Ziyang Luo

Ziyang Luo, Can Xu, Pu Zhao et al. · microsoft-research

Code Large Language Models (Code LLMs), such as StarCoder, have demonstrated exceptional performance in code-related tasks. However, most existing models are solely pre-trained on extensive raw code data without instruction fine-tuning. In this paper, we introduce WizardCoder, which empowers Code LLMs with complex instruction fine-tuning, by adapting the Evol-Instruct method to the domain of code. Through comprehensive experiments on four prominent code generation benchmarks, namely HumanEval, HumanEval+, MBPP, and DS-1000, we unveil the exceptional capabilities of our model. It surpasses all other open-source Code LLMs by a substantial margin. Moreover, our model even outperforms the largest closed LLMs, Anthropic's Claude and Google's Bard, on HumanEval and HumanEval+. Our code, model weights, and data are public at https://github.com/nlpxucan/WizardLM

Ziyang Luo, Nian Liu, Wangbo Zhao et al.

Salient object detection (SOD) and camouflaged object detection (COD) are related yet distinct binary mapping tasks. These tasks involve multiple modalities, sharing commonalities and unique cues. Existing research often employs intricate task-specific specialist models, potentially leading to redundancy and suboptimal results. We introduce VSCode, a generalist model with novel 2D prompt learning, to jointly address four SOD tasks and three COD tasks. We utilize VST as the foundation model and introduce 2D prompts within the encoder-decoder architecture to learn domain and task-specific knowledge on two separate dimensions. A prompt discrimination loss helps disentangle peculiarities to benefit model optimization. VSCode outperforms state-of-the-art methods across six tasks on 26 datasets and exhibits zero-shot generalization to unseen tasks by combining 2D prompts, such as RGB-D COD. Source code has been available at https://github.com/Sssssuperior/VSCode.

Yuwei Zhao, Ziyang Luo, Yuchen Tian et al.

Recent advancements in large language models (LLMs) have showcased impressive code generation capabilities, primarily evaluated through language-to-code benchmarks. However, these benchmarks may not fully capture a model's code understanding abilities. We introduce CodeJudge-Eval (CJ-Eval), a novel benchmark designed to assess LLMs' code understanding abilities from the perspective of code judging rather than code generation. CJ-Eval challenges models to determine the correctness of provided code solutions, encompassing various error types and compilation issues. By leveraging a diverse set of problems and a fine-grained judging system, CJ-Eval addresses the limitations of traditional benchmarks, including the potential memorization of solutions. Evaluation of 12 well-known LLMs on CJ-Eval reveals that even state-of-the-art models struggle, highlighting the benchmark's ability to probe deeper into models' code understanding abilities. Our codes and benchmark are available at \url{https://github.com/CodeLLM-Research/CodeJudge-Eval}.

Ziyang luo, Pu Zhao, Can Xu et al. · microsoft-research

Image-text retrieval (ITR) is a task to retrieve the relevant images/texts, given the query from another modality. The conventional dense retrieval paradigm relies on encoding images and texts into dense representations using dual-stream encoders, however, it faces challenges with low retrieval speed in large-scale retrieval scenarios. In this work, we propose the lexicon-weighting paradigm, where sparse representations in vocabulary space are learned for images and texts to take advantage of the bag-of-words models and efficient inverted indexes, resulting in significantly reduced retrieval latency. A crucial gap arises from the continuous nature of image data, and the requirement for a sparse vocabulary space representation. To bridge this gap, we introduce a novel pre-training framework, Lexicon-Bottlenecked Language-Image Pre-Training (LexLIP), that learns importance-aware lexicon representations. This framework features lexicon-bottlenecked modules between the dual-stream encoders and weakened text decoders, allowing for constructing continuous bag-of-words bottlenecks to learn lexicon-importance distributions. Upon pre-training with same-scale data, our LexLIP achieves state-of-the-art performance on two benchmark ITR datasets, MSCOCO and Flickr30k. Furthermore, in large-scale retrieval scenarios, LexLIP outperforms CLIP with a 5.5 ~ 221.3X faster retrieval speed and 13.2 ~ 48.8X less index storage memory.

Ziyang Luo, Yadong Xi, Jing Ma et al. · bytedance

Since 2017, the Transformer-based models play critical roles in various downstream Natural Language Processing tasks. However, a common limitation of the attention mechanism utilized in Transformer Encoder is that it cannot automatically capture the information of word order, so explicit position embeddings are generally required to be fed into the target model. In contrast, Transformer Decoder with the causal attention masks is naturally sensitive to the word order. In this work, we focus on improving the position encoding ability of BERT with the causal attention masks. Furthermore, we propose a new pre-trained language model DecBERT and evaluate it on the GLUE benchmark. Experimental results show that (1) the causal attention mask is effective for BERT on the language understanding tasks; (2) our DecBERT model without position embeddings achieve comparable performance on the GLUE benchmark; and (3) our modification accelerates the pre-training process and DecBERT w/ PE achieves better overall performance than the baseline systems when pre-training with the same amount of computational resources.

Nian Liu, Ziyang Luo, Ni Zhang et al.

While previous CNN-based models have exhibited promising results for salient object detection (SOD), their ability to explore global long-range dependencies is restricted. Our previous work, the Visual Saliency Transformer (VST), addressed this constraint from a transformer-based sequence-to-sequence perspective, to unify RGB and RGB-D SOD. In VST, we developed a multi-task transformer decoder that concurrently predicts saliency and boundary outcomes in a pure transformer architecture. Moreover, we introduced a novel token upsampling method called reverse T2T for predicting a high-resolution saliency map effortlessly within transformer-based structures. Building upon the VST model, we further propose an efficient and stronger VST version in this work, i.e. VST++. To mitigate the computational costs of the VST model, we propose a Select-Integrate Attention (SIA) module, partitioning foreground into fine-grained segments and aggregating background information into a single coarse-grained token. To incorporate 3D depth information with low cost, we design a novel depth position encoding method tailored for depth maps. Furthermore, we introduce a token-supervised prediction loss to provide straightforward guidance for the task-related tokens. We evaluate our VST++ model across various transformer-based backbones on RGB, RGB-D, and RGB-T SOD benchmark datasets. Experimental results show that our model outperforms existing methods while achieving a 25% reduction in computational costs without significant performance compromise. The demonstrated strong ability for generalization, enhanced performance, and heightened efficiency of our VST++ model highlight its potential.

Hongzhan Lin, Pengyao Yi, Jing Ma et al.

The spread of rumors along with breaking events seriously hinders the truth in the era of social media. Previous studies reveal that due to the lack of annotated resources, rumors presented in minority languages are hard to be detected. Furthermore, the unforeseen breaking events not involved in yesterday's news exacerbate the scarcity of data resources. In this work, we propose a novel zero-shot framework based on prompt learning to detect rumors falling in different domains or presented in different languages. More specifically, we firstly represent rumor circulated on social media as diverse propagation threads, then design a hierarchical prompt encoding mechanism to learn language-agnostic contextual representations for both prompts and rumor data. To further enhance domain adaptation, we model the domain-invariant structural features from the propagation threads, to incorporate structural position representations of influential community response. In addition, a new virtual response augmentation method is used to improve model training. Extensive experiments conducted on three real-world datasets demonstrate that our proposed model achieves much better performance than state-of-the-art methods and exhibits a superior capacity for detecting rumors at early stages.

Gailun Zeng, Ziyang Luo, Hongzhan Lin et al.

The ability of critique is vital for models to self-improve and serve as reliable AI assistants. While extensively studied in language-only settings, multimodal critique of Large Multimodal Models (LMMs) remains underexplored despite their growing capabilities in tasks like captioning and visual reasoning. In this work, we introduce MM-CRITIC, a holistic benchmark for evaluating the critique ability of LMMs across multiple dimensions: basic, correction, and comparison. Covering 8 main task types and over 500 tasks, MM-CRITIC collects responses from various LMMs with different model sizes and is composed of 4471 samples. To enhance the evaluation reliability, we integrate expert-informed ground answers into scoring rubrics that guide GPT-4o in annotating responses and generating reference critiques, which serve as anchors for trustworthy judgments. Extensive experiments validate the effectiveness of MM-CRITIC and provide a comprehensive assessment of leading LMMs' critique capabilities under multiple dimensions. Further analysis reveals some key insights, including the correlation between response quality and critique, and varying critique difficulty across evaluation dimensions. Our code is available at https://github.com/MichealZeng0420/MM-Critic.

Zixin Chen, Hongzhan Lin, Kaixin Li et al.

The proliferation of memes on social media necessitates the capabilities of multimodal Large Language Models (mLLMs) to effectively understand multimodal harmfulness. Existing evaluation approaches predominantly focus on mLLMs' detection accuracy for binary classification tasks, which often fail to reflect the in-depth interpretive nuance of harmfulness across diverse contexts. In this paper, we propose MemeArena, an agent-based arena-style evaluation framework that provides a context-aware and unbiased assessment for mLLMs' understanding of multimodal harmfulness. Specifically, MemeArena simulates diverse interpretive contexts to formulate evaluation tasks that elicit perspective-specific analyses from mLLMs. By integrating varied viewpoints and reaching consensus among evaluators, it enables fair and unbiased comparisons of mLLMs' abilities to interpret multimodal harmfulness. Extensive experiments demonstrate that our framework effectively reduces the evaluation biases of judge agents, with judgment results closely aligning with human preferences, offering valuable insights into reliable and comprehensive mLLM evaluations in multimodal harmfulness understanding. Our code and data are publicly available at https://github.com/Lbotirx/MemeArena.

Zhiwei Yang, Jing Ma, Hechang Chen et al.

Existing fake news detection methods aim to classify a piece of news as true or false and provide veracity explanations, achieving remarkable performances. However, they often tailor automated solutions on manual fact-checked reports, suffering from limited news coverage and debunking delays. When a piece of news has not yet been fact-checked or debunked, certain amounts of relevant raw reports are usually disseminated on various media outlets, containing the wisdom of crowds to verify the news claim and explain its verdict. In this paper, we propose a novel Coarse-to-fine Cascaded Evidence-Distillation (CofCED) neural network for explainable fake news detection based on such raw reports, alleviating the dependency on fact-checked ones. Specifically, we first utilize a hierarchical encoder for web text representation, and then develop two cascaded selectors to select the most explainable sentences for verdicts on top of the selected top-K reports in a coarse-to-fine manner. Besides, we construct two explainable fake news datasets, which are publicly available. Experimental results demonstrate that our model significantly outperforms state-of-the-art baselines and generates high-quality explanations from diverse evaluation perspectives.

Zirui Zhao, Jun Hao Liew, Yan Yang et al.

GUI Process Automation (GPA) is a lightweight but general vision-based Robotic Process Automation (RPA), which enables fast and stable process replay with only a single demo. Addressing the fragility of traditional RPA and the non-deterministic risks of current vision language model-based GUI agents, GPA introduces three core benefits: (1) Robustness via Sequential Monte Carlo-based localization to handle rescaling and detection uncertainty; (2) Deterministic and Reliability safeguarded by readiness calibration; and (3) Privacy through fast, fully local execution. This approach delivers the adaptability, robustness, and security required for enterprise workflows. It can also be used as an MCP/CLI tool by other agents with coding capabilities so that the agent only reasons and orchestrates while GPA handles the GUI execution. We conducted a pilot experiment to compare GPA with Gemini 3 Pro (with CUA tools) and found that GPA achieves higher success rate with 10 times faster execution speed in finishing long-horizon GUI tasks.

Ruichao Yang, Wei Gao, Xiaobin Zhu et al.

Multimodal misinformation poses an escalating challenge that often evades traditional detectors, which are opaque black boxes and fragile against new manipulation tactics. We present Probabilistic Concept Graph Reasoning (PCGR), an interpretable and evolvable framework that reframes multimodal misinformation detection (MMD) as structured and concept-based reasoning. PCGR follows a build-then-infer paradigm, which first constructs a graph of human-understandable concept nodes, including novel high-level concepts automatically discovered and validated by multimodal large language models (MLLMs), and then applies hierarchical attention over this concept graph to infer claim veracity. This design produces interpretable reasoning chains linking evidence to conclusions. Experiments demonstrate that PCGR achieves state-of-the-art MMD accuracy and robustness to emerging manipulation types, outperforming prior methods in both coarse detection and fine-grained manipulation recognition.

Yuchen Tian, Weixiang Yan, Qian Yang et al. · berkeley, mila

Large Language Models (LLMs) have made significant progress in code generation, offering developers groundbreaking automated programming support. However, LLMs often generate code that is syntactically correct and even semantically plausible, but may not execute as expected or fulfill specified requirements. This phenomenon of hallucinations in the code domain has not been systematically explored. To advance the community's understanding and research on this issue, we introduce the concept of code hallucinations and propose a classification method for code hallucination based on execution verification. We categorize code hallucinations into four main types: mapping, naming, resource, and logic hallucinations, with each category further divided into different subcategories to understand and address the unique challenges faced by LLMs in code generation with finer granularity. Additionally, we present a dynamic detection algorithm called CodeHalu designed to detect and quantify code hallucinations. We also introduce the CodeHaluEval benchmark, which includes 8,883 samples from 699 tasks, to systematically and quantitatively evaluate code hallucinations. By evaluating 17 popular LLMs using this benchmark, we reveal significant differences in their accuracy and reliability in code generation, offering detailed insights for further improving the code generation capabilities of LLMs. The CodeHalu benchmark and code are publicly available at https://github.com/yuchen814/CodeHalu.

Kaixin Li, Yuchen Tian, Qisheng Hu et al.

Programming often involves converting detailed and complex specifications into code, a process during which developers typically utilize visual aids to more effectively convey concepts. While recent developments in Large Multimodal Models have demonstrated remarkable abilities in visual reasoning and mathematical tasks, there is little work on investigating whether these models can effectively interpret visual elements for code generation. To this end, we present MMCode, the first multi-modal coding dataset for evaluating algorithmic problem-solving skills in visually rich contexts. MMCode contains 3,548 questions and 6,620 images collected from real-world programming challenges harvested from 10 code competition websites, presenting significant challenges due to the extreme demand for reasoning abilities. Our experiment results show that current state-of-the-art models struggle to solve these problems. The results highlight the lack of powerful vision-code models, and we hope MMCode can serve as an inspiration for future works in this domain. The data and code are publicly available at https://github.com/likaixin2000/MMCode.

Shidong Cao, Hongzhan Lin, Yuxuan Gu et al.

Chain-of-Thought (CoT) reasoning improves multi-step mathematical problem solving in large language models but remains vulnerable to exposure bias and error accumulation, as early mistakes propagate irreversibly through autoregressive decoding. In this work, we propose DiffCoT, a diffusion-styled CoT framework that reformulates CoT reasoning as an iterative denoising process. DiffCoT integrates diffusion principles at the reasoning-step level via a sliding-window mechanism, enabling unified generation and retrospective correction of intermediate steps while preserving token-level autoregression. To maintain causal consistency, we further introduce a causal diffusion noise schedule that respects the temporal structure of reasoning chains. Extensive experiments on three multi-step CoT reasoning benchmarks across diverse model backbones demonstrate that DiffCoT consistently outperforms existing CoT preference optimization methods, yielding improved robustness and error-correction capability in CoT reasoning.

Long Li, Shuichen Ji, Ziyang Luo et al.

We present the first unified framework that jointly handles three operationally heterogeneous saliency tasks, eg, SOD, CoSOD, and SIS, by casting each as a Chain-of-Thought (CoT) reasoning process in a Vision-Language Model (VLM) to bridge task heterogeneity. CoT training follows a two-stage paradigm: Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL). To enhance CoT quality in RL, we propose Confidence-Guided Policy Optimization (CGPO), a lightweight single-sample algorithm that leverages the discrepancy between reward and model confidence as a per-sample advantage signal. This design naturally focuses updates on informative responses while eliminating group sampling, thereby addressing GRPO's key limitations: confidence-agnostic learning, signal dilution, and prohibitive computational overhead. We also introduce an "output-to-reasoning" strategy to construct high-fidelity SFT data that ensures logical consistency with ground-truth masks. Experiments show our model matches or outperforms specialized SOTA methods and strong closed-source VLMs across all tasks, especially achieving an S-measure of 0.899 on CoCA for CoSOD, surpassing the prior best by 8.0 percentage points, despite using far less training data.

Ziyang Luo, Zhiqi Shen, Wenzhuo Yang et al.

The Model Context Protocol has emerged as a transformative standard for connecting large language models to external data sources and tools, rapidly gaining adoption across major AI providers and development platforms. However, existing benchmarks are overly simplistic and fail to capture real application challenges such as long-horizon reasoning and large, unfamiliar tool spaces. To address this critical gap, we introduce MCP-Universe, the first comprehensive benchmark specifically designed to evaluate LLMs in realistic and hard tasks through interaction with real-world MCP servers. Our benchmark encompasses 6 core domains spanning 11 different MCP servers: Location Navigation, Repository Management, Financial Analysis, 3D Design, Browser Automation, and Web Searching. To ensure rigorous evaluation, we implement execution-based evaluators, including format evaluators for agent format compliance, static evaluators for time-invariant content matching, and dynamic evaluators that automatically retrieve real-time ground truth for temporally sensitive tasks. Through extensive evaluation of leading LLMs, we find that even SOTA models such as GPT-5 (43.72%), Grok-4 (33.33%) and Claude-4.0-Sonnet (29.44%) exhibit significant performance limitations. In addition, our benchmark poses a significant long-context challenge for LLM agents, as the number of input tokens increases rapidly with the number of interaction steps. Moreover, it introduces an unknown-tools challenge, as LLM agents often lack familiarity with the precise usage of the MCP servers. Notably, enterprise-level agents like Cursor cannot achieve better performance than standard ReAct frameworks. Beyond evaluation, we open-source our extensible evaluation framework with UI support, enabling researchers and practitioners to seamlessly integrate new agents and MCP servers while fostering innovation in the rapidly evolving MCP ecosystem.

Taishi Nakamura, Mayank Mishra, Simone Tedeschi et al. · ibm-research, stanford

Pretrained language models are an integral part of AI applications, but their high computational cost for training limits accessibility. Initiatives such as Bloom and StarCoder aim to democratize access to pretrained models for collaborative community development. Despite these efforts, such models encounter challenges such as limited multilingual capabilities, risks of catastrophic forgetting during continual pretraining, and the high costs of training models from scratch, alongside the need to align with AI safety standards and regulatory frameworks. This paper presents Aurora-M, a 15B parameter multilingual open-source model trained on English, Finnish, Hindi, Japanese, Vietnamese, and code. Continually pretrained from StarCoderPlus on 435B additional tokens, Aurora-M surpasses 2T tokens in total training token count. It is the first open-source multilingual model fine-tuned on human-reviewed safety instructions, thus aligning its development not only with conventional red-teaming considerations, but also with the specific concerns articulated in the Biden-Harris Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence. We evaluate Aurora-M across a wide range of tasks and languages, showcasing its robustness against catastrophic forgetting and its superior performance in multilingual settings, particularly in safety evaluations. We open-source Aurora-M and its variants to encourage responsible open-source development of large language models at https://huggingface.co/aurora-m.

Zixin Chen, Hongzhan Lin, Kaixin Li et al.

The proliferation of multimodal memes in the social media era demands that multimodal Large Language Models (mLLMs) effectively understand meme harmfulness. Existing benchmarks for assessing mLLMs on harmful meme understanding rely on accuracy-based, model-agnostic evaluations using static datasets. These benchmarks are limited in their ability to provide up-to-date and thorough assessments, as online memes evolve dynamically. To address this, we propose AdamMeme, a flexible, agent-based evaluation framework that adaptively probes the reasoning capabilities of mLLMs in deciphering meme harmfulness. Through multi-agent collaboration, AdamMeme provides comprehensive evaluations by iteratively updating the meme data with challenging samples, thereby exposing specific limitations in how mLLMs interpret harmfulness. Extensive experiments show that our framework systematically reveals the varying performance of different target mLLMs, offering in-depth, fine-grained analyses of model-specific weaknesses. Our code is available at https://github.com/Lbotirx/AdamMeme.

Ziyang Luo, Nian Liu, Junwei Han

Omni-modal Large Language Models (Omni-MLLMs) promise a unified integration of diverse sensory streams. However, recent evaluations reveal a critical performance paradox: unimodal baselines frequently outperform joint multimodal inference. We trace this perceptual fragility to the static fusion topologies universally employed by current models, identifying two structural pathologies: positional bias in sequential inputs and alignment traps in interleaved formats, which systematically distort attention regardless of task semantics. To resolve this functional rigidity, we propose Chain of Modality (CoM), an agentic framework that transitions multimodal fusion from passive concatenation to dynamic orchestration. CoM adaptively orchestrates input topologies, switching among parallel, sequential, and interleaved pathways to neutralize structural biases. Furthermore, CoM bifurcates cognitive execution into two task-aligned pathways: a streamlined ``Direct-Decide'' path for direct perception and a structured ``Reason-Decide'' path for analytical auditing. Operating in either a training-free or a data-efficient SFT setting, CoM achieves robust and consistent generalization across diverse benchmarks.

Rao Fu, Ziyang Luo, Hongzhan Lin et al.

Recent advancements in large multimodal models (LMMs) have showcased impressive code generation capabilities, primarily evaluated through image-to-code benchmarks. However, these benchmarks are limited to specific visual programming scenarios where the logic reasoning and the multimodal understanding capacities are split apart. To fill this gap, we propose ScratchEval, a novel benchmark designed to evaluate the visual programming reasoning ability of LMMs. ScratchEval is based on Scratch, a block-based visual programming language widely used in children's programming education. By integrating visual elements and embedded programming logic, ScratchEval requires the model to process both visual information and code structure, thereby comprehensively evaluating its programming intent understanding ability. Our evaluation approach goes beyond the traditional image-to-code mapping and focuses on unified logical thinking and problem-solving abilities, providing a more comprehensive and challenging framework for evaluating the visual programming ability of LMMs. ScratchEval not only fills the gap in existing evaluation methods, but also provides new insights for the future development of LMMs in the field of visual programming. Our benchmark can be accessed at https://github.com/HKBUNLP/ScratchEval .

Xin Chen, Junchao Wu, Shu Yang et al.

Detecting content generated by large language models (LLMs) is crucial for preventing misuse and building trustworthy AI systems. Although existing detection methods perform well, their robustness in out-of-distribution (OOD) scenarios is still lacking. In this paper, we hypothesize that, compared to features used by existing detection methods, the internal representations of LLMs contain more comprehensive and raw features that can more effectively capture and distinguish the statistical pattern differences between LLM-generated texts (LGT) and human-written texts (HWT). We validated this hypothesis across different LLMs and observed significant differences in neural activation patterns when processing these two types of texts. Based on this, we propose RepreGuard, an efficient statistics-based detection method. Specifically, we first employ a surrogate model to collect representation of LGT and HWT, and extract the distinct activation feature that can better identify LGT. We can classify the text by calculating the projection score of the text representations along this feature direction and comparing with a precomputed threshold. Experimental results show that RepreGuard outperforms all baselines with average 94.92% AUROC on both in-distribution (ID) and OOD scenarios, while also demonstrating robust resilience to various text sizes and mainstream attacks. Data and code are publicly available at: https://github.com/NLP2CT/RepreGuard

Jinghao Luo, Yuchen Tian, Chuxue Cao et al.

Large Language Model (LLM)-based agents have fundamentally reshaped artificial intelligence by integrating external tools and planning capabilities. While memory mechanisms have emerged as the architectural cornerstone of these systems, current research remains fragmented, oscillating between operating system engineering and cognitive science. This theoretical divide prevents a unified view of technological synthesis and a coherent evolutionary perspective. To bridge this gap, this survey proposes a novel evolutionary framework for LLM agent memory mechanisms, formalizing the development process into three stages: Storage (trajectory preservation), Reflection (trajectory refinement), and Experience (trajectory abstraction). We first formally define these three stages before analyzing the three core drivers of this evolution: the necessity for long-range consistency, the challenges in dynamic environments, and the ultimate goal of continual learning. Furthermore, we specifically explore two transformative mechanisms in the frontier Experience stage: proactive exploration and cross-trajectory abstraction. By synthesizing these disparate views, this work offers robust design principles and a clear roadmap for the development of next-generation LLM agents.

Shengkang Wang, Hongzhan Lin, Ziyang Luo et al.

Large vision-language models (LVLMs) have significantly improved multimodal reasoning tasks, such as visual question answering and image captioning. These models embed multimodal facts within their parameters, rather than relying on external knowledge bases to store factual information explicitly. However, the content discerned by LVLMs may deviate from factuality due to inherent bias or incorrect inference. To address this issue, we introduce MFC-Bench, a rigorous and comprehensive benchmark designed to evaluate the factual accuracy of LVLMs across three stages of verdict prediction for MFC: Manipulation, Out-of-Context, and Veracity Classification. Through our evaluation on MFC-Bench, we benchmarked a dozen diverse and representative LVLMs, uncovering that current models still fall short in multimodal fact-checking and demonstrate insensitivity to various forms of manipulated content. We hope that MFC-Bench could raise attention to the trustworthy AI potentially assisted by LVLMs in the future. The MFC-Bench and accompanying resources are publicly accessible at https://github.com/wskbest/MFC-Bench, contributing to ongoing research in the multimodal fact-checking field.

Zesen Cheng, Sicong Leng, Hang Zhang et al.

In this paper, we present the VideoLLaMA 2, a set of Video Large Language Models (Video-LLMs) designed to enhance spatial-temporal modeling and audio understanding in video and audio-oriented tasks. Building upon its predecessor, VideoLLaMA 2 incorporates a tailor-made Spatial-Temporal Convolution (STC) connector, which effectively captures the intricate spatial and temporal dynamics of video data. Additionally, we integrate an Audio Branch into the model through joint training, thereby enriching the multimodal understanding capabilities of the model by seamlessly incorporating audio cues. Comprehensive evaluations on multiple-choice video question answering (MC-VQA), open-ended video question answering (OE-VQA), and video captioning (VC) tasks demonstrate that VideoLLaMA 2 consistently achieves competitive results among open-source models and even gets close to some proprietary models on several benchmarks. Furthermore, VideoLLaMA 2 exhibits reasonable improvements in audio-only and audio-video question-answering (AQA & OE-AVQA) benchmarks over existing models. These advancements underline VideoLLaMA 2's superior performance in multimodal comprehension, setting a new standard for intelligent video analysis systems. All models are public to facilitate further research.

Ziyang Luo, Can Xu, Pu Zhao et al.

Large Language Models (LLMs) have significantly advanced natural language processing (NLP) with their impressive language understanding and generation capabilities. However, their performance may be suboptimal for domain-specific tasks that require specialized knowledge due to limited exposure to the related data. Additionally, the lack of transparency of most state-of-the-art (SOTA) LLMs, which can only be accessed via APIs, impedes further fine-tuning with domain custom data. Moreover, providing private data to the LLMs' owner leads to data privacy problems. To address these challenges, we propose the novel Parametric Knowledge Guiding (PKG) framework, which equips LLMs with a knowledge-guiding module to access relevant knowledge without altering the LLMs' parameters. Our PKG is based on open-source "white-box" language models, allowing offline memory of any knowledge that LLMs require. We demonstrate that our PKG framework can enhance the performance of "black-box" LLMs on a range of domain knowledge-intensive tasks that require factual (+7.9%), tabular (+11.9%), medical (+3.0%), and multimodal (+8.1%) knowledge.

Kaixin Li, Ziyang Meng, Hongzhan Lin et al.

Recent advancements in Multi-modal Large Language Models (MLLMs) have led to significant progress in developing GUI agents for general tasks such as web browsing and mobile phone use. However, their application in professional domains remains under-explored. These specialized workflows introduce unique challenges for GUI perception models, including high-resolution displays, smaller target sizes, and complex environments. In this paper, we introduce ScreenSpot-Pro, a new benchmark designed to rigorously evaluate the grounding capabilities of MLLMs in high-resolution professional settings. The benchmark comprises authentic high-resolution images from a variety of professional domains with expert annotations. It spans 23 applications across five industries and three operating systems. Existing GUI grounding models perform poorly on this dataset, with the best model achieving only 18.9%. Our experiments reveal that strategically reducing the search area enhances accuracy. Based on this insight, we propose ScreenSeekeR, a visual search method that utilizes the GUI knowledge of a strong planner to guide a cascaded search, achieving state-of-the-art performance with 48.1% without any additional training. We hope that our benchmark and findings will advance the development of GUI agents for professional applications. Code, data and leaderboard can be found at https://gui-agent.github.io/grounding-leaderboard.

Yuhao Yang, Yue Wang, Dongxu Li et al.

Digital agents for automating tasks across different platforms by directly manipulating the GUIs are increasingly important. For these agents, grounding from language instructions to target elements remains a significant challenge due to reliance on HTML or AXTree inputs. In this paper, we introduce Aria-UI, a large multimodal model specifically designed for GUI grounding. Aria-UI adopts a pure-vision approach, eschewing reliance on auxiliary inputs. To adapt to heterogeneous planning instructions, we propose a scalable data pipeline that synthesizes diverse and high-quality instruction samples for grounding. To handle dynamic contexts in task performing, Aria-UI incorporates textual and text-image interleaved action histories, enabling robust context-aware reasoning for grounding. Aria-UI sets new state-of-the-art results across offline and online agent benchmarks, outperforming both vision-only and AXTree-reliant baselines. We release all training data and model checkpoints to foster further research at https://ariaui.github.io.

Inception Labs, Samar Khanna, Siddhant Kharbanda et al. · deepmind, microsoft-research

We present Mercury, a new generation of commercial-scale large language models (LLMs) based on diffusion. These models are parameterized via the Transformer architecture and trained to predict multiple tokens in parallel. In this report, we detail Mercury Coder, our first set of diffusion LLMs designed for coding applications. Currently, Mercury Coder comes in two sizes: Mini and Small. These models set a new state-of-the-art on the speed-quality frontier. Based on independent evaluations conducted by Artificial Analysis, Mercury Coder Mini and Mercury Coder Small achieve state-of-the-art throughputs of 1109 tokens/sec and 737 tokens/sec, respectively, on NVIDIA H100 GPUs and outperform speed-optimized frontier models by up to 10x on average while maintaining comparable quality. We discuss additional results on a variety of code benchmarks spanning multiple languages and use-cases as well as real-world validation by developers on Copilot Arena, where the model currently ranks second on quality and is the fastest model overall. We also release a public API at https://platform.inceptionlabs.ai/ and free playground at https://chat.inceptionlabs.ai

Hongzhan Lin, Ziyang Luo, Bo Wang et al.

The exponential growth of social media has profoundly transformed how information is created, disseminated, and absorbed, exceeding any precedent in the digital age. Regrettably, this explosion has also spawned a significant increase in the online abuse of memes. Evaluating the negative impact of memes is notably challenging, owing to their often subtle and implicit meanings, which are not directly conveyed through the overt text and image. In light of this, large multimodal models (LMMs) have emerged as a focal point of interest due to their remarkable capabilities in handling diverse multimodal tasks. In response to this development, our paper aims to thoroughly examine the capacity of various LMMs (e.g., GPT-4o) to discern and respond to the nuanced aspects of social abuse manifested in memes. We introduce the comprehensive meme benchmark, GOAT-Bench, comprising over 6K varied memes encapsulating themes such as implicit hate speech, sexism, and cyberbullying, etc. Utilizing GOAT-Bench, we delve into the ability of LMMs to accurately assess hatefulness, misogyny, offensiveness, sarcasm, and harmful content. Our extensive experiments across a range of LMMs reveal that current models still exhibit a deficiency in safety awareness, showing insensitivity to various forms of implicit abuse. We posit that this shortfall represents a critical impediment to the realization of safe artificial intelligence. The GOAT-Bench and accompanying resources are publicly accessible at https://goatlmm.github.io/, contributing to ongoing research in this vital field.

Yan Yang, Dongxu Li, Yutong Dai et al.

Graphical user interface (GUI) agents autonomously complete tasks across platforms (\eg, Linux) by sequentially decomposing user instructions into action proposals that iteratively interact with visual elements in the evolving environment. However, two main challenges arise: i) planning (\ie, the action proposal sequence) under expansive action space, where selecting an appropriate plan is non-trivial, as many valid ones may exist; ii) accurately grounding actions in complex and high-resolution interfaces, \ie, precisely interacting with visual targets. This paper investigates the aforementioned challenges with our \textbf{G}UI \textbf{T}est-time Scaling \textbf{A}gent, namely GTA1. First, we conduct test-time scaling to select the most appropriate action proposal: at each step, multiple candidate proposals are sampled and evaluated and selected by a judge model. It trades off computation for better decision quality by concurrent sampling. Second, we propose a model that improves grounding of the selected action proposals to its corresponding visual elements. Our key insight is that reinforcement learning (RL) facilitates grounding through inherent objective alignments, rewarding successful clicks on interface elements. Experimentally, GTA1 achieves state-of-the-art performance on both grounding and agent task execution benchmarks. The code and models are released here.

Hongzhan Lin, Ziyang Luo, Jing Ma et al.

The age of social media is rife with memes. Understanding and detecting harmful memes pose a significant challenge due to their implicit meaning that is not explicitly conveyed through the surface text and image. However, existing harmful meme detection approaches only recognize superficial harm-indicative signals in an end-to-end classification manner but ignore in-depth cognition of the meme text and image. In this paper, we attempt to detect harmful memes based on advanced reasoning over the interplay of multimodal information in memes. Inspired by the success of Large Language Models (LLMs) on complex reasoning, we first conduct abductive reasoning with LLMs. Then we propose a novel generative framework to learn reasonable thoughts from LLMs for better multimodal fusion and lightweight fine-tuning, which consists of two training stages: 1) Distill multimodal reasoning knowledge from LLMs; and 2) Fine-tune the generative framework to infer harmfulness. Extensive experiments conducted on three meme datasets demonstrate that our proposed approach achieves superior performance than state-of-the-art methods on the harmful meme detection task.

Jianzhao Huang, Hongzhan Lin, Ziyan Liu et al.

The proliferation of Internet memes in the age of social media necessitates effective identification of harmful ones. Due to the dynamic nature of memes, existing data-driven models may struggle in low-resource scenarios where only a few labeled examples are available. In this paper, we propose an agency-driven framework for low-resource harmful meme detection, employing both outward and inward analysis with few-shot annotated samples. Inspired by the powerful capacity of Large Multimodal Models (LMMs) on multimodal reasoning, we first retrieve relative memes with annotations to leverage label information as auxiliary signals for the LMM agent. Then, we elicit knowledge-revising behavior within the LMM agent to derive well-generalized insights into meme harmfulness. By combining these strategies, our approach enables dialectical reasoning over intricate and implicit harm-indicative patterns. Extensive experiments conducted on three meme datasets demonstrate that our proposed approach achieves superior performance than state-of-the-art methods on the low-resource harmful meme detection task.

Ziyang Luo, Haoning Wu, Dongxu Li et al.

Large multimodal models (LMMs) with advanced video analysis capabilities have recently garnered significant attention. However, most evaluations rely on traditional methods like multiple-choice questions in benchmarks such as VideoMME and LongVideoBench, which are prone to lack the depth needed to capture the complex demands of real-world users. To address this limitation-and due to the prohibitive cost and slow pace of human annotation for video tasks-we introduce VideoAutoArena, an arena-style benchmark inspired by LMSYS Chatbot Arena's framework, designed to automatically assess LMMs' video analysis abilities. VideoAutoArena utilizes user simulation to generate open-ended, adaptive questions that rigorously assess model performance in video understanding. The benchmark features an automated, scalable evaluation framework, incorporating a modified ELO Rating System for fair and continuous comparisons across multiple LMMs. To validate our automated judging system, we construct a 'gold standard' using a carefully curated subset of human annotations, demonstrating that our arena strongly aligns with human judgment while maintaining scalability. Additionally, we introduce a fault-driven evolution strategy, progressively increasing question complexity to push models toward handling more challenging video analysis scenarios. Experimental results demonstrate that VideoAutoArena effectively differentiates among state-of-the-art LMMs, providing insights into model strengths and areas for improvement. To further streamline our evaluation, we introduce VideoAutoBench as an auxiliary benchmark, where human annotators label winners in a subset of VideoAutoArena battles. We use GPT-4o as a judge to compare responses against these human-validated answers. Together, VideoAutoArena and VideoAutoBench offer a cost-effective, and scalable framework for evaluating LMMs in user-centric video analysis.

Hongzhan Lin, Zixin Chen, Ziyang Luo et al.

Social media abounds with multimodal sarcasm, and identifying sarcasm targets is particularly challenging due to the implicit incongruity not directly evident in the text and image modalities. Current methods for Multimodal Sarcasm Target Identification (MSTI) predominantly focus on superficial indicators in an end-to-end manner, overlooking the nuanced understanding of multimodal sarcasm conveyed through both the text and image. This paper proposes a versatile MSTI framework with a coarse-to-fine paradigm, by augmenting sarcasm explainability with reasoning and pre-training knowledge. Inspired by the powerful capacity of Large Multimodal Models (LMMs) on multimodal reasoning, we first engage LMMs to generate competing rationales for coarser-grained pre-training of a small language model on multimodal sarcasm detection. We then propose fine-tuning the model for finer-grained sarcasm target identification. Our framework is thus empowered to adeptly unveil the intricate targets within multimodal sarcasm and mitigate the negative impact posed by potential noise inherently in LMMs. Experimental results demonstrate that our model far outperforms state-of-the-art MSTI methods, and markedly exhibits explainability in deciphering sarcasm as well.

Chuyi Kong, Ziyang Luo, Hongzhan Lin et al.

The advanced role-playing capabilities of Large Language Models (LLMs) have enabled rich interactive scenarios, yet existing research in social interactions neglects hallucination while struggling with poor generalizability and implicit character fidelity judgments. To bridge this gap, motivated by human behaviour, we introduce a generalizable and explicit paradigm for uncovering interactive patterns of LLMs across diverse worldviews. Specifically, we first define interactive hallucination through stance transfer, then construct SHARP, a benchmark built by extracting relations from commonsense knowledge graphs and utilizing LLMs' inherent hallucination properties to simulate multi-role interactions. Extensive experiments confirm our paradigm's effectiveness and stability, examine the factors that influence these metrics, and challenge conventional hallucination mitigation solutions. More broadly, our work reveals a fundamental limitation in popular post-training methods for role-playing LLMs: the tendency to obscure knowledge beneath style, resulting in monotonous yet human-like behaviors - interactive hallucination.

Ziyang Luo, Nian Liu, Xuguang Yang et al.

Audio-Visual Segmentation (AVS) faces a fundamental challenge of effectively aligning audio and visual modalities. While recent approaches leverage foundation models to address data scarcity, they often rely on single-modality knowledge or combine foundation models in an off-the-shelf manner, failing to address the cross-modal alignment challenge. In this paper, we present TAViS, a novel framework that \textbf{couples} the knowledge of multimodal foundation models (ImageBind) for cross-modal alignment and a segmentation foundation model (SAM2) for precise segmentation. However, effectively combining these models poses two key challenges: the difficulty in transferring the knowledge between SAM2 and ImageBind due to their different feature spaces, and the insufficiency of using only segmentation loss for supervision. To address these challenges, we introduce a text-bridged design with two key components: (1) a text-bridged hybrid prompting mechanism where pseudo text provides class prototype information while retaining modality-specific details from both audio and visual inputs, and (2) an alignment supervision strategy that leverages text as a bridge to align shared semantic concepts within audio-visual modalities. Our approach achieves superior performance on single-source, multi-source, semantic datasets, and excels in zero-shot settings.

Yuchen Tian, Ruiyuan Huang, Xuanwu Wang et al.

Large Language Models (LLMs) for formal theorem proving have shown significant promise, yet they often lack generalizability and are fragile to even minor transformations of problem statements. To address this limitation, we introduce a novel data augmentation pipeline designed to enhance model robustness from two perspectives: symmetry and difficulty. From the symmetry perspective, we propose two complementary methods: EvolAST, an Abstract Syntax Tree (AST) based approach that targets syntactic symmetry to generate semantically equivalent problem variants, and EvolDomain, which leverages LLMs to address semantic symmetry by translating theorems across mathematical domains. From the difficulty perspective, we propose EvolDifficulty, which uses carefully designed evolutionary instructions to guide LLMs in generating new theorems with a wider range of difficulty. We then use the evolved data to train EvolProver, a 7B-parameter non-reasoning theorem prover. EvolProver establishes a new state-of-the-art (SOTA) on FormalMATH-Lite with a 53.8% pass@32 rate, surpassing all models of comparable size, including reasoning-based models. It also sets new SOTA records for non-reasoning models on MiniF2F-Test (69.8% pass@32), Ineq-Comp-Seed (52.2% pass@32), and Ineq-Comp-Transformed (34.0% pass@32). Ablation studies further confirm our data augmentation pipeline's effectiveness across multiple benchmarks.

Yuchen Tian, Kaixin Li, Hao Chen et al.

Large Language Models (LLMs) have recently demonstrated strong capabilities in translating natural language into database queries, especially when dealing with complex graph-structured data. However, real-world queries often contain inherent ambiguities, and the interconnected nature of graph structures can amplify these challenges, leading to unintended or incorrect query results. To systematically evaluate LLMs on this front, we propose a taxonomy of graph-query ambiguities, comprising three primary types: Attribute Ambiguity, Relationship Ambiguity, and Attribute-Relationship Ambiguity, each subdivided into Same-Entity and Cross-Entity scenarios. We introduce AmbiGraph-Eval, a novel benchmark of real-world ambiguous queries paired with expert-verified graph query answers. Evaluating 9 representative LLMs shows that even top models struggle with ambiguous graph queries. Our findings reveal a critical gap in ambiguity handling and motivate future work on specialized resolution techniques.

Ziyang Luo, Nian Liu, Fahad Shahbaz Khan et al.

Reference Audio-Visual Segmentation (Ref-AVS) tasks challenge models to precisely locate sounding objects by integrating visual, auditory, and textual cues. Existing methods often lack genuine semantic understanding, tending to memorize fixed reasoning patterns. Furthermore, jointly training for reasoning and segmentation can compromise pixel-level precision. To address these issues, we introduce AURORA, a novel framework designed to enhance genuine reasoning and language comprehension in reference audio-visual segmentation. We employ a structured Chain-of-Thought (CoT) prompting mechanism to guide the model through a step-by-step reasoning process and introduce a novel segmentation feature distillation loss to effectively integrate these reasoning abilities without sacrificing segmentation performance. To further cultivate the model's genuine reasoning capabilities, we devise a further two-stage training strategy: first, a ``corrective reflective-style training" stage utilizes self-correction to enhance the quality of reasoning paths, followed by reinforcement learning via Group Reward Policy Optimization (GRPO) to bolster robustness in challenging scenarios. Experiments demonstrate that AURORA achieves state-of-the-art performance on Ref-AVS benchmarks and generalizes effectively to unreferenced segmentation.

Junran Guo, Tonglin Mu, Keyuan Li et al.

Detecting small objects, such as drones, over long distances presents a significant challenge with broad implications for security, surveillance, environmental monitoring, and autonomous systems. Traditional imaging-based methods rely on high-resolution image acquisition, but are often constrained by range, power consumption, and cost. In contrast, data-driven single-photon-single-pixel light detection and ranging (\text{D\textsuperscript{2}SP\textsuperscript{2}-LiDAR}) provides an imaging-free alternative, directly enabling target identification while reducing system complexity and cost. However, its detection range has been limited to a few hundred meters. Here, we introduce a novel integration of residual neural networks (ResNet) with \text{D\textsuperscript{2}SP\textsuperscript{2}-LiDAR}, incorporating a refined observation model to extend the detection range to 5~\si{\kilo\meter} in an intracity environment while enabling high-accuracy identification of drone poses and types. Experimental results demonstrate that our approach not only outperforms conventional imaging-based recognition systems, but also achieves 94.93\% pose identification accuracy and 97.99\% type classification accuracy, even under weak signal conditions with long distances and low signal-to-noise ratios (SNRs). These findings highlight the potential of imaging-free methods for robust long-range detection of small targets in real-world scenarios.

Yuxi Sun, Wei Gao, Jing Ma et al.

With the rise and widespread use of Large Language Models (LLMs), ensuring their safety is crucial to prevent harm to humans and promote ethical behaviors. However, directly assessing value valence (i.e., support or oppose) by leveraging large-scale data training is untrustworthy and inexplainable. We assume that emulating humans to rely on social norms to make moral decisions can help LLMs understand and predict moral judgment. However, capturing human values remains a challenge, as multiple related norms might conflict in specific contexts. Consider norms that are upheld by the majority and promote the well-being of society are more likely to be accepted and widely adopted (e.g., "don't cheat,"). Therefore, it is essential for LLM to identify the appropriate norms for a given scenario before making moral decisions. To this end, we introduce a novel moral judgment approach called \textit{ClarityEthic} that leverages LLMs' reasoning ability and contrastive learning to uncover relevant social norms for human actions from different perspectives and select the most reliable one to enhance judgment accuracy. Extensive experiments demonstrate that our method outperforms state-of-the-art approaches in moral judgment tasks. Moreover, human evaluations confirm that the generated social norms provide plausible explanations that support the judgments. This suggests that modeling human moral judgment with the emulating humans moral strategy is promising for improving the ethical behaviors of LLMs.

Hongzhan Lin, Ziyang Luo, Wei Gao et al.

The age of social media is flooded with Internet memes, necessitating a clear grasp and effective identification of harmful ones. This task presents a significant challenge due to the implicit meaning embedded in memes, which is not explicitly conveyed through the surface text and image. However, existing harmful meme detection methods do not present readable explanations that unveil such implicit meaning to support their detection decisions. In this paper, we propose an explainable approach to detect harmful memes, achieved through reasoning over conflicting rationales from both harmless and harmful positions. Specifically, inspired by the powerful capacity of Large Language Models (LLMs) on text generation and reasoning, we first elicit multimodal debate between LLMs to generate the explanations derived from the contradictory arguments. Then we propose to fine-tune a small language model as the debate judge for harmfulness inference, to facilitate multimodal fusion between the harmfulness rationales and the intrinsic multimodal information within memes. In this way, our model is empowered to perform dialectical reasoning over intricate and implicit harm-indicative patterns, utilizing multimodal explanations originating from both harmless and harmful arguments. Extensive experiments on three public meme datasets demonstrate that our harmful meme detection approach achieves much better performance than state-of-the-art methods and exhibits a superior capacity for explaining the meme harmfulness of the model predictions.

Ziyang Luo, Zhipeng Hu, Yadong Xi et al.

Image Captioning is a traditional vision-and-language task that aims to generate the language description of an image. Recent studies focus on scaling up the model size and the number of training data, which significantly increase the cost of model training. Different to these heavy-cost models, we introduce a lightweight image captioning framework (I-Tuning), which contains a small number of trainable parameters. We design a novel I-Tuning cross-attention module to connect the non-trainable pre-trained language decoder GPT2 and vision encoder CLIP-ViT. Since most parameters are not required to be updated during training, our framework is lightweight and fast. Experimental results conducted on three image captioning benchmarks reveal that our framework achieves comparable or better performance than the large-scale baseline systems. But our models contain up to 10 times fewer trainable parameters and require much fewer data for training compared with state-of-the-art baselines.

Ziyang Luo, Yadong Xi, Rongsheng Zhang et al.

Image Captioning is a fundamental task to join vision and language, concerning about cross-modal understanding and text generation. Recent years witness the emerging attention on image captioning. Most of existing works follow a traditional two-stage training paradigm. Before training the captioning models, an extra object detector is utilized to recognize the objects in the image at first. However, they require sizeable datasets with fine-grained object annotation for training the object detector, which is a daunting task. In addition, the errors of the object detectors are easy to propagate to the following captioning models, degenerating models' performance. To alleviate such defects, we propose a frustratingly simple but highly effective end-to-end image captioning framework, Visual Conditioned GPT (VC-GPT), by connecting the pre-trained visual encoder (CLIP-ViT) and language decoder (GPT2). Different from the vanilla connection method that directly inserts the cross-attention modules into GPT2, we come up with a self-ensemble cross-modal fusion mechanism that comprehensively considers both the single- and cross-modal knowledge. As a result, we do not need extra object detectors for model training. Experimental results conducted on three popular image captioning benchmarks (MSCOCO, Flickr30k and NoCaps) demonstrate that our VC-GPT achieves either the best or the second-best performance across all evaluation metrics over extensive baseline systems.

Meichun Jiao, Ziyang Luo

Gender bias in word embeddings gradually becomes a vivid research field in recent years. Most studies in this field aim at measurement and debiasing methods with English as the target language. This paper investigates gender bias in static word embeddings from a unique perspective, Chinese adjectives. By training word representations with different models, the gender bias behind the vectors of adjectives is assessed. Through a comparison between the produced results and a human-scored data set, we demonstrate how gender bias encoded in word embeddings differentiates from people's attitudes.

Ziyang Luo

With the success of pre-trained language models in recent years, more and more researchers focus on opening the "black box" of these models. Following this interest, we carry out a qualitative and quantitative analysis of constituency grammar in attention heads of BERT and RoBERTa. We employ the syntactic distance method to extract implicit constituency grammar from the attention weights of each head. Our results show that there exist heads that can induce some grammar types much better than baselines, suggesting that some heads act as a proxy for constituency grammar. We also analyze how attention heads' constituency grammar inducing (CGI) ability changes after fine-tuning with two kinds of tasks, including sentence meaning similarity (SMS) tasks and natural language inference (NLI) tasks. Our results suggest that SMS tasks decrease the average CGI ability of upper layers, while NLI tasks increase it. Lastly, we investigate the connections between CGI ability and natural language understanding ability on QQP and MNLI tasks.

Ziyang Luo, Artur Kulmizev, Xiaoxi Mao

In this work, we demonstrate that the contextualized word vectors derived from pretrained masked language model-based encoders share a common, perhaps undesirable pattern across layers. Namely, we find cases of persistent outlier neurons within BERT and RoBERTa's hidden state vectors that consistently bear the smallest or largest values in said vectors. In an attempt to investigate the source of this information, we introduce a neuron-level analysis method, which reveals that the outliers are closely related to information captured by positional embeddings. We also pre-train the RoBERTa-base models from scratch and find that the outliers disappear without using positional embeddings. These outliers, we find, are the major cause of anisotropy of encoders' raw vector spaces, and clipping them leads to increased similarity across vectors. We demonstrate this in practice by showing that clipped vectors can more accurately distinguish word senses, as well as lead to better sentence embeddings when mean pooling. In three supervised tasks, we find that clipping does not affect the performance.