A novel training-free method for high-resolution remote sensing image synthesis that dynamically adapts positional embeddings during denoising, outperforming existing baselines.

A novel frequency-domain fusion framework that integrates radar observations with weather foundation model forecasts to significantly extend precipitation nowcasting horizons.

A novel framework for high-fidelity novel view synthesis through refractive water surfaces by jointly modeling the water surface and the scene beneath using 3D Gaussian ray tracing.

EnterpriseLab is a full-stack platform enabling enterprises to develop and deploy specialized, cost-effective AI agents that match frontier model performance while ensuring data sovereignty.

WorldCache accelerates video generation by intelligently reusing intermediate model features, achieving significant speedups with minimal quality loss.

A framework for precise clue hunting in long videos by integrating query relevance with intrinsic video structure, improving MLLM performance.

A unified training approach for latent diffusion models that simplifies the training pipeline and improves performance across modalities.

A VLM-guided latent world model framework that combines dense-frame dynamics with long-horizon semantic guidance for improved trajectory prediction.

DualCoT-VLA enhances robotic action planning by enabling parallel visual and linguistic reasoning for complex multi-step tasks, achieving state-of-the-art performance.

A framework for precise, text-instructed 3D scene editing using scene-graph reasoning to improve spatial understanding and layout consistency.

Accelerate and reduce memory usage for high-rank LoRA adaptations in large language models with optimized kernels and factored norm computations.

Repurpose geometric foundation models to accelerate and improve multi-view image generation for novel view synthesis.

A novel three-stage framework for one-step video super-resolution that significantly improves visual quality and efficiency by unifying distribution matching and adversarial supervision.

A generative framework for unsupervised optical flow learning that synthesizes perfectly aligned frame-flow data, eliminating the need for human annotations and achieving state-of-the-art results.

A foundation suite for universal dexterous hand control using egocentric human videos, enabling cross-hand generalization and reducing reliance on costly robot demonstrations.

A hierarchical policy for dexterous robotic drumming, trained in simulation with sim-to-real transfer, demonstrating complex in-hand, contact-rich, and long-horizon manipulation.

A new dataset and benchmark for detecting social groups in real-world, first-person video, enabling more socially intelligent AI agents.

MemDLM enhances Diffusion Language Models by embedding a simulated denoising process into training, leading to faster convergence, lower loss, and emergent in-weight retrieval capabilities for improved long-context understanding.

A generative recommendation system that optimizes for both user engagement and ad revenue by integrating bid awareness directly into the generation process.

Leveraging transformer-based semantic segmentation of multispectral point clouds for precise riverine land cover mapping and environmental monitoring.

Benchmarking state-of-the-art deep learning models for aerial LiDAR point cloud semantic segmentation to identify optimal solutions for real-world acquisition conditions.

A verifiable reward model that enforces fine-grained spatial consistency in text-to-image generation, improving accuracy and controllability.

Dyadic is a no-code web platform enabling scalable research into human-human and human-AI conversations with multi-modal support and real-time monitoring.

A novel benchmarking framework for time series forecasting that rigorously evaluates model robustness to non-stationarity and noise, outperforming foundation models with a specialized probabilistic generative architecture.

A novel distillation technique that significantly improves the generation quality of parallel language models, enabling faster and more efficient text generation.

Automate LLM quality and security assessments using LLMs as judges, achieving high correlation with human evaluations.

A simple, prompt-engineered data augmentation method that significantly improves LLM knowledge in specialized domains, outperforming complex baselines.

A new benchmark and evaluation framework to assess the interactive response capabilities of 4D world models, addressing a critical gap in current AI research.

A neural framework that retargets human motion to humanoid robots, eliminating physical artifacts and accelerating control policy training.

A parameter-efficient module that significantly boosts the performance of computational pathology AI models by optimizing the transformation of patch features.

PAM is a unified engine for generating realistic hand-object interaction videos, improving existing methods and enabling sim-to-real applications.

A framework that uses visual feedback to iteratively refine text layouts, ensuring aesthetic and readable designs.

Leverage advanced LLMs with RAG and agentic feedback to significantly improve quantum code generation, outperforming specialized models without costly fine-tuning.

A cross-modal learning framework enhances robot navigation by inferring depth from grayscale images when depth sensors fail, ensuring robust performance in challenging environments.

An augmented reality system that guides robotic ultrasound for spine procedures, improving accuracy and efficiency.

A closed-loop verbal reinforcement learning framework for interpretable and adaptive task-level robotic planning, leveraging LLMs and VLM for symbolic policy refinement.

ACPO is a novel alignment mechanism for vision-language models that prevents hallucinations by asymmetrically constraining preference optimization, leading to improved performance on benchmark tasks.

This research demonstrates that LLMs can be trained to actively use internal confidence signals to regulate their behavior, paving the way for more reliable and autonomous AI agents.

A vision-only UAV navigation system that accurately predicts location and heading from aerial and satellite imagery, enabling GNSS-denied operation.

An agent framework that creates its own tools for Earth Observation tasks, outperforming specialized agents with fewer resources.

A framework for adaptive failure discovery in robot policy deployment, enabling efficient risk management and safe industrial integration.

Develop patient-specific respiratory motion models using physics-regularized implicit neural representations for improved radiotherapy precision.

Generate short videos from text queries to precisely ground moments in long videos, improving retrieval accuracy and efficiency.

This research proposes a novel method to analyze and improve LLM reasoning by focusing on the direction of reinforcement learning updates, enabling test-time extrapolation and more efficient training.

Leveraging diverse LLM personas to predict narrative similarity by embracing interpretive plurality, outperforming single-ground-truth approaches.

Develop a real-time 3D motion capture tool that uses Articulated Gaussian Splatting for free-moving scenarios.

A graph-based memory system that enhances clinical reasoning agents by organizing and reusing past experiences, significantly improving accuracy.

A principled defense framework for vision-language models that enhances safety against jailbreak attacks without degrading performance on benign inputs.

P-Flow simplifies visual effects generation using cutting-edge prompting techniques.

A framework for improving enterprise AI agents using offline reinforcement learning and digital twins to overcome data limitations and enhance reasoning.

This research compares world action models and vision-language-action models for robot control, demonstrating superior robustness of world action models in challenging scenarios and providing insights for future development.

This research provides a controlled comparison of autoregressive and masked diffusion language models, releasing code and checkpoints to enable the development of more diverse and fluent text generation systems.

A novel, interpretable algorithm for classifying complex time series data that significantly outperforms existing state-of-the-art models.

A Bayesian framework for adapting 3D vision-language models to domain shifts using online distribution learning.

Enhance existing vision models with 3D spatial understanding by leveraging language-guided reasoning and LLMs.

A unified framework for modeling and solving the kinematics of underground mining robots in virtual environments, enabling real-time performance and robustness for training, planning, and digital-twin applications.

FontCrafter enables high-fidelity artistic font creation by using visual elements as style references and offering fine-grained control over glyph shape and texture.

A taxonomy-aware language-audio model and dataset for species recognition and trait inference from animal vocalizations, outperforming existing methods.

Develops a novel multifidelity approach for Gaussian Process Regression to create more accurate and cost-effective surrogate models from scarce high-fidelity data, augmented by cheaper low-fidelity data.

A dual-stage defense framework for safe text-to-image generation that intervenes at both textual and visual stages to capture and attenuate unsafe content.

Reconstruct translucent 3D objects from images with a novel Gaussian-based pipeline that models subsurface scattering for improved detail and real-time rendering.

Leveraging braid theory for more accurate multi-agent trajectory prediction in autonomous vehicles, improving safety and reducing computational overhead.

An open-source, large-scale, and durable quadruped robot with multimodal perception, built using readily available sheet metal welding and machining.

A novel framework for real-world image restoration that leverages test-time scaling with flow matching models to achieve state-of-the-art performance.

A framework and benchmark for automatically detecting consistency between scientific papers and their code implementations, starting with bioinformatics.

A specialized multi-modal LLM for additive manufacturing, achieving over 90% accuracy on domain-specific knowledge tasks.

ROM is a lightweight, real-time system that mitigates overthinking in large language models, reducing response length and improving efficiency without retraining the backbone.

A framework to detect emerging climate change disinformation narratives by reframing it as a retrieval task, outperforming traditional methods on high-variance narratives.

VP-VLA decouples high-level reasoning from low-level execution in robotics using structured visual prompts, improving spatial precision and robustness.

A lightweight hybrid Mamba-Transformer model for efficient and high-performing 3D medical image segmentation, reducing parameters and FLOPs significantly.

A novel method for non-parametric conditional distribution estimation that provides narrower prediction intervals with guaranteed coverage, outperforming existing conformal methods on benchmarks.

A novel network for RGB-D salient object detection that uses superpixels to enhance global and local feature extraction, outperforming state-of-the-art methods.

A diffusion model that generates complex 3D CAD designs by understanding hierarchical structure, overcoming limitations of existing Transformer-based methods for long command sequences.

A novel AI method using boosted decision trees to achieve state-of-the-art, scalable, and robust power flow analysis for distribution systems.

A systematic recipe for scaling Reinforcement Learning to build advanced LLM agents for complex, long-horizon tasks.

Achieve state-of-the-art medical text summarization with significantly reduced computational cost using parameter-efficient fine-tuning techniques like LoRA.

A plug-and-play module for Video-LLMs that drastically reduces computational costs by intelligently compressing visual tokens, enabling faster inference and lower memory usage without retraining.

A new approach to 3D object detection using open-vocabulary models for semantic grouping in diverse environments.

A real-time, lightweight geolocalization system that iteratively refines location hypotheses for safe autonomous navigation in GPS-denied areas.

A new dataset and baseline models for Tunisian dialect spoken language understanding to unlock task-oriented dialogue systems for under-resourced languages.

A robust AI-generated image detection framework using feature distillation and a multi-expert ensemble to combat deepfakes in real-world conditions.

A new benchmark and evaluation protocol to diagnose and improve attribute binding in multi-subject image generation, addressing a key failure mode in current generative models.

Leveraging chronological patient imaging data with contrastive learning to dramatically reduce expert annotation needs for disease severity assessment.

A camera-agnostic method for pruning 3D Gaussian splats using descriptor-based evidence, enabling efficient storage and transmission of 3D scene data.

A geometrically regularized NeRF model that significantly reduces overfitting artifacts in satellite imagery for more accurate 3D reconstructions.

A novel method for efficient and generalized online adaptation of AI models to changing data distributions, enabling robust performance in real-world scenarios.

A multimodal RAG system for ophthalmology that retrieves and reasons over guideline images to improve clinical decision support, outperforming existing models on challenging cases.

A latent representation learning framework for generating synthetic hyperspectral images, enabling faster and more accurate remote sensing data simulation and analysis.

A medical image segmentation framework that ensures global anatomical plausibility through structure-aware adaptation and hypergraph-based validation, outperforming existing methods.

A mask-free AI framework for adaptive video subtitle removal that achieves end-to-end inference and strong zero-shot generalization across multiple languages.

A family of self-supervised learning models for multi-dialectal Arabic speech processing that achieves state-of-the-art performance on dialect identification.

A probabilistic planning framework for real-time object search in dynamic indoor environments, leveraging scene priors and vision-language models to improve efficiency and success rates.

A lightweight fusion model that significantly improves interactive text-to-image retrieval by adaptively combining multi-modal feedback, outperforming existing methods with minimal parameter increase.

Adaptive LoRA ranks for personalized image generation that balances performance and memory by dynamically adjusting layer importance.

Integrate advanced learning-based stereo matching into satellite imagery pipelines to generate more accurate and detailed digital surface models.

A parameter-efficient UNet model using vector quantization and mixed kernel convolutions for improved precipitation nowcasting, with publicly available code.

Optimize LLM reasoning by jointly evolving prompts and policies to efficiently learn from challenging examples.

Develops a universal physical patch attack for infrared vision systems that bypasses adversarial defenses with no online computation.

A novel reinforcement learning approach for video generation that stabilizes alignment by constraining exploration to the data manifold, achieving superior quality and reward maximization.

Develops an adversarial camouflage technique to protect user privacy against facial recognition systems by maximizing recognition error across multiple architectures.

A novel distillation framework for efficient and high-fidelity few-step video generation, overcoming oversaturation and temporal collapse.

A novel sim-to-real method for humanoid locomotion that injects state-dependent joint torque perturbations to improve policy robustness in real-world deployments.

Develop adaptive AI explanations for complex domains using agentic personas trained with significantly reduced feedback.

Reduce NLP model evaluation costs by up to 95% by intelligently selecting the most informative test samples for annotation.

A comprehensive benchmark for PPG-based clinical prediction, establishing baselines for multi-class heart rhythm classification and physiological parameter estimation, with strong generalizability and insights into subgroup performance.

A lightweight adapter that significantly boosts multivariate time series forecasting performance by capturing complex channel correlations, designed for easy integration with existing foundation models.

A vision-language dataset and benchmark for generalizable steel surface defect detection, enabling explainable and transferable AI models.

A framework for infrared and visible image fusion that drastically reduces data acquisition costs by enabling training on unaligned image pairs, achieving comparable performance to methods requiring 100x more data.

Automated image data augmentation that dynamically adjusts augmentation strength using control theory, eliminating manual policy engineering for new vision tasks.

A framework that transfers hypertension prediction knowledge from expensive MRI scans to low-cost retinal fundus images using a clinical similarity graph, enabling better screening with unpaired data.

A cascade-free Mandarin visual speech recognition system that improves accuracy and efficiency by jointly aligning multiple intermediate representations.

An active MRI acquisition framework using anatomical token uncertainty to significantly accelerate scans and improve image quality.

Unlock controllable image editing and semantic understanding by unifying generative models and vision encoders through a shared Gaussian latent space.

Dynamically predict optimal exposure times for burst photography to achieve state-of-the-art image restoration quality, validated on real-world camera systems.

A training-free method to efficiently sample from model feature invariances using pretrained diffusion models, enabling rapid validation of model behavior.

A novel memory-efficient boundary map representation for large-scale 3D occupancy grid mapping in robotics, with open-source code available.

A framework for more efficient and effective multimodal reasoning by dynamically integrating and precisely representing visual information, significantly reducing token consumption.

A framework that uses checklist-grounded reinforcement learning to enable LLMs to systematically evolve and refine scientific ideas based on fine-grained feedback.

FISFormer replaces self-attention in Transformers with a fuzzy inference system for more accurate, robust, and interpretable time series forecasting.

A collaborative navigation framework for heterogeneous robots that requires no training or simulation, enabling real-world deployment of robot cooperation.

A new benchmark and dataset for abductive event reasoning to enable LLMs to infer direct causes of real-world events from evidence.

Synthesize structured table data to significantly boost LLM long-context reasoning capabilities, improving performance on both in-domain and out-of-domain tasks.

Leveraging large language models to guide incremental learning for imbalanced datasets, improving performance on rare classes.

A novel data-free model merging technique that leverages Fisher Information to significantly improve long-to-short reasoning in LLMs, outperforming existing methods without calibration data.

A novel learning-based method to efficiently reduce visual tokens in large vision-language models for multi-turn question answering, improving inference costs without sacrificing accuracy.

A framework for multi-agent negotiation that infers user preferences to achieve consensus in complex planning scenarios.

A deterministic method to detect hallucinations in medical AI by analyzing token-level confidence and visual evidence, offering a computationally efficient and self-contained solution.

We expose fundamental vulnerabilities in visual-language AI by demonstrating 'mirage reasoning' where models hallucinate visual understanding, and introduce a solution for robust, vision-grounded evaluation.

A visual affordance-based framework for bimanual robotic manipulation that anticipates and facilitates complex preparatory actions for goal-directed tasks.

A neuroscience-inspired Bayesian network for fast, energy-efficient, and uncertainty-aware reliability analysis of complex engineering systems.

A training-free multi-agent LLM framework that generates interpretable, domain-specific weather captions by fusing textual gradients from specialized agents.

A new dense evaluation paradigm for robotic auditing using process reward models to provide fine-grained insights beyond binary success rates.

A novel multimodal AI system that accurately identifies who said what and when in complex conversations by analyzing high-frequency visual cues, overcoming limitations of current models.

TAMTRL improves long-context LLM performance by providing fine-grained, teacher-aligned rewards during multi-turn memory updates, overcoming temporal credit assignment challenges.

A plug-and-play module that adapts image deraining models to unseen real-world conditions using unpaired data and synthesized pseudo-rain, significantly improving performance and training speed.

A modality- and task-agnostic federated learning framework for heterogeneous medical imaging that unifies diverse downstream tasks by leveraging frequency-domain insights.

A federated learning framework for trustworthy medical AI that ensures privacy and reliable uncertainty quantification across diverse healthcare data.

A profile-free framework for next app prediction using multi-hop session graph learning to capture evolving user intent, outperforming baselines in real-world scenarios.

A federated learning framework that enables privacy-preserving cross-view video understanding by aligning representations and reducing communication overhead.

A fully sparse object detection system for event cameras that dramatically reduces memory and storage requirements while maintaining high accuracy for detecting fast-moving objects.

A test-time adaptation framework for multi-object tracking that leverages memory and experience to improve performance under distribution shifts.

A novel MRI segmentation network that leverages data-driven spatial priors to precisely identify brain tumors, outperforming existing methods with a compact model.

A computationally efficient, training-free method for detecting AI-generated images with significantly improved accuracy over state-of-the-art.

Reconstruct dynamic 360° objects from single videos with improved 3D geometry and occluded region handling.

AdaEdit offers a plug-and-play framework for training-free, text-guided image editing that adaptively balances source feature preservation with synthesized content generation, outperforming existing methods on key metrics.

AgenticRec optimizes end-to-end recommender agent decision-making for improved ranking accuracy by integrating tool use and refining user preferences.

A multimodal time series anomaly detection model that aligns semantic information across time and text to identify critical system failures.

A generative framework for robustly reassembling large numbers of 3D fragments into complete shapes, outperforming existing methods in challenging scenarios.

A Bayesian framework for compliance monitoring that infers rule activation and drift from noisy data, offering significant speedups over traditional methods.

A novel method to improve the reliability of LLMs in retrieval-augmented generation by accurately detecting hallucinations through induction-aware uncertainty estimation.

An overfitting-aware algorithm that optimizes multi-task SFT by dynamically adjusting data mixtures, improving model performance and efficiency.

A predictive scheduling system for heterogeneous LLM clusters that optimizes end-to-end latency and task performance in multi-agent workflows.

Develop a multimodal AI tool for cardiac diagnosis, leveraging state-of-the-art vision-language models.

Real-time weather classification using a fusion of LiDAR, RADAR, and camera data for autonomous vehicles.

A streamlined architecture that speeds up audio-video generative models with state-of-the-art performance.

A post-training method to enable spoken dialogue models to control response duration for improved voice assistant interaction.

This research reveals gender discrimination in voice-enabled LLMs and proposes pitch manipulation as a mitigation strategy to ensure equitable AI accessibility.

Adapting self-supervised speech models to detect speech impairments across languages for Parkinson's disease patients.

A Bayesian regime switching model for early detection of internal bleeding in ICU patients using vital signs and lab trends.

A multimodal survival analysis model that generates evidence-based prognoses using locally deployable LLMs, addressing privacy and computational constraints in healthcare.

A computationally efficient classifier for safety-critical applications that provides actionable uncertainty bounds, suitable for real-time medical monitoring.

A novel VAE adaptation technique that enables higher resolution image generation for diffusion models with significantly reduced token counts and faster inference.

A unified agent framework for real-time streaming video understanding and embodied intelligence with long-term multimodal memory and proactive interaction.

Leveraging LLMs as tuning experts within a Bayesian optimization loop to significantly improve FDM 3D print configuration selection and reduce failure rates.

A physics-informed masking technique for Earth observation foundation models that significantly improves predictive accuracy and label efficiency.

Enhancing accuracy in hyperbolic vision-language models through uncertainty-guided alignment.

A new family of sparse retrieval models for codebases that aims to improve LLM-based software engineering systems.

A new safety-oriented dataset to detect and block harmful LLM outputs, improving model security and robustness.

Develop an advanced registration tool for aligning 3D models using Gaussian splatting techniques.

This research explores the potential for language models to encode individual neural responses, suggesting a path towards personalized brain-computer interfaces.

A smartwatch system for fine-grained badminton stroke evaluation and quality prediction.

Benchmarking recurrent event-based object detection for industrial multi-class recognition to improve performance over non-recurrent baselines.

An AI system that dynamically tunes visual odometry parameters for robots based on real-time image analysis, improving tracking and reducing computational cost.

A reinforcement learning framework for generating biologically plausible virtual cells to accelerate drug discovery.

A novel optimization method for memory-efficient LLM fine-tuning that reduces variance and speeds up convergence using adaptive curvature guidance.

A new 'Mixture-Greedy' strategy for generative AI model selection outperforms traditional UCB methods by leveraging intrinsic exploration, leading to faster convergence and better performance without explicit exploration bonuses.

An agentic diagnostic system for pheochromocytomas and paragangliomas that automates risk stratification and detects hereditary syndromes by integrating multimodal data and providing auditable reasoning.

Develop Lipschitz-continuous amplitude modifiers for robust audio signal processing, enhancing applications like speech dereverberation.

This research analyzes prompt injection vulnerabilities in AI-assisted development tools, identifying security gaps and providing guidance for building safer AI workflows.

An AI-driven socio-technical framework to optimize regional economic flow and mitigate under-vibrancy by analyzing spending and sentiment data.

A federated learning framework for real-time, privacy-preserving in-network attack detection on resource-constrained IoT devices.

A unified framework for understanding and generating human motion, text, and images by treating motion as a continuous modality.

A novel method for visualizing machine learning decision boundaries by transforming data into Shapley space, leading to more compact and interpretable decision zones.

This research benchmarks self-supervised learning with texture descriptors for face analysis, finding backbone performance is task-dependent.

Leveraging LLMs for document-level machine translation by creating and filtering synthetic parallel data, overcoming data scarcity and generation issues.

Leveraging determinantal point processes to intelligently curate training data for machine learning interatomic potentials, reducing computational costs and improving model accuracy.

Design systems that present multi-AI advice to improve human decision-making without increasing conformity pressure.

A novel threshold ML-DSA construction enabling one-round online signing with high success rates, overcoming theoretical limitations in cryptographic schemes.

A novel generative adversarial approach to improve knowledge distillation between large language models with different tokenizers, showing modest gains in text generation quality.

A framework that improves the accuracy of soft robot simulations across different shapes by learning a transferable corrective dynamics field.

Enables precise 6D robotic scanning of curved tissue surfaces for improved medical imaging by eliminating reliance on image registration.

A novel activation function that bridges smooth neural network training with ReLU-compatible deployment pipelines by learning a controllable 'hardness' parameter.

A framework for explaining the decision-making process of multi-objective reinforcement learning agents by attributing trajectory segments to specific objectives.

A new dataset of Burmese handwritten digits with code and benchmark results to enable research and development in optical character recognition for underrepresented scripts.

A framework for optimizing energy efficiency in edge device inference by dynamically scaling voltage and frequency based on operator sparsity.

An optimization algorithm for efficient agent routing in dynamic environments with moving targets and obstacles.

A video diffusion model generates climate simulations conditioned on key metrics to bridge theoretical frameworks and improve tropical atmosphere prediction.

A new causal discovery algorithm significantly reduces the number of conditional independence tests required, improving efficiency for learning causal relationships from data.

A novel two-stage framework for coronary artery segmentation in CTA images, combining multi-view deformable convolution with visual Mamba to improve accuracy and efficiency.

A multimodal dataset and pipeline for analyzing comedic timing using language, gesture, and audience laughter.

This research explores how explicit object pointing in Large Vision-Language Models improves zero-shot counting accuracy and generalization by encoding spatial information.

A unified framework for uncertainty quantification in generative models for scientific imaging, improving reliability and detecting out-of-distribution cases.

Develops a framework for self-evolving coding agents to automate and accelerate vehicle aerodynamic drag prediction through surrogate model discovery.

This research identifies a novel vulnerability in multimodal LLMs using comic-based jailbreaks, highlighting the need for more robust safety alignment.

A new primitive for structured reasoning provenance in AI agents enables population-level behavioral analytics beyond traditional debugging tools.

This research analyzes LLM memorization across multiple model families to uncover universal patterns and model-specific behaviors, aiming for a fundamental understanding of how LLMs retain information.

This paper provides a comprehensive review of security threats and defenses for Retrieval-Augmented Generation (RAG) systems, aiming to foster the development of robust and trustworthy RAG applications.

Leveraging TLS certificate and domain features to detect phishing domains in the Danish .dk namespace.

A security auditing toolkit for LLM tool integration protocols to detect and mitigate over-privileged capabilities.

A framework to audit LLM benchmarks for contamination sensitivity, providing confidence scores for model evaluations.

A novel deep learning framework for adaptive wireless broadcasting that optimizes image quality, transmission rate, and processing complexity for heterogeneous edge devices.

This research clarifies how vision-language models process spatial relationships, identifying two key mechanisms within their architecture.

Develop an Uncertainty Guided Tree Search tool for enhancing exploration in reinforcement learning tasks.

This paper provides a theoretical analysis of confidence-based decoding strategies for diffusion language models to improve sampling efficiency.

A theoretical framework for improving external forecasts through online post-processing to minimize cumulative losses.

A new randomized method for unbiased gradient extrapolation in variational inequalities.

A novel trainable activation function that interpolates between non-linearity and linearity to improve training efficiency and robustness in deep neural networks.

A framework for progressively formalizing natural language into machine-actionable knowledge graphs.

This research investigates the failure of topic-matched contrast baselines in removing refusal behavior from language models, suggesting a need for revised methodologies in abliteration research.

This research investigates how overparameterization and priors reshape Bayesian neural network posteriors, offering theoretical insights into their geometric properties.

This research analyzes disengagements in an open-source Level 4 autonomous driving system to identify robustness issues missed by standard metrics.

A new framework for measuring concentration in weighted networks that accounts for both weight distribution and network topology.

Optimizing waypoint passage times for multi-agent systems on predefined routes to improve efficiency and avoid collisions.

A theoretical proposal for a decentralized banking application using Ethereum smart contracts to enhance security and transparency in traditional banking.

A framework for optimizing Mixture-of-Experts LLM architectures by establishing new scaling laws based on compute and parameter constraints.

This research empirically analyzes LLM responses to moral dilemmas to understand if they exhibit genuine moral reasoning or merely mimic it through alignment training, revealing a 'moral ventriloquism' phenomenon.

A novel binary operator and constant can generate all scientific calculator functions, enabling gradient-based symbolic regression for exact function recovery from data.

A non-cryptographic voting system allowing voters to verify election integrity using only basic arithmetic and physical receipts, restoring public trust in democratic processes.

A novel theoretical framework for generating smooth, waypoint-interpolating paths for autonomous and industrial robots.

This paper analyzes the function and distribution of interrogative sentences in French news articles to understand how they structure discourse and foreground specific actors.

This paper surveys novel interoperability solutions for on-chain finance, proposing metrics and models for empirical research.

A novel amortized Bayesian inference approach for fast and scalable parameter estimation in complex dynamical systems like Kuramoto models.

A review of government guidance for smart home cybersecurity incidents reveals a lack of structured incident response support for users.

This paper proposes a financial market for AI tokens to hedge against compute cost volatility for enterprises.

A depth-recurrent Transformer architecture designed for improved compositional generalization in tasks requiring variable-depth reasoning.

A framework for safe, real-time trajectory planning that accounts for unknown disturbances.

A statistical method for multiclass classification that accounts for noisy labels using empirical likelihood, offering theoretical guarantees.

A theoretical framework for optimizing generative reinforcement learning policies by reformulating proximal policy optimization in path space.

A framework for auditing the cybersecurity risks of Internet of Things devices in corporate environments.

A theoretical cryptographic scheme for hierarchical secret sharing using the Chinese Remainder Theorem.

A theoretical cryptographic scheme for secure secret sharing with improved information rates.

This paper theoretically analyzes the nuances of temporal difference errors in deep reinforcement learning, revealing potential performance impacts in deep differential RL methods.

A new model for managing tacit and explicit knowledge using generative AI, introducing the concept of 'Digital Fragmented Knowledge'.

A novel representation framework for symbolic regression that encodes expression DAGs as strings and computes a pruned canonical string to collapse equivalent representations.

A theoretical framework analyzing the structural demands of different reasoning types to understand limitations of current AI approaches.

This paper theorizes and quantifies the risk of cognitive agency surrender due to frictionless AI interfaces, proposing 'Scaffolded Cognitive Friction' as a technical prerequisite for AI governance and societal cognitive resilience.

A theoretical framework for graph foundation models using Riemannian geometry to capture complex structural patterns.

This paper analyzes the philosophical underpinnings of representation genesis in AI, identifying structural limitations in current theories.

This paper explores theoretical bridges between encryption schemes, inspired by homomorphic encryption, with security guarantees based on existing schemes.