Papers

While Multimodal Large Language Models (MLLMs) excel on general visual tasks, their capacity to comprehend complex dynamic systems remains a critical open question. Such systems, governed by physical laws, explicit rules, and multi-agent interactions, form the fabric of the real world. To facilitate a systematic diagnosis of current MLLM limitations, we introduce BasketVision, a new benchmark that leverages professional basketball as a microcosm for these dynamic environments. BasketVision probes model capabilities across seven dimensions—spanning perception, reasoning, and prediction—through 6,000 curated, bilingual questions from professional game data. An automated data generation pipeline underpins the benchmark, ensuring both scalability and fine-grained precision. Our evaluation of 23 leading models reveals a chasm between machine and human cognition: human experts attain 96.34% accuracy, while the premier model, GPT-4o, achieves only 63.15%. The analysis pinpoints spatial reasoning as a persistent bottleneck and uncovers specific patterns of task specialization. BasketVision thus serves as a crucial apparatus for charting the frontiers of MLLMs and steering future work toward more robust reasoning in dynamic visual worlds.

Materials define the eras of human civilization, yet the design of novel materials is fundamentally constrained by the immense chemical space, which renders traditional enumeration-screening methodology computationally prohibitive and inefficient. This paper introduces a paradigm shift towards insight-exploration-validation, enabling an intelligent and evolutionary exploration of material design pathways. To actualize this paradigm, we propose MatEvolve, a synergistic symbolic–LLM agent that reconceptualizes material design as a closed-loop, programmatic evolution task. Central to MatEvolve is a novel symbolic formalism, Material Edit Language, which empowers the agent to programmatically take chemical operations. The exploration trajectory is directed by a multifaceted guidance strategy, comprising a dynamic knowledge injection mechanism and a two-stage exploration strategy that balances broad exploration and deep optimization. Furthermore, a multi-objective fitness landscape ensures directional and efficient navigational guidance. These integrated strategies contribute to a 32.2% improvement over direct material structure modification. Crucially, comparisons demonstrate that our insight-exploration-validation paradigm outperforms the traditional enumeration-screening approach by 33.6%, highlighting its superior efficacy in navigating vast design spaces.

Endoscopic video super-resolution (EVSR) seeks to reconstruct high-resolution frames from low-resolution endoscopic video, a task critical for enhancing clinical visualization of fine anatomical details. However, EVSR is uniquely challenging due to rapid camera motion, non-rigid tissue deformation, specular highlights, and frequent occlusions, which undermine the effectiveness of both conventional CNN-based and transformer-based models. To address these issues, we propose a novel EVSR framework that leverages the Receptance Weighted Key Value (RWKV) architecture for efficient long-range temporal modeling. To further adapt to the highly non-stationary and diverse content of endoscopic scenes, we introduce a Dynamic Group-wise Shift mechanism that adaptively composes spatial kernels based on local appearance and motion, enabling robust implicit alignment and detail restoration without explicit motion estimation. Our approach integrates these innovations into both temporal and spatial modules, achieving a strong balance between global context modeling and local adaptability. Extensive experiments on a synthetic endoscopic video dataset demonstrate that our method achieves consistently strong performance, maintaining small yet stable advantages over recent CNN- and transformer-based baselines in quantitative comparisons.

Endoscopic video super-resolution (EVSR) seeks to reconstruct high-resolution frames from low-resolution endoscopic video, a task critical for enhancing clinical visualization of fine anatomical details. However, EVSR is uniquely challenging due to rapid camera motion, non-rigid tissue deformation, specular highlights, and frequent occlusions, which undermine the effectiveness of both conventional CNN-based and transformer-based models. To address these issues, we propose a novel EVSR framework that leverages the Receptance Weighted Key Value (RWKV) architecture for efficient long-range temporal modeling. To further adapt to the highly non-stationary and diverse content of endoscopic scenes, we introduce a Dynamic Group-wise Shift mechanism that adaptively composes spatial kernels based on local appearance and motion, enabling robust implicit alignment and detail restoration without explicit motion estimation. Our approach integrates these innovations into both temporal and spatial modules, achieving a strong balance between global context modeling and local adaptability. Extensive experiments on a synthetic endoscopic video dataset demonstrate that our method achieves consistently strong performance, maintaining small yet stable advantages over recent CNN- and transformer-based baselines in quantitative comparisons.

Artificial intelligence (AI) has demonstrated impressive progress in mathematical reasoning, yet its integration into the practice of mathematical research remains limited. In this study, we investigate how the AI Mathematician (AIM) system can operate as a research partner rather than a mere problem solver. Focusing on a challenging problem in homogenization theory, we analyze the autonomous reasoning trajectories of AIM and incorporate targeted human interventions to structure the discovery process. Through iterative decomposition of the problem into tractable subgoals, selection of appropriate analytical methods, and validation of intermediate results, we reveal how human intuition and machine computation can complement one another. This collaborative paradigm enhances the reliability, transparency, and interpretability of the resulting proofs, while retaining human oversight for formal rigor and correctness. The approach leads to a complete and verifiable proof, and more broadly, demonstrates how systematic human-AI co-reasoning can advance the frontier of mathematical discovery.

The escalating volume of scientific literature necessitates efficient methods for identifying foundational works that significantly inform new research. This paper addresses the Paper Source Tracing (PST) problem, which aims to quantify the influence of cited references on a focal paper, assigning importance weights to its most salient sources. To this end, we propose a novel multi-agent ensemble architecture for PST, integrating Deepseek-R1-250528, GPT-5-2025-08-07, and Gemini-2.5-pro. Our system employs a robust pipeline, featuring advanced XML parsing, empirically optimized prompt engineering with counterfactual reasoning and multi-role Socratic dialogue, and a sophisticated multi-agent integration strat- egy. This strategy utilizes weighted model predictions, intelligent default scoring, and a consistency penalty mechanism to derive precise source paper identifica- tions. Our method becomes a strong tuning-free baseline for the PST problem that does not require feature engineering. Our method also achieves top-ranked results when combined with feature engineering techinques. This work highlights the efficacy of multi-agent ensembles and advanced prompt engineering for com- plex academic information tracing tasks.

This paper addresses the critical issue of enhancing peer review quality at AI conferences by implementing anonymized feedback and adaptive reward systems. The growing volume of conference submissions and limited reviewer accountability result in inconsistent review quality, bias, and a lack of transparency, posing significant challenges to the integrity of AI research. Our proposed solution involves a dynamic feedback loop that anonymizes and aggregates feedback to minimize biases, coupled with an adaptive reward system to motivate reviewers while preserving the integrity of the review process. Utilizing sentiment analysis, feedback is processed to detect and mitigate potential biases, enhancing the fairness and efficacy of peer reviews. Experiments conducted using a logistic regression model on the Yelp Polarity dataset demonstrate a significant improvement in sentiment classification accuracy, from 54.1\% to 83.4\%, indicating the effectiveness of our anonymized feedback loop. However, the bias detection score of 0.0 across all runs highlights the need for further refinement in bias mitigation. Our method's scalability and adaptability across various conference settings are supported by its successful implementation in sentiment analysis tasks. Overall, this study provides a robust framework for enhancing the accountability and quality of peer reviews, with implications for future research aimed at integrating advanced bias detection and mitigation techniques.

The field of generative modeling in computer vision has been propelled significantly forward by methods such as Generative Adversarial Networks (GANs) and diffusion models; however, challenges like balancing image fidelity and diversity alongside incorporating class-specific details persist. These traditional approaches often exhibit limitations in adaptability and computational efficiency. This paper introduces Reinforced Adaptive Diffusion Networks (RAD-Nets), a novel generative framework that synergizes diffusion processes with reinforcement learning to enhance image synthesis through dynamic parameter optimization. The core innovation lies in integrating a Reinforced Learning Layer and an Adaptive Feedback Mechanism, which employ real-time feedback to iteratively refine outputs. The Multi-Objective Optimization module within RAD-Nets specifically targets the concurrent enhancement of image quality, diversity, and class fidelity, addressing the issues found in static optimization techniques. Empirical evaluations demonstrate that RAD-Nets outperform existing generative models on standard benchmarks like CIFAR-10 and CelebA, achieving superior metrics in quality and diversity without compromising fidelity. By focusing on class-conditional image synthesis, RAD-Nets also demonstrate significant improvements in class-specific feature representation, marking a substantial advancement over conventional generative modeling frameworks.

The research focuses on the development of a novel hierarchical and adaptive recommendation system that addresses the dual challenge of personalization and fairness in cross-domain environments. Traditional recommendation systems have struggled to effectively integrate diverse user interactions and adapt to rapidly evolving user preferences while maintaining fairness. The proposed solution leverages three core innovations: cross-domain collaborative filtering, meta-reinforcement learning, and fairness-aware mechanisms. By synthesizing data from multiple domains, the system constructs enriched user profiles that inform a meta-reinforcement learning framework, enhancing adaptability to user behavior changes. Additionally, fairness-aware mechanisms are incorporated to mitigate biases and ensure equitable content distribution. This integrated approach aims to resolve key challenges in recommendation systems, namely the precise prediction of preferences and the equitable treatment of diverse user groups. Empirical evaluations demonstrate that the proposed methodology not only improves recommendation accuracy but also enhances fairness metrics, thereby fostering a balanced and inclusive recommendation landscape.

This paper addresses challenges in unsupervised representation learning, particularly in high-fidelity image generation and domain adaptability across diverse data modalities. Current frameworks such as GANs and VQ-VAE have shown promise but face limitations in maintaining consistent performance across variable data distributions without significant supervision. To overcome these challenges, we propose a Multi-Modal Vector Quantized Variational AutoEncoder (VQ-VAE) integrated with Self-Supervised Learning (SSL). Our innovative approach incorporates a harmonizer module within the VQ-VAE architecture, which aligns and transforms data representations across multiple modalities. By leveraging self-supervised learning techniques, the model iteratively refines its parameters, enhancing both image reconstruction quality and adaptability to new domains with minimal supervision. The proposed framework processes CIFAR-10 datasets to facilitate structured data integration, employing advanced standardization and batching techniques for optimal performance. Empirical evaluations reveal substantial improvements in image reconstruction fidelity and domain adaptability compared to standard VQ-VAE models, corroborated by metrics such as PSNR, SSIM, and FID. The seamless integration of modality-specific feature extraction and embedding generalization within our framework demonstrates the potential to advance unsupervised learning paradigms. Our contribution establishes a robust solution, optimizing the generative process, and expanding applicability in real-world scenarios characterized by unlabeled, multi-modal datasets.

In recent years, personalized recommendation systems have become integral to enhancing user experiences on digital platforms, yet challenges remain in effectively integrating causal inference with adaptive learning mechanisms and semantic alignment. Traditional systems predominantly rely on correlation-based models, often overlooking the dynamic causal relationships within user interaction data that could enhance recommendation precision and contextual relevance. This paper addresses these gaps by presenting a novel framework that synergizes causal inference using structural equation models and causal diagrams, adaptive learning algorithms via a refined hybrid multi-armed bandit strategy, and semantic content mapping with advanced natural language processing techniques such as Latent Dirichlet Allocation and BERT-based embeddings. Through this integrated approach, our method dynamically adjusts recommendations to align with user preferences and adapt to context changes. Empirical evaluation demonstrates our method's superiority in achieving higher accuracy and relevance in personalized content delivery compared to existing models. The findings underscore the potential of our framework to significantly improve recommendation cohesion and user satisfaction, marking a substantial advancement in the field of contextual personalization.

In the domain of neural architectures for generative models, the emergence of diffusion processes and flow-based transformations has revolutionized image synthesis, traditionally dominated by Generative Adversarial Networks and Variational Autoencoders. These novel techniques have been pivotal in enhancing image fidelity and stability, fundamental for robust image generation tasks. The Adaptive Diffusion-Latent Flow Model (ADLFM) addresses the challenges of scalability and parameter optimization inherent in high-dimensional generative frameworks by integrating diffusion processes with invertible flow-based transformations. This hybrid model enhances fidelity and stability by harnessing adaptive and adversarial mechanisms. ADLFM's architecture leverages innovative invertible latent flow transformations to ensure reversibility and structural coherence in latent spaces, while an Adaptive Diffusion Network refines latent features through context-adaptive noise scheduling. To enrich output diversity and robustness, an Adversarial Regularization Structure mitigates mode collapse through competitive generator-discriminator dynamics. Empirical evaluations reveal a substantial improvement in inception scores, indicating enhanced image synthesis quality with limited data resources. Furthermore, the model's synergistic integration of adaptive and adversarial strategies leads to a significant reduction in synthesis errors, maintaining high fidelity in generated images. These findings underscore the potential of ADLFM as a formidable engine for high-quality image synthesis, effectively addressing the complexities of diverse generative scenarios.

In the field of generative modeling, generating high-fidelity class-conditional images remains challenging despite advancements in methodologies. Traditional approaches such as Generative Adversarial Networks, variational autoencoders, and diffusion models have improved image synthesis but still face limitations in efficiency and adaptability, especially when deploying flow-based models. This paper presents a novel Unified Generative Framework with Dynamic Adaptation, which integrates flow-based and diffusion models enhanced by reinforcement learning to address these challenges. The proposed framework consists of five key components: Flow-Diffusion Integration, Reinforced Adaptive Learning, Multi-Scale Processing, Conditional Generation, and Dynamic Resource Management. Together, these components enable dynamic parameter adjustments, efficient resource use, and the generation of class-specific images with structural coherence across various scales. Our results, validated on the CIFAR-10 dataset, demonstrate significant improvements in image fidelity and diversity, establishing a new standard for scalable class-conditional image generation. The framework showcases the successful combination of deterministic and stochastic modeling techniques, providing an adaptive solution for real-time applications and highlighting the potential for broader deployment across diverse datasets.

This paper addresses the challenge of handling incomplete high-dimensional datasets, a significant issue in domains such as healthcare and finance where missing data undermines predictive accuracy. Current methods struggle with datasets exhibiting over 30\% missing values, especially when missingness is non-random and complex. To tackle this, we propose a hybrid approach that combines variational methods with importance weighting, introducing a dynamic weighting strategy that adjusts according to data complexity and missingness patterns. This strategy is implemented through an alternating algorithm that balances variational updates with importance weight recalibrations, maintaining computational efficiency while capturing diverse missingness mechanisms. Our experimental evaluation, conducted on the IMDb dataset using a shallow MLP model, demonstrates that our method significantly outperforms traditional techniques, achieving validation accuracies up to 84.65\% with corresponding F1 scores of 0.8505. These results confirm the robustness and adaptability of our approach, showcasing its potential to improve score matching performance on incomplete high-dimensional data. Our contributions include the development of a flexible latent variable model and a novel dynamic weighting strategy, offering a scalable solution applicable to critical sectors like healthcare and finance.

Large Language Models (LLMs) typically operate as passive responders, limiting their effectiveness in multi-turn interactions where users have complex, evolving intents. This research introduces COLLAB LLM, a novel training framework that leverages a collaborative simulation to estimate the long-term impact of responses through Multiturn-aware Rewards (MR). By applying reinforcement learning with these rewards, COLLAB LLM encourages active intent discovery and insight- ful suggestions from the model, thereby transforming the nature of user-LLM interactions. We propose a multiturn interaction benchmark that includes three challenging tasks, such as collaborative document creation. Preliminary results indicate that COLLAB LLM outperforms traditional models with an average of 18.5% higher task performance and 46.3% improved interactivity, as rated by LLM judges. Furthermore, a large user study with 201 participants revealed an increase in user satisfaction by 17.6% and a reduction in time spent by 10.4%. This research aims to pave the way for more engaging and efficient AI-driven conversations.

As machine learning models are increasingly deployed in critical applications, the need for reliable uncertainty quantification becomes paramount. Traditional conformal prediction methods provide distribution-free guarantees but often lack the flexibility to accommodate varying user risk preferences. This paper intro- duces an innovative framework that merges Bayesian quadrature with conformal prediction, allowing for the incorporation of user-specified risk preferences into uncertainty estimates. By modeling the posterior distribution of potential losses and adapting prediction sets based on individual risk thresholds, this approach en- hances the relevance and utility of uncertainty quantification in practical scenar- ios. Through empirical validation across multiple datasets, we demonstrate that the proposed method achieves lower failure rates and more informative prediction intervals compared to standard conformal prediction techniques.

This proposal investigates the integration of deep learning techniques with score matching to address the challenge of missing data in high-dimensional settings. Current methodologies primarily focus on traditional statistical approaches, leav- ing a significant gap in exploring the potential of neural networks in this con- text. We propose a novel framework that combines score matching with generative deep learning models, allowing for the effective estimation of score functions even when data is partially missing. Our approach not only leverages the capacity of deep learning to capture complex patterns but also provides robust performance across various datasets. We will validate the framework through a series of ex- periments involving both real and synthetic datasets, emphasizing applications in healthcare and social sciences. By doing so, we aim to push the boundaries of score matching methods and enhance their applicability in practical scenarios.

In high-stakes domains such as climate science and epidemiology, achieving robust uncertainty quantification (UQ) in spatio-temporal models is crucial due to the significant impact on public safety and resource management. Existing frequentist and Bayesian approaches often fall short in capturing the complex uncertainties inherent in high-dimensional, dynamic environments. This paper introduces a novel Bayesian Quadrature-Conformal Prediction framework that integrates the probabilistic richness of Bayesian quadrature with the distribution-free guarantees of conformal prediction, aiming to enhance both accuracy and interpretability of UQ. Our method employs hierarchical Bayesian modeling and advanced sampling techniques such as Hamiltonian Monte Carlo and variational inference to address the computational challenges posed by Bayesian approaches, ensuring efficiency without compromising accuracy. Empirical evaluation on the MNIST dataset demonstrates significant improvements in Conformal Prediction Error Rates across multiple runs, evidencing our framework's capability to provide more nuanced and reliable uncertainty estimates compared to traditional methods. This work sets a new benchmark for uncertainty quantification in spatio-temporal models, promising advancements in predictive accuracy and decision-making for critical applications.

This paper addresses the computational challenges inherent in training Masked Diffusion Models (MDMs) for discrete generative tasks, which are crucial for applications like game development and biomedical modeling. The importance of this research lies in the need for efficient and scalable generative models across various AI applications. However, MDMs face significant difficulties due to computationally intractable subproblems that limit scalability, coupled with the challenge of optimizing the decoding process in non-causally ordered tasks without sacrificing performance. We propose a dual-pronged solution: an optimization framework using batch sampling to reduce the computational complexity during training and an adaptive learning mechanism that dynamically adjusts the decoding order during inference. This approach improves both training efficiency and inference flexibility. Our experimental evaluation on the MNIST dataset demonstrates a notable improvement in performance, achieving an average accuracy of 95.53\% and maintaining an average inference time of 7.39 seconds, surpassing the performance of traditional autoregressive models. These results validate that our method significantly reduces computational overhead while maintaining high accuracy, setting a new benchmark for MDMs in discrete generative tasks. The contributions of this study include the introduction of innovative optimization techniques and a comprehensive framework that enhances MDM applicability with fewer parameters and increased efficiency.