In Korean medical LLMs, domain-trained models underperform General LLMs of comparable parameter scale. We attribute this *paradoxical failure* to a *cascading two-bottleneck* structure: tokenizer over-fragmentation (*Fragmentation bottleneck, B1*) compounded by a learning-stage representation-alignment failure (*Alignment bottleneck, B2*). Using the Korean medical domain as a case study, we diagnose this paradox and propose **VocMoE**, which is designed to address both bottlenecks: DVE selects domain vocabulary from a medical BERT via an information-theoretic score, and MoDE uses the DVE output (the partition of token ids) as the vocabulary-conditioned, deterministic routing key of the MoE FFN. In this view, the vocabulary partition itself serves as a routing prior, shifting routing from hidden-state learning to vocabulary identity. MediKo-9B-A4B reaches simultaneous SOTA on KorMedMCQA (**57.5**, +14.8 pt), KMMLU Medical, and HAE-RAE, while MediKo-4B-A1B (active 1.69B) surpasses Qwen3-8B (8B) and KULLM3 (10.7B) on KorMedMCQA, updating the active-parameter efficiency frontier.

RL-trained search agents interleave retrieval with reasoning to answer knowledge-intensive questions, achieving strong end-to-end accuracy on multi-hop QA benchmarks. Yet end-to-end accuracy conflates two distinct abilities: searching for relevant documents and reading what they contain. Existing evaluations do not measure whether the agent reads the documents it retrieves. We design a controlled evaluation that force-injects gold-bearing documents into each agent’s retrieval pipeline, isolating reading from retrieval. Across five RL search agents on three multi-hop QA benchmarks, up to 46.2% of samples are answered incorrectly even when the gold answer is literally present in the retrieved documents—retrieval succeeds but reading fails. We term this phenomenon Retrieval Myopia (RM). The agent’s interaction format is itself a substantial cause: simply re-presenting the same documents to the same model in a plain QA format, with no weight updates, recovers up to 48.9% of RM cases. Guided by this diagnosis, we propose CWFT (Confidence-Weighted Format Transfer), a training-free pipeline that anchors the agent’s prediction against k confidence-weighted clean-format re-reads. CWFT consistently improves accuracy across all evaluated system–dataset configurations on three multi-hop QA benchmarks.

Recent advances in Multimodal Large Language Models (MLLMs) have improved image recognition and reasoning, but video-related tasks remain challenging due to memory constraints from dense frame processing. Existing Video Moment Retrieval (VMR) methodologies rely on sparse frame sampling, risking potential information loss, especially in lengthy videos. We propose SMORE (See MORE, store less), a framework that enhances memory efficiency while maintaining high information resolution. SMORE (1) uses query-guided captions to encode semantics aligned with user intent, (2) applies query-aware importance modulation to highlight relevant segments, and (3) adaptively compresses frames to preserve key content while reducing redundancy. This enables efficient video understanding without exceeding memory budgets. Experimental validation reveals that SMORE achieves state-of-the-art performance on QVHighlights, Charades-STA, and ActivityNet-Captions benchmarks.

A calibrated Vision-Language Model (VLM) can repeatedly self-correct, say "Wait, I should recheck," arrive at the wrong answer, and still report high confidence. We find that this occurs because verbalized confidence is largely trajectory-independent in the VLMs and calibration methods we evaluate. We examine this through three complementary lenses: content variation, token masking, and the model’s own hesitation markers. We show that confidence is insufficiently sensitive to what the reasoning trajectory actually contains, and that calibration training can paradoxically worsen this disconnect. Since existing metrics like ECE and AUROC cannot detect this problem, we propose the Trajectory-Grounding Score (TGS) in two complementary forms: TGS-self, which compares confidence with and without access to the model’s own trajectory, and TGS-pair, which tests whether the model assigns higher confidence to correct trajectories than to flawed ones along the vision, reasoning, and answer axes. We propose TGS-Bench, a model-agnostic suite spanning 10 benchmarks with controlled good/bad trajectory pairs, and show that conventional calibration rankings diverge from trajectory-grounding rankings, exposing a blind spot in current evaluation practice.

Multi-domain image-to-image translation requires grounding semantic differences expressed in natural language prompts into corresponding visual transformations, while preserving unrelated structural and semantic content. Existing methods struggle to maintain structural integrity and provide fine-grained, attribute-specific control, especially when multiple domains are involved. We propose LACE (Language-grounded Attribute-Controllable Translation), built on two components: (1) a GLIP-Adapter that fuses global semantics with local structural features to preserve consistency, and (2) a Multi-Domain Control Guidance mechanism that explicitly grounds the semantic delta between source and target prompts into per-attribute translation vectors, aligning linguistic semantics with domain-level visual changes. Together, these modules enable compositional multi-domain control with independent strength modulation for each attribute. Experiments on CelebA(Dialog) and BDD100K demonstrate that LACE achieves high visual fidelity, structural preservation, and interpretable domain-specific control, surpassing prior baselines. This positions LACE as a cross-modal content generation framework bridging language semantics and controllable visual translation. Code will be publicly available.

Vision language models (VLMs) demonstrate strong zero-shot performance, but often perpetuate social stereotypes in person-centric queries, yielding skewed demographic distributions. Current debiasing methods apply uniform bias corrections across all input queries regardless of their bias sensitivity, creating a fundamental fairness–utility trade-off. Strong debiasing distorts semantically meaningful information in bias-insensitive queries, while weak debiasing fails to mitigate stereotypes in bias-sensitive ones. This one-size-fits-all approach hampers simultaneously achieving high utility on bias-insensitive queries and fairness on bias-sensitive queries. We introduce Reward-Gated Test-Time Adaptation (RG-TTA), a reinforcement learning-based test-time adaptation framework that selectively applies debiasing based on input sensitivity. RG-TTA adaptively triggers fairness regularization based on the bias sensitivity of each input during test-time policy adaptation, while focusing exclusively on optimizing cross-modal alignment for bias-insensitive inputs. Experiments on fairness benchmarks (e.g., FairFace, UTKFace) demonstrate substantial bias reduction while simultaneously improving zero-shot utility, resolving the trade-off of uniform debiasing.

Video-to-text summarization remains underexplored in terms of comprehensive evaluation methods. Traditional n-gram overlap-based metrics and recent large language model (LLM)-based approaches depend heavily on human-written reference summaries, limiting their practicality and sensitivity to nuanced semantic aspects. In this paper, we propose QEVA, a reference-free metric evaluating candidate summaries directly against source videos through multimodal question answering. QEVA assesses summaries along three clear dimensions: Coverage, Factuality, and Temporal Coherence. We also introduce MLVU(VS)-Eval, a new annotated benchmark derived from the MLVU dataset, comprising 800 summaries generated from 200 videos using state-of-the-art video-language multimodal models. This dataset establishes a transparent and consistent framework for evaluation. Experimental results demonstrate that QEVA shows higher correlation with human judgments compared to existing approaches, as measured by Kendall’s 𝜏_b, 𝜏_c, and Spearman’s 𝜌. We hope that our benchmark and metric will facilitate meaningful progress in video-to-text summarization research and provide valuable insights for the development of future evaluation methods.

Automated Audio Captioning (AAC) aims to generate natural language descriptions of audio content, enabling machines to interpret and communicate complex acoustic scenes. However, current AAC datasets often suffer from short and simplistic captions, limiting model expressiveness and semantic depth. To address this, we introduce VggCaps, a new multi-modal dataset that pairs audio with corresponding video and leverages large language models (LLMs) to generate rich, descriptive captions. VggCaps significantly outperforms existing benchmarks in caption length, lexical diversity, and human-rated quality. Furthermore, we propose Multi2Cap, a novel AAC framework that learns audio-visual representations through a AV-grounding module during pre-training and reconstructs visual semantics using audio alone at inference. This enables visually grounded captioning in audio-only scenarios. Experimental results on Clotho and AudioCaps demonstrate that Multi2Cap achieves state-of-the-art performance across multiple metrics, validating the effectiveness of cross-modal supervision and LLM-based generation in advancing AAC.

Video-grounded Dialogue (VGD) aims to answer questions regarding a given multi-modal input comprising video, audio, and dialogue history. Although there have been numerous efforts in developing VGD systems to improve the quality of their responses, existing systems are competent only to incorporate the information in the video and text and tend to struggle in extracting the necessary information from the audio when generating appropriate responses to the question. The VGD system seems to be deaf, and thus, we coin this symptom of current systems’ ignoring audio data as a deaf response. To overcome the deaf response problem, Hearing Enhanced Audio Response (HEAR) framework is proposed to perform sensible listening by selectively attending to audio whenever the question requires it. The HEAR framework enhances the accuracy and audibility of VGD systems in a model-agnostic manner. HEAR is validated on VGD datasets (i.e., AVSD@DSTC7 and AVSD@DSTC8) and shows effectiveness with various VGD systems.

Video-grounded Dialogue (VGD) aims to decode an answer sentence to a question regarding a given video and dialogue context. Despite the recent success of multi-modal reasoning to generate answer sentences, existing dialogue systems still suffer from a text hallucination problem, which denotes indiscriminate text-copying from input texts without an understanding of the question. This is due to learning spurious correlations from the fact that answer sentences in the dataset usually include the words of input texts, thus the VGD system excessively relies on copying words from input texts by hoping those words to overlap with ground-truth texts. Hence, we design Text Hallucination Mitigating (THAM) framework, which incorporates Text Hallucination Regularization (THR) loss derived from the proposed information-theoretic text hallucination measurement approach. Applying THAM with current dialogue systems validates the effectiveness on VGD benchmarks (i.e., AVSD@DSTC7 and AVSD@DSTC8) and shows enhanced interpretability.