Archive, newest first. Filter by topic, impact and company →
This paper introduces IntraShuffler, a middleware framework that enhances privacy in heterogeneous differential privacy federated learning by mitigating inference risks.
This paper introduces a method for certifying the safety of belief-space neural safety filters in interactive robotics, enhancing performance under uncertainty.
This paper introduces SubFit, a novel method for post-training compression of LLMs that operates at the submodule level, enhancing efficiency and performance.
This paper introduces HERO'S JOURNEY, a benchmark for evaluating rule induction in episodic tasks, revealing limitations in current LLM capabilities.
LongLive-RAG introduces a retrieval-augmented framework for autoregressive long video generation, mitigating error accumulation and identity drift.
This paper introduces MDA, a mixture-density approach for depth estimation that mitigates flying-point artifacts near object boundaries.
This paper introduces AFUN, an affordance foundation model that enhances functionality understanding for robot manipulation in diverse environments.
This paper introduces Script-Normalized WER (SN-WER), a novel metric for evaluating ASR systems in multilingual contexts with script variations.
This paper presents a novel three-stage machine learning approach for self-harm surveillance using emergency department triage notes, enhancing detection accuracy.
This paper introduces SimSD, a speculative decoding method for diffusion language models, enhancing decoding throughput while preserving generation quality.
This paper introduces SkillHarm, a benchmark for evaluating skill-based attacks in agents, highlighting vulnerabilities and proposing an automated attack construction pipeline.
This paper introduces a framework for quantifying agent traits through skill file diffs, achieving high accuracy in behavioral trajectory tracking.
This paper introduces LL-Bench, a benchmark for evaluating large-scale generative models on low-level vision tasks, and proposes LL-Score for improved quality assessment.
This paper presents a mask-conditioned latent diffusion model for augmenting TEM image datasets, enhancing defect detection and classification performance.
This paper introduces SafeSteer, a novel on-policy distillation method for aligning large language models with human safety values while preserving general capabilities.
This paper investigates biases in financial large language models (LLMs) towards Bitcoin, revealing internal representations that influence portfolio decisions.
This paper introduces Self-Adaptive Monotonic Normalization (SAMN), a hyperparameter-friendly method for improving long-tailed recognition in deep learning.
This paper introduces FigSIM, a novel dataset for analyzing suicide memes, focusing on severity levels and figurative language, to enhance content moderation.
This paper introduces Moment-Video, a benchmark for evaluating the temporal fidelity of video MLLMs in understanding momentary visual events.
This paper introduces Drifting Preference Optimization (DrPO), an innovative method for preference finetuning one-step text-to-image generators efficiently.
ToolFG introduces a novel MLLM-based framework for fine-grained image classification, integrating tool use for enhanced reliability and grounding.
This paper presents Optimal Mixture Transport (OMT), a scalable framework for optimal transport of mixture models with theoretical stability guarantees.
This paper explores the use of LLMs to extract ADHD-related signals from Turkish teacher narratives, revealing insights beyond traditional rating scales.
This paper reviews advancements in generative models and multimodal learning for inverse materials design, emphasizing closed-loop workflows and evaluation practices.
This paper introduces CRAM, a novel framework for Multimodal Continual Instruction Tuning that enhances parameter efficiency and mitigates catastrophic forgetting.
This paper introduces an LLM-agent framework for enhancing time series forecasting by integrating contextual business insights into the prediction process.
This paper introduces a triage methodology for monitoring agentic systems, focusing on structural defects that obscure task-level error detection.
This paper introduces FRANZ, a framework for auditing LLM response framing, highlighting the importance of communicative characteristics in subjective queries.
This paper investigates the limitations of congruence-based architectures for classifying symmetric positive-definite matrices, revealing expressivity constraints.
This paper presents RASER, a cost-effective routing mechanism for multi-hop question answering that optimizes retrieval actions based on recoverability.
