainting model is trained solely on the dialog reconstruction task, resulting in the generation of questions with low contextual relevance due to insufficient learning of question-answer alignment. To overcome this limitation, we propose a novel framework called Dialogizer, which has the capability to automatically generate ConvQA datasets with high contextual relevance from textual sources. The framework incorporates two training tasks: question-answer matching (QAM) and topic-aware dialog generation (TDG). Moreover, re-ranking is conducted during the inference phase based on the contextual relevance of the generated questions. Using our framework, we produce four ConvQA datasets by utilizing documents from multiple domains as the primary source. Through automatic evaluation using diverse metrics, as well as human evaluation, we validate that our proposed framework exhibits the ability to generate datasets of higher quality compared to the baseline dialog inpainting model.

Kyungmin Min, Minbeom Kim, Kang-il Lee et al.

Detailed image captioning demands both factual grounding and fine-grained coverage, yet existing methods have struggled to achieve them simultaneously. We address this tension with Reflective Note-Guided Captioning (ReflectCAP), where a multi-agent pipeline analyzes what the target large vision-language model (LVLM) consistently hallucinates and what it systematically overlooks, distilling these patterns into reusable guidelines called Structured Reflection Notes. At inference time, these notes steer the captioning model along both axes -- what to avoid and what to attend to -- yielding detailed captions that jointly improve factuality and coverage. Applying this method to 8 LVLMs spanning the GPT-4.1 family, Qwen series, and InternVL variants, ReflectCAP reaches the Pareto frontier of the trade-off between factuality and coverage, and delivers substantial gains on CapArena-Auto, where generated captions are judged head-to-head against strong reference models. Moreover, ReflectCAP offers a more favorable trade-off between caption quality and compute cost than model scaling or existing multi-agent pipelines, which incur 21--36\% greater overhead. This makes high-quality detailed captioning viable under real-world cost and latency constraints.

Nakyeong Yang, Taegwan Kang, Jungkyu Choi et al.

Instruction-following language models often show undesirable biases. These undesirable biases may be accelerated in the real-world usage of language models, where a wide range of instructions is used through zero-shot example prompting. To solve this problem, we first define the bias neuron, which significantly affects biased outputs, and prove its existence empirically. Furthermore, we propose a novel and practical bias mitigation method, CRISPR, to eliminate bias neurons of language models in instruction-following settings. CRISPR automatically determines biased outputs and categorizes neurons that affect the biased outputs as bias neurons using an explainability method. Experimental results demonstrate the effectiveness of our method in mitigating biases under zero-shot instruction-following settings without losing the model's task performance and existing knowledge. The experimental results reveal the generalizability of our method as it shows robustness under various instructions and datasets. Surprisingly, our method can mitigate the bias in language models by eliminating only a few neurons (at least three).

Junseok Kim, Nakyeong Yang, Kyomin Jung

Recent studies demonstrate that prompting a role-playing persona to an LLM improves reasoning capability. However, assigning an adequate persona is difficult since LLMs are extremely sensitive to assigned prompts; thus, inaccurately defined personas sometimes hinder LLMs and degrade their reasoning capabilities. In this paper, we first investigate the potential negative impact of injecting persona into language models. Furthermore, we propose a novel framework, Jekyll \& Hyde, which ensembles the outcomes of both role-playing and neutral prompts to enhance the robustness of reasoning ability. Specifically, Jekyll \& Hyde predicts an appropriate persona using an LLM when defining the role-playing prompt. Then, Jekyll \& Hyde collects two potential solutions from role-playing and neutral prompts and selects a better solution using the LLM evaluator. The experimental analysis demonstrates that role-playing prompts sometimes distract LLMs, degrading their reasoning abilities in 7 out of 12 datasets in llama3. Meanwhile, Jekyll \& Hyde improve reasoning capabilities by selecting better choices among the potential solutions on twelve widely-used natural language reasoning datasets. In addition, we reveal that assigning LLM-generated personas obtains more stable results than handcrafted personas.

Kyeongman Park, Minbeom Kim, Kyomin Jung

Creative story generation has long been a goal of NLP research. While existing methodologies have aimed to generate long and coherent stories, they fall significantly short of human capabilities in terms of diversity and character depth. To address this, we introduce a novel story generation framework called CCI (Character-centric Creative story generation via Imagination). CCI features two modules for creative story generation: IG (Image-Guided Imagination) and MW (Multi-Writer model). In the IG module, we utilize a text-to-image model to create visual representations of key story elements, such as characters, backgrounds, and main plots, in a more novel and concrete manner than text-only approaches. The MW module uses these story elements to generate multiple persona-description candidates and selects the best one to insert into the story, thereby enhancing the richness and depth of the narrative. We compared the stories generated by CCI and baseline models through statistical analysis, as well as human and LLM evaluations. The results showed that the IG and MW modules significantly improve various aspects of the stories' creativity. Furthermore, our framework enables interactive multi-modal story generation with users, opening up new possibilities for human-LLM integration in cultural development. Project page : https://www.2024cci.p-e.kr/

Yunah Jang, Megha Sundriyal, Kyomin Jung et al.

Adaptive Retrieval-Augmented Generation (RAG) promises accuracy and efficiency by dynamically triggering retrieval only when needed and is widely used in practice. However, real-world queries vary in surface form even with the same intent, and their impact on Adaptive RAG remains under-explored. We introduce the first large-scale benchmark of diverse yet semantically identical query variations, combining human-written and model-generated rewrites. Our benchmark facilitates a systematic evaluation of Adaptive RAG robustness by examining its key components across three dimensions: answer quality, computational cost, and retrieval decisions. We discover a critical robustness gap, where small surface-level changes in queries dramatically alter retrieval behavior and accuracy. Although larger models show better performance, robustness does not improve accordingly. These findings reveal that Adaptive RAG methods are highly vulnerable to query variations that preserve identical semantics, exposing a critical robustness challenge.

Junseok Kim, Nakyeong Yang, Kyungmin Min et al.

Self-Consistency improves reasoning reliability through multi-sample aggregation, but incurs substantial inference cost. Adaptive self-consistency methods mitigate this issue by adjusting the sampling budget; however, they rely on count-based stopping rules that treat all responses equally, often leading to unnecessary sampling. We propose Reliability-Aware Adaptive Self-Consistency (ReASC), which addresses this limitation by reframing adaptive sampling from response counting to evidence sufficiency, leveraging response-level confidence for principled information aggregation. ReASC operates in two stages: a single-sample decision stage that resolves instances confidently answerable from a single response, and a reliability-aware accumulation stage that aggregates responses by jointly leveraging their frequency and confidence. Across five models and four datasets, ReASC consistently achieves the best accuracy-cost trade-off compared to existing baselines, yielding improved inference efficiency across model scales from 3B to 27B parameters. As a concrete example, ReASC reduces inference cost by up to 70\% relative to self-consistency while preserving accuracy on GSM8K using Gemma-3-4B-it.

Kyeongman Park, Nakyeong Yang, Kyomin Jung

A human author can write any length of story without losing coherence. Also, they always bring the story to a proper ending, an ability that current language models lack. In this work, we present the LongStory for coherent, complete, and length-controlled long story generation. LongStory introduces two novel methodologies: (1) the long and short-term contexts weight calibrator (CWC) and (2) long story structural positions (LSP). The CWC adjusts weights for long-term context Memory and short-term context Cheating, acknowledging their distinct roles. The LSP employs discourse tokens to convey the structural positions of a long story. Trained on three datasets with varied average story lengths, LongStory outperforms other baselines, including the strong story generator Plotmachine, in coherence, completeness, relevance, and repetitiveness. We also perform zero-shot tests on each dataset to assess the model's ability to predict outcomes beyond its training data and validate our methodology by comparing its performance with variants of our model.

Kang-il Lee, Segwang Kim, Kyomin Jung

The problem of spurious programs is a longstanding challenge when training a semantic parser from weak supervision. To eliminate such programs that have wrong semantics but correct denotation, existing methods focus on exploiting similarities between examples based on domain-specific knowledge. In this paper, we propose a domain-agnostic filtering mechanism based on program execution results. Specifically, for each program obtained through the search process, we first construct a representation that captures the program's semantics as execution results under various inputs. Then, we run a majority vote on these representations to identify and filter out programs with significantly different semantics from the other programs. In particular, our method is orthogonal to the program search process so that it can easily augment any of the existing weakly supervised semantic parsing frameworks. Empirical evaluations on the Natural Language Visual Reasoning and WikiTableQuestions demonstrate that applying our method to the existing semantic parsers induces significantly improved performances.

Dohyung Kim, Minbeom Kim, Jeonghye Kim et al.

Reward-maximizing RL methods enhance the reasoning performance of LLMs, but often reduce the diversity among outputs. Recent works address this issue by adopting GFlowNets, training LLMs to match a target distribution while jointly learning its partition function. In contrast to prior works that treat this partition function solely as a normalizer, we reinterpret it as a per-prompt expected-reward (i.e., online accuracy) signal, leveraging this unused information to improve sample efficiency. Specifically, we first establish a theoretical relationship between the partition function and per-prompt accuracy estimates. Building on this key insight, we propose Partition Function-Guided RL (PACED-RL), a post-training framework that leverages accuracy estimates to prioritize informative question prompts during training, and further improves sample efficiency through an accuracy estimate error-prioritized replay. Crucially, both components reuse information already produced during GFlowNet training, effectively amortizing the compute overhead into the existing optimization process. Extensive experiments across diverse benchmarks demonstrate strong performance improvements over GRPO and prior GFlowNet approaches, highlighting PACED-RL as a promising direction for a more sample efficient distribution-matching training for LLMs.

Minbeom Kim, Lesly Miculicich, Bhavana Dalvi Mishra et al.

As AI agents move from chat interfaces to systems that read private data, call tools, and execute multi-step workflows, guardrails become a last line of defense against concrete deployment harms. In these settings, guardrail failures are no longer merely answer-quality errors: they can leak secrets, authorize unsafe actions, or block legitimate work. The hardest failures are often contextual: whether an action is acceptable depends on local privacy norms, organizational policies, and user expectations that resist pre-deployment specification. This creates a practical gap: guardrails must adapt to their own operating environments, yet deployment feedback is typically limited to sparse, noisy user-reported failures, and repeated fine-tuning is often impractical. To address this gap, we propose LiSA (Lifelong Safety Adaptation), a conservative policy induction framework that improves a fixed base guardrail through structured memory. LiSA converts occasional failures into reusable policy abstractions so that sparse reports can generalize beyond individual cases, adds conflict-aware local rules to prevent overgeneralization in mixed-label contexts, and applies evidence-aware confidence gating via a posterior lower bound, so that memory reuse scales with accumulated evidence rather than empirical accuracy alone. Across PrivacyLens+, ConFaide+, and AgentHarm, LiSA consistently outperforms strong memory-based baselines under sparse feedback, remains robust under noisy user feedback even at 20% label-flip rates, and pushes the latency--performance frontier beyond backbone model scaling. Ultimately, LiSA offers a practical path to secure AI agents against the unpredictable long tail of real-world edge risks.

Yerin Hwang, Dongryeol Lee, Taegwan Kang et al.

Large language models (LLMs) are known to produce varying responses depending on prompt phrasing, indicating that subtle guidance in phrasing can steer their answers. However, the impact of this framing bias on LLM-based evaluation, where models are expected to make stable and impartial judgments, remains largely underexplored. Drawing inspiration from the framing effect in psychology, we systematically investigate how deliberate prompt framing skews model judgments across four high-stakes evaluation tasks. We design symmetric prompts using predicate-positive and predicate-negative constructions and demonstrate that such framing induces significant discrepancies in model outputs. Across 14 LLM judges, we observe clear susceptibility to framing, with model families showing distinct tendencies toward agreement or rejection. These findings suggest that framing bias is a structural property of current LLM-based evaluation systems, underscoring the need for framing-aware protocols.

Minwoo Lee, Hyukhun Koh, Minsung Kim et al.

In machine translation, the problem of ambiguously gendered input has been pointed out, where the gender of an entity is not available in the source sentence. To address this ambiguity issue, the task of controlled translation that takes the gender of the ambiguous entity as additional input have been proposed. However, most existing works have only considered a simplified setup of one target gender for input. In this paper, we tackle controlled translation in a more realistic setting of inputs with multiple entities and propose Gender-of-Entity (GoE) prompting method for LLMs. Our proposed method instructs the model with fine-grained entity-level gender information to translate with correct gender inflections. By utilizing four evaluation benchmarks, we investigate the controlled translation capability of LLMs in multiple dimensions and find that LLMs reach state-of-the-art performance in controlled translation. Furthermore, we discover an emergence of gender interference phenomenon when controlling the gender of multiple entities. Finally, we address the limitations of existing gender accuracy evaluation metrics and propose leveraging LLMs as an evaluator for gender inflection in machine translation.

Dongryeol Lee, Yerin Hwang, Taegwan Kang et al.

While large language models (LLMs) are increasingly used as automatic judges for question answering (QA) and other reference-conditioned evaluation tasks, little is known about their ability to adhere to a provided reference. We identify a critical failure mode of such reference-based LLM QA evaluation: when the provided reference conflicts with the judge model's parametric knowledge, the resulting scores become unreliable, substantially degrading evaluation fidelity. To study this phenomenon systematically, we introduce a controlled swapped-reference QA framework that induces reference-belief conflicts. Specifically, we replace the reference answer with an incorrect entity and construct diverse pairings of original and swapped references with correspondingly aligned candidate answers. Surprisingly, grading reliability drops sharply under swapped references across a broad set of judge models. We empirically show that this vulnerability is driven by judges' over-reliance on parametric knowledge, leading judges to disregard the given reference under conflict. Finally, we find that this failure persists under common prompt-based mitigation strategies, highlighting a fundamental limitation of LLM-as-a-judge evaluation and motivating reference-based protocols that enforce stronger adherence to the provided reference.

Kyeongman Park, Minha Jhang, Kyomin Jung

Modern generative models still lack human-level creativity, particularly in multi-branch diversity. Prior approaches to address this problem often incur heavy computation or strong dependency on model architecture. Therefore, we introduce UAG(Universal Avoidance Generation), a model-agnostic and computationally efficient generation strategy that penalizes similarity among previously generated outputs. Thus, UAG can enhance multi-branch diversity across both diffusion and transformer models, with minimal additional computation. In experiments, our method achieves up to 1.9 times higher diversity, runs 4.4 times faster, and requires only 1/64 of the FLOPs compared to state-of-the-art methods. The full code is https://anonymous.4open.science/r/2026_ACL_Universal/.

Junseok Kim, Nakyeong Yang, Kyomin Jung

Role-play prompting is known to steer the behavior of language models by injecting a persona into the prompt, improving their zero-shot reasoning capabilities. However, such improvements are inconsistent across different tasks or instances. This inconsistency suggests that zero-shot and role-play prompting may offer complementary strengths rather than one being universally superior. Building on this insight, we propose Persona Switch, a novel decoding method that dynamically combines the benefits of both prompting strategies. Our method proceeds step-by-step, selecting the better output between zero-shot and role-play prompting at each step by comparing their output confidence, as measured by the logit gap. Experiments with widely-used LLMs demonstrate that Persona Switch consistently outperforms competitive baselines, achieving up to 5.13% accuracy improvement. Furthermore, we show that output confidence serves as an informative measure for selecting the more reliable output.

Kang-il Lee, Jahyun Koo, Seunghyun Yoon et al.

We introduce transductive program synthesis, a new formulation of the program synthesis task that explicitly leverages test inputs during synthesis. While prior approaches to program synthesis--whether based on natural language descriptions or input-output examples--typically aim to generalize from training examples, they often struggle with robustness, especially in real-world settings where training examples are limited and test inputs involve various edge cases. To address this, we propose a novel framework that improves robustness by treating synthesis as an active learning over a finite hypothesis class defined by programs' outputs. We use an LLM to predict outputs for selected test inputs and eliminate inconsistent hypotheses, where the inputs are chosen via a greedy maximin algorithm to minimize the number of LLM queries required. We evaluate our approach on four benchmarks: Playgol, MBPP+, 1D-ARC, and programmatic world modeling on MiniGrid. We demonstrate that our method significantly improves program synthesis in both accuracy and efficiency. We release our code at https://github.com/klee972/SYNTRA.

Hwanhee Lee, Seunghyun Yoon, Franck Dernoncourt et al.

In the automatic evaluation of generative question answering (GenQA) systems, it is difficult to assess the correctness of generated answers due to the free-form of the answer. Especially, widely used n-gram similarity metrics often fail to discriminate the incorrect answers since they equally consider all of the tokens. To alleviate this problem, we propose KPQA-metric, a new metric for evaluating the correctness of GenQA. Specifically, our new metric assigns different weights to each token via keyphrase prediction, thereby judging whether a generated answer sentence captures the key meaning of the reference answer. To evaluate our metric, we create high-quality human judgments of correctness on two GenQA datasets. Using our human-evaluation datasets, we show that our proposed metric has a significantly higher correlation with human judgments than existing metrics. The code is available at https://github.com/hwanheelee1993/KPQA.

Sangmook Lee, Dohyung Kim, Hyukhun Koh et al.

Recent advances in Large Language Models (LLMs) - particularly model scaling and test-time techniques - have greatly enhanced the reasoning capabilities of language models at the expense of higher inference costs. To lower inference costs, prior works train router models or deferral mechanisms that allocate easy queries to a small, efficient model, while forwarding harder queries to larger, more expensive models. However, these trained router models often lack robustness under domain shifts and require expensive data synthesis techniques such as Monte Carlo rollouts to obtain sufficient ground-truth routing labels for training. In this work, we propose Confidence-Guided Stepwise Model Routing for Cost-Efficient Reasoning (STEER), a domain-agnostic framework that performs fine-grained, step-level routing between smaller and larger LLMs without utilizing external models. STEER leverages confidence scores from the smaller model's logits prior to generating a reasoning step, so that the large model is invoked only when necessary. Extensive evaluations using different LLMs on a diverse set of challenging benchmarks across multiple domains such as Mathematical Reasoning, Multi-Hop QA, and Planning tasks indicate that STEER achieves competitive or enhanced accuracy while reducing inference costs (up to +20% accuracy with 48% less FLOPs compared to solely using the larger model on AIME), outperforming baselines that rely on trained external modules. Our results establish model-internal confidence as a robust, domain-agnostic signal for model routing, offering a scalable pathway for efficient LLM deployment.

Seongho Joo, Hyukhun Koh, Kyomin Jung

With the rapid advancement in deep generative models, recent neural Text-To-Speech(TTS) models have succeeded in synthesizing human-like speech. There have been some efforts to generate speech with various prosody beyond monotonous prosody patterns. However, previous works have several limitations. First, typical TTS models depend on the scaled sampling temperature for boosting the diversity of prosody. Speech samples generated at high sampling temperatures often lack perceptual prosodic diversity, which can adversely affect the naturalness of the speech. Second, the diversity among samples is neglected since the sampling procedure often focuses on a single speech sample rather than multiple ones. In this paper, we propose DPP-TTS: a text-to-speech model based on Determinantal Point Processes (DPPs) with a prosody diversifying module. Our TTS model is capable of generating speech samples that simultaneously consider perceptual diversity in each sample and among multiple samples. We demonstrate that DPP-TTS generates speech samples with more diversified prosody than baselines in the side-by-side comparison test considering the naturalness of speech at the same time.

Hyukhun Koh, Dohyung Kim, Minwoo Lee et al.

In the pursuit of developing Large Language Models (LLMs) that adhere to societal standards, it is imperative to detect the toxicity in the generated text. The majority of existing toxicity metrics rely on encoder models trained on specific toxicity datasets, which are susceptible to out-of-distribution (OOD) problems and depend on the dataset's definition of toxicity. In this paper, we introduce a robust metric grounded on LLMs to flexibly measure toxicity according to the given definition. We first analyze the toxicity factors, followed by an examination of the intrinsic toxic attributes of LLMs to ascertain their suitability as evaluators. Finally, we evaluate the performance of our metric with detailed analysis. Our empirical results demonstrate outstanding performance in measuring toxicity within verified factors, improving on conventional metrics by 12 points in the F1 score. Our findings also indicate that upstream toxicity significantly influences downstream metrics, suggesting that LLMs are unsuitable for toxicity evaluations within unverified factors.

Dongryeol Lee, Yerin Hwang, Yongil Kim et al.

In line with the principle of honesty, there has been a growing effort to train large language models (LLMs) to generate outputs containing epistemic markers. However, evaluation in the presence of epistemic markers has been largely overlooked, raising a critical question: Could the use of epistemic markers in LLM-generated outputs lead to unintended negative consequences? To address this, we present EMBER, a benchmark designed to assess the robustness of LLM-judges to epistemic markers in both single and pairwise evaluation settings. Our findings, based on evaluations using EMBER, reveal that all tested LLM-judges, including GPT-4o, show a notable lack of robustness in the presence of epistemic markers. Specifically, we observe a negative bias toward epistemic markers, with a stronger bias against markers expressing uncertainty. This suggests that LLM-judges are influenced by the presence of these markers and do not focus solely on the correctness of the content.

Kyungmin Min, Minbeom Kim, Kang-il Lee et al.

Large Vision-Language Models (LVLMs) demonstrate impressive capabilities in generating detailed and coherent responses from visual inputs. However, they are prone to generate hallucinations due to an over-reliance on language priors. To address this issue, we investigate the language priors in LVLMs and make two key observations: (1) Even when predicting the tokens associated with image-related part-of-speech (POS), models increasingly rely on linguistic priors as the token sequences grow, thereby amplifying hallucinations. (2) Methods that directly calibrate LVLM's output distribution to mitigate language priors can lead to a degradation in text quality or even exacerbate hallucinations. Based on these findings, we propose a novel method, Summary-Guided Decoding (SumGD). This method naturally encourages the model to focus more on image information by reducing the text context through summaries, while controlling only the image-related POS tokens to maintain text quality. Through experiments, we demonstrate that SumGD achieves state-of-the-art performance on object hallucination benchmarks. Furthermore, in terms of the trade-off between precision and recall, SumGD achieves Pareto optimality among the existing methods. Lastly, we observe that although existing methods struggle to balance the reduction of object hallucinations with maintaining text quality, SumGD demonstrates robustness in handling this challenge.

Yerin Hwang, Yongil Kim, Jahyun Koo et al.

Despite the fact that large language models (LLMs) show exceptional skill in instruction following tasks, this strength can turn into a vulnerability when the models are required to disregard certain instructions. Instruction-following tasks typically involve a clear task description and input text containing the target data to be processed. However, when the input itself resembles an instruction, confusion may arise, even if there is explicit prompting to distinguish between the task instruction and the input. We refer to this phenomenon as instructional distraction. In this paper, we introduce a novel benchmark, named DIM-Bench, specifically designed to assess LLMs' performance under instructional distraction. The benchmark categorizes real-world instances of instructional distraction and evaluates LLMs across four instruction tasks: rewriting, proofreading, translation, and style transfer -- alongside five input tasks: reasoning, code generation, mathematical reasoning, bias detection, and question answering. Our experimental results reveal that even the most advanced LLMs are susceptible to instructional distraction, often failing to accurately follow user intent in such cases.

Minbeom Kim, Kang-il Lee, Seongho Joo et al.

Personalized alignments for individual users have been a long-standing goal in large language models (LLMs). We introduce Drift, a novel framework that personalizes LLMs at decoding time with implicit user preferences. Traditional Reinforcement Learning from Human Feedback (RLHF) requires thousands of annotated examples and expensive gradient updates. In contrast, Drift personalizes LLMs in a training-free manner, using only a few dozen examples to steer a frozen model through efficient preference modeling. Our approach models user preferences as a composition of predefined, interpretable attributes and aligns them at decoding time to enable personalized generation. Experiments on both a synthetic persona dataset (Perspective) and a real human-annotated dataset (PRISM) demonstrate that Drift significantly outperforms RLHF baselines while using only 50-100 examples. Our results and analysis show that Drift is both computationally efficient and interpretable.

Minbeom Kim, Hwanhee Lee, Joonsuk Park et al.

As the integration of large language models into daily life is on the rise, there is a clear gap in benchmarks for advising on subjective and personal dilemmas. To address this, we introduce AdvisorQA, the first benchmark developed to assess LLMs' capability in offering advice for deeply personalized concerns, utilizing the LifeProTips subreddit forum. This forum features a dynamic interaction where users post advice-seeking questions, receiving an average of 8.9 advice per query, with 164.2 upvotes from hundreds of users, embodying a collective intelligence framework. Therefore, we've completed a benchmark encompassing daily life questions, diverse corresponding responses, and majority vote ranking to train our helpfulness metric. Baseline experiments validate the efficacy of AdvisorQA through our helpfulness metric, GPT-4, and human evaluation, analyzing phenomena beyond the trade-off between helpfulness and harmlessness. AdvisorQA marks a significant leap in enhancing QA systems for providing personalized, empathetic advice, showcasing LLMs' improved understanding of human subjectivity.

Kang-il Lee, Hyukhun Koh, Dongryeol Lee et al.

Inductive reasoning - the process of inferring general rules from a small number of observations - is a fundamental aspect of human intelligence. Recent works suggest that large language models (LLMs) can engage in inductive reasoning by sampling multiple hypotheses about the rules and selecting the one that best explains the observations. However, due to the IID sampling, semantically redundant hypotheses are frequently generated, leading to significant wastage of compute. In this paper, we 1) demonstrate that increasing the temperature to enhance the diversity is limited due to text degeneration issue, and 2) propose a novel method to improve the diversity while maintaining text quality. We first analyze the effect of increasing the temperature parameter, which is regarded as the LLM's diversity control, on IID hypotheses. Our analysis shows that as temperature rises, diversity and accuracy of hypotheses increase up to a certain point, but this trend saturates due to text degeneration. To generate hypotheses that are more semantically diverse and of higher quality, we propose a novel approach inspired by human inductive reasoning, which we call Mixture of Concepts (MoC). When applied to several inductive reasoning benchmarks, MoC demonstrated significant performance improvements compared to standard IID sampling and other approaches.

Jeonghye Kim, Sojeong Rhee, Minbeom Kim et al.

Recent advances in LLM agents have largely built on reasoning backbones like ReAct, which interleave thought and action in complex environments. However, ReAct often produces ungrounded or incoherent reasoning steps, leading to misalignment between the agent's actual state and goal. Our analysis finds that this stems from ReAct's inability to maintain consistent internal beliefs and goal alignment, causing compounding errors and hallucinations. To address this, we introduce ReflAct, a novel backbone that shifts reasoning from merely planning next actions to continuously reflecting on the agent's state relative to its goal. By explicitly grounding decisions in states and enforcing ongoing goal alignment, ReflAct dramatically improves strategic reliability. This design delivers substantial empirical gains: ReflAct surpasses ReAct by 27.7% on average, achieving a 93.3% success rate in ALFWorld. Notably, ReflAct even outperforms ReAct with added enhancement modules (e.g., Reflexion, WKM), showing that strengthening the core reasoning backbone is key to reliable agent performance.

Nakyeong Yang, Minsung Kim, Seunghyun Yoon et al.

Various studies have attempted to remove sensitive or private knowledge from a language model to prevent its unauthorized exposure. However, prior studies have overlooked the complex and interconnected nature of knowledge, where related knowledge must be carefully examined. Specifically, they have failed to evaluate whether an unlearning method faithfully erases interconnected knowledge that should be removed, retaining knowledge that appears relevant but exists in a completely different context. To resolve this problem, we first define a new concept called superficial unlearning, which refers to the phenomenon where an unlearning method either fails to erase the interconnected knowledge it should remove or unintentionally erases irrelevant knowledge. Based on the definition, we introduce a new benchmark, FaithUn, to analyze and evaluate the faithfulness of unlearning in real-world knowledge QA settings. Furthermore, we propose a novel unlearning method, KLUE, which updates only knowledge-related neurons to achieve faithful unlearning. KLUE identifies knowledge neurons using an explainability method and updates only those neurons using selected unforgotten samples. Experimental results demonstrate that widely-used unlearning methods fail to ensure faithful unlearning, while our method shows significant effectiveness in real-world QA unlearning.

Yerin Hwang, Yongil Kim, Yunah Jang et al.

Despite advancements in on-topic dialogue systems, effectively managing topic shifts within dialogues remains a persistent challenge, largely attributed to the limited availability of training datasets. To address this issue, we propose Multi-Passage to Dialogue (MP2D), a data generation framework that automatically creates conversational question-answering datasets with natural topic transitions. By leveraging the relationships between entities in a knowledge graph, MP2D maps the flow of topics within a dialogue, effectively mirroring the dynamics of human conversation. It retrieves relevant passages corresponding to the topics and transforms them into dialogues through the passage-to-dialogue method. Through quantitative and qualitative experiments, we demonstrate MP2D's efficacy in generating dialogue with natural topic shifts. Furthermore, this study introduces a novel benchmark for topic shift dialogues, TS-WikiDialog. Utilizing the dataset, we demonstrate that even Large Language Models (LLMs) struggle to handle topic shifts in dialogue effectively, and we showcase the performance improvements of models trained on datasets generated by MP2D across diverse topic shift dialogue tasks.

Seongho Joo, Kyungmin Min, Jahyun Koo et al.

Despite the great advancement of Language modeling in recent days, Large Language Models (LLMs) such as GPT3 are notorious for generating non-factual responses, so-called "hallucination" problems. Existing methods for detecting and alleviating this hallucination problem require external resources or the internal state of LLMs, such as the output probability of each token. Given the LLM's restricted external API availability and the limited scope of external resources, there is an urgent demand to establish the Black-Box approach as the cornerstone for effective hallucination detection. In this work, we propose a simple black-box hallucination detection metric after the investigation of the behavior of LLMs under expression of uncertainty. Our comprehensive analysis reveals that LLMs generate consistent responses when they present factual responses while non-consistent responses vice versa. Based on the analysis, we propose an efficient black-box hallucination detection metric with the expression of uncertainty. The experiment demonstrates that our metric is more predictive of the factuality in model responses than baselines that use internal knowledge of LLMs.

Jiwon Moon, Yerin Hwang, Dongryeol Lee et al.

With the growing use of large language models(LLMs) as evaluators, their application has expanded to code evaluation tasks, where they assess the correctness of generated code without relying on reference implementations. While this offers scalability and flexibility, it also raises a critical, unresolved question: Can LLM judges fairly and robustly evaluate semantically equivalent code with superficial variations? Functionally correct code often exhibits variations-such as differences in variable names, comments, or formatting-that should not influence its correctness. Yet, whether LLM judges can reliably handle these variations remains unclear. We present the first comprehensive study of this issue, defining six types of potential bias in code evaluation and revealing their systematic impact on LLM judges. Across five programming languages and multiple LLMs, we empirically demonstrate that all tested LLM judges are susceptible to both positive and negative biases, resulting in inflated or unfairly low scores. Moreover, we observe that LLM judges remain vulnerable to these biases even when prompted to generate test cases before scoring, highlighting the need for more robust code evaluation methods.

Kyeongman Park, Seongho Joo, Kyomin Jung

We introduce MultiActor-Audiobook, a zero-shot approach for generating audiobooks that automatically produces consistent, expressive, and speaker-appropriate prosody, including intonation and emotion. Previous audiobook systems have several limitations: they require users to manually configure the speaker's prosody, read each sentence with a monotonic tone compared to voice actors, or rely on costly training. However, our MultiActor-Audiobook addresses these issues by introducing two novel processes: (1) MSP (**Multimodal Speaker Persona Generation**) and (2) LSI (**LLM-based Script Instruction Generation**). With these two processes, MultiActor-Audiobook can generate more emotionally expressive audiobooks with a consistent speaker prosody without additional training. We compare our system with commercial products, through human and MLLM evaluations, achieving competitive results. Furthermore, we demonstrate the effectiveness of MSP and LSI through ablation studies.

Dongryeol Lee, Minwoo Lee, Kyungmin Min et al.

Recently, directly using large language models (LLMs) has been shown to be the most reliable method to evaluate QA models. However, it suffers from limited interpretability, high cost, and environmental harm. To address these, we propose to use soft EM with entity-driven answer set expansion. Our approach expands the gold answer set to include diverse surface forms, based on the observation that the surface forms often follow particular patterns depending on the entity type. The experimental results show that our method outperforms traditional evaluation methods by a large margin. Moreover, the reliability of our evaluation method is comparable to that of LLM-based ones, while offering the benefits of high interpretability and reduced environmental harm.

Yerin Hwang, Dongryeol Lee, Taegwan Kang et al.

As large language models take on growing roles as automated evaluators in practical settings, a critical question arises: Can individuals persuade an LLM judge to assign unfairly high scores? This study is the first to reveal that strategically embedded persuasive language can bias LLM judges when scoring mathematical reasoning tasks, where correctness should be independent of stylistic variation. Grounded in Aristotle's rhetorical principles, we formalize seven persuasion techniques (Majority, Consistency, Flattery, Reciprocity, Pity, Authority, Identity) and embed them into otherwise identical responses. Across six math benchmarks, we find that persuasive language leads LLM judges to assign inflated scores to incorrect solutions, by up to 8% on average, with Consistency causing the most severe distortion. Notably, increasing model size does not substantially mitigate this vulnerability. Further analysis demonstrates that combining multiple persuasion techniques amplifies the bias, and pairwise evaluation is likewise susceptible. Moreover, the persuasive effect persists under counter prompting strategies, highlighting a critical vulnerability in LLM-as-a-Judge pipelines and underscoring the need for robust defenses against persuasion-based attacks.

Yerin Hwang, Dongryeol Lee, Kyungmin Min et al.

Recently, large vision-language models (LVLMs) have emerged as the preferred tools for judging text-image alignment, yet their robustness along the visual modality remains underexplored. This work is the first study to address a key research question: Can adversarial visual manipulations systematically fool LVLM judges into assigning unfairly inflated scores? We define potential image induced biases within the context of T2I evaluation and examine how these biases affect the evaluations of LVLM judges. Moreover, we introduce a novel, fine-grained, multi-domain meta-evaluation benchmark named FRAME, which is deliberately constructed to exhibit diverse score distributions. By introducing the defined biases into the benchmark, we reveal that all tested LVLM judges exhibit vulnerability across all domains, consistently inflating scores for manipulated images. Further analysis reveals that combining multiple biases amplifies their effects, and pairwise evaluations are similarly susceptible. Moreover, we observe that visual biases persist under prompt-based mitigation strategies, highlighting the vulnerability of current LVLM evaluation systems and underscoring the urgent need for more robust LVLM judges.

Jahyun Koo, Yerin Hwang, Yongil Kim et al.

Despite the success of Large Language Models (LLMs), they still face challenges related to high inference costs and memory requirements. To address these issues, Knowledge Distillation (KD) has emerged as a popular method for model compression, with student-generated outputs (SGOs) as training data being particularly notable for reducing the mismatch between training and inference. However, SGOs often produce noisy and biased sequences, which can lead to misguidance from the teacher model, especially in long sequences. To mitigate these challenges, we propose SWITCH (Studying WIth TeaCHer for Knowledge Distillation), a novel approach that strategically incorporates the teacher model during the student's sequence generation. SWITCH identifies discrepancies between the token probabilities of the teacher and student models, allowing the teacher to intervene selectively, particularly in long sequences that are more prone to teacher misguidance. Extensive experimental results across three model families and five instruction-following datasets show that SWITCH surpasses traditional KD methods, particularly excelling in the generation of long sequential data.

Minsung Kim, Dong-Kyum Kim, Jea Kwon et al.

Large language models often encounter conflicts between in-context knowledge retrieved at inference time and parametric knowledge acquired during pretraining. Models that accept external knowledge uncritically are vulnerable to misinformation, whereas models that adhere rigidly to parametric knowledge fail to benefit from retrieval. Despite the widespread adoption of retrieval-augmented generation, we still lack a systematic understanding of what shapes knowledge-arbitration strategies during training. This gap risks producing pretrained models with undesirable arbitration behaviors and, consequently, wasting substantial computational resources after the pretraining budget has already been spent. To address this problem, we present the first controlled study of how training conditions influence models' use of in-context and parametric knowledge, and how they arbitrate between them. We train transformer-based language models on a synthetic biographies corpus while systematically controlling various conditions. Our experiments reveal that intra-document repetition of facts fosters the development of both parametric and in-context capabilities. Moreover, training on a corpus that contains inconsistent information or distributional skew encourages models to develop robust strategies for leveraging parametric and in-context knowledge. Rather than viewing these non-ideal properties as artifacts to remove, our results indicate that they are important for learning robust arbitration. These insights offer concrete, empirical guidance for pretraining models that harmoniously integrate parametric and in-context knowledge.

Nakyeong Yang, Dong-Kyum Kim, Jea Kwon et al.

Large language models trained on web-scale data can memorize private or sensitive knowledge, raising significant privacy risks. Although some unlearning methods mitigate these risks, they remain vulnerable to "relearning" during subsequent training, allowing a substantial portion of forgotten knowledge to resurface. In this paper, we show that widely used unlearning methods cause shallow alignment: instead of faithfully erasing target knowledge, they generate spurious unlearning neurons that amplify negative influence to hide it. To overcome this limitation, we introduce Ssiuu, a new class of unlearning methods that employs attribution-guided regularization to prevent spurious negative influence and faithfully remove target knowledge. Experimental results confirm that our method reliably erases target knowledge and outperforms strong baselines across two practical retraining scenarios: (1) adversarial injection of private data, and (2) benign attack using an instruction-following benchmark. Our findings highlight the necessity of robust and faithful unlearning methods for safe deployment of language models.

Seongho Joo, Hyukhun Koh, Kyomin Jung

Large Language Models (LLMs) have demonstrated remarkable capabilities across diverse tasks, but their potential misuse for harmful purposes remains a significant concern. To strengthen defenses against such vulnerabilities, it is essential to investigate universal jailbreak attacks that exploit intrinsic weaknesses in the architecture and learning paradigms of LLMs. In response, we propose \textbf{H}armful \textbf{P}rompt \textbf{La}undering (HaPLa), a novel and broadly applicable jailbreaking technique that requires only black-box access to target models. HaPLa incorporates two primary strategies: 1) \textit{abductive framing}, which instructs LLMs to infer plausible intermediate steps toward harmful activities, rather than directly responding to explicit harmful queries; and 2) \textit{symbolic encoding}, a lightweight and flexible approach designed to obfuscate harmful content, given that current LLMs remain sensitive primarily to explicit harmful keywords. Experimental results show that HaPLa achieves over 95% attack success rate on GPT-series models and 70% across all targets. Further analysis with diverse symbolic encoding rules also reveals a fundamental challenge: it remains difficult to safely tune LLMs without significantly diminishing their helpfulness in responding to benign queries.

Seongho Joo, Hyukhun Koh, Kyomin Jung

In-context learning (ICL) in Large Language Models (LLMs) has shown remarkable performance across various tasks without requiring fine-tuning. However, recent studies have highlighted the risk of private data leakage through the prompt in ICL, especially when LLMs are exposed to malicious attacks. While differential privacy (DP) provides strong privacy guarantees, it often significantly reduces the utility of in-context learning (ICL). To address this challenge, we incorporate task-related public data into the ICL framework while maintaining the DP guarantee. Based on this approach, we propose a private in-context learning algorithm that effectively balances privacy protection and model utility. Through experiments, we demonstrate that our approach significantly improves the utility of private ICL with the assistance of public data. Additionally, we show that our method is robust against membership inference attacks, demonstrating empirical privacy protection.