Multimodal large language models (MLLMs) are increasingly deployed in open-ended, real-world environments where inputs are messy, underspecified, and not always trustworthy. Unlike curated benchmarks, these settings frequently involve instructions that reference missing objects or contradictory facts, rely on ambiguous cues, or request infeasible actions. In such cases, success hinges not merely on task execution, but on the model’s ability to detect when something is silently wrong. This paper presents a systematic analysis of how current MLLMs handle such underspecified and misspecified scenarios: cases where flaws must be inferred from context rather than explicitly stated. Using a curated diagnostic suite spanning four categories of real-world failure modes, we evaluate nine MLLMs, including o3 and GPT-4o, and find that models often fail to surface hidden issues, even when they possess the necessary perceptual and reasoning skills. Explicit prompting reveals that the underlying capabilities exist but are frequently suppressed in favor of user compliance.We further show that simple inference-time interventions, such as cautious persona prompting and, in particular, requiring a clarifying question, can substantially recover performance. Our findings highlight a persistent gap between reasoning competence and behavioral compliance in current MLLMs, and suggest practical strategies for making these systems more trustworthy in underconstrained environments.

Large Reasoning Models (LRMs) introduce a new generation paradigm of explicitly reasoning before answering, leading to remarkable improvements in complex tasks. However, they pose great safety risks against harmful queries and adversarial attacks. While recent mainstream safety efforts on LRMs, supervised fine-tuning (SFT), improve safety performance, we find that SFT-aligned models struggle to generalize to unseen jailbreak prompts. After thorough investigation, we identify a safety aha moment that can activate safety reasoning and lead to a safe response. This aha moment typically appears in the ‘key sentence’ that follows models’ query understanding process and can indicate whether the model will proceed safely. Based on these insights, we propose SafeKey, including two complementary objectives to better activate the safety aha-moment in the key sentence: (1) a Dual-Path Safety Head to enhance the safety signal in the model’s internal representations before the key sentence, and (2) a Query-Mask Modeling objective to improve the models’ attention on its query understanding, which has important safety hints. Experiments across multiple safety benchmarks demonstrate that our methods significantly improve safety generalization to a wide range of jailbreak attacks and out-of-distribution harmful prompts, lowering the harmfulness rate by 9.6%, while maintaining general abilities. Our analysis reveals how SafeKey enhances safety by reshaping internal attention and improving the quality of hidden representations.

Graphical User Interface (GUI) action grounding, mapping language instructions to actionable elements on GUI screens, is important for assisting users in interactive tutorials, task automation, accessibility support, etc. Most recent works of GUI action grounding use large GUI datasets to fine-tune Multimodal Large Language Models (MLLMs). However, the fine-tuning data is inherently limited to specific GUI environments, leading to significant performance degradation in novel environments due to the generalization challenges in the GUI domain. Therefore, we argue that GUI action grounding models should be further aligned with novel environments before deployment to optimize their performance. To address this, we first propose GUI-Bee, an MLLM-based autonomous agent, to collect high-quality, environment-specific data through exploration and then continuously fine-tune GUI grounding models with the collected data. To ensure the GUI action grounding models generalize to various screens within the target novel environment after the continuous fine-tuning, we equip GUI-Bee with a novel Q-value-Incentive In-Context Reinforcement Learning (Q-ICRL) algorithm that optimizes exploration efficiency and exploration data quality. In the experiment, we introduce NovelScreenSpot to test how well the data can help align GUI action grounding models to novel environments. Furthermore, we conduct an ablation study to validate the Q-ICRL method in enhancing the efficiency of GUI-Bee.

Large Language Models (LLMs) have demonstrated emergent common-sense reasoning and Theory of Mind (ToM) capabilities, making them promising candidates for developing coordination agents. This study introduces the LLM-Coordination Benchmark, a novel benchmark for analyzing LLMs in the context of Pure Coordination Settings, where agents must cooperate to maximize gains. Our benchmark evaluates LLMs through two distinct tasks. The first is Agentic Coordination, where LLMs act as proactive participants in four pure coordination games. The second is Coordination Question Answering (CoordQA), which tests LLMs on 198 multiple-choice questions across these games to evaluate three key abilities: Environment Comprehension, ToM Reasoning, and Joint Planning. Results from Agentic Coordination experiments reveal that LLM-Agents excel in multi-agent coordination settings where decision-making primarily relies on environmental variables but face challenges in scenarios requiring active consideration of partners’ beliefs and intentions. The CoordQA experiments further highlight significant room for improvement in LLMs’ Theory of Mind reasoning and joint planning capabilities. Zero-Shot Coordination (ZSC) experiments in the Agentic Coordination setting demonstrate that LLM agents, unlike RL methods, exhibit robustness to unseen partners. These findings indicate the potential of LLMs as Agents in pure coordination setups and underscore areas for improvement.

Existing Multimodal Large Language Models (MLLMs) are predominantly trained and tested on consistent visual-textual inputs, leaving open the question of whether they can handle inconsistencies in real-world, layout-rich content. To bridge this gap, we propose the Multimodal Inconsistency Reasoning (MMIR) benchmark to assess MLLMs’ ability to detect and reason about semantic mismatches in artifacts such as webpages, presentation slides, and posters. MMIR comprises 534 challenging samples, each containing synthetically injected errors across five reasoning-heavy categories: Factual Contradiction, Identity Misattribution, Contextual Mismatch, Quantitative Discrepancy, and Temporal/Spatial Incoherence. We evaluate eight state-of-the-art MLLMs, showing that models with dedicated multimodal reasoning capabilities, such as o1, substantially outperform their counterparts while open-source models remain particularly vulnerable to inconsistency errors. Detailed error analyses further show that models excel in detecting inconsistencies confined to a single modality, particularly in text, but struggle with cross-modal conflicts and complex layouts. Probing experiments reveal that single-modality prompting, including Chain-of-Thought (CoT) and Set-of-Mark (SoM) methods, yields marginal gains, revealing a key bottleneck in cross-modal reasoning. Our findings highlight the need for advanced multimodal reasoning and point to future research on multimodal inconsistency.

Large Multimodal Models (LMMs) have demonstrated impressive performance on existing medical Visual Question Answering (Med-VQA) benchmarks. However, high reported accuracy does not necessarily reflect their true diagnostic reliability in clinical settings. This study reveals that state-of-the-art models perform worse than random guessing on medical diagnosis questions when subjected to simple Probing Evaluation for Medical Diagnosis (ProbMed). ProbMed challenges models through probing evaluation and procedural diagnosis. Particularly, probing evaluation features pairing ground-truth questions with adversarial counterparts that feature negated and hallucinated attributes, while procedural diagnosis requires reasoning across various dimensions for each image, including modality recognition, organ identification, clinical findings, abnormalities, and positional grounding. Our evaluation reveals that even top-performing models like GPT-4o, GPT-4V, and Gemini Pro perform worse than random guessing on specialized diagnostic questions, indicating significant limitations in handling fine-grained medical inquiries. Furthermore, our ablation study on open-source models (e.g., LLaVA, LLaVA-Med, and Med-Flamingo) identifies poor visual understanding as a primary bottleneck—a limitation that can be partially mitigated by incorporating visual descriptions generated by GPT-4o, resulting in an average performance improvement of 9.44%. These findings underscore the urgent need for more robust evaluation methods and domain-specific expertise to ensure the reliability of LMMs in high-stakes medical applications.

Oversensitivity occurs when language models defensively reject prompts that are actually benign. This behavior not only disrupts user interactions but also obscures the boundary between harmful and harmless content. Existing benchmarks rely on static datasets that degrade over time as models evolve, leading to data contamination and diminished evaluative power. To address this, we develop a framework that dynamically generates model-specific challenging datasets, capturing emerging defensive patterns and aligning with each model’s unique behavior. Building on this approach, we construct OverBench, a benchmark that aggregates these datasets across diverse LLM families, encompassing 450,000 samples from 25 models. OverBench provides a dynamic and evolving perspective on oversensitivity, allowing for continuous monitoring of defensive triggers as models advance, highlighting vulnerabilities that static datasets overlook.

Graphical User Interfaces (GUIs) are central to our interaction with digital devices and growing efforts have been made to build models for various GUI understanding tasks. However, these efforts largely overlook an important GUI-referring task: screen reading based on user-indicated points, which we name the Screen Point-and-Read (ScreenPR) task. Currently, this task is predominantly handled by rigid accessible screen reading tools, in great need of new models driven by advancements in Multimodal Large Language Models (MLLMs). In this paper, we propose a Tree-of-Lens (ToL) agent, utilizing a novel ToL grounding mechanism, to address the ScreenPR task. Based on the input point coordinate and the corresponding GUI screenshot, our ToL agent constructs a Hierarchical Layout Tree. Based on the tree, our ToL agent not only comprehends the content of the indicated area but also articulates the layout and spatial relationships between elements. Such layout information is crucial for accurately interpreting information on the screen, distinguishing our ToL agent from other screen reading tools. We also thoroughly evaluate the ToL agent against other baselines on a newly proposed ScreenPR benchmark, which includes GUIs from mobile, web, and operating systems. Last but not least, we test the ToL agent on mobile GUI navigation tasks, demonstrating its utility in identifying incorrect actions along the path of agent execution trajectories. Code and data: https://screen-point-and-read.github.io.

Embodied agents have achieved prominent performance in following human instructions to complete tasks. However, the potential of providing instructions informed by texts and images to assist humans in completing tasks remains underexplored. To uncover this capability, we present the multimodal procedural planning (MPP) task, in which models are given a high-level goal and generate plans of paired text-image steps, providing more complementary and informative guidance than unimodal plans. The key challenges of MPP are to ensure the informativeness, temporal coherence,and accuracy of plans across modalities. To tackle this, we propose Text-Image Prompting (TIP), a dual-modality prompting method that jointly leverages zero-shot reasoning ability in large language models (LLMs) and compelling text-to-image generation ability from diffusion-based models. TIP improves the interaction in the dual modalities using Text-to-Image Bridge and Image-to-Text Bridge, allowing LLMs to guide the textual-grounded image plan generation and leveraging the descriptions of image plans to ground the textual plan reversely. To address the lack of relevant datasets, we collect WIKIPLAN and RECIPEPLAN as a testbed for MPP. Our results show compelling human preferences and automatic scores against unimodal and multimodal baselines on WIKIPLAN and RECIPEPLAN in terms of informativeness, temporal coherence, and plan accuracy.

Large language models (LLMs) capable of casual conversation have recently become widely available. We hypothesize that users of conversational systems want a more personalized experience, and existing work shows that users are highly receptive to personalized questions (PQs). Question Generation tasks, however, focus on factual questions from textual excerpts. To create a PQ generator, we first identify over 400 real user interests by anonymously aggregating ~39K user models. We then populate prompt templates with these 400 interests and use an LLM to generate PQs customized to user interests. The result is PerQs, a novel corpus of ~19K question/answer pairs. We evaluate PerQs at scale in the unique context of the Alexa Prize. Our results show significant positive effects on perceived conversation quality. We then fine-tune, deploy, and evaluate PerQy, a neural model that generates PQs in real-time. When evaluated against several competitive LLM baselines, PerQy produced the most natural and engaging responses.