Vision-Language Models (VLMs) have recently demonstrated remarkable capabilities in comprehending complex visual content. However, the mechanisms underlying how VLMs process visual information remain largely unexplored. In this paper, we conduct a thorough empirical analysis, focusing on attention modules across layers. We reveal several key insights about how these models process visual data: (i) the internal representation of the query tokens (e.g., representations of "describe the image"), is utilized by VLMs to store global image information; we demonstrate that these models generate surprisingly descriptive responses solely from these tokens, without direct access to image tokens. (ii) Cross-modal information flow is predominantly influenced by the middle layers (approximately 25% of all layers), while early and late layers contribute only marginally.(iii) Fine-grained visual attributes and object details are directly extracted from image tokens in a spatially localized manner, i.e., the generated tokens associated with a specific object or attribute attend strongly to their corresponding regions in the image. We propose novel quantitative evaluation to validate our observations, leveraging real-world complex visual scenes. Finally, we demonstrate the potential of our findings in facilitating efficient visual processing in state-of-the-art VLMs.

Building Graphical User Interface (GUI) assistants holds significant promise for enhancing human workflow productivity. While most agents are language-based, relying on closed-source API with text-rich meta-information (e.g., HTML or accessibility tree), they show limitations in perceiving UI visuals as humans do, highlighting the need for GUI visual agents. In this work, we develop a vision-language-action model in digital world, namely ShowUI, which features the following innovations: (i) UI-Guided Visual Token Selection to reduce computational costs by formulating screenshots as an UI connected graph, adaptively identifying their redundant relationship and serve as the criteria for token selection during self-attention blocks; (ii) Interleaved Vision-Language-Action Streaming that flexibly unifies diverse needs within GUI tasks, enabling effective management of visual-action history in navigation or pairing multi-turn query-action sequences per screenshot to enhance training efficiency; (iii) Small-scale High-quality GUI Instruction-following Datasets by careful data curation and employing a resampling strategy to address significant data type imbalances. With above components, ShowUI, a lightweight 2B model using 256K data, achieves a strong 75.1% accuracy in zero-shot screenshot grounding. Its UI-guided token selection further reduces 33% of redundant visual tokens during training and speeds up the performance by 1.4x. Navigation experiments across web Mind2Web, mobile AITW, and online MiniWob environments further underscore the effectiveness and potential of our model in advancing GUI visual agents. The models are available at https://github.com/showlab/ShowUI.

Vision-language generative reward models (VL-GenRMs) play a crucial role in aligning and evaluating multimodal AI systems, yet their own evaluation remains under-explored. Current assessment methods primarily rely on AI-annotated preference labels from traditional VL tasks, which can introduce biases and often fail to effectively challenge state-of-the-art models. To address these limitations, we introduce VL-RewardBench, a comprehensive benchmark spanning general multimodal queries, visual hallucination detection, and complex reasoning tasks. Through our AI-assisted annotation pipeline combining sample selection with human verification, we curate 1,250 high-quality examples specifically designed to probe model limitations. Comprehensive evaluation across 16 leading large vision-language models, demonstrates VL-RewardBench's effectiveness as a challenging testbed, where even GPT-4o achieves only 65.4% accuracy, and state-of-the-art open-source models such as Qwen2-VL-72B, struggle to surpass random-guessing. Importantly, performance on VL-RewardBench strongly correlates (Pearson's r > 0.9) with MMMU-Pro accuracy using Best-of-N sampling with VL-GenRMs. Analysis experiments uncover three critical insights for improving VL-GenRMs: (i) models predominantly fail at basic visual perception tasks rather than reasoning tasks; (ii) inference-time scaling benefits vary dramatically by model capacity; and (iii) training VL-GenRMs to learn to judge substantially boosts judgment capability (+14.7% accuracy for a 7B VL-GenRM). We believe VL-RewardBench along with the experimental insights will become a valuable resource for advancing VL-GenRMs.

Across all fields of academic study, experts cite their sources when sharing information. While large language models (LLMs) excel at synthesizing information, they do not provide reliable citation to sources, making it difficult to trace and verify the origins of the information they present. In contrast, search engines make sources readily accessible to users and place the burden of synthesizing information on the user. Through a survey, we find that users prefer search engines over LLMs for high-stakes queries, where concerns regarding information provenance outweigh the perceived utility of LLM responses. To examine the interplay between verifiability and utility of information-sharing tools, we introduce the extractive-abstractive spectrum, in which search engines and LLMs are extreme endpoints encapsulating multiple unexplored intermediate operating points. Search engines are extractive because they respond to queries with snippets of sources with links (citations) to the original webpages. LLMs are abstractive because they address queries with answers that synthesize and logically transform relevant information from training and in-context sources without reliable citation. We define five operating points that span the extractive-abstractive spectrum and conduct human evaluations on seven systems across four diverse query distributions that reflect real-world QA settings: web search, language simplification, multi-step reasoning, and medical advice. As outputs become more abstractive, we find that perceived utility improves by as much as 200%, while the proportion of properly cited sentences decreases by as much as 50% and users take up to 3 times as long to verify cited information. Our findings recommend distinct operating points for domain-specific LLM systems and our failure analysis informs approaches to high-utility LLM systems that empower users to verify information.

Large language models (LLMs) often reflect real-world biases, leading to efforts to mitigate these effects and make the models unbiased. Achieving this goal requires defining clear criteria for an unbiased state, with any deviation from these criteria considered biased. Some studies define an unbiased state as equal treatment across diverse demographic groups, aiming for balanced outputs from LLMs. However, differing perspectives on equality and the importance of pluralism make it challenging to establish a universal standard. Alternatively, other approaches propose using fact-based criteria for more consistent and objective evaluations, though these methods have not yet been fully applied to LLM bias assessments. Thus, there is a need for a metric with objective criteria that offers a distinct perspective from equality-based approaches. Motivated by this need, we introduce a novel metric to assess bias using fact-based criteria and real-world statistics. In this paper, we conducted a human survey demonstrating that humans tend to perceive LLM outputs more positively when they align closely with real-world demographic distributions. Evaluating various LLMs with our proposed metric reveals that model bias varies depending on the criteria used, highlighting the need for multi-perspective assessment.

This paper introduces a novel method, referred to as "hashing", which involves masking potentially bias-inducing words in large language models (LLMs) with hash-like meaningless identifiers to reduce cognitive biases and reliance on external knowledge. The method was tested across three sets of experiments involving a total of 490 prompts. Statistical analysis using chi-square tests showed significant improvements in all tested scenarios, which covered LLama, ChatGPT, Copilot, Gemini and Mixtral models. In the first experiment, hashing decreased the fallacy rate in a modified version of the "Linda" problem aimed at evaluating susceptibility to cognitive biases. In the second experiment, it improved LLM results on the frequent itemset extraction task. In the third experiment, we found hashing is also effective when the Linda problem is presented in a tabular format rather than text, indicating that the technique works across various input representations. Overall, the method was shown to improve bias reduction and incorporation of external knowledge. Despite bias reduction, hallucination rates were inconsistently reduced across types of LLM models. These findings suggest that masking bias-inducing terms can improve LLM performance, although its effectiveness is model- and task-dependent.

Automated radiology report generation (R2Gen) has advanced significantly, introducing challenges in accurate evaluation due to its complexity. Traditional metrics often fall short by relying on rigid word-matching or focusing only on pathological entities, leading to inconsistencies with human assessments. To bridge this gap, we introduce ER2Score, an automatic evaluation metric designed specifically for R2Gen. Our metric utilizes a reward model, guided by our margin-based reward enforcement loss, along with a tailored training data design that enables customization of evaluation criteria to suit user-defined needs. It not only scores reports according to user-specified criteria but also provides detailed sub-scores, enhancing interpretability and allowing users to adjust the criteria between different aspects of reports. Leveraging GPT-4, we designed an easy-to-use data generation pipeline, enabling us to produce extensive training data based on two distinct scoring systems, each containing reports of varying quality along with corresponding scores. These GPT-generated reports are then paired as accepted and rejected samples through our pairing rule to train an LLM towards our fine-grained reward model, which assigns higher rewards to the report with high quality. Our reward-control loss enables this model to simultaneously output multiple individual rewards corresponding to the number of evaluation criteria, with their summation as our final ER2Score. Our experiments demonstrate ER2Score's heightened correlation with human judgments and superior performance in model selection compared to traditional metrics. Notably, our model provides both an overall score and individual scores for each evaluation item, enhancing interpretability. We also demonstrate its flexible training across various evaluation systems.

2D Matryoshka Training is an advanced embedding representation training approach designed to train an encoder model simultaneously across various layer-dimension setups. This method has demonstrated higher effectiveness in Semantic Text Similarity (STS) tasks over traditional training approaches when using sub-layers for embeddings. Despite its success, discrepancies exist between two published implementations, leading to varied comparative results with baseline models. In this reproducibility study, we implement and evaluate both versions of 2D Matryoshka Training on STS tasks and extend our analysis to retrieval tasks. Our findings indicate that while both versions achieve higher effectiveness than traditional Matryoshka training on sub-dimensions, and traditional full-sized model training approaches, they do not outperform models trained separately on specific sub-layer and sub-dimension setups. Moreover, these results generalize well to retrieval tasks, both in supervised (MSMARCO) and zero-shot (BEIR) settings. Further explorations of different loss computations reveals more suitable implementations for retrieval tasks, such as incorporating full-dimension loss and training on a broader range of target dimensions. Conversely, some intuitive approaches, such as fixing document encoders to full model outputs, do not yield improvements. Our reproduction code is available at https://github.com/ielab/2DMSE-Reproduce.

AIGC images are prevalent across various fields, yet they frequently suffer from quality issues like artifacts and unnatural textures. Specialized models aim to predict defect region heatmaps but face two primary challenges: (1) lack of explainability, failing to provide reasons and analyses for subtle defects, and (2) inability to leverage common sense and logical reasoning, leading to poor generalization. Multimodal large language models (MLLMs) promise better comprehension and reasoning but face their own challenges: (1) difficulty in fine-grained defect localization due to the limitations in capturing tiny details; and (2) constraints in providing pixel-wise outputs necessary for precise heatmap generation. To address these challenges, we propose HEIE: a novel MLLM-Based Hierarchical Explainable image Implausibility Evaluator. We introduce the CoT-Driven Explainable Trinity Evaluator, which integrates heatmaps, scores, and explanation outputs, using CoT to decompose complex tasks into subtasks of increasing difficulty and enhance interpretability. Our Adaptive Hierarchical Implausibility Mapper synergizes low-level image features with high-level mapper tokens from LLMs, enabling precise local-to-global hierarchical heatmap predictions through an uncertainty-based adaptive token approach. Moreover, we propose a new dataset: Expl-AIGI-Eval, designed to facilitate interpretable implausibility evaluation of AIGC images. Our method demonstrates state-of-the-art performance through extensive experiments.

We present Buffer Anytime, a framework for estimation of depth and normal maps (which we call geometric buffers) from video that eliminates the need for paired video--depth and video--normal training data. Instead of relying on large-scale annotated video datasets, we demonstrate high-quality video buffer estimation by leveraging single-image priors with temporal consistency constraints. Our zero-shot training strategy combines state-of-the-art image estimation models based on optical flow smoothness through a hybrid loss function, implemented via a lightweight temporal attention architecture. Applied to leading image models like Depth Anything V2 and Marigold-E2E-FT, our approach significantly improves temporal consistency while maintaining accuracy. Experiments show that our method not only outperforms image-based approaches but also achieves results comparable to state-of-the-art video models trained on large-scale paired video datasets, despite using no such paired video data.