Archive, newest first. Filter by topic, impact and company →
This paper presents DiffUNet^2, a conditional diffusion model for bidirectional prediction and interactive visual analytics in scientific data exploration.
This paper introduces High-Confidence Causally Aligned Training (HICAT) to enhance adversarial training by addressing overfitting to spurious correlations.
This paper explores knowledge editing in masked diffusion models, revealing performance discrepancies compared to autoregressive models and proposing a corrective method.
This paper introduces a contrastive learning framework for approximate k-coloring in graphs, enhancing generalization across varying graph sizes and distributions.
GARDEN introduces a novel RGB-only framework for reconstructing 3D environments by leveraging gravity as a physical prior for scene factorization.
This paper introduces the Visual STAte Tracking benchmark (VSTAT) to evaluate visual state tracking capabilities in Multimodal Large Language Models (MLLMs).
This paper presents a model for predicting the diffusion of scientific concepts, focusing on quantum computing, using a co-occurrence network and LightGBM.
This paper introduces Hedge-Bench, a benchmark for evaluating AI agents on complex financial reasoning tasks, addressing gaps in existing evaluation methods.
This paper introduces continuous-time Adam optimization for analog Ising machines, enhancing performance on optimization benchmarks like Max-Cut.
This paper introduces PatchScene, a diffusion-based framework for large-scale LiDAR scene completion, enhancing geometric accuracy and temporal consistency.
This paper introduces Bootstrap Your Generator (ByG), a novel framework for unpaired visual editing using flow matching, enhancing scalability in generative models.
This paper introduces NetKV, a network-aware scheduling method that optimizes Time to First Token in disaggregated LLM inference by considering network costs.
SparseStreet introduces a novel compression framework for 3D Gaussian Splatting, optimizing street scene simulation by reducing Gaussian primitives while maintaining fidelity.
This paper presents a study protocol to investigate how configuration mechanisms in agentic AI coding tools influence build-versus-buy decisions.
This paper introduces scTranslation, a benchmark for evaluating single-cell multi-omics modality translation, addressing gaps in systematic evaluation.
This paper introduces MAdam, a novel optimizer that addresses systematic gaps in multi-objective optimization when using Adam.
This paper presents a novel attention-based denoising model for diffusion-weighted imaging, addressing noise challenges in low-acquisition-time scans.
This paper elucidates the theoretical role of momentum in Muon, demonstrating its function as a spectral filter that enhances gradient updates.
This paper introduces Agent libOS, a runtime for long-running LLM agents that emphasizes capability control and auditing in agent operations.
This paper introduces PROVE, a framework for training LLMs in multi-step tool use through realistic environments and innovative reward mechanisms.
RealClawBench introduces a benchmark framework derived from real developer-agent sessions, enhancing realism in evaluating agent capabilities.
CoralBay introduces a self-supervised learning framework for 3D CT imaging, enhancing representation learning for medical tasks through a novel architecture.
This paper introduces Reveal-IG, a novel path-based attribution method that enhances feature importance explanations by utilizing structured probe distributions.
This paper introduces a benchmark for evaluating the reasoning structures of large reasoning models, revealing insights beyond traditional accuracy metrics.
This paper introduces a novel framework for evaluating encoder roles in multi-encoder vision-language models, revealing insights into optimal configurations.
This paper presents a decision-theoretic framework for incrementality measurement in advertising systems under privacy-induced signal degradation.
This paper presents a novel approach for generating narrative summaries from multi-modal tracking data for remote family members of older adults using LLMs.
This paper presents a method for visual instruction tuning that enhances multimodal integration in Large Language Models by embedding visual features in intermediate layers.
This paper introduces a training-free mixture-of-agents framework for multi-document summarization, leveraging LLMs and knowledge graphs to enhance performance.
This paper introduces Taiji, a novel framework for optimizing LLM-enhanced recommendation systems by addressing semantic-ID alignment and reward trade-offs.
