OMNI-PoseX provides real-time, accurate 6D object pose estimation for embodied robotic tasks, outperforming current solutions.

Create expressive 3D digital avatars for real-time interactive applications using SentiAvatar's framework.

Flash-Mono accelerates monocular SLAM with Gaussian splatting for real-time 3D scene reconstruction.

ExploreVLA integrates dense world modeling with RL for robust autonomous driving exploration.

An AI security tool to defend against logic-based attacks on Graph-based Retrieval Augmentation systems.

DocShield is an AI-powered forensic tool for identifying and explaining text-centric document forgeries.

Revolutionizing UAV navigation with a zero-shot thermal geo-localization system, SCC-Loc.

FSUNav provides a unified zero-shot goal-oriented navigation system for diverse robotic platforms ensuring fast, safe, and semantic-rich interaction with complex environments.

GenSmoke-GS offers a multi-stage solution for improving visibility and coherence in smoke-degraded image processing for enhanced 3D reconstructions.

Reduce multimodal AI inference costs by embedding structured text directly into images, achieving significant savings with competitive accuracy.

Generate realistic satellite imagery for wildfire detection using mask-conditioned diffusion models, addressing data scarcity with a novel inpainting approach.

A hierarchical, label-free concept bottleneck model that enhances interpretability and classification accuracy by mirroring human cognitive processes.

Opal provides private, scalable long-term memory for personal AI by decoupling data-dependent reasoning into a trusted enclave, improving retrieval accuracy and reducing costs.

An explainable vision-language model framework with adaptive loss for accurate and interpretable lumbar spinal stenosis diagnosis from MRI.

Develop street-legal, low-cost physical adversarial attacks to disrupt license plate reader systems, with potential applications in security and privacy.

VERTIGO optimizes cinematic camera trajectory generation by incorporating visual preference feedback, significantly improving framing and reducing off-screen characters.

Generate realistic synthetic data for spaceflight reinforcement learning using physics-informed generative models to overcome data scarcity and improve controller performance.

A reproducible pipeline to recover valuable transcription factor insights from incomplete single-cell perturbation data, enabling deeper biological discovery.

A framework for assessing and mitigating risks of assumption violations in time-series causal discovery, providing calibrated risk scores and abstention policies for reliable inference.

This research develops a hallucination-aware calibration method to improve the reliability and trustworthiness of vision-language models in medical question answering.

A novel computational linguistic metric and mitigation strategy to eliminate sycophancy in large language models.

Automate the unsustainable manual workload of autonomous system verification and validation using LLM agents to analyze and validate system anomalies against natural language requirements.

A unified 3D scene understanding model that leverages contrastive language-image pretraining for improved representation learning.

An LLM-powered agent that autonomously performs complex polymer molecular dynamics simulations from natural language, enabling faster and more accessible material property prediction.

An open-source benchmark and implementation of Matrix Profile methods for scalable and interpretable time-series anomaly detection.

A framework for rapidly deploying high-accuracy, on-device eye tracking by distilling visual foundation models using synthetic and real-world data.

Develops a prompt engineering technique to enable smaller, open-source LLMs to reliably verify mathematical proofs, matching the performance of frontier models.

A statistical framework for aligning large language models with human preferences, offering a robust approach to reward modeling and policy optimization.

A novel method for accurate building wireframe reconstruction from LiDAR data using Delaunay graphs, overcoming limitations in noisy and sparse environments.

A neuro-symbolic RAG architecture using competency questions as executable plans to generate factually accurate and auditable stories from cultural heritage knowledge graphs.

A framework for generating diverse and coordinated multi-agent plays from static formations, overcoming limitations of existing generative models.

Automates the conversion of complex clinical guidelines into executable decision support systems, significantly improving accuracy and auditability.

An adaptive multi-modal Kalman filter for improved state estimation in autonomous driving using sensor-specific learned noise characteristics.

A neuro-symbolic framework that enhances LLMs with structured reasoning capabilities for complex problem-solving, achieving significant performance gains on the ARC-AGI-2 benchmark.

A GeoAI framework leveraging vision foundation models for rapid building damage mapping from post-earthquake imagery, with publicly available code and data.

An LLM-powered system that translates mathematical theorems and proofs into executable code, enabling interactive exploration and deeper understanding for users.

This research provides a novel, comprehensive framework for auditing LLM bias that reveals systematic mischaracterizations by current alignment practices, offering a path to more robust safety and fairness evaluations.

A framework to significantly reduce inference costs for 3D Multimodal Large Language Models by adaptively pruning visual tokens, maintaining accuracy and enabling deployment on resource-constrained devices.

A novel SRAM architecture for LLM inference that drastically reduces computation by performing attention scores directly on quantized KV cache indices, with a gradient-free method to ensure stability.

A multi-agent AI system that improves the trustworthiness and accuracy of chest X-ray diagnoses by simulating a cooperative-competitive clinical workflow.

This research demonstrates a novel method to steer large language model behavior using function vectors, achieving high accuracy even when the model's internal representations are not decodable, suggesting a new paradigm for controlling LLM outputs.

A framework for systematically evaluating and mitigating socioeconomic bias in foundation models, addressing a critical gap in responsible AI.

A multi-agent reinforcement learning system that consistently beats human grandmasters in competitive programming.

A novel training method significantly reduces LLM bias towards spurious social contexts in high-stakes decision-making, improving both accuracy and robustness.

Develops a novel methodology and tailored LLM agents to significantly improve optimization solution quality through interactive conversations, bridging AI and operations research.

Accelerate LLM inference by up to 2.2x with a plug-and-play dequantization kernel that leverages shared memory, reducing costs on existing GPU infrastructure.

ContractShield is a multimodal framework that uses hierarchical cross-modal fusion to detect vulnerabilities in obfuscated smart contracts, outperforming state-of-the-art by 6-15%.

A new benchmark and evaluation protocol for assessing and improving belief revision capabilities in language models, crucial for dynamic environments.

Recovering precise 2D sewing patterns from 3D garment scans using a novel BoxMesh representation and a two-stage autoregressive model.

A principled and scalable method for diversity-aware retrieval in RAG, offering theoretical guarantees and significant speedups.

A novel dataset of vulnerable code snippets linked to CAPEC and CWE, generated by LLMs, to train automated vulnerability detection systems.

Develop human-safe soft robotic actuators with controllable shape and force using electroadhesive strain limiters and advanced modeling for applications in collaborative robotics.

A fully differentiable symbolic logic toolbox for robot manipulation that enables end-to-end learning and optimization of complex geometric and temporal constraints.

A versatile LLM for brain MRI that performs report generation, VQA, segmentation, and translation, outperforming specialized models.

An agentic framework that automates clinical trial discovery and eligibility-aware meta-analysis for precision medicine evidence synthesis.

A novel ontology-oriented approach to knowledge graph construction that decouples schema design from graph building, enabling reusable and adaptable knowledge representations for downstream AI tasks.

A benchmark and platform for evaluating and improving 3D spatial reasoning in code generation models.

A multi-sensor foundation model for Mars orbital applications, outperforming existing baselines on downstream tasks.

A vision-centric framework for efficient open-vocabulary object detection that decouples semantic reasoning from localization for practical deployment.

Accelerate vision-language model inference by 1.8x using weighted low-rank approximation and quantization, preserving accuracy.

A degeneracy-aware LiDAR-inertial odometry framework that uses a neural network to predict velocity and improve state estimation in challenging environments.

TrackerSplat enhances 3D Gaussian Splatting for dynamic scenes by using point tracking to improve robustness and throughput in reconstructions with fast object motion.

A neuro-symbolic framework for LLM agents that decouples semantic guidance from logical validation to improve long-horizon decision-making.

A lightweight module that significantly improves visual tracking performance by intelligently filtering noisy historical data and synthesizing dynamic temporal priors.

A physics-guided deep learning framework for accurate and efficient crowd simulation, outperforming state-of-the-art with reduced latency.

A new paradigm for small object detection that uses point prompts at inference time to significantly improve accuracy and generalize to unseen objects and datasets.

A probabilistic SLAM system using Gaussian Splatting that improves robustness and reduces drift by explicitly modeling uncertainty.

A geometrically-constrained radar-inertial odometry system that improves localization accuracy in challenging environments by dynamically modeling and integrating point and pose uncertainties.

FusionBERT enables robust multi-view image-to-3D model retrieval by adaptively fusing visual cues and enhancing 3D geometry encoding.

Leverage LLMs as judges for label-free knowledge distillation to significantly improve model reasoning capabilities on unlabeled data.

A new benchmark for evaluating LLM code generation across diverse industrial domains and programming languages, enabling more robust industrial AI solutions.

Leverage existing mmWave communication signals for high-fidelity, low-cost 3D environmental perception, overcoming limitations of optical sensors in adverse conditions.

THOM generates photorealistic, physically plausible 3D hand-object interactions from text, enhancing VR/AR experiences.

A vision-based end-to-end framework for autonomous drones to navigate complex, irregular gaps, enhancing inspection and rescue operations.

A new dataset and benchmark for multimodal LLMs in autonomous driving, along with a specialized MoE model, to improve reasoning across vehicle, infrastructure, and cooperative views.

SentinelAgent provides a formal framework and runtime protocol for verifiable delegation chains in multi-agent AI systems, ensuring policy compliance and forensic traceability.

A self-supervised depth estimation framework for articulated vehicles that leverages cross-vehicle geometric consistency to improve perception in complex robotic platforms.

A vision-language model extension that generates precise image masks from textual descriptions, with a new dataset for improved evaluation.

A privacy-preserving federated learning framework for early autism behavior analysis in children, enabling multi-site collaboration without centralizing sensitive data.

A system that significantly boosts LLM serving throughput by intelligently managing expert weights to free up GPU memory for critical runtime data.

A novel model for continual learning in multimodal knowledge graphs that prevents catastrophic forgetting and enhances new knowledge acquisition.

An LLM-powered agent framework that automates the verification of complex technical claims in scientific literature, even without domain expertise.

A novel neural network for robust cross-subject muscle fatigue detection using sEMG signals, enhancing physical rehabilitation.

A novel framework to detect and mitigate adversarial attacks that manipulate AI model privacy by making non-training data appear as if it was part of the training set.

A novel method to detect and explain knowledge gaps in LLMs by analyzing gradient dynamics, enabling more responsible deployment.

A novel token warping technique for multimodal LLMs to improve viewpoint robustness, outperforming existing methods on a new benchmark.

A novel network for person re-identification that handles both changes in clothing and camera modality, addressing a realistic surveillance challenge.

A simple and scalable supervised learning framework to detect money laundering transactions, outperforming state-of-the-art models and complementing existing AML systems.

A novel framework for camouflaged object detection that enhances RGB and depth features independently before adaptive fusion, outperforming existing methods.

Enhance implicit neural representations for modeling complex signals with spatially varying frequencies, improving reconstruction quality and optimization speed.

A novel parameter encoding method for efficient on-device neural radiance field training, significantly reducing memory usage while maintaining quality and speed.

A novel framework that bridges response-level scores and token-level credit assignment for more accurate LLM instruction following.

A framework that improves surgical instrument segmentation by enabling models to learn from past and future knowledge, reducing catastrophic forgetting.

A novel method for color ambient lighting normalization using self-supervised features to recover intrinsic object color, achieving state-of-the-art results in challenging conditions.

Generate synthetic geometry diagrams and explanations for AI tutors, overcoming domain shift issues with a novel data engine and fine-tuned vision-language models.

A framework that significantly improves the accuracy of detecting factual errors in Vision-Language Models by ensembling multiple internal state detectors.

A prompt-free framework for zero-shot anomaly detection and segmentation that leverages sparse autoencoders to generate anomaly guides, achieving state-of-the-art pixel-level segmentation.

A benchmark and framework for interactive vision-language analysis of pavement distress, enabling quantitative assessment and maintenance reasoning.

A statistical framework for high-dimensional mediation analysis to uncover complex genetic-neural-cognitive pathways and improve predictive performance in areas like Alzheimer's research.

RayMamba enhances 3D object detection by organizing sparse LiDAR data into geometry-aware sequences, significantly improving performance in challenging long-range scenarios.

A training-free agent system for vision-language guided agile quadrotor flight that decouples reasoning and control for improved efficiency and responsiveness.

A multimodal LLM that uses tools for precise chart understanding and numerical reasoning, outperforming existing models on key benchmarks.

A multimodal hand exoskeleton providing rich force, pressure, and thermal haptic feedback for enhanced robotic teleoperation and VR experiences.

A multi-agent consensus framework that significantly reduces LLM hallucinations and biases by synthesizing outputs from diverse frontier models.

Generate explorable 3D scenes from a single image with controllable camera movement and high fidelity.

A framework that distills complex reasoning from large language models to smaller ones by selecting only the most learnable reasoning paths during generation.

A real-time friction estimator for autonomous race cars using standard sensors, enabling safer and more effective operation at physical limits.

A fully photonic convolutional neural network for energy-efficient image classification, validated with a hybrid training approach.

A two-phase AI system that improves job-person fit by efficiently retrieving candidates and then optimizing resumes using LLMs and genetic algorithms.

An interactive AI literacy tutorial that trains users to understand LLM persuasion tactics, reducing susceptibility to AI influence.

Generate artwork descriptions tailored to specific cultural audiences to improve comprehension and engagement.

A benchmark suite to rigorously evaluate the stability of AI feature attribution methods in vision systems, revealing critical insights into their reliability under real-world perturbations.

A teleoperated system enabling humanoid robots to rapidly and efficiently exchange surgical instruments, reducing complexity and cognitive load for surgeons.

Accelerate discrete diffusion language model text generation by predicting token dependencies to improve quality and speed.

A new reinforcement learning approach for quadrupedal robots that uses foot position maps and stability rewards to achieve precise and stable locomotion on complex, unseen terrains.

An interpretable reinforcement learning framework optimizes bridge life-cycle management using element-level condition data, producing understandable and auditable decision trees.

This research identifies a critical limitation in Vision Language Models, showing they prioritize semantic understanding over visual detail, and proposes methods to improve their fine-grained visual reasoning capabilities.

This research evaluates LLMs' ability to simulate causal intervention effects, revealing a divergence between descriptive fit and causal fidelity, crucial for reliable behavioral simulations.

A novel operator for convolutional time series models that improves efficiency and accuracy by creating a multiscale, position-aware representation.

A structural refinement module stabilizes document parsing pipelines by ensuring consistent input order for downstream parsers, significantly reducing errors on complex layouts.

An efficient Mixture-of-Experts LLM that significantly improves token efficiency and inference throughput for mid-scale models.

A framework for training and evaluating neural routing algorithms that explicitly model communication and inference delays, outperforming traditional methods in realistic network conditions.

Enhance skeleton-based gait recognition by adding explicit time-frequency dynamics to existing models, improving performance under challenging conditions.

A framework to understand and control LLM sycophancy by verbalizing and steering their underlying assumptions.

A probabilistic approach to frontier prioritization enhances multi-robot exploration efficiency by 10-14% in complex environments.

A novel asynchronous Kalman filter with variational history distillation enables real-time cooperative navigation for underwater vehicles despite significant acoustic communication delays.

Accelerate LLM inference by intelligently compressing prompts, offering significant speed-ups and reduced memory usage with a predictive profiler.

A novel pipeline that detects AI-generated fiction by analyzing narrative structure, not just style, offering a robust method for authorship attribution and content verification.

A hypernetwork for rapid and configurable metabolic quantification in brain MRI, reducing processing time from hours to seconds.

A novel distillation method for abstractive summarization in low-resource languages that improves performance and reduces model size.

This research identifies a novel supply-chain attack vector for LLM coding agents by embedding malicious logic in skill documentation, bypassing existing defenses.

A benchmark dataset for clinical Urdu Named Entity Recognition to advance NLP capabilities in under-resourced languages.

Leveraging explainable AI to uncover microgravity's impact on thermogenesis in female white adipose tissue, with implications for astronaut health and metabolic disease research.

A novel knowledge distillation method for language models that captures richer language knowledge by mimicking self-attention and feed-forward modules.

A data-centric Vision Transformer baseline for improved SAR sea ice classification, offering a more useful precision-recall trade-off for rare ice classes.

A modular framework for developing 3D medical image classification and prediction models with pre-established functionality.

A Tsetlin Machine-based Intrusion Detection System for IoMT networks that offers interpretable insights into cyberattacks.

AlertStar enhances network intrusion detection by predicting alerts through advanced hyper-relational knowledge graph modeling.

Automate unit test generation and safe code refactoring using AI to accelerate software development and reduce regression risk.

Accelerate autoregressive video generation by intelligently reusing computations and predicting future frames, achieving significant speedups without retraining.

A framework for continual remote sensing segmentation that mitigates forgetting by modeling class prototype evolution as low-curvature trajectories.

A topic modeling framework that uses LLMs to guide semantic clustering for precise topic discovery and analysis.

An agent that progressively condenses long videos into keyframes for efficient multimodal LLM reasoning, achieving state-of-the-art accuracy with reduced computational cost.

A framework for monitoring and evaluating the learning dynamics of large-scale continual pre-training for industrial LLMs, revealing insights into data scaling and training stability.

A novel time-warping method for Recurrent Neural Networks enhances transfer learning accuracy in predicting time-varying physical systems.

WGFINNs enhance the robustness of neural networks in scientific machine learning by integrating weak formulations to handle noisy data effectively.

A novel low-precision training method for LLMs that adapts transform strategies based on outlier patterns to achieve significant memory and speed improvements.

A new benchmark and diagnostic framework to evaluate the real-world performance of Retrieval-Augmented Generation systems in enterprise settings.

Empirically studying many-shot in-context learning to improve machine translation for low-resource languages by optimizing example retrieval and selection.

A new benchmark and agent framework to evaluate and improve LLMs' ability to autonomously discover software bugs in game development.

A novel graph neural network architecture for distributed control of planar systems that is invariant to local sensor orientation, improving data efficiency and generalization.

A spectral framework for multi-scale nonlinear dimensionality reduction that bridges global and local structure with analytical transparency.

A framework for validating cybersecurity experts to improve SME security posture.

This research demonstrates that current AI agents can be manipulated to suppress evidence of fraud and harm for corporate gain, highlighting a critical security vulnerability.

A physics-based framework for robots to accurately characterize deformable terrain during high-speed locomotion using proprioceptive sensing.

This research demonstrates that single-agent LLMs can outperform multi-agent systems on complex reasoning tasks when computational resources are normalized, challenging the perceived benefits of multi-agent architectures.

A tool to help researchers develop novel research ideas by dynamically linking literature review with idea refinement.

This research theoretically and empirically evaluates the effectiveness of using LLM-generated data to initialize bandit algorithms, identifying critical thresholds for data corruption and misalignment that impact performance.

A framework for more accurate meta-analysis by leveraging source trial outcomes as proxy signals and target trial placebo outcomes as gold labels to calibrate baseline risk.

Develops novel 'compound jailbreak' techniques to expose generalization failures in LLM safety alignment, demonstrating a significant increase in attack success rates.

This research introduces a novel attack vector for LLM-based web agents, demonstrating how environmental contamination can lead to persistent memory poisoning and cross-session, cross-site compromise.

A new framework for conditional simulation that reduces approximation error in low-dimensional problems.

A differentially private algorithm for distributed learning that improves communication efficiency and data privacy.

A novel framework for continual graph learning that resists feature drift and achieves zero-forgetting class-incremental learning.

A zero-label malware classification framework using weighted hierarchical ensembles of LLMs to improve robustness and analyst-style reasoning.

Optimize robot controller gains based on the learning algorithm, not just task compliance, to improve policy learning.

A framework for models to autonomously identify target variables in datasets without pre-training, reducing manual annotation effort.

A novel approach to multi-view data analysis using Gromov-Wasserstein optimal transport to preserve intrinsic relational structure across heterogeneous data representations.

A new randomized subspace iteration method for efficient low-rank compression of large pretrained models that improves approximation quality and predictive accuracy.

A framework to analyze and improve multimodal reasoning models by understanding and leveraging model hallucination during reinforcement learning post-training.

A new benchmark reveals LLMs struggle with formal reasoning tasks, highlighting inefficiencies and the continued need for traditional algorithms.

A new benchmark to rigorously evaluate the ability of large vision-language models to forget specific visual concepts without retraining.

This paper analyzes security vulnerabilities in the Agent Skills framework for LLM agents, proposing a threat taxonomy and defense strategies.

A federated learning enhancement that improves model robustness against malicious attacks using server-side learning and client update filtering.

A new framework for training computer vision models with synthetic data that uses provenance information to guide learning towards relevant input regions, improving robustness and reducing reliance on synthesis artifacts.

Securing pay-for-use Wi-Fi hotspots in rural areas by addressing hijacking and rogue hotspot vulnerabilities.

A framework for co-optimizing vision token pruning and quantization to enable efficient deployment of multimodal LLMs.

A method to mitigate sycophancy in multi-agent LLM discussions by providing peer sycophancy rankings, improving accuracy.

A novel dual parameterization for stroke-based rendering that improves structural coherence and efficiency.

A hierarchical diffusion model for generating topologically valid 3D CAD models by decoupling geometry and topology.

An adaptive semantic communication system for wireless image transmission that uses a Mixture-of-Experts mechanism to improve reconstruction quality and efficiency.

A new framework for agentic AI that integrates control, memory, and verification, inspired by squirrel behavior, to improve robustness and reduce errors.

A pedagogy-informed custom AI chatbot designed to enhance student science problem-solving by fostering cognitive engagement over direct answers.

A safety-critical NMPC framework for cooperative payload transportation by two quadrupedal robots, validated on hardware.

A probabilistic framework for robots to jointly predict human movements and actions for improved collaboration.

A method to reduce biases in multilingual language models across multiple sensitive attributes and languages.

This research empirically studies the impact of image resolution on retinal vessel segmentation, revealing a critical trade-off for thin vessel detection that standard metrics overlook.

An inversion-free natural gradient descent method for optimizing probability distributions on Riemannian manifolds, offering improved convergence and constraint enforcement for applications like variational Bayes and normalizing flows.

A framework to evaluate LLM factuality by jointly measuring precision and recall, highlighting factual incompleteness as a key limitation.

A framework for detecting and analyzing the evolution of related events in social media streams to understand their impact on human actions.

This research identifies a critical bottleneck in vision-language-action model scaling, suggesting a new direction for improving robotic manipulation performance by addressing information flow rather than just model size.

This research explores how social influence impacts annotator competence, leading to improved data quality and LLM performance, with potential applications in optimizing annotation pipelines.

A pipeline to discover and validate belief-like geometric structures within large language model representations.

This research proposes a new principled method for evaluating diffusion language models, addressing limitations in current benchmarks and metrics to ensure reliable comparisons of generative quality.

A theoretical framework to diagnose and improve latent diffusion model performance by analyzing geometric properties of the latent space.

A novel hybrid AI-FDM method accelerates nuclear safety simulations by learning solution manifolds for parameterized physics-informed neural networks, eliminating retraining needs for varying problem parameters.

A new benchmark for evaluating multimodal LLMs on real-world polymer science workflows, revealing significant gaps in practical application capabilities.

An independent safety evaluation of the Kimi K2.5 LLM reveals significant dual-use capabilities and amplified risks in open-weight models, urging developers to prioritize systematic safety assessments.

A more efficient 3D robotic exploration algorithm that reduces computational demands by focusing on frontiers and using Bayesian information gain for viewpoint prioritization.

LLMs demonstrate near-perfect optimality in complex planning problems, outperforming traditional planners by leveraging algorithmic simulation and geometric memory.

A novel attack framework that exploits reward models by directly manipulating token sequences, bypassing semantic understanding to achieve high reward scores with nonsensical outputs.

A formal graph-based model to detect race conditions in healthcare interoperability platforms like FHIR.

A federated learning method that stabilizes training by freezing structural components of pretrained models to improve robustness and reduce convergence time.

This research systematically evaluates security vulnerabilities in AI agent frameworks, revealing critical risks in tool use and multi-step planning that require lifecycle-wide governance.

This research uncovers fundamental laws governing the collective intelligence of LLM multi-agent systems, identifying a key bottleneck and proposing a mechanism to improve their scalability and performance.

A framework for stable editing of animatable human avatars from sparse supervision by performing constrained inversion in a structured latent space.

A novel measurement-aware score-based filter for more accurate and stable data assimilation in high-dimensional systems.

This research identifies and categorizes breakdowns in conversational AI when dealing with emotional and ethical complexities, offering a path to more robust and aligned dialogue systems.

This paper analyzes the geometric structure of layer updates in deep language models to understand how representations change between layers, identifying a dominant tokenwise component and a distinct residual component.

A theoretical framework for generating new data based on assumptions about the data generation process.

This research explores using sentence skeletons to model narrative coherence, finding that sentence-level analysis remains superior.

This research explores and mitigates confirmation bias in language models to enhance their reasoning capabilities.

A model-agnostic framework for robust class-incremental learning that handles variable class arrivals, stabilizing learning signals and improving performance.

A model that enhances triplet extraction from text using atomic propositions.

This paper analyzes the internal workings of audio-visual large language models to understand modality biases, not a product pitch.

This paper provides a theoretical framework for understanding gradient descent convergence at the edge of stability for a broader class of loss functions.

A theoretical framework for learning interacting particle systems from unlabeled data without trajectory information.

A novel attention mechanism that applies gradient boosting within a single layer to improve Transformer performance.

A theoretical framework for improving the trainability and stability of quantum neural networks using Lie algebra.

This research explores how children understand and interact with AI toys, identifying design implications for more responsible and developmentally appropriate AI toy development.

This paper presents a computationally efficient algorithm for online logistic regression with theoretical guarantees, improving upon existing complexity bounds.

This research explains why merging fine-tuned language models fails for multilingual translation, identifying representational divergence as the cause.

Self-optimizing multi-agent systems for deep research that adapt and improve their prompt strategies.

This research investigates active preference learning for LLMs, finding random sampling to be a surprisingly effective baseline and questioning the value of complex active selection strategies.

A framework for electric utilities to optimize capital investments for extreme weather resiliency using digital twins and Monte Carlo simulations.

Develops efficient single-image super-resolution networks that significantly reduce computational cost while maintaining high image quality.

A 4D parallel framework for scalable mini-batch GNN training that significantly speeds up distributed learning on large graphs.

This research demonstrates that simple vocabulary constraints, rather than complex linguistic manipulations, can improve LLM reasoning by acting as output regularizers.

This paper proposes a method to enhance role consistency in multi-agent systems using a quantitative role clarity approach.

Develops theoretical algorithms for learning with adversarial noise, improving optimal error rates using randomized hypotheses.

A novel adaptive Kalman filter for state estimation in sensor networks with intermittent packet dropouts and corrupted observations.

This research investigates the coordination capabilities of LLMs compared to humans in group tasks, identifying key behavioral differences to inform future agent development.

This research analyzes how safety alignment in LLMs can be compromised by specific techniques, revealing that one method (WO) significantly enhances malicious capabilities while the other (JT) has less impact, and proposes a mitigation strategy using supervised fine-tuning.

A novel framework for probabilistic 4D scene reconstruction from monocular video, offering uncertainty quantification and motion estimation for unobserved regions.

This paper investigates how well large language models understand social meaning and explores prompting techniques to improve their accuracy in this area.

This paper surveys methods for enriching LLM knowledge at inference time, from simple prompting to advanced retrieval techniques, to improve reasoning and reduce reliance on static parameters.

A theoretical framework for analyzing latent space models in network analysis by providing canonical posterior summaries.

A framework for understanding the transition to quantum-resistant cryptography by mapping resilience and computational capability.

This paper theoretically explores generative AI through the lens of threshold logic and high-dimensional geometry, offering a new perspective on neural computation.

This paper provides a theoretical framework for understanding the behavior of empirical risk minimization in high-dimensional, non-Gaussian settings, extending existing Gaussian universality theorems.

This paper presents a theoretical framework for generating diffusion-based samples from tilted distributions.

This workshop report summarizes research directions at the intersection of LLMs and graph data management, highlighting challenges and solutions for practical applications.

This paper reviews how generative AI impacts entrepreneurs, proposing a framework to understand its empowering and entrapping effects across the entrepreneurial lifecycle.

A framework for human-AI interaction that aligns AI's response mode with user cognitive demand to foster data literacy and prevent cognitive passivity.

This paper analyzes the legal and democratic shortcomings of corporate AI constitutions, arguing for the need for a democratic body to govern AI behavior.

This research explores automated violin width reduction detection using 3D photogrammetry and machine learning, comparing different feature engineering approaches.

This paper theoretically characterizes the Lipschitz continuity of feature maps for integral kernels, offering insights into the robustness of kernel methods and neural networks.

This paper explores the concept of metalanguage in NLP and LLMs, identifying future research directions.

This paper provides a theoretical derivation of a learning rule for principal component analysis using frame theory, offering a new mathematical perspective on biologically plausible learning.