As AI assistants serve millions of users daily, evaluating user experience (UX) beyond general model capability has become increasingly important. We present \textbf{UXBench}, the first user-centric benchmark grounded in real user feedback signals for evaluating preference alignment and dialogue generation. The benchmark consists of three interconnected tasks, UX Judge, UX Eval, and UX Recovery, with 7,400 test instances extracted from over 70K interaction logs of a mainstream Chinese AI assistant. The dataset closely reflects real user distributions, covering 8 scenarios, 83 domains, and diverse failure patterns that pose severe challenges. Extensive experiments on 26 frontier language models provide novel insights into how well models perceive user experience and how improvements in model capability contribute to better dialogue engagement. Through analyses of model behavior and performance gaps, we document six important findings, demonstrating that user feedback prediction is a learnable capability and revealing different aspects that influence user experience. UXBench establishes a new evaluation landscape and calls for greater attention to tailored UX optimization, contributing toward a user-centric scaling law for the development of successful AI assistants. The full project is released at https://github.com/mengze-hong/UXBench.
LLM agents increasingly support research planning and execution, where judgments about current and anticipated research activity inform decisions. Yet whether they can infer changes across a whole field from the partial evidence they retrieve through online scholarly search remains unclear. We introduce Research Attention Prediction (RAP), a literature search benchmark for evaluating this ability. We build a local search environment from 356,357 dated arXiv papers and construct two complementary evaluation tasks. The benchmark comprises a Core set of 2,160 questions across 54 AI/ML fields and an Extended set covering 269 fields. Our evaluation and analysis of LLM agents show that search substantially improves stronger agents' recognition of recent changes, while forecasting gains remain limited or negative. Agents update forecasts across dates and adjust their investigations to forecasting goals, but these adaptations do not reliably yield accurate predictions of future shifts. Even for recent changes, final judgments can diverge from what the retrieved papers indicate. These findings show that agents can recognise some recent changes, yet reliably anticipating future research activity remains difficult despite changes in their answers and investigations.
Synthetic image attribution aims at identifying the generator responsible for a given AI-generated image. Training-free reference-based attribution methods are easily scalable, since newly emerging generators can be incorporated by adding source-specific references rather than retraining a task-specific classifier. Their performance depends on two coupled factors: the representation space used for comparison and the way source-specific references are constructed. However, the interaction between these two factors remains largely unexplored. In this paper, we provide a controlled analysis of this interaction using references and off-the-shelf pretrained representations. We study representations extracted from different layers of CLIP and DINOv2, along with three reference selection strategies with varying semantic constraints: arbitrary, semantically aligned, and resynthesis-based references. Our results show that attribution accuracy generally peaks at intermediate representation levels, indicating that source-discriminative cues are more accessible before strong semantic abstraction dominates. We further show that intermediate representations are not completely semantically neutral, making reference selection critical: semantically constrained references reduce query-reference mismatch and improve attribution, especially under limited reference budgets. Resynthesis is useful in low-reference regimes, while semantically aligned references provide a better accuracy-cost trade-off when a moderate-sized reference pool is available. Our findings show that training-free reference-based attribution should be understood as the interaction between where images are compared, how the reference set is constructed, and how many references are available. Code and datasets are available at https://github.com/meiling-fdu/RefPrint.
As large language models (LLMs) advance, concerns about their misconduct in complex social contexts intensify. Existing research overlooked the systematic understanding and assessment of their criminal capability in realistic interactions. We propose a unified framework PRISON, to quantify LLMs' criminal potential across five dimensions: False Statements, Frame-Up, Psychological Manipulation, Emotional Disguise, and Moral Disengagement. Using structured crime scenarios adapted from classic films, we evaluate both criminal potential and anti-crime ability of LLMs via role-play. Results show that state-of-the-art LLMs frequently exhibit emergent criminal tendencies, such as proposing misleading statements or evasion tactics, even without explicit instructions. Moreover, when placed in a detective role, models recognize deceptive behavior with only 41% accuracy on average, revealing a striking mismatch between conducting and detecting criminal behavior. These findings underscore the urgent need for adversarial robustness, behavioral alignment, and safety mechanisms before broader LLM deployment.
Simulating collisions of deformable objects is a fundamental yet challenging task due to the complexity of modeling solid mechanics and multi-body interactions. Existing data-driven methods often suffer from lack of equivariance to physical symmetries, inadequate handling of collisions, and limited scalability. Here we introduce EqCollide, the first end-to-end equivariant neural fields simulator for deformable objects and their collisions. We propose an equivariant encoder to map object geometry and velocity into latent control points. A subsequent equivariant Graph Neural Network-based Neural Ordinary Differential Equation models the interactions among control points via collision-aware message passing. To reconstruct velocity fields, we query a neural field conditioned on control point features, enabling continuous and resolution-independent motion predictions. Experimental results show that EqCollide achieves accurate, stable, and scalable simulations across diverse object configurations, and our model achieves 24.34% to 35.82% lower rollout MSE even compared with the best-performing baseline model. Furthermore, our model could generalize to more colliding objects and extended temporal horizons, and stay robust to input transformed with group action.
Large language models (LLMs) are increasingly expected to tackle complex tasks, driven by their expanding applications and users' growing proficiency in crafting sophisticated prompts. However, as the number of explicitly stated requirements increases (particularly more than 10 constraints), LLMs often struggle to accurately follow such complex instructions. To address this challenge, we propose RECAST, a novel framework for synthesizing datasets where each example incorporates far more constraints than those in existing benchmarks. These constraints are extracted from real-world prompt-response pairs to ensure practical relevance. RECAST enables automatic verification of constraint satisfaction via rule-based validators for quantitative constraints and LLM-based validators for qualitative ones. Using this framework, we construct RECAST-30K, a large-scale, high-quality dataset comprising 30k instances spanning 15 constraint types. Experimental results demonstrate that models fine-tuned on RECAST-30K show substantial improvements in following complex instructions. Moreover, the verifiability provided by RECAST enables the design of reward functions for reinforcement learning, which further boosts model performance on complex and challenging tasks.
Physics-informed neural networks (PINNs), owing to their mesh-free nature, offer a powerful approach for solving high-dimensional partial differential equations (PDEs) in complex geometries, including irregular domains. This capability effectively circumvents the challenges of mesh generation that traditional numerical methods face in high-dimensional or geometrically intricate settings. While recent studies have extended PINNs to manifolds, the theoretical foundations remain scarce. Existing theoretical analyses of PINNs in Euclidean space often rely on smoothness assumptions for the solutions. However, recent empirical evidence indicates that PINNs may struggle to approximate solutions with low regularity, such as those arising from nonlinear hyperbolic equations. In this paper, we develop a framework for PINNs tailored to the efficient approximation of weak solutions, particularly nonlinear hyperbolic equations defined on manifolds. We introduce a novel weak PINN (wPINN) formulation on manifolds that leverages the well-posedness theory to approximate entropy solutions of geometry-compatible hyperbolic conservation laws on manifolds. Employing tools from approximation theory, we establish a convergence analysis of the algorithm, including an analysis of approximation errors for time-dependent entropy solutions. This analysis provides insight into the accumulation of approximation errors over long time horizons. Notably, the network complexity depends only on the intrinsic dimension, independent of the ambient space dimension. Our results match the minimax rate in the d-dimensional Euclidean space, demonstrating that PINNs can alleviate the curse of dimensionality in the context of low-dimensional manifolds. Finally, we validate the performance of the proposed wPINN framework through numerical experiments, confirming its ability to efficiently approximate entropy solutions on manifolds.
Recent image-to-video generation methods have demonstrated success in enabling control over one or two visual elements, such as camera motion or object motion. However, these methods are unable to offer control over multiple visual elements due to limitations in data and network efficacy. In this paper, we introduce VidCRAFT3, a novel framework for precise image-to-video generation that enables control over camera motion, object motion, and lighting direction simultaneously. VidCRAFT3 integrates three core components: Image2Cloud generates 3D point cloud from a reference image; ObjMotionNet encodes sparse object trajectories using multi-scale optical flow features; and Spatial Triple-Attention Transformer incorporates lighting direction embeddings via parallel cross-attention modules. Additionally, we introduce the VideoLightingDirection dataset, providing synthetic yet realistic video clips with accurate per-frame lighting direction annotations, effectively mitigating the lack of annotated real-world datasets. We further adopt a three-stage training strategy, ensuring robust learning even without joint multi-element annotations. Extensive experiments show that VidCRAFT3 produces high-quality video content, outperforming state-of-the-art methods in control granularity and visual coherence. Code and data will be publicly available.
3D vision and spatial reasoning have long been recognized as preferable for accurately perceiving our three-dimensional world, especially when compared with traditional visual reasoning based on 2D images. Due to the difficulties in collecting high-quality 3D data, research in this area has only recently gained momentum. With the advent of powerful large language models (LLMs), multi-modal LLMs for 3D vision have been developed over the past few years. However, most of these models focus primarily on the vision encoder for 3D data. In this paper, we propose converting existing densely activated LLMs into mixture-of-experts (MoE) models, which have proven effective for multi-modal data processing. In addition to leveraging these models' instruction-following capabilities, we further enable embodied task planning by attaching a diffusion head, Pose-DiT, that employs a novel rectified flow diffusion scheduler. Experimental results on 3D question answering and task-planning tasks demonstrate that our 3D-MoE framework achieves improved performance with fewer activated parameters.
Large Multimodal Models (LMMs) have made significant breakthroughs with the advancement of instruction tuning. However, while existing models can understand images and videos at a holistic level, they still struggle with instance-level understanding that requires a more nuanced comprehension and alignment. Instance-level understanding is crucial, as it focuses on the specific elements that we are most interested in. Excitingly, existing works find that the state-of-the-art LMMs exhibit strong instance understanding capabilities when provided with explicit visual cues. Motivated by this, we introduce an automated annotation pipeline assisted by GPT-4o to extract instance-level information from images and videos through explicit visual prompting for instance guidance. Building upon this pipeline, we proposed Inst-IT, a solution to enhance LMMs in Instance understanding via explicit visual prompt Instruction Tuning. Inst-IT consists of a benchmark to diagnose multimodal instance-level understanding, a large-scale instruction-tuning dataset, and a continuous instruction-tuning training paradigm to effectively enhance spatial-temporal instance understanding capabilities of existing LMMs. Experimental results show that, with the boost of Inst-IT, our models not only achieve outstanding performance on Inst-IT Bench but also demonstrate significant improvements across various generic image and video understanding benchmarks. This highlights that our dataset not only boosts instance-level understanding but also strengthens the overall capabilities of generic image and video comprehension.
Skip connection is an essential ingredient for modern deep models to be deeper and more powerful. Despite their huge success in normal scenarios (state-of-the-art classification performance on natural examples), we investigate and identify an interesting property of skip connections under adversarial scenarios, namely, the use of skip connections allows easier generation of highly transferable adversarial examples. Specifically, in ResNet-like models (with skip connections), we find that using more gradients from the skip connections rather than the residual modules according to a decay factor during backpropagation allows one to craft adversarial examples with high transferability. The above method is termed as Skip Gradient Method (SGM). Although starting from ResNet-like models in vision domains, we further extend SGM to more advanced architectures, including Vision Transformers (ViTs) and models with length-varying paths and other domains, i.e. natural language processing. We conduct comprehensive transfer attacks against various models including ResNets, Transformers, Inceptions, Neural Architecture Search, and Large Language Models (LLMs). We show that employing SGM can greatly improve the transferability of crafted attacks in almost all cases. Furthermore, considering the big complexity for practical use, we further demonstrate that SGM can even improve the transferability on ensembles of models or targeted attacks and the stealthiness against current defenses. At last, we provide theoretical explanations and empirical insights on how SGM works. Our findings not only motivate new adversarial research into the architectural characteristics of models but also open up further challenges for secure model architecture design. Our code is available at https://github.com/mo666666/SGM.
Deep neural networks (DNNs) are known to be sensitive to adversarial input perturbations, leading to a reduction in either prediction accuracy or individual fairness. To jointly characterize the susceptibility of prediction accuracy and individual fairness to adversarial perturbations, we introduce a novel robustness definition termed robust accurate fairness. Informally, robust accurate fairness requires that predictions for an instance and its similar counterparts consistently align with the ground truth when subjected to input perturbations. We propose an adversarial attack approach dubbed RAFair to expose false or biased adversarial defects in DNN, which either deceive accuracy or compromise individual fairness. Then, we show that such adversarial instances can be effectively addressed by carefully designed benign perturbations, correcting their predictions to be accurate and fair. Our work explores the double-edged sword of input perturbations to robust accurate fairness in DNN and the potential of using benign perturbations to correct adversarial instances.
Building upon the impressive success of CLIP (Contrastive Language-Image Pretraining), recent pioneer works have proposed to adapt the powerful CLIP to video data, leading to efficient and effective video learners for open-vocabulary action recognition. Inspired by that humans perform actions in diverse environments, our work delves into an intriguing question: Can CLIP-based video learners effectively generalize to video domains they have not encountered during training? To answer this, we establish a CROSS-domain Open-Vocabulary Action recognition benchmark named XOV-Action, and conduct a comprehensive evaluation of five state-of-the-art CLIP-based video learners under various types of domain gaps. The evaluation demonstrates that previous methods exhibit limited action recognition performance in unseen video domains, revealing potential challenges of the cross-domain open-vocabulary action recognition task. In this paper, we focus on one critical challenge of the task, namely scene bias, and accordingly contribute a novel scene-aware video-text alignment method. Our key idea is to distinguish video representations apart from scene-encoded text representations, aiming to learn scene-agnostic video representations for recognizing actions across domains. Extensive experiments demonstrate the effectiveness of our method. The benchmark and code will be available at https://github.com/KunyuLin/XOV-Action/.
This paper considers stochastic first-order algorithms for minimax optimization under Polyak--{\L}ojasiewicz (PL) conditions. We propose SPIDER-GDA for solving the finite-sum problem of the form $\min_x \max_y f(x,y)\triangleq \frac{1}{n} \sum_{i=1}^n f_i(x,y)$, where the objective function $f(x,y)$ is $\mu_x$-PL in $x$ and $\mu_y$-PL in $y$; and each $f_i(x,y)$ is $L$-smooth. We prove SPIDER-GDA could find an $\epsilon$-optimal solution within ${\mathcal O}\left((n + \sqrt{n}\,\kappa_x\kappa_y^2)\log (1/\epsilon)\right)$ stochastic first-order oracle (SFO) complexity, which is better than the state-of-the-art method whose SFO upper bound is ${\mathcal O}\big((n + n^{2/3}\kappa_x\kappa_y^2)\log (1/\epsilon)\big)$, where $\kappa_x\triangleq L/\mu_x$ and $\kappa_y\triangleq L/\mu_y$. For the ill-conditioned case, we provide an accelerated algorithm to reduce the computational cost further. It achieves $\tilde{{\mathcal O}}\big((n+\sqrt{n}\,\kappa_x\kappa_y)\log^2 (1/\epsilon)\big)$ SFO upper bound when $\kappa_y \gtrsim \sqrt{n}$. Our ideas also can be applied to the more general setting that the objective function only satisfies PL condition for one variable. Numerical experiments validate the superiority of proposed methods.
Natural language instruction following is paramount to enable collaboration between artificial agents and human beings. Natural language-conditioned reinforcement learning (RL) agents have shown how natural languages' properties, such as compositionality, can provide a strong inductive bias to learn complex policies. Previous architectures like HIGhER combine the benefit of language-conditioning with Hindsight Experience Replay (HER) to deal with sparse rewards environments. Yet, like HER, HIGhER relies on an oracle predicate function to provide a feedback signal highlighting which linguistic description is valid for which state. This reliance on an oracle limits its application. Additionally, HIGhER only leverages the linguistic information contained in successful RL trajectories, thus hurting its final performance and data-efficiency. Without early successful trajectories, HIGhER is no better than DQN upon which it is built. In this paper, we propose the Emergent Textual Hindsight Experience Replay (ETHER) agent, which builds on HIGhER and addresses both of its limitations by means of (i) a discriminative visual referential game, commonly studied in the subfield of Emergent Communication (EC), used here as an unsupervised auxiliary task and (ii) a semantic grounding scheme to align the emergent language with the natural language of the instruction-following benchmark. We show that the referential game's agents make an artificial language emerge that is aligned with the natural-like language used to describe goals in the BabyAI benchmark and that it is expressive enough so as to also describe unsuccessful RL trajectories and thus provide feedback to the RL agent to leverage the linguistic, structured information contained in all trajectories. Our work shows that EC is a viable unsupervised auxiliary task for RL and provides missing pieces to make HER more widely applicable.
Large Language Models(LLMs) have demonstrated remarkable performance across various natural language processing tasks; however, how to comprehensively and accurately assess their performance becomes an urgent issue to be addressed. This paper introduces GAOKAO-Bench, an intuitive benchmark that employs questions from the Chinese GAOKAO examination as test samples, including both subjective and objective questions. To align with human examination methods, we design a method based on zero-shot settings to evaluate the performance of LLMs. With human evaluation, we obtain the converted total score of LLMs, including GPT-4, ChatGPT and ERNIE-Bot.Our findings reveal that LLMs have achieved competitive scores in Chinese GAOKAO examination, while they exhibit significant performance disparities across various subjects. We also use LLMs to grade the subjective questions, and find that model scores achieve a moderate level of consistency with human scores. In conclusion, this research contributes a robust evaluation benchmark for future large language models and offers valuable insights into the advantages and limitations of such models.
Transformer architectures have exhibited promising performance in various autonomous driving applications in recent years. On the other hand, its dedicated hardware acceleration on portable computational platforms has become the next critical step for practical deployment in real autonomous vehicles. This survey paper provides a comprehensive overview, benchmark, and analysis of Transformer-based models specifically tailored for autonomous driving tasks such as lane detection, segmentation, tracking, planning, and decision-making. We review different architectures for organizing Transformer inputs and outputs, such as encoder-decoder and encoder-only structures, and explore their respective advantages and disadvantages. Furthermore, we discuss Transformer-related operators and their hardware acceleration schemes in depth, taking into account key factors such as quantization and runtime. We specifically illustrate the operator level comparison between layers from convolutional neural network, Swin-Transformer, and Transformer with 4D encoder. The paper also highlights the challenges, trends, and current insights in Transformer-based models, addressing their hardware deployment and acceleration issues within the context of long-term autonomous driving applications.