GONDAR identifies and exploits Java vulnerabilities through a novel LLM-assisted fuzzing framework, significantly outperforming existing tools.

Harness Sony AI's 'Woosh' for groundbreaking, high-quality sound effects generation for multimedia solutions.

PLUME enhances universal multimodal retrieval by embedding latent reasoning for faster, more efficient inference.

NearID offers a robust identity verification system that isolates identity signals for enhanced personalization and image editing.

An AI framework for automating the construction of Jiangnan gardens using aesthetic principles, facilitating virtual tourism and digital heritage preservation.

A unified vision-language-action system that enhances autonomous driving by decoupling spatial perception and semantic reasoning.

Embodied UAV tracking system leveraging vision-language-action models for dynamic real-world scenarios.

HyVGGT-VO delivers real-time dense visual odometry using a hybrid framework for efficient 3D mapping and pose estimation.

Sven is a novel optimization algorithm that significantly outperforms Adam and LBFGS on regression tasks by efficiently approximating natural gradient descent, offering faster convergence and lower loss.

A training-free framework that enhances multimodal LLMs' ability to identify and utilize relevant visual and textual evidence for knowledge-intensive question answering.

UQ-SHRED quantifies uncertainty in reconstructing complex spatiotemporal fields from sparse sensor data using a novel distributional regression approach.

A large-scale, expert-validated multimodal dataset to benchmark and improve AI's ability to verify scientific claims against visual evidence, revealing significant gaps in current state-of-the-art models.

A parameter-efficient medical vision-language model that achieves near full fine-tuning performance with significantly fewer trainable parameters and reduced catastrophic forgetting.

A novel framework for interpretable and bias-aware reward modeling in LLMs that significantly improves preference learning accuracy.

A distributed graph modeling framework to detect complex money laundering patterns with reduced false positives, validated on real and synthetic datasets.

A Vision Transformer model and expanded dataset for early detection of risky tackles in American football, enabling coach-centered injury prevention.

Improve human pose estimation for extreme sports by fine-tuning models on a novel synthetic dataset.

A data-free method to merge AI models, inheriting capabilities without needing original training data, outperforming existing techniques.

A framework that uses evolutionary algorithms to fuse deepfake speech detectors, achieving state-of-the-art accuracy with significantly reduced system complexity.

A framework for building highly robust and stable AI models for early collision prediction in autonomous driving, significantly reducing reliability risks.

A novel bootstrapping method for vision transformers that quantifies attention score uncertainty, leading to more interpretable and sparse attention maps for image analysis.

A massively parallel PyTorch library for exact inference of multivariate Hawkes processes, enabling analysis at unprecedented scales.

Accelerate ontology engineering by using LLMs to semi-automatically elicit competency questions from domain experts through an iterative feedback loop.

A retrieval-augmented generation framework that leverages a large corpus of procedural knowledge to significantly improve language model reasoning capabilities on complex tasks.

A physics-informed AI framework that drastically reduces the need for expensive simulation data by pretraining on unlabeled physics parameters, enabling faster and more accurate multiphysics surrogate models.

An agentic system that automatically drafts content moderation policies from minimal seed information, outperforming existing methods and expert-written policies.

Generate prototype images from text to enable zero-shot open-vocabulary semantic segmentation for 3D lidar data, outperforming existing methods.

This research quantifies and mitigates cognitive biases in LLMs by identifying and manipulating internal representations, offering a path to more reliable AI decision-making.

AffordTissue predicts tool-action specific safe interaction regions on tissue for surgical automation, outperforming general vision-language models.

Leveraging LLMs to predict future academic collaborations by analyzing author profiles, outperforming traditional methods and identifying novel connections.

A training-free AI pipeline for precise First Point of Contact localization in American football practice videos, enabling biomechanical analysis without labeled data.

A novel framework for open-vocabulary 3D scene understanding that uses sparse voxel rasterization and foundation models to overcome spatial and semantic ambiguities, achieving state-of-the-art performance.

A conformal prediction framework for multi-turn LLM reasoning that guarantees accuracy while reducing cost and latency.

An adaptive fusion strategy for vision and torque sensors in robotic manipulation that significantly improves success rates in contact-rich tasks.

Develop a cost-efficient LLM benchmarking tool that predicts model performance on unseen tasks with 85% cost reduction.

EgoFlow generates physically consistent 6DoF object trajectories from egocentric video by combining a Mamba-Transformer-Perceiver architecture with gradient-guided flow matching, outperforming existing methods in accuracy and realism.

A new benchmark and evaluation framework to rigorously test the safety of AI agents operating with elevated privileges in realistic professional environments.

A specialized guardrail model and benchmark to mitigate food safety risks in large language models.

A multi-agent, human-in-the-loop framework that leverages LLMs and structured AI to accelerate scientific discovery by automating literature analysis and insight extraction.

Fine-tune LLMs to solve complex optimization problems by detecting infeasibility and accelerating search, outperforming existing models.

A tool to detect and mitigate overconfidence in LLM responses by identifying and intervening on specific internal circuits.

This research introduces a structured reward system for video multimodal large language models to improve factual and temporal grounding, reducing hallucinations and enhancing faithfulness in video understanding.

A novel method to reduce LLM hallucinations in scientific literature analysis by leveraging peer document context, improving data extraction accuracy and streamlining research workflows.

Personalize assistive robots using natural language feedback for users with paralysis, reducing user fatigue and ensuring safety.

A novel transformer architecture for self-driving that models agent behaviors with SE(2)-equivariance at a reduced computational cost.

A novel, low-latency LLM guardrail that uses token-level logits to detect jailbreaks with significantly reduced resource overhead.

Adapt powerful closed-box AI models like GPT-4o efficiently and affordably by priming a local proxy, drastically reducing API calls and improving performance.

An AI agent that autonomously learns and refines workflows for complex optical device design, reducing the need for specialized expertise.

A unified framework for 3D reconstruction and generation from sparse views, improving fidelity and completeness.

A hierarchical reinforcement learning framework for generating high-fidelity, privacy-preserving synthetic clinical data that significantly improves downstream classifier performance.

A formal verification platform for AI agents in finance, ensuring mathematically verifiable compliance with regulatory mandates.

An LLM-based agentic framework that automates the generation and optimization of high-performance GPU kernels through iterative refinement and execution-based validation.

A two-stage fine-tuning recipe that significantly improves the code generation and execution capabilities of open-weight LLMs for software engineering tasks.

A lightweight post-hoc reranker that decomposes and corrects for biases in long-tailed classification, improving accuracy on rare classes.

A deterministic benchmark and dataset to rigorously evaluate and improve the reliability of AI agents in handling tool misuse and recovery.

Develops a vision-guided control system for autonomous navigation of magnetic millirobots in cardiac flow for targeted drug delivery.

Unlock legacy ECG image data for automated cardiovascular diagnostics with a self-supervised framework that bridges the performance gap with signal-based analysis.

A production-derived benchmark for evaluating AI coding agents, enabling more realistic performance assessment and driving improvements in agent capabilities.

A role-based LLM framework automates structured information extraction from complex healthy food policies, overcoming common LLM limitations like hallucinations and misclassifications.

A benchmark for evaluating LLM agents in industrial maintenance tasks, revealing significant gaps in current capabilities.

A model merging framework adapts large language models for clinical applications by preserving instruction-following ability and domain expertise, offering a scalable solution for resource-constrained healthcare.

A universal computational thermal imaging framework that overcomes ghosting for high-fidelity night vision, enabling applications from autonomous navigation to healthcare.

An efficient prototype-based semantic segmentation framework for low-altitude UAV imagery that balances performance and computational efficiency.

ZEUS accelerates diffusion model inference by up to 3.2x using a novel second-order predictor without architectural changes or training, maintaining perceptual quality.

A novel domain transfer framework for retinal vessel segmentation that achieves state-of-the-art performance by leveraging latent vascular similarity and iterative co-optimization.

DeltaMem is an agentic memory management system that uses reinforcement learning to significantly improve persona memory performance in conversational AI, outperforming existing product-level baselines.

A unified framework for monocular dynamic scene reconstruction that learns 3D motion through self-correction, improving stability and accuracy.

A diffusion-based policy for mobile manipulation that enables efficient, reactive, and multimodal action generation with self-correction.

A new benchmark and evaluation framework to rigorously test and improve the grounded spatio-temporal reasoning capabilities of video multimodal large language models.

Develops a novel framework for supportive AI dialogue agents that prioritizes user autonomy while maintaining helpfulness, addressing relational risks like dependency and coercion.

A novel learning paradigm that aligns gradients between tabular and image data to improve multimodal fusion performance.

A new semantic evaluation metric for open-vocabulary human-object interaction detection that aligns better with human judgment, enabling more scalable and flexible assessment of AI models.

A framework that enables LLMs to accurately model complex nonlinear optimization problems by decomposing them into solver-compatible elements, with a new benchmark to drive progress.

A robust open-set test-time adaptation method that mitigates the ID-OOD tradeoff for improved model reliability in shifting environments.

A two-phase framework that jointly trains LLMs to solve reasoning problems and refine their own answers, significantly improving accuracy on mathematical benchmarks.

A novel AI approach that learns denoised EEG graph structures and informative representations for improved seizure detection, offering clinically meaningful insights.

Accelerate LLM training and reduce unfaithful reasoning by intelligently filtering training data using influence scores.

A novel framework for precise robot localization using text descriptions by leveraging hierarchical geometric alignments and outperforming state-of-the-art by 19%.

SteerFlow enables high-fidelity, model-agnostic image editing by steering generative flow trajectories to preserve source details and achieve complex multi-turn edits.

A compact, reasoning-driven AI model for grounded claim verification that achieves state-of-the-art accuracy with significantly fewer parameters.

Leveraging synthetic EEG data and a novel cognitive energy metric to build adaptive human-machine systems that respond in real-time to user cognitive states.

An end-to-end system for estimating shared attention by simultaneously detecting groups of people and their focus points, outperforming existing methods.

Swift-SVD offers a theoretically optimal and practically efficient method for compressing Large Language Models, achieving significant speedups in compression time.

A multimodal model that generates accurate and executable code from charts using reinforcement learning and self-correction.

CRIT provides a novel dataset and benchmark for training Vision-Language Models to perform complex, multi-hop reasoning across text and visual information, addressing hallucination and improving grounding.

A diffusion-guided defense system that injects adversarial perturbations into latent space to shield facial identities from GAN and diffusion-based deepfakes, offering robust protection in both white-box and black-box scenarios.

A post-hoc label shift estimation method that incrementally updates priors for any black-box probabilistic classifier, outperforming state-of-the-art.

A novel approach to integrate diverse textual inputs for autonomous driving, improving driving performance and safety through causal scene narration and runtime supervision.

This research introduces a novel routing mechanism for diffusion language models that significantly improves training efficiency and performance by adaptively allocating computational resources based on denoising steps, with available code for implementation.

POCO enhances generative policy learning by maintaining stability and efficiency in robotic manipulation through posterior optimization.

A video frame interpolation model that uses bidirectional cycle consistency to ensure temporal accuracy and reversibility, outperforming existing methods without added inference cost.

An adaptive questionnaire framework using LLMs and alternative data to personalize insurance underwriting, improve user experience, and reduce fraud.

Automate the diagnosis of rubric quality issues in LLM benchmarks to improve evaluation reliability.

A new framework for adaptive open-loop planning in POMDPs with formal performance guarantees, enabling faster and safer decision-making under uncertainty.

Develops a lightweight, fabric-based pneumatic haptic interface for realistic tactile feedback in VR/AR and teleoperation.

A new multimodal dataset bridging RF signals with visual and lidar data to enable AI-driven wireless system design and RF-based perception.

A simplified approach to 3D avatar reconstruction by optimizing the body model rather than scaling network complexity, achieving state-of-the-art results with less computational overhead.

A novel planning algorithm for partially observable environments that intelligently filters information to improve decision-making efficiency.

A general-purpose pipeline for analyzing complex human pose data to extract meaningful behavioral insights.

Train language models to use test-time compute for improved latent generalization and deductive reasoning.

A pattern-guided approach to query reformulation that elicits and applies explicit patterns to constrain LLM-based query generation for improved retrieval.

A benchmark to evaluate and mitigate novel privacy risks in human-centered AI agent social networks.

This research optimizes web agent performance by adaptively selecting observation representations based on model capability and token budget, and incorporating historical context.

A novel method to integrate powerful pre-trained visual encoders into autoregressive generative models, improving performance and unifying understanding and generation.

A variational LSTM with augmented inputs and Monte Carlo dropout to quantify both aleatoric and epistemic uncertainty in high-dimensional nonlinear dynamic structural systems, reducing computational burden.

A cost-aware evaluation framework for selecting the optimal LLM for receipt-item categorization on AWS Bedrock.

A novel dynamic fusion strategy for multi-modal brain network analysis that adaptively processes samples for improved performance.

A framework for LLM agents to autonomously generate complex, multi-file skills for professional tasks, improving performance and reducing manual effort.

This research analyzes the impact of AI-generated content on online platforms, revealing a scale-over-preference dynamic and advocating for AIGC-sensitive distribution algorithms.

A parameter-efficient fine-tuning method that adapts underutilized model subspaces for more efficient knowledge acquisition in LLMs.

A robotic dog that can dynamically switch between skills in real-time, controlled by natural language instructions.

A transformer-based digital twin system for wind-induced structural response forecasting and early warning of structural changes in bridges.

A BERT-based system for nuanced toxic language detection in Bulgarian that preserves access to essential information.

A knowledge distillation framework that uses RGBI imagery to infer high-resolution forest structure data, making detailed ecosystem monitoring more accessible and scalable.

A novel semi-supervised domain adaptation method for medical image severity classification that aligns class-specific rank score distributions.

An AI system that accurately attributes spelling errors to dyslexic writers, with a strong focus on ethical deployment guidelines for educational contexts.

A lightweight dual-loss training framework that significantly improves fine-grained action segmentation quality with minimal architectural changes, making complex models more practical.

A self-supervised learning framework that improves medical image classification accuracy on imbalanced and scarce datasets.

An AI system that optimizes non-pharmaceutical interventions for infectious disease simulations to minimize societal disruption.

A diffusion-based knowledge distillation framework to improve collaborative perception in autonomous systems by actively recovering from sensor and communication corruptions.

Leverage payment network analysis to provide leading indicators of economic change and improve real-time forecasting accuracy, especially during disruptions.

A framework for compressing large language models without retraining by accounting for input distribution shifts and refining transformer blocks end-to-end.

A risk-bounded planning algorithm for coordinated UAV-UGV teams that uses a UGV as a mobile charging station in stochastic environments.

A two-stage framework predicts GPU resource utilization and power consumption in HPC systems for efficient scheduling and power management.

An interactive web-based system for visualizing and understanding the end-to-end inference pipeline of Vision Transformers.

A conversational AI system that integrates text and images to improve learning outcomes in STEM education.

A novel framework for zero-shot skeleton-based action recognition that leverages conditional variational autoencoders and a confidence-aware energy loss to improve accuracy without explicit skeleton-text alignment.

An adaptive prompting framework that leverages emotional tone to subtly improve LLM performance on specific tasks.

A generic strategy and improved heuristic for learning state machines from streaming data, implemented in an open-source library and demonstrating effectiveness in runtime, memory, and quality.

A neural-symbolic framework to robustly discover physical laws from data, addressing bias inheritance in scientific modeling.

A curated set of benchmark problems and datasets for evaluating machine learning on photoplethysmography signals to quantify uncertainty in medical applications.

A new benchmark for evaluating binary function similarity detection models to uncover critical generalization gaps in software security.

A novel regularization technique for vision-language-action models to improve sample efficiency and generalization in robotic manipulation by exploiting the inherent neighborhood of feasible actions.

A novel reinforcement learning algorithm that robustly handles noisy temporal difference errors for more stable and efficient learning.

Leveraging AI and LLMs to automate and accelerate hardware security verification processes for complex systems.

ModTrans enables the use of real-world machine learning models within distributed training simulators, bridging the gap between ML development and system research.

This paper proposes guidelines and a proof-of-concept for reproducible and explainable evaluations of AI agents in software engineering by making agent trajectories publicly accessible.

This research analyzes the reliability and failure modes of AI medication decision systems to mitigate risks and improve patient safety in healthcare.

This research optimizes citation granularity in attributed generation models to improve attribution quality and answer correctness, finding that intermediate granularities perform best.

A framework combining reinforcement learning and model predictive control for practical and scalable policy synthesis in complex control tasks.

A platform that uses LLM agents to automate social science research by simulating human behavior and generating reports.

A framework for training drone geo-localization models without relying on satellite imagery, enabling deployment in GPS-denied or data-restricted environments.

A survey of methods for estimating correspondences between deformed 3D shapes, highlighting recent advances and future research directions.

A novel randomized coordinate descent framework for efficient optimization over probability measures, offering significant speedups over traditional methods for machine learning and mean-field modeling.

A prototype generator for creating realistic event streams to benchmark and improve streaming process mining algorithms.

This research identifies and quantifies a critical security flaw in shared-state LLM agents where benign user interactions can unintentionally corrupt outcomes for other users, requiring new artifact-level defenses.

This research investigates the perceptual differences in voice cloning for standard versus accented Chinese speech, revealing that accent significantly impacts perceived identity and intelligibility.

A new offline reinforcement learning framework that explicitly accounts for estimation error in transition dynamics to improve policy optimization.

A statistical method to identify causal direct effects in the presence of unmeasured confounding, applicable to vaccine studies.

A diagnostic framework for generative simulation and inverse problems that ensures precise statistical agreement with ground-truth data, applicable across various scientific and financial domains.

A system that scales reasoning token budgets for competitive programming using RL and parallel thinking to significantly improve performance on hard problems.

This research compares gradient-free Evolution Strategies with gradient-based RL for LLM fine-tuning, revealing distinct solution geometries that impact knowledge preservation.

An ML-accelerated framework to optimize energy system design by bridging architectural and operational performance gaps, reducing high-fidelity model evaluations.

This research analyzes claim verification datasets to reveal their limitations in testing complex reasoning, suggesting improvements for more robust AI evaluation.

A scalable methodology to evaluate and improve skin tone accuracy and fairness in virtual human rendering pipelines.

A novel attention mechanism that improves language model training stability and efficiency by jointly evolving queries and keys.

A framework for generating 3D adversarial textures to test and improve the robustness of vision-language-action models in robotic manipulation.

A novel inference parallelization strategy for LLMs that improves performance by offloading MoE weights and enabling independent GPU execution.

Develops a communication pipeline for efficient transmission of grid-map data for intelligent vehicles.

A model-based approach to improve precision and control in robot-assisted 3D printing by identifying system parameters.

A new corpus and evaluation framework for tackling extreme class imbalance in scientific text classification, enabling more robust NLP models for specialized domains.

A framework for physics-driven 4D dynamic simulation that uses dual-domain motion supervision to improve physical fidelity and reduce computational cost.

This research investigates the accuracy and uncertainty estimation of LLMs for Bayesian elicitation, finding that model size is key but reasoning effort is not, and that statistical correction is needed for reliable decision-making.

Develops a Koopman operator-based adaptive control system for turbofan engines to improve robustness and thrust response.

An adaptive reinforcement learning algorithm that maintains control performance in systems with changing dynamics by intelligently managing historical data.

This research reveals a fundamental geometric reasoning gap in current Vision-Language Models, highlighting a need for improved spatial invariance in future multimodal systems.

A SAT-based framework for optimizing the 2D cutting stock problem, outperforming existing solvers on benchmark instances.

A framework for achieving compliance with the UK Cyber Security and Resilience Bill using Zero Trust Architecture.

An RL agent that uses smaller language models to suggest actions only when it's uncertain, improving out-of-distribution performance without retraining.

A structured engineering method to verify Operational Design Domain coverage for safety-critical AI systems, meeting aviation certification standards.

A Graph Neural Network approach to improve data association in multi-object tracking for autonomous vehicles.

An auction-based framework for reinforcement learning agents to dynamically adapt to changing objectives by bidding for action execution.

A novel method for training multi-class classifiers using graded plausibility supervision, improving predictive performance by projecting predictions onto admissible probability distributions.

An agentic AI pipeline automates strategic asset allocation by having specialized agents generate, critique, and improve portfolio construction methods, guided by an investment policy statement.

A physics-informed Transformer reconstructs wireless channel frequency responses from fragmented spectrum data, outperforming classical methods in accuracy.

A framework for robust multimodal sensing that synthesizes proxy tokens for missing modalities to ensure complete fusion.

Leveraging neural networks for faster and more accurate design of optical reflectors to precisely shape light.

A novel method for robustly solving inverse kinematics in multi-segment continuum robots, improving convergence rates and reducing iteration counts.

A neurosymbolic model enhances news classification robustness by combining text embeddings with symbolic features like genre, topic, and persuasion techniques.

Develops algorithms for online decision-making with noisy observations by integrating learning and decision-making.

Develops multi-agent architectures for video recommendation systems that leverage foundation models and LLMs for more precise and explainable content delivery.

This research identifies and proposes solutions for fundamental limitations of diffusion models in generating discrete data, with validated improvements across text, code, and protein generation.

A novel PAC-Bayesian framework for robustly estimating individualized treatment rules by accounting for reward uncertainty, leading to improved treatment regime selection.

A LiDAR perception framework for mobile robots in dynamic production environments that improves 6D pose estimation and multi-object tracking using synthetic data and a novel detection method.

A new framework for regression fairness that targets specific distribution tails, offering more nuanced and context-sensitive interventions.

A Bayesian federated learning framework for private multimodal survival prediction with uncertainty estimates.

A hybrid framework combining predictive and heuristic methods to optimize cloud infrastructure costs and response times.

A scalable approach for country-wide mapping of forest greenness anomalies using satellite data to detect disturbances.

A language-agnostic method to measure prosodic contrast in speech representations using a novel ABX task and a released dataset.

A framework to improve the continuous execution quality of robot path planning by bridging discrete planning and inverse kinematics.

A method to merge diverse language models from different architectures into a single, more capable model.

A novel dual-branch architecture for chest X-ray classification that improves multi-label and zero-shot performance by integrating projection-specific models and contrastive learning.

A novel autoencoder learns position-independent image representations for faster, direct image interpolation and synthesis.

Develops a multimodal transformer model for improved short-term mortality prediction in heart failure patients by combining clinical text and structured EHR data.

A hybrid physics-informed neural network and finite-volume method for simulating UAV traffic in complex wind fields.

A tutorial on Bayesian Optimization for accelerating scientific discovery by automating the hypothesize-experiment-refine cycle.

A theoretical framework for estimating counterfactual gradients in adaptive inverse reinforcement learning using Malliavin calculus.

This paper identifies and formalizes safety, security, and cognitive risks in world models, proposing mitigations for critical AI deployments.

A new framework for multi-agent systems that uses a shared world representation for improved consistency and stability.

This research analyzes the continuous and broad-based nature of AI automation in labor market tasks, contrasting with abrupt 'crashing waves' of capability surges.

This research identifies critical privacy and security vulnerabilities in digital devices used by incarcerated individuals, highlighting the need for user-centric design and policy reform.

This research investigates internal mechanisms within language models for entity-centric factual questions, identifying specific neurons responsible for entity recall.

Develops new scaling laws for LLM pretraining that optimize for end-to-end compute budgets, including inference costs, leading to overtrained models with improved performance.

This research quantitatively demonstrates that word usage patterns in Hindi preserve etymological signals, distinguishing between Sanskrit and Perso-Arabic origins of synonyms.

A novel framework for differentially private count mechanisms that balances distribution accuracy with count accuracy and runtime.

A method to mitigate reward hacking in LLMs by penalizing shortcut behaviors during training.

A novel geometric design for continuum robots that significantly improves distal stiffness for enhanced precision in constrained environments.

Develops theoretical guarantees for conformal risk control under non-monotonic loss functions, improving stability in practical applications.

A new framework for understanding and evaluating LLM output diversity across different task objectives, revealing trade-offs between safety, representation, and creativity.

This research reveals a fundamental limitation in unlearning concepts from diffusion models using only natural language instructions, suggesting a need for new intervention methods.

This research investigates the detrimental interaction between specific normalization layers and optimizers in LLM training, revealing failure modes that can be mitigated through parameter adjustments or EMA blending.

Enhances active learning for regression by weighting features to improve sample selection accuracy.

A novel behavior-based system to profile AI agent temperaments across four key axes, independent of model capability.

A data-driven framework using a recurrent neural network for reconstructing magnetohydrodynamic flows in fusion reactor blankets from sparse measurements.

A new algorithm for stochastic shortest path problems that significantly reduces computation by intelligently pruning actions based on heuristics, leading to faster problem solving.

This research identifies teacher needs and AI usage patterns in Indonesian education to inform the development of context-appropriate AI tools and policies.

A brain-inspired neural cellular automata with attention for improved robustness and learning speed in self-organization tasks.

Investigating the theoretical underpinnings of how transfer learning and deep ensembles improve psychiatric disorder classification.

This study evaluates LLM-generated translations of radiology reports, finding significant discrepancies between LLM and human expert evaluations, highlighting the continued need for human oversight in medical contexts.

A novel framework for 3D relative robot localization using angle and self-displacement measurements, addressing noise and optimization challenges.

This paper explores the theoretical capacity regions of different reinforcement learning controllers for urban traffic networks, comparing them to a classical baseline.

A new mathematical framework for parameterizing layer connectivity in large language models that scales efficiently and improves expressivity.

Learning spatial structure directly from raw radar data for automotive perception.

A theoretical framework to measure the optimality of word and gesture order based on swap distance minimization.

A self-gated post-training framework for autonomous knowledge acquisition in language models, learning only what the model doesn't know and sharpening existing knowledge to reduce hallucinations.

A novel evaluation protocol for facial action unit detection that quantifies noise and improves robustness, revealing that many reported gains may be artifacts of the testing method.

Adapting large language models for vision tasks through a novel bridge training method.

This paper theoretically analyzes the benefits of depth in Recurrent Neural Networks, showing it enhances memory capacity and expressivity.

Develops an ontology to improve semantic interoperability for scientific experimental data, addressing inconsistencies in existing standards.

A theoretical framework for clustering heteroscedastic Gaussian data with a novel cost function and Wald kernel, leading to a new algorithm called CENTRE-X.

A novel recurrent latent modeling approach for improved long-horizon sequential data representation and generalization.

A new framework for knowledge graphs that treats domain as a core part of representation, enabling disambiguation and cross-domain relations.

A deep learning approach for more accurate 3D channel estimation in MIMO systems.

This research investigates whether Large Language Models exhibit mentalizing behavior when teaching, using cognitive models to analyze their decision-making processes in a simulated learning environment.

A novel differentiable framework for unsupervised prototype-based representation learning that unifies feature extraction and cluster assignment.

Develop resilient clinical AI systems by applying ontology-aware design patterns to mitigate data distortion and semantic drift.

A novel training strategy for LLMs that balances in-context and in-weights learning by using contrastive context sampling to improve performance and prevent label copying.

A framework for validating simulated counterfactual outcomes in causal inference by introducing a digital twin and a hierarchical validation architecture.

This research investigates the disconnect between how convolutional neural networks represent textures and human perception, suggesting current object recognition models are insufficient for understanding texture perception.

This paper proposes a novel method for analyzing typing pauses in translation to understand cognitive processes, but lacks immediate commercial application.

This research maps the historical evolution of symbolic meaning in Persian poetry, identifying recurring themes and their changing relationships over centuries.

A protocol for proving secure connections in quantum key distribution networks without revealing network topology.

A theoretical framework for self-organizing transformer architectures that minimize computational resources by dynamically determining their structure during training.

A formal logic system for reasoning about context-dependent and temporal information.

Develops theoretical generalization bounds for multiple operator learning architectures, specifically MNO networks, to understand sample complexity for unseen operator instances.

This paper surveys the concept of latent space in language models, exploring its foundations, evolution, mechanisms, abilities, and future outlook.

This research investigates the emergence of linguistic structure within self-supervised speech models, analyzing layerwise patterns and learning trajectories.

This paper argues that generative AI should shift data science education to focus on uniquely human reasoning skills, rather than automating them.

Develops theoretical communication schemes for multi-armed bandit problems with noisy actuation, linking to channel capacity.