Seunghyun Yoon

Jaemin Cho, Seunghyun Yoon, Ajinkya Kale et al. · allen-ai

Modern image captioning models are usually trained with text similarity objectives. However, since reference captions in public datasets often describe the most salient common objects, models trained with text similarity objectives tend to ignore specific and detailed aspects of an image that distinguish it from others. Toward more descriptive and distinctive caption generation, we propose using CLIP, a multimodal encoder trained on huge image-text pairs from web, to calculate multimodal similarity and use it as a reward function. We also propose a simple finetuning strategy of the CLIP text encoder to improve grammar that does not require extra text annotation. This completely eliminates the need for reference captions during the reward computation. To comprehensively evaluate descriptive captions, we introduce FineCapEval, a new dataset for caption evaluation with fine-grained criteria: overall, background, object, relations. In our experiments on text-to-image retrieval and FineCapEval, the proposed CLIP-guided model generates more distinctive captions than the CIDEr-optimized model. We also show that our unsupervised grammar finetuning of the CLIP text encoder alleviates the degeneration problem of the naive CLIP reward. Lastly, we show human analysis where the annotators strongly prefer the CLIP reward to the CIDEr and MLE objectives according to various criteria. Code and Data: https://github.com/j-min/CLIP-Caption-Reward

Mihir Parmar, Hanieh Deilamsalehy, Franck Dernoncourt et al.

Extractive summarization plays a pivotal role in natural language processing due to its wide-range applications in summarizing diverse content efficiently, while also being faithful to the original content. Despite significant advancement achieved in extractive summarization by Large Language Models (LLMs), these summaries frequently exhibit incoherence. An important aspect of the coherent summary is its readability for intended users. Although there have been many datasets and benchmarks proposed for creating coherent extractive summaries, none of them currently incorporate user intent to improve coherence in extractive summarization. Motivated by this, we propose a systematically created human-annotated dataset consisting of coherent summaries for five publicly available datasets and natural language user feedback, offering valuable insights into how to improve coherence in extractive summaries. We utilize this dataset for aligning LLMs through supervised fine-tuning with natural language human feedback to enhance the coherence of their generated summaries. Preliminary experiments with Falcon-40B and Llama-2-13B show significant performance improvements (~10% Rouge-L) in terms of producing coherent summaries. We further utilize human feedback to benchmark results over instruction-tuned models such as FLAN-T5 which resulted in several interesting findings. Data and source code are available at https://github.com/Mihir3009/Extract-AI.

Minh Nguyen, Franck Dernoncourt, Seunghyun Yoon et al.

We introduce an approach to identifying speaker names in dialogue transcripts, a crucial task for enhancing content accessibility and searchability in digital media archives. Despite the advancements in speech recognition, the task of text-based speaker identification (SpeakerID) has received limited attention, lacking large-scale, diverse datasets for effective model training. Addressing these gaps, we present a novel, large-scale dataset derived from the MediaSum corpus, encompassing transcripts from a wide range of media sources. We propose novel transformer-based models tailored for SpeakerID, leveraging contextual cues within dialogues to accurately attribute speaker names. Through extensive experiments, our best model achieves a great precision of 80.3\%, setting a new benchmark for SpeakerID. The data and code are publicly available here: \url{https://github.com/adobe-research/speaker-identification}

Hyewon Choi, Yejun Yoon, Seunghyun Yoon et al.

This study investigates how fake news uses a thumbnail for a news article with a focus on whether a news article's thumbnail represents the news content correctly. A news article shared with an irrelevant thumbnail can mislead readers into having a wrong impression of the issue, especially in social media environments where users are less likely to click the link and consume the entire content. We propose to capture the degree of semantic incongruity in the multimodal relation by using the pretrained CLIP representation. From a source-level analysis, we found that fake news employs a more incongruous image to the main content than general news. Going further, we attempted to detect news articles with image-text incongruity. Evaluation experiments suggest that CLIP-based methods can successfully detect news articles in which the thumbnail is semantically irrelevant to news text. This study contributes to the research by providing a novel view on tackling online fake news and misinformation. Code and datasets are available at https://github.com/ssu-humane/fake-news-thumbnail.

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.

Mohamed Elmoghany, Liangbing Zhao, Xiaoqian Shen et al. · allen-ai

Generating long-form storytelling videos with consistent visual narratives remains a significant challenge in video synthesis. We present a novel framework, dataset, and a model that address three critical limitations: background consistency across shots, seamless multi-subject shot-to-shot transitions, and scalability to hour-long narratives. Our approach introduces a background-consistent generation pipeline that maintains visual coherence across scenes while preserving character identity and spatial relationships. We further propose a transition-aware video synthesis module that generates smooth shot transitions for complex scenarios involving multiple subjects entering or exiting frames, going beyond the single-subject limitations of prior work. To support this, we contribute with a synthetic dataset of 10,000 multi-subject transition sequences covering underrepresented dynamic scene compositions. On VBench, InfinityStory achieves the highest Background Consistency (88.94), highest Subject Consistency (82.11), and the best overall average rank (2.80), showing improved stability, smoother transitions, and better temporal coherence.

Hyunjae Kim, Jaehyo Yoo, Seunghyun Yoon et al.

Most weakly supervised named entity recognition (NER) models rely on domain-specific dictionaries provided by experts. This approach is infeasible in many domains where dictionaries do not exist. While a phrase retrieval model was used to construct pseudo-dictionaries with entities retrieved from Wikipedia automatically in a recent study, these dictionaries often have limited coverage because the retriever is likely to retrieve popular entities rather than rare ones. In this study, we present a novel framework, HighGEN, that generates NER datasets with high-coverage pseudo-dictionaries. Specifically, we create entity-rich dictionaries with a novel search method, called phrase embedding search, which encourages the retriever to search a space densely populated with various entities. In addition, we use a new verification process based on the embedding distance between candidate entity mentions and entity types to reduce the false-positive noise in weak labels generated by high-coverage dictionaries. We demonstrate that HighGEN outperforms the previous best model by an average F1 score of 4.7 across five NER benchmark datasets.

Khalil Mrini, Harpreet Singh, Franck Dernoncourt et al.

Current medical question answering systems have difficulty processing long, detailed and informally worded questions submitted by patients, called Consumer Health Questions (CHQs). To address this issue, we introduce a medical question understanding and answering system with knowledge grounding and semantic self-supervision. Our system is a pipeline that first summarizes a long, medical, user-written question, using a supervised summarization loss. Then, our system performs a two-step retrieval to return answers. The system first matches the summarized user question with an FAQ from a trusted medical knowledge base, and then retrieves a fixed number of relevant sentences from the corresponding answer document. In the absence of labels for question matching or answer relevance, we design 3 novel, self-supervised and semantically-guided losses. We evaluate our model against two strong retrieval-based question answering baselines. Evaluators ask their own questions and rate the answers retrieved by our baselines and own system according to their relevance. They find that our system retrieves more relevant answers, while achieving speeds 20 times faster. Our self-supervised losses also help the summarizer achieve higher scores in ROUGE, as well as in human evaluation metrics. We release our code to encourage further research.

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.

Swetha Eppalapally, Daksh Dangi, Chaithra Bhat et al.

Question-answering for domain-specific applications has recently attracted much interest due to the latest advancements in large language models (LLMs). However, accurately assessing the performance of these applications remains a challenge, mainly due to the lack of suitable benchmarks that effectively simulate real-world scenarios. To address this challenge, we introduce two product question-answering (QA) datasets focused on Adobe Acrobat and Photoshop products to help evaluate the performance of existing models on domain-specific product QA tasks. Additionally, we propose a novel knowledge-driven RAG-QA framework to enhance the performance of the models in the product QA task. Our experiments demonstrated that inducing domain knowledge through query reformulation allowed for increased retrieval and generative performance when compared to standard RAG-QA methods. This improvement, however, is slight, and thus illustrates the challenge posed by the datasets introduced.

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.

Thang M. Pham, Seunghyun Yoon, Trung Bui et al.

While contextualized word embeddings have been a de-facto standard, learning contextualized phrase embeddings is less explored and being hindered by the lack of a human-annotated benchmark that tests machine understanding of phrase semantics given a context sentence or paragraph (instead of phrases alone). To fill this gap, we propose PiC -- a dataset of ~28K of noun phrases accompanied by their contextual Wikipedia pages and a suite of three tasks for training and evaluating phrase embeddings. Training on PiC improves ranking models' accuracy and remarkably pushes span-selection (SS) models (i.e., predicting the start and end index of the target phrase) near-human accuracy, which is 95% Exact Match (EM) on semantic search given a query phrase and a passage. Interestingly, we find evidence that such impressive performance is because the SS models learn to better capture the common meaning of a phrase regardless of its actual context. SotA models perform poorly in distinguishing two senses of the same phrase in two contexts (~60% EM) and in estimating the similarity between two different phrases in the same context (~70% EM).

Zhongfen Deng, Seunghyun Yoon, Trung Bui et al.

Aspect-based meeting transcript summarization aims to produce multiple summaries, each focusing on one aspect of content in a meeting transcript. It is challenging as sentences related to different aspects can mingle together, and those relevant to a specific aspect can be scattered throughout the long transcript of a meeting. The traditional summarization methods produce one summary mixing information of all aspects, which cannot deal with the above challenges of aspect-based meeting transcript summarization. In this paper, we propose a two-stage method for aspect-based meeting transcript summarization. To select the input content related to specific aspects, we train a sentence classifier on a dataset constructed from the AMI corpus with pseudo-labeling. Then we merge the sentences selected for a specific aspect as the input for the summarizer to produce the aspect-based summary. Experimental results on the AMI corpus outperform many strong baselines, which verifies the effectiveness of our proposed method.

Meryem M'hamdi, Jonathan May, Franck Dernoncourt et al.

Multilingual semantic search is the task of retrieving relevant contents to a query expressed in different language combinations. This requires a better semantic understanding of the user's intent and its contextual meaning. Multilingual semantic search is less explored and more challenging than its monolingual or bilingual counterparts, due to the lack of multilingual parallel resources for this task and the need to circumvent "language bias". In this work, we propose an alignment approach: MAML-Align, specifically for low-resource scenarios. Our approach leverages meta-distillation learning based on MAML, an optimization-based Model-Agnostic Meta-Learner. MAML-Align distills knowledge from a Teacher meta-transfer model T-MAML, specialized in transferring from monolingual to bilingual semantic search, to a Student model S-MAML, which meta-transfers from bilingual to multilingual semantic search. To the best of our knowledge, we are the first to extend meta-distillation to a multilingual search application. Our empirical results show that on top of a strong baseline based on sentence transformers, our meta-distillation approach boosts the gains provided by MAML and significantly outperforms naive fine-tuning methods. Furthermore, multilingual meta-distillation learning improves generalization even to unseen languages.

Bo Ni, Haowei Fu, Qinwen Ge et al.

As Large Language Models (LLMs) advance, personalization has become a key mechanism for tailoring outputs to individual user needs. However, most existing methods rely heavily on dense interaction histories, making them ineffective in cold-start scenarios where such data is sparse or unavailable. While external signals (e.g., content of similar users) can offer a potential remedy, leveraging them effectively remains challenging: raw context is often noisy, and existing methods struggle to reason over heterogeneous data sources. To address these issues, we introduce PAT (Personalization with Aligned Trajectories), a reasoning framework for cold-start LLM personalization. PAT first retrieves information along two complementary trajectories: writing-style cues from stylistically similar users and topic-specific context from preference-aligned users. It then employs a reinforcement learning-based, iterative dual-reasoning mechanism that enables the LLM to jointly refine and integrate these signals. Experimental results across real-world personalization benchmarks show that PAT consistently improves generation quality and alignment under sparse-data conditions, establishing a strong solution to the cold-start personalization problem.

Linzi Xing, Quan Tran, Fabian Caba et al.

Video topic segmentation unveils the coarse-grained semantic structure underlying videos and is essential for other video understanding tasks. Given the recent surge in multi-modal, relying solely on a single modality is arguably insufficient. On the other hand, prior solutions for similar tasks like video scene/shot segmentation cater to short videos with clear visual shifts but falter for long videos with subtle changes, such as livestreams. In this paper, we introduce a multi-modal video topic segmenter that utilizes both video transcripts and frames, bolstered by a cross-modal attention mechanism. Furthermore, we propose a dual-contrastive learning framework adhering to the unsupervised domain adaptation paradigm, enhancing our model's adaptability to longer, more semantically complex videos. Experiments on short and long video corpora demonstrate that our proposed solution, significantly surpasses baseline methods in terms of both accuracy and transferability, in both intra- and cross-domain settings.

Daeun Lee, Subhojyoti Mukherjee, Branislav Kveton et al.

Streaming video understanding requires models not only to process temporally incoming frames, but also to anticipate user intention for realistic applications like AR glasses. While prior streaming benchmarks evaluate temporal reasoning, none measure whether MLLMs can interpret or leverage human gaze signals within a streaming setting. To fill this gap, we introduce StreamGaze, the first benchmark designed to evaluate how effectively MLLMs use gaze for temporal and proactive reasoning in streaming videos. StreamGaze introduces gaze-guided past, present, and proactive tasks that comprehensively evaluate streaming video understanding. These tasks assess whether models can use real-time gaze to follow shifting attention and infer user intentions from only past and currently observed frames. To build StreamGaze, we develop a gaze-video QA generation pipeline that aligns egocentric videos with raw gaze trajectories via fixation extraction, region-specific visual prompting, and scanpath construction. This pipeline produces spatio-temporally grounded QA pairs that closely reflect human perceptual dynamics. Across all StreamGaze tasks, we observe substantial performance gaps between state-of-the-art MLLMs and human performance, revealing fundamental limitations in gaze-based temporal reasoning, intention modeling, and proactive prediction. We further provide detailed analyses of gaze-prompting strategies, reasoning behaviors, and task-specific failure modes, offering deeper insight into why current MLLMs struggle and what capabilities future models must develop. All data and code will be publicly released to support continued research in gaze-guided streaming video understanding.

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.

Jaeyoon Jung, Yejun Yoon, Seunghyun Yoon et al.

This paper describes VILLAIN, a multimodal fact-checking system that verifies image-text claims through prompt-based multi-agent collaboration. For the AVerImaTeC shared task, VILLAIN employs vision-language model agents across multiple stages of fact-checking. Textual and visual evidence is retrieved from the knowledge store enriched through additional web collection. To identify key information and address inconsistencies among evidence items, modality-specific and cross-modal agents generate analysis reports. In the subsequent stage, question-answer pairs are produced based on these reports. Finally, the Verdict Prediction agent produces the verification outcome based on the image-text claim and the generated question-answer pairs. Our system ranked first on the leaderboard across all evaluation metrics. The source code is publicly available at https://github.com/ssu-humane/VILLAIN.

Ali Modarressi, Hanieh Deilamsalehy, Franck Dernoncourt et al.

Recent large language models (LLMs) support long contexts ranging from 128K to 1M tokens. A popular method for evaluating these capabilities is the needle-in-a-haystack (NIAH) test, which involves retrieving a "needle" (relevant information) from a "haystack" (long irrelevant context). Extensions of this approach include increasing distractors, fact chaining, and in-context reasoning. However, in these benchmarks, models can exploit existing literal matches between the needle and haystack to simplify the task. To address this, we introduce NoLiMa, a benchmark extending NIAH with a carefully designed needle set, where questions and needles have minimal lexical overlap, requiring models to infer latent associations to locate the needle within the haystack. We evaluate 13 popular LLMs that claim to support contexts of at least 128K tokens. While they perform well in short contexts (<1K), performance degrades significantly as context length increases. At 32K, for instance, 11 models drop below 50% of their strong short-length baselines. Even GPT-4o, one of the top-performing exceptions, experiences a reduction from an almost-perfect baseline of 99.3% to 69.7%. Our analysis suggests these declines stem from the increased difficulty the attention mechanism faces in longer contexts when literal matches are absent, making it harder to retrieve relevant information. Even models enhanced with reasoning capabilities or CoT prompting struggle to maintain performance in long contexts. We publicly release the dataset and evaluation code at https://github.com/adobe-research/NoLiMa.

Yifan Li, Seunghyun Yoon, Viet Dac Lai et al.

Vision Transformers (ViTs) based vision foundation models (VFMs) have achieved remarkable performance across diverse vision tasks, but suffer from quadratic complexity that limits scalability to long sequences. Existing linear attention approaches for ViTs are typically trained from scratch, requiring substantial computational resources, while linearization-based methods developed for large language model decoders do not transfer well to ViTs. To address these challenges, we propose ViT-AdaLA, a novel framework for effectively adapting and transferring prior knowledge from VFMs to linear attention ViTs. ViT-AdaLA consists of three stages: attention alignment, feature alignment, and supervised fine-tuning. In the attention alignment stage, we align vanilla linear attention with the original softmax-based attention in each block to approximate the behavior of softmax attention. However, residual approximation errors inevitably accumulate across layers. We mitigate this by fine-tuning the linearized ViT to align its final-layer features with a frozen softmax VFM teacher. Finally, the adapted prior knowledge is transferred to downstream tasks through supervised fine-tuning. Extensive experiments on classification and segmentation tasks demonstrate the effectiveness and generality of ViT-AdaLA over various state-of-the-art linear attention counterpart.

Dang Nguyen, Viet Dac Lai, Seunghyun Yoon et al.

Existing LLM agent systems typically select actions from a fixed and predefined set at every step. While this approach is effective in closed, narrowly scoped environments, it presents two major challenges for real-world, open-ended scenarios: (1) it significantly restricts the planning and acting capabilities of LLM agents, and (2) it requires substantial human effort to enumerate and implement all possible actions, which is impractical in complex environments with a vast number of potential actions. To address these limitations, we propose an LLM agent framework that can dynamically create and compose actions as needed. In this framework, the agent interacts with its environment by generating and executing programs written in a general-purpose programming language. Moreover, generated actions are accumulated over time for future reuse. Our extensive experiments across multiple benchmarks show that this framework significantly improves flexibility and outperforms prior methods that rely on a fixed action set. Notably, it enables LLM agents to adapt and recover in scenarios where predefined actions are insufficient or fail due to unforeseen edge cases. Our code can be found in https://github.com/adobe-research/dynasaur.

Yejun Yoon, Jaeyoon Jung, Seunghyun Yoon et al.

To tackle the AVeriTeC shared task hosted by the FEVER-24, we introduce a system that only employs publicly available large language models (LLMs) for each step of automated fact-checking, dubbed the Herd of Open LLMs for verifying real-world claims (HerO). For evidence retrieval, a language model is used to enhance a query by generating hypothetical fact-checking documents. We prompt pretrained and fine-tuned LLMs for question generation and veracity prediction by crafting prompts with retrieved in-context samples. HerO achieved 2nd place on the leaderboard with the AVeriTeC score of 0.57, suggesting the potential of open LLMs for verifying real-world claims. For future research, we make our code publicly available at https://github.com/ssu-humane/HerO.

Saehyung Lee, Seunghyun Yoon, Trung Bui et al.

Multimodal large language models (MLLMs) excel at generating highly detailed captions but often produce hallucinations. Our analysis reveals that existing hallucination detection methods struggle with detailed captions. We attribute this to the increasing reliance of MLLMs on their generated text, rather than the input image, as the sequence length grows. To address this issue, we propose a multiagent approach that leverages LLM-MLLM collaboration to correct given captions. Additionally, we introduce an evaluation framework and a benchmark dataset to facilitate the systematic analysis of detailed captions. Our experiments demonstrate that our proposed evaluation method better aligns with human judgments of factuality than existing metrics and that existing approaches to improve the MLLM factuality may fall short in hyper-detailed image captioning tasks. In contrast, our proposed method significantly enhances the factual accuracy of captions, even improving those generated by GPT-4V. Finally, we highlight a limitation of VQA-centric benchmarking by demonstrating that an MLLM's performance on VQA benchmarks may not correlate with its ability to generate detailed image captions. Our code and data are available at https://github.com/adobe-research/CapMAS.

Dawood Wasif, Terrence J. Moore, Seunghyun Yoon et al.

Autonomous vehicles must remain safe and effective when encountering rare long-tailed scenarios or cyber-physical intrusions during driving. We present RAIL, a risk-aware human-in-the-loop framework that turns heterogeneous runtime signals into calibrated control adaptations and focused learning. RAIL fuses three cues (curvature actuation integrity, time-to-collision proximity, and observation-shift consistency) into an Intrusion Risk Score (IRS) via a weighted Noisy-OR. When IRS exceeds a threshold, actions are blended with a cue-specific shield using a learned authority, while human override remains available; when risk is low, the nominal policy executes. A contextual bandit arbitrates among shields based on the cue vector, improving mitigation choices online. RAIL couples Soft Actor-Critic (SAC) with risk-prioritized replay and dual rewards so that takeovers and near misses steer learning while nominal behavior remains covered. On MetaDrive, RAIL achieves a Test Return (TR) of 360.65, a Test Success Rate (TSR) of 0.85, a Test Safety Violation (TSV) of 0.75, and a Disturbance Rate (DR) of 0.0027, while logging only 29.07 training safety violations, outperforming RL, safe RL, offline/imitation learning, and prior HITL baselines. Under Controller Area Network (CAN) injection and LiDAR spoofing attacks, it improves Success Rate (SR) to 0.68 and 0.80, lowers the Disengagement Rate under Attack (DRA) to 0.37 and 0.03, and reduces the Attack Success Rate (ASR) to 0.34 and 0.11. In CARLA, RAIL attains a TR of 1609.70 and TSR of 0.41 with only 8000 steps.

Zelin Wan, Han Jun Yoon, Nithin Alluru et al.

We present DASH (Deception-Augmented Shared mental model for Human-machine teaming), a novel framework that enhances mission resilience by embedding proactive deception into Shared Mental Models (SMM). Designed for mission-critical applications such as surveillance and rescue, DASH introduces "bait tasks" to detect insider threats, e.g., compromised Unmanned Ground Vehicles (UGVs), AI agents, or human analysts, before they degrade team performance. Upon detection, tailored recovery mechanisms are activated, including UGV system reinstallation, AI model retraining, or human analyst replacement. In contrast to existing SMM approaches that neglect insider risks, DASH improves both coordination and security. Empirical evaluations across four schemes (DASH, SMM-only, no-SMM, and baseline) show that DASH sustains approximately 80% mission success under high attack rates, eight times higher than the baseline. This work contributes a practical human-AI teaming framework grounded in shared mental models, a deception-based strategy for insider threat detection, and empirical evidence of enhanced robustness under adversarial conditions. DASH establishes a foundation for secure, adaptive human-machine teaming in contested environments.

Vardhan Dongre, Joseph Hsieh, Viet Dac Lai et al.

Large language models can follow complex instructions in a single turn, yet over long multi-turn interactions they often lose the thread of instructions, persona, and rules. This degradation has been measured behaviorally but not mechanistically explained. We propose a channel-transition account: goal-defining tokens become less accessible through attention, while goal-related information may persist in residual representations. We introduce the Goal Accessibility Ratio (GAR), measuring attention from generated tokens to task-defining goal tokens, and combine it with sliding-window ablations and residual-stream probes. When attention to instructions closes, what survives reveals architecture. Across architectures, the transition yields qualitatively distinct failure modes: some models preserve goal-conditioned behavior at vanishing attention, others fail despite decodable residual goal information, and the layer at which this encoding emerges varies from 2 to 27. A within-model causal ablation that force-closes the attention channel in Mistral collapses recall from near-perfect to 11% on a 20-fact retention task and raises persona-constraint violations above an adversarial-pressure baseline without user pressure, with both effects emerging at the predictable crossover turn. Linear probes recover per-episode recall outcomes from residual representations with AUC up to 0.99 across all four primary architectures, while input embeddings remain at chance. Across architectures and model scales, the gap between attention loss and residual decodability predicts whether goal-conditioned behavior survives channel closure. We contribute GAR as a diagnostic, the channel-transition framework as a controlled mechanistic account, and a parametric prediction of failure timing under windowed attention closure.

Yejun Yoon, Jaeyoon Jung, Seunghyun Yoon et al.

This paper presents HerO 2, Team HUMANE's system for the AVeriTeC shared task at the FEVER-25 workshop. HerO 2 is an enhanced version of HerO, the best-performing open-source model from the previous year's challenge. It improves evidence quality through document summarization and answer reformulation, optimizes veracity prediction via post-training quantization under computational constraints, and enhances overall system performance by integrating updated language model (LM) backbones. HerO 2 ranked second on the leaderboard while achieving the shortest runtime among the top three systems, demonstrating both high efficiency and strong potential for real-world fact verification. The code is available at https://github.com/ssu-humane/HerO2.

Yejun Yoon, Seunghyun Yoon, Kunwoo Park

This paper addresses the critical challenge of assessing the representativeness of news thumbnail images, which often serve as the first visual engagement for readers when an article is disseminated on social media. We focus on whether a news image represents the actors discussed in the news text. To serve the challenge, we introduce NewsTT, a manually annotated dataset of 1000 news thumbnail images and text pairs. We found that the pretrained vision and language models, such as BLIP-2, struggle with this task. Since news subjects frequently involve named entities or proper nouns, the pretrained models could have a limited capability to match news actors' visual and textual appearances. We hypothesize that learning to contrast news text with its counterfactual, of which named entities are replaced, can enhance the cross-modal matching ability of vision and language models. We propose CFT-CLIP, a contrastive learning framework that updates vision and language bi-encoders according to the hypothesis. We found that our simple method can boost the performance for assessing news thumbnail representativeness, supporting our assumption. Code and data can be accessed at https://github.com/ssu-humane/news-images-acl24.

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.

Hyounghun Kim, Doo Soon Kim, Seunghyun Yoon et al.

Demand for image editing has been increasing as users' desire for expression is also increasing. However, for most users, image editing tools are not easy to use since the tools require certain expertise in photo effects and have complex interfaces. Hence, users might need someone to help edit their images, but having a personal dedicated human assistant for every user is impossible to scale. For that reason, an automated assistant system for image editing is desirable. Additionally, users want more image sources for diverse image editing works, and integrating an image search functionality into the editing tool is a potential remedy for this demand. Thus, we propose a dataset of an automated Conversational Agent for Image Search and Editing (CAISE). To our knowledge, this is the first dataset that provides conversational image search and editing annotations, where the agent holds a grounded conversation with users and helps them to search and edit images according to their requests. To build such a system, we first collect image search and editing conversations between pairs of annotators. The assistant-annotators are equipped with a customized image search and editing tool to address the requests from the user-annotators. The functions that the assistant-annotators conduct with the tool are recorded as executable commands, allowing the trained system to be useful for real-world application execution. We also introduce a generator-extractor baseline model for this task, which can adaptively select the source of the next token (i.e., from the vocabulary or from textual/visual contexts) for the executable command. This serves as a strong starting point while still leaving a large human-machine performance gap for useful future work. Our code and dataset are publicly available at: https://github.com/hyounghk/CAISE

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.

Dang Nguyen, Jian Chen, Yu Wang et al.

Graphical User Interface (GUI) agents, powered by Large Foundation Models, have emerged as a transformative approach to automating human-computer interaction. These agents autonomously interact with digital systems or software applications via GUIs, emulating human actions such as clicking, typing, and navigating visual elements across diverse platforms. Motivated by the growing interest and fundamental importance of GUI agents, we provide a comprehensive survey that categorizes their benchmarks, evaluation metrics, architectures, and training methods. We propose a unified framework that delineates their perception, reasoning, planning, and acting capabilities. Furthermore, we identify important open challenges and discuss key future directions. Finally, this work serves as a basis for practitioners and researchers to gain an intuitive understanding of current progress, techniques, benchmarks, and critical open problems that remain to be addressed.

Dawit Mureja Argaw, Seunghyun Yoon, Fabian Caba Heilbron et al.

Long-form video content constitutes a significant portion of internet traffic, making automated video summarization an essential research problem. However, existing video summarization datasets are notably limited in their size, constraining the effectiveness of state-of-the-art methods for generalization. Our work aims to overcome this limitation by capitalizing on the abundance of long-form videos with dense speech-to-video alignment and the remarkable capabilities of recent large language models (LLMs) in summarizing long text. We introduce an automated and scalable pipeline for generating a large-scale video summarization dataset using LLMs as Oracle summarizers. By leveraging the generated dataset, we analyze the limitations of existing approaches and propose a new video summarization model that effectively addresses them. To facilitate further research in the field, our work also presents a new benchmark dataset that contains 1200 long videos each with high-quality summaries annotated by professionals. Extensive experiments clearly indicate that our proposed approach sets a new state-of-the-art in video summarization across several benchmarks.

Joonho Yang, Seunghyun Yoon, Byeongjeong Kim et al.

Through the advent of pre-trained language models, there have been notable advancements in abstractive summarization systems. Simultaneously, a considerable number of novel methods for evaluating factual consistency in abstractive summarization systems has been developed. But these evaluation approaches incorporate substantial limitations, especially on refinement and interpretability. In this work, we propose highly effective and interpretable factual inconsistency detection method metric Factual Inconsistency Detection by Zoom-in Summary and Zoom-out Document for abstractive summarization systems that is based on fine-grained atomic facts decomposition. Moreover, we align atomic facts decomposed from the summary with the source document through adaptive granularity expansion. These atomic facts represent a more fine-grained unit of information, facilitating detailed understanding and interpretability of the summary's factual inconsistency. Experimental results demonstrate that our proposed factual consistency checking system significantly outperforms existing systems.

Thang M. Pham, Peijie Chen, Tin Nguyen et al.

CLIP-based classifiers rely on the prompt containing a {class name} that is known to the text encoder. Therefore, they perform poorly on new classes or the classes whose names rarely appear on the Internet (e.g., scientific names of birds). For fine-grained classification, we propose PEEB - an explainable and editable classifier to (1) express the class name into a set of text descriptors that describe the visual parts of that class; and (2) match the embeddings of the detected parts to their textual descriptors in each class to compute a logit score for classification. In a zero-shot setting where the class names are unknown, PEEB outperforms CLIP by a huge margin (~10x in top-1 accuracy). Compared to part-based classifiers, PEEB is not only the state-of-the-art (SOTA) on the supervised-learning setting (88.80% and 92.20% accuracy on CUB-200 and Dogs-120, respectively) but also the first to enable users to edit the text descriptors to form a new classifier without any re-training. Compared to concept bottleneck models, PEEB is also the SOTA in both zero-shot and supervised-learning settings.

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.

Yu Xia, Subhojyoti Mukherjee, Zhouhang Xie et al.

Active Learning (AL) has been a powerful paradigm for improving model efficiency and performance by selecting the most informative data points for labeling and training. In recent active learning frameworks, Large Language Models (LLMs) have been employed not only for selection but also for generating entirely new data instances and providing more cost-effective annotations. Motivated by the increasing importance of high-quality data and efficient model training in the era of LLMs, we present a comprehensive survey on LLM-based Active Learning. We introduce an intuitive taxonomy that categorizes these techniques and discuss the transformative roles LLMs can play in the active learning loop. We further examine the impact of AL on LLM learning paradigms and its applications across various domains. Finally, we identify open challenges and propose future research directions. This survey aims to serve as an up-to-date resource for researchers and practitioners seeking to gain an intuitive understanding of LLM-based AL techniques and deploy them to new applications.

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.

Hyunjae Kim, Seunghyun Yoon, Trung Bui et al.

Contrastive language-image pre-training (CLIP) models have demonstrated considerable success across various vision-language tasks, such as text-to-image retrieval, where the model is required to effectively process natural language input to produce an accurate visual output. However, current models still face limitations in dealing with linguistic variations in input queries, such as paraphrases, making it challenging to handle a broad range of user queries in real-world applications. In this study, we introduce a straightforward fine-tuning approach to enhance the representations of CLIP models for paraphrases. Our approach involves a two-step paraphrase generation process, where we automatically create two categories of paraphrases from web-scale image captions by leveraging large language models. Subsequently, we fine-tune the CLIP text encoder using these generated paraphrases while freezing the image encoder. Our resulting model, which we call ParaCLIP, exhibits significant improvements over baseline CLIP models across various tasks, including paraphrased retrieval (with rank similarity scores improved by up to 2.0% and 5.6%), Visual Genome Relation and Attribution, as well as seven semantic textual similarity tasks.

Junuk Cha, Mengwei Ren, Krishna Kumar Singh et al.

We present a lighting-aware image editing pipeline that, given a portrait image and a text prompt, performs single image relighting. Our model modifies the lighting and color of both the foreground and background to align with the provided text description. The unbounded nature in creativeness of a text allows us to describe the lighting of a scene with any sensory features including temperature, emotion, smell, time, and so on. However, the modeling of such mapping between the unbounded text and lighting is extremely challenging due to the lack of dataset where there exists no scalable data that provides large pairs of text and relighting, and therefore, current text-driven image editing models does not generalize to lighting-specific use cases. We overcome this problem by introducing a novel data synthesis pipeline: First, diverse and creative text prompts that describe the scenes with various lighting are automatically generated under a crafted hierarchy using a large language model (*e.g.,* ChatGPT). A text-guided image generation model creates a lighting image that best matches the text. As a condition of the lighting images, we perform image-based relighting for both foreground and background using a single portrait image or a set of OLAT (One-Light-at-A-Time) images captured from lightstage system. Particularly for the background relighting, we represent the lighting image as a set of point lights and transfer them to other background images. A generative diffusion model learns the synthesized large-scale data with auxiliary task augmentation (*e.g.,* portrait delighting and light positioning) to correlate the latent text and lighting distribution for text-guided portrait relighting.

Hyunji Lee, Seunghyun Yoon, Yunjae Won et al.

Large language models (LLMs) are often supplemented with external knowledge to provide information not encoded in their parameters or to reduce hallucination. In such cases, we expect the model to generate responses by grounding its response in the provided external context. However, prior work has shown that simply appending context at inference time does not ensure grounded generation. To address this, we propose Context-INformed Grounding Supervision (CINGS), a post-training supervision in which the model is trained with relevant context prepended to the response, while computing the loss only over the response tokens and masking out the context. Our experiments demonstrate that models trained with CINGS exhibit stronger grounding in both textual and visual domains compared to standard instruction-tuned models. In the text domain, CINGS outperforms other training methods across 11 information-seeking datasets and is complementary to inference-time grounding techniques. In the vision-language domain, replacing a vision-language model's LLM backbone with a CINGS-trained model reduces hallucinations across four benchmarks and maintains factual consistency throughout the generated response. This improved grounding comes without degradation in general downstream performance. Finally, we analyze the mechanism underlying the enhanced grounding in CINGS and find that it induces a shift in the model's prior knowledge and behavior, implicitly encouraging greater reliance on the external context.

Subhojyoti Mukherjee, Viet Dac Lai, Raghavendra Addanki et al.

Offline reinforcement learning (RL) is a variant of RL where the policy is learned from a previously collected dataset of trajectories and rewards. In our work, we propose a practical approach to offline RL with large language models (LLMs). We recast the problem as reward-weighted fine-tuning, which can be solved using similar techniques to supervised fine-tuning (SFT). To showcase the value of our approach, we apply it to learning short-horizon question-answering policies of a fixed length, where the agent reasons about potential answers or asks clarifying questions. Our work stands in a stark contrast to state-of-the-art methods in this domain, based on SFT and direct preference optimization, which have additional hyper-parameters and do not directly optimize for rewards. We compare to them empirically, and report major gains in both optimized rewards and language quality.

Joonho Yang, Seunghyun Yoon, Hwan Chang et al.

Large Language Models (LLMs) have significantly advanced text generation capabilities, including tasks like summarization, often producing coherent and fluent outputs. However, faithfulness to source material remains a significant challenge due to the generation of hallucinations. While extensive research focuses on detecting and reducing these inaccuracies, less attention has been paid to the positional distribution of hallucination within generated text, particularly in long outputs. In this work, we investigate where hallucinations occur in LLM-based long response generation, using long document summarization as a key case study. Focusing on the challenging setting of long context-aware long response generation, we find a consistent and concerning phenomenon: hallucinations tend to concentrate disproportionately in the latter parts of the generated long response. To understand this bias, we explore potential contributing factors related to the dynamics of attention and decoding over long sequences. Furthermore, we investigate methods to mitigate this positional hallucination, aiming to improve faithfulness specifically in the concluding segments of long outputs.

Thang M. Pham, Phat T. Nguyen, Seunghyun Yoon et al.

While small language models (SLMs) show promises for mobile deployment, their real-world performance and applications on smartphones remains underexplored. We present SlimLM, a series of SLMs optimized for document assistance tasks on mobile devices. Through extensive experiments on a Samsung Galaxy S24, we identify the optimal trade-offs between model size (ranging from 125M to 7B parameters), context length, and inference time for efficient on-device processing. SlimLM is pre-trained on SlimPajama-627B and fine-tuned on DocAssist, our constructed dataset for summarization, question answering and suggestion tasks. Our smallest model demonstrates efficient performance on S24, while larger variants offer enhanced capabilities within mobile constraints. We evaluate SlimLM against existing SLMs, showing comparable or superior performance and offering a benchmark for future research in on-device language models. We also provide an Android application, offering practical insights into SLM deployment. Our findings provide valuable insights and illuminate the capabilities of running advanced language models on high-end smartphones, potentially reducing server costs and enhancing privacy through on-device processing.

Jiyeon Kim, Hyunji Lee, Dylan Zhou et al.

LLMs operating in dynamic real-world contexts often encounter knowledge that evolves continuously or emerges incrementally. To remain accurate and effective, models must adapt to newly arriving information on the fly. We introduce Online Adaptation to Continual Knowledge Streams(OAKS) to evaluate this capability, establishing a benchmark for online adaptation over streaming, continually updating knowledge. Specifically, the benchmark is structured as a sequence of fine-grained context chunks where facts change dynamically across time intervals. OAKS comprises two datasets: OAKS-BABI and OAKS-Novel, where individual facts evolve multiple times across context chunks. These datasets include dense annotations to measure whether models track changes accurately. Evaluating 14 models with varied inference approaches, we observe significant limitations in current methodologies. Both state-of-the-art models and agentic memory systems fail to adapt robustly on OAKS, demonstrating delays in state-tracking and susceptibility to distraction within streaming environments.

Reuben A. Luera, Ryan Rossi, Franck Dernoncourt et al.

In an ideal design pipeline, user interface (UI) design is intertwined with user research to validate decisions, yet studies are often resource-constrained during early exploration. Recent advances in multimodal large language models (MLLMs) offer a promising opportunity to act as early evaluators, helping designers narrow options before formal testing. Unlike prior work that emphasizes user behavior in narrow domains such as e-commerce with metrics like clicks or conversions, we focus on subjective user evaluations across varied interfaces. We investigate whether MLLMs can mimic human preferences when evaluating individual UIs and comparing them. Using data from a crowdsourcing platform, we benchmark GPT-4o, Claude, and Llama across 30 interfaces and examine alignment with human judgments on multiple UI factors. Our results show that MLLMs approximate human preferences on some dimensions but diverge on others, underscoring both their potential and limitations in supplementing early UX research.

Hyunji Lee, Franck Dernoncourt, Trung Bui et al.

In a real-world corpus, knowledge frequently recurs across documents but often contains inconsistencies due to ambiguous naming, outdated information, or errors, leading to complex interrelationships between contexts. Previous research has shown that language models struggle with these complexities, typically focusing on single factors in isolation. We classify these relationships into four types: distracting, ambiguous, counterfactual, and duplicated. Our analysis reveals that no single approach effectively addresses all these interrelationships simultaneously. Therefore, we introduce Context Organizer (CORG), a framework that organizes multiple contexts into independently processed groups. This design allows the model to efficiently find all relevant answers while ensuring disambiguation. CORG consists of three key components: a graph constructor, a reranker, and an aggregator. Our results demonstrate that CORG balances performance and efficiency effectively, outperforming existing grouping methods and achieving comparable results to more computationally intensive, single-context approaches.

Yejun Yoon, Jaeyoon Jung, Seunghyun Yoon et al.

Query expansion methods powered by large language models (LLMs) have demonstrated effectiveness in zero-shot retrieval tasks. These methods assume that LLMs can generate hypothetical documents that, when incorporated into a query vector, enhance the retrieval of real evidence. However, we challenge this assumption by investigating whether knowledge leakage in benchmarks contributes to the observed performance gains. Using fact verification as a testbed, we analyze whether the generated documents contain information entailed by ground-truth evidence and assess their impact on performance. Our findings indicate that, on average, performance improvements consistently occurred for claims whose generated documents included sentences entailed by gold evidence. This suggests that knowledge leakage may be present in fact-verification benchmarks, potentially inflating the perceived performance of LLM-based query expansion methods.

Jeonghyun Park, Ingeol Baek, Seunghyun Yoon et al.

Real-world Multi-hop Question Answering (QA) often involves ambiguity that is inseparable from the reasoning process itself. This ambiguity creates a distinct challenge, where multiple reasoning paths emerge from a single question, each requiring independent resolution. Since each sub-question is ambiguous, the model must resolve ambiguity at every step. Thus, answering a single question requires handling multiple layers of ambiguity throughout the reasoning chain. We find that current Large Language Models (LLMs) struggle in this setting, typically exploring wrong reasoning paths and producing incomplete answers. To facilitate research on multi-hop ambiguity, we introduce MultI-hop Reasoning with AmbiGuity Evaluation for Illusory Questions (MIRAGE), a benchmark designed to analyze and evaluate this challenging intersection of ambiguity interpretation and multi-hop reasoning. MIRAGE contains 1,142 high-quality examples of ambiguous multi-hop questions, categorized under a taxonomy of syntactic, general, and semantic ambiguity, and curated through a rigorous multi-LLM verification pipeline. Our experiments reveal that even state-of-the-art models struggle on MIRAGE, confirming that resolving ambiguity combined with multi-step inference is a distinct and significant challenge. To establish a robust baseline, we propose CLarifying Ambiguity with a Reasoning and InstructiON (CLARION), a multi-agent framework that significantly outperforms existing approaches on MIRAGE, paving the way for more adaptive and robust reasoning systems.