EgoNav enables humanoid robots to autonomously navigate diverse environments using human walking data, bypassing traditional robot-specific data collection.

YieldSAT provides high-resolution crop yield predictions using a multimodal dataset to improve agricultural productivity efficiently.

Building a CI tool to save over $490K annually in infrastructure costs by optimizing performance regression testing with risk-aware batch strategies.

Unbrowse transforms web interaction for agents by converting redundant browser discoveries into a shared API index, vastly improving speed and efficiency.

BloClaw is a robust AI operating system for scientific discovery, enhancing data visualization and computational workflows with state-driven interfaces.

HippoCamp is a benchmark evaluating digital assistants' capabilities in managing personal file systems for enhanced user-specific reasoning.

Build robust coding tool harnesses utilizing GPT-OSS for improved integrated AI development environments.

CliffSearch offers an AI-driven, evolutionary framework that enhances scientific algorithm discovery by ensuring correctness and originality through agent-based mutation and review processes.

RELISH enhances LLM-based regression by iteratively refining latent states for precise numerical predictions.

SYNTHONY selects optimal deep tabular generative models based on dataset stress profiles and user intent, outperforming LLM selectors.

An unsupervised, domain-agnostic framework for fine-grained semantic segmentation that discovers masks in complex scenes using feature calibration and self-supervised upsampling.

An LLM-based essay scoring system that identifies and corrects systematic biases in grading, improving accuracy with minimal data.

An efficient video anomaly detection method using Mamba, CNN, and Transformer modules for grayscale to RGB reconstruction.

A conditional decoding strategy that significantly improves multimodal LLM safety by predicting a binary safety token before response generation, without external classifiers or modality-specific fine-tuning.

An unsupervised neural network for enhancing and unwrapping velocity fields in 4D Flow MRI data.

An AI framework that uses a mathematical solver as a verifier to automate optimization modeling from natural language, reducing the need for costly process supervision and enabling cross-solver generalization.

A framework for investigating how fault localization granularity impacts automatic code repair at the repository scale, offering a proof of concept for optimizing this process.

A novel attack framework for multi-object tracking systems that exploits query propagation vulnerabilities to cause track terminations and identity switches.

An AI system that reconstructs authentic LiDAR point clouds under jamming attacks by analyzing full-waveform data.

A diffusion-based framework for generating camera raw images from text or sRGB inputs, enabling scalable synthetic data for downstream vision tasks.

A unified architecture metamodel for LLM-oriented applications to automate and standardize the software development lifecycle.

Real-time local wind inference using LiDAR and deep learning enables robust autonomous navigation for aerial robots, improving energy efficiency and obstacle avoidance.

Label-efficient underwater species classification using semi-supervised learning on frozen foundation model embeddings, requiring no model training.

An AI model that uses dynamic graphs and adaptive feature selection from RGB-D data for more accurate indoor scene recognition.

An adaptive LLM serving router that enforces budgets, adapts to non-stationary conditions, and allows hot-swapping of models.

Multi-Group Policy Optimization (MUPO) incentivizes divergent thinking in Vision-Language Models to overcome diversity collapse and improve reasoning capabilities.

Context-aware transfer adaptation for trajectory prediction in autonomous driving, significantly reducing prediction error in new geographic domains.

A lightweight AI model that restores image quality for under-display cameras by adaptively processing regions based on estimated uncertainty.

An agentic framework that iteratively synthesizes and repairs formal software specifications using LLMs and verification feedback to ensure correctness and completeness.

Accelerate mixed-integer programming solvers by using interpretable, sparse learning models for branching decisions, outperforming GPU-accelerated methods.

Introducing Lévy-Flows, a new class of normalizing flow models that capture heavy-tailed financial data for improved risk management and density estimation.

GUIDE is a reinforcement learning framework that provides behavioral recommendations to complement automated insulin delivery systems for Type 1 Diabetes management.

A new task and pipeline for understanding social interactions from raw audio and vision, with a focus on identity attribution and reasoning.

A framework for generating statistically consistent subsurface velocity models from sparse well logs and seismic images, enabling data-efficient seismic imaging and inversion.

A reproducible taxonomy of programming languages by resource availability to guide dataset curation and tier-aware LLM evaluation for code generation.

This research enhances Graph Question Answering by improving how graph structures are encoded into LLMs, achieving competitive results with a novel multi-token pooling approach and LoRA stabilization.

A parameter-efficient deep learning model for single-cell segmentation that matches larger models' performance and generalizes well to unseen data.

A fine-tuning framework that explicitly reduces the modality gap in Vision-Language Models by aligning geometric and distributional structures for improved cross-modal tasks.

Transformer-based models trained on 7T MRI data for automated detection of multiple sclerosis lesions, outperforming classical methods and releasing code for research.

A novel attention mechanism for Transformers that improves training stability, performance, and robustness to data corruptions and adversarial attacks, with applications in vision and beyond.

A frequency-aware framework for epileptic seizure detection using GCNs on EEG signals, offering improved interpretability and diagnostic precision.

A plug-in quality assurance module for ECG digitization that uses VLMs and domain-specific tools for improved accuracy.

Agent Q-Mix is a reinforcement learning framework that optimizes LLM agent communication topology for improved accuracy and token efficiency in complex tasks.

A text-free image-to-image translation framework for medical imaging, effectively reducing artifacts with competitive performance.

A neural-assisted framework for rapid and accurate initial heading estimation in autonomous ocean platforms, significantly reducing alignment time and improving accuracy.

EvolveTool-Bench evaluates LLM-generated tool libraries as software artifacts, revealing quality risks invisible to task-only evaluation and enabling better tool development.

Leveraging distributional semantics to uncover relationships between Polish word phonology, morphology, and meaning, with potential for language learning tools.

A novel diversity-aware distillation objective for LLMs that improves performance and the fidelity-diversity trade-off.

A training-free method called First Logit Boosting to mitigate object hallucination in Large Vision-Language Models with negligible inference overhead.

An LLM code generation system that executes code concurrently with generation, reducing end-to-end latency by up to 55%.

A hierarchical pre-training framework that enhances vision-language alignment by enabling structured feature fusion between vision encoders and LLMs.

RAGShield provides a five-layer defense-in-depth framework for government RAG systems, using supply chain provenance verification to prevent knowledge base poisoning attacks.

A Shapley-guided neural repair approach that uses derivative-free optimization to localize and fix defects like backdoors, adversarial attacks, and unfairness in DNNs.

MAC-Attention accelerates LLM long-context decoding by reusing prior attention computations, reducing latency and KV accesses while maintaining fidelity.

mmAnomaly is a multi-modal framework combining mmWave radar and RGBD input for robust anomaly detection in non-visual scenarios, achieving up to 94% F1 score.

A dual-stream Transformer architecture for illumination-invariant TIR-LiDAR person tracking in autonomous robots.

A memory control framework for LLM agents that mimics human selective forgetting to reduce interference and latency.

A modified retrieval model that integrates LLM-derived relevance signals to identify fine-grained evidence cues with minimal latency overhead.

A novel Mixture-of-Experts management system for LLMs that learns 'molecular memory' for faster domain adaptation, promising significant cost and energy savings.

RegFormer is a transferable relational grounding module for efficient and accurate weakly-supervised Human-Object Interaction detection that learns spatial cues for instance-level reasoning.

HarassGuard: A privacy-preserving vision-language model system for detecting physical harassment in social VR using only visual input.

A dataset and evaluation framework for fact-checking in ten African languages, addressing critical information gaps.

Adapt pre-trained video reasoning models to new domains at test-time without labels using reinforcement learning, outperforming state-of-the-art.

Auto-Regressive Gaussian Splatting (ARGS) enables parallel, multi-scale 3D object generation with controllable detail and visual fidelity.

An LLM-powered multi-agent system for end-to-end housing consultation that significantly outperforms baselines in accuracy and reliability.

A comparative study of LoRA/QLoRA and preference optimization for mental health text classification, providing a practical framework for choosing effective training strategies.

KG-CMI is a knowledge graph enhanced framework for medical visual question answering that improves accuracy and supports free-form answers.

A Vision Transformer-based tactile perception model that generalizes to new robot hand sensors, reducing data collection and retraining needs for scalable robotic manipulation.

A novel Mixture-of-Experts model that eliminates centralized routing for improved scalability and robustness in LLMs.

Develops adaptive cascade algorithms for streaming, per-partition execution of large language models in data warehouses to reduce inference costs.

A training-free co-salient object detection pipeline leveraging SAM and DINO for improved generalization and performance.

A new speech-to-text translation dataset for Central Kurdish derived from TED talks, enabling improved translation performance through orthographic standardization.

A framework for data-efficient robot imitation learning by scaling camera views during demonstration collection to improve generalization.

TALENT is a parameter-efficient tuning framework for referring image segmentation that addresses non-target activation issues to improve accuracy.

A unified medical lesion segmentation framework that uses task-aware adapters and prototype-guided decoders to achieve state-of-the-art performance across diverse imaging modalities and lesion types.

A framework for predicting task-level performance of LLM agents in coding benchmarks, enabling better task design and agent evaluation.

Fine-tune quantized small language models on limited human data to create deterministic, highly aligned evaluators and annotators.

A streaming Vision-Geometry-Action model for end-to-end autonomous driving that jointly outputs dense geometry and trajectory planning.

A pipeline to generate enriched surgical video datasets for fine-grained spatial-temporal understanding, improving vision-language model performance.

An LLM-assisted static analysis framework for auditing code for quantum-vulnerable cryptography and scoring migration risk.

A frequency-guided framework for robust contactless physiological monitoring from facial videos, outperforming existing methods under challenging conditions.

Behavioral Score Diffusion offers a model-free, training-free trajectory planner for robotics that uses kernel-based score estimation from data, outperforming retrieval methods.

An asymmetric actor-critic framework that uses a powerful proprietary LLM as an actor and a smaller open-source critic for runtime supervision to improve reliability in multi-turn agent interactions.

ReMoGen generates real-time human reactions to interactions using modular learning and segment-level refinement.

A novel framework for prostate cancer localization using only T2-weighted MRI by leveraging diffusion-weighted imaging as a privileged latent modality during training.

Gradient-based data valuation significantly improves curriculum learning for game-theoretic motion planners by identifying crucial training scenarios.

Introduces novel orthogonal learners for robust estimation of heterogeneous long-term treatment effects, particularly in settings with limited data overlap.

A shoe-mounted sensor system for accurate Sit-to-Stand transition detection and duration measurement to assess fall risk in older adults.

A training-free method using Canonical Correlation Analysis to select and reduce dimensionality of image representations, improving downstream performance.

A framework for LLM-based agents to selectively forget sensitive or outdated knowledge, enabling controlled unlearning with a natural language interface.

An inverse-free approach for sparse variational Gaussian processes that uses only matrix multiplications for improved stability and speed.

BioCOMPASS integrates biomarkers and treatment information into transformer models to improve the generalizability of immunotherapy response prediction.

A weakly-supervised paradigm for training feed-forward 3D reconstruction models using only relative depths and sparse correspondences.

An AI decoding method for diffusion language models that balances generation quality with exploration for improved reasoning.

A hierarchical planning and control framework for geometry-aware robot navigation in complex environments using MILP-MPC and Minkowski-based CBFs.

A domain adaptation framework for brain metastases segmentation across institutions using VAE-MMD and nnU-Net.

A white-box membership inference attack for LLMs that uses gradient-induced feature drift to detect training data.

A dual-action fabric-based soft robotic glove with customized actuators for ergonomic hand rehabilitation, showing reduced muscle activity and improved grasp patterns.

A distributed AI system for explainable and adaptable regulation that models stakeholder preferences to ensure justice and legitimacy.

A lightweight, prompt-guided adapter for CLIP to perform monocular depth estimation with minimal supervision.

An AI-enabled street-view analysis pipeline for extracting building elevation data to improve regional flood risk assessment.

Language models can predict their refusal behavior with high accuracy, enabling confidence-based routing for safety-critical applications.

A method to speed up neural network classification for millions of classes by reducing label prediction complexity.

A method to adapt text LLMs to speech by inserting and training new transformer layers, minimizing text capability degradation.

VibeGuard is a pre-publish security gate for AI-generated code, addressing blind spots in artifact hygiene, packaging, and supply-chain risk.

Adaptive Parallel Monte Carlo Tree Search with negative early exit and boosting mechanism reduces p99 latency and improves throughput for LLM reasoning without sacrificing accuracy.

A lightweight functional shape representation using Gaussian Process mixture models for efficient and accurate 3D geometry encoding.

Leveraging 3D foundation model priors to improve viewpoint robustness in end-to-end autonomous driving trajectory planning.

Decomposing multi-LLM pipeline gains into re-solving, scaffolding, and content correction to inform targeted pipeline designs.

A new wordplay game benchmark to evaluate the social intelligence and reasoning capabilities of AI agents.

LightGuard enhances WiFi security by offloading cryptographic key establishment to a physically confined LiFi channel.

A standardized benchmark dataset and evaluation toolkit for privacy-preserving video action recognition, enabling nuanced analysis of the privacy-utility trade-off.

StretchBot: A neuro-symbolic framework for adaptive guidance in assistive robots, combining multimodal perception with LLM reasoning for context-aware adjustments.

A secure and efficient inventory control system using NFC tags to replace vulnerable barcodes for retail applications.

A method to disentangle lexical confounds from semantic superposition in large language models, improving word sense disambiguation and knowledge editing.

Two transfer learning methods for nonparametric Bayesian networks to improve learning performance with scarce data.

A theory-grounded benchmark and Bayesian inference framework for multi-dimensional emotion understanding in rich textual contexts.

This work engineers structured features from founder career data to predict startup success, outperforming LLM baselines and diagnosing dataset limitations for future improvements.

Evaluating frontier AI models for research sabotage when deployed as coding assistants, finding some models refuse safety-relevant tasks.

A case-adaptive multi-agent system that dynamically assembles specialist panels for clinical prediction, improving accuracy and transparency.

An infinite-horizon ergodic controller using kernel mean embeddings for long-duration coverage tasks.

A data-driven framework for personalized darts training that uses biomechanical analysis and motion modeling to provide targeted feedback.

A compact, lightweight wearable haptic device that simulates distinct edge and surface contact feedback for enhanced object manipulation in VR.

A theoretical framework for multi-agent reinforcement learning in partially observable games using internal states and a policy gradient method, with a proven convergence bound.

A decoupled generative framework for dynamic multi-objective optimization that uses wavelet transform and sparse dictionary learning to track moving Pareto fronts with zero-shot inference.

Collaborative AI agents and critics for fault detection and cause analysis in network telemetry, leveraging private cost functions and foundation models.

Mine-JEPA is an in-domain self-supervised learning pipeline for side-scan sonar mine classification, outperforming larger foundation models in data-scarce maritime imagery.

A novel reinforcement learning algorithm that improves transfer learning by optimizing successor features with full gradient updates, offering theoretical convergence guarantees and empirical performance gains.

An influence-guided framework that optimizes synthetic data generation for LLMs by using target model feedback to adapt data curation rubrics.

Interpreting LLM failures on character counting tasks to reveal internal reasoning and output suppression mechanisms.

A benchmark and analysis of continual learning for remote sensing vision-language models to address catastrophic forgetting.

A unified framework for Gaussian Splatting that integrates microfacet material models and ray tracing to jointly model specular reflection and transmittance for realistic novel view synthesis.

A Japanese benchmark dataset to evaluate social bias in LLM reasoning based on attribution theory, sensitive to cultural nuances.

KD-ML integrates numeric data with granular knowledge landmarks to build more accurate models, outperforming data-driven approaches on physics-governed benchmarks.

A training-free approach using tracking-enhanced prompts for complex video object segmentation.

A novel extension of the Bradley-Terry model for multiplayer games with an adapted algorithm for ranking.

Lightweight and practical encrypted face recognition with GPU acceleration, reducing memory and improving runtime.

GPT-NL Public Corpus is a large, permissively licensed dataset of Dutch language resources for LLM pre-training.

An automated multi-agent framework for generating exploits against known vulnerabilities in black-box web applications.

A novel divide-and-conquer approach for scalable and accurate differentially private query answering and synthetic data generation.

An automated peer review system that incorporates auxiliary context like figures and external scholarly signals using reinforcement learning.

An uncertainty-aware scheduling metric for LLM inference that significantly reduces latency and improves throughput by accounting for output length variability.

Stochastic Attention, inspired by fruit fly connectomes, enhances efficient attention mechanisms by introducing random routing for improved expressivity and receptive field growth.

Discovering novel data shuffling strategies for stochastic optimization using LLM-guided program evolution to improve convergence.

A cybersecurity risk assessment framework adapted for resource-constrained CubeSat missions, offering proportionate guidance for mission designers and regulators.

An efficient and expressive text-to-speech system using only state-space models for conditioning, reducing parameters and improving throughput.

A novel network for animal activity recognition that customizes features and calibrates classifiers to improve accuracy across all behaviors, even with imbalanced data.

A framework for understanding the trade-offs between pretraining and retrieval for language models, guiding optimal data allocation.

An interactive route planning algorithm for users with accessibility needs that prioritizes user feedback for efficient optimization.

A human-in-the-loop curriculum for computer science education that teaches students to control objective drift in LLM-assisted workflows.

HALIDE: A hierarchical apprenticeship learning framework that leverages imperfect student demonstrations to infer evolving rewards and improve pedagogical decisions.

A playbook for using GenAI tutors to improve student learning in software engineering by grounding them in course knowledge.

A deep learning framework accelerates high-dimensional well control optimization in stress-sensitive reservoirs by using a surrogate model to reduce computational cost by orders of magnitude.

A reinforcement learning framework that autonomously selects solvers for chemistry integration to reduce computational cost while maintaining accuracy.

A generative super-resolution framework for weather forecasts that enhances spatial resolution post-processing.

A 3D-printed, anthropomorphic robot hand with an integrated wrist and carpal tunnel design that enables more versatile and human-like manipulation.

A benchmark for evaluating large multimodal models on the complex task of converting PCB schematic diagrams into structured, spatially weighted netlist graphs.

Enhancing REST API fuzzing with novel oracles for access policy violations and injection attacks, generating executable test cases.

A deep learning framework using MRI-derived brain anatomy to predict early Alzheimer's disease conversion from neuropsychiatric symptoms.

A novel Tucker diffusion model for generating structured high-dimensional tensor data with theoretical advantages over vectorized approaches.

Voice features can predict health deterioration in chronic heart failure patients, offering a non-invasive alternative to current monitoring methods.

A new method to evaluate LLM performance during training, reducing evaluation time from hours to minutes.

A collaborative planning strategy using Multi-Agent PPO to coordinate heterogeneous robotic teams for efficient extraterrestrial exploration.

Improving robotic manipulation robustness under distribution shift by combining deep reinforcement learning with bounded extremum seeking.

This research quantifies the impact of JavaScript obfuscation on static code analysis tools, revealing significant weaknesses in current security pipelines.

Brainstacks is a modular architecture for continual multi-domain LLM fine-tuning using frozen MoE-LoRA adapter stacks that enable cross-domain cognitive capability composition.

Generating sub-meter lunar digital elevation models from high-resolution orbital imagery using an open-source photogrammetry pipeline.

EmbedPart is an embedding-driven graph partitioning approach that accelerates scalable distributed GNN training by clustering node embeddings.

A theoretical framework and empirical validation for improving AI weather prediction by focusing on the entire learning pipeline, not just architecture.

PaperRecon is a new evaluation framework for AI-written papers, disentangling presentation quality from hallucination risks.

A hierarchical generative framework for graph generation that reduces computational cost and generation time.

A library for computing and analyzing multiple protein representations to provide a comparative framework for molecular dynamics simulations.

Demonstrates that current multimodal language models fail to detect spatial inconsistencies in 3D scenes, highlighting a gap in their physical understanding.

An attention-based recurrent network with gradual modality dropout for multicentric thrombus segmentation in 3D brain scans.

This research identifies a new adversarial failure mode in spiking neural network object detectors where detection count is preserved but accuracy collapses.

A differentiable framework for improved fiber recovery in diffusion MRI by fitting explicit multi-compartment models over spatial patches.

This research analyzes how deep networks favor simpler data, revealing a consistent pattern across various models and datasets, but lacks a clear product path.

A multi-agent LLM framework simulating supportive behavioral health dialogue through coordinated, role-differentiated agents with continuous safety auditing.

A method for training a single model to be robust across multiple quantization formats for elastic inference.

A tuning-free GNN prompting framework for cross-graph adaptation without retraining.

A system-level characterization of MPC and FHE for privacy-preserving machine learning, evaluating performance, energy, and cost across various scenarios.

A novel approach to automatic speech recognition that leverages large language models for contextual reasoning and user-guided transcription.

A unified architecture for recommendation systems that unifies mainstream scaling blocks and optimizes scaling efficiency.

Investigating how personality traits moderate the effectiveness of AI-mediated workplace negotiation coaching, suggesting tailored interventions for different user profiles.

Physical Imitation Learning distills RL policies into passive body dynamics for energy-efficient quadruped locomotion, reducing control effort through parallel elastic joints.

A novel stochastic gradient tracking method designed to maintain convergence properties in distributed optimization networks with Byzantine agents.

A framework for assessing programming skills in Scratch using fuzzy clustering, inspired by CEFR levels, to identify learning gaps and guide curriculum design.

Investigating the relationship between score saturation and issue discovery in LLM-based agent evaluations to understand the optimal number of agents needed.

An exploratory study on whether large language models can self-correct in medical question answering, finding inconsistent benefits.

A post-hoc method to suppress non-semantic information in image representations for improved zero-shot performance.

A hierarchical framework for autonomous motion planning and control under unknown nonlinear dynamics using piecewise-affine models and graph-based reachability.

A certificate-driven approach for closed-loop multi-agent path finding that improves scalability and global guarantees by incorporating inheritable factorization.

A framework for visualizing high-dimensional data by embracing the multiplicity of embeddings, leading to more interpretable and robust representations.

Utilize transformer models with program slices to improve accuracy in detecting software vulnerabilities.

Offline constrained reinforcement learning from human feedback with multiple preference oracles for performance and safety trade-offs.

A parameter-free routing mechanism for Mixture-of-Experts models that eliminates the need for a learned router.

A theoretical framework for causal effect estimation that improves overlap conditions using novel 'deconfounding scores' for high-dimensional data.

A novel framework for multi-agent collaboration that uses active shared perception to overcome non-stationarity and accelerate convergence.

This thesis explores AI safety by automating circuit discovery, removing dangerous behaviors with Latent Adversarial Training, and analyzing agentic misalignment in frontier models.

A decision-centric framework for LLM systems that separates control decisions from generation for improved reliability and diagnosability.

A novel algorithm for dueling bandits over continuous action spaces with Lipschitz structure, achieving logarithmic space complexity.

A generalized scenario theory for multi-criteria data-driven decision making that provides more accurate robustness certificates.

This survey analyzes the integration of Large Language Models into the Abuse Detection Lifecycle, covering label generation, detection, review, and governance, while highlighting challenges and future directions.

Compares neural and polynomial operator surrogates for parametric PDEs, finding no universally superior method and highlighting the importance of matching methodology to problem characteristics.

Identifies a critical feature norm threshold that predicts the onset of neural collapse in deep networks.

A new diagnostic framework using Lagrangian Descriptors to evaluate the phase space integrity of neural network models for Hamiltonian dynamics.

Utilizing Lead Zirconate Titanate as a physical reservoir for improved handwritten digit classification.

A multi-agent LLM governance system for safe, two-timescale reinforcement learning in SDN-IoT defense.

Organizes and analyzes the literature on agentic tool use in large language models across three paradigms, highlighting challenges and evolutionary views.

A novel controller for nonlinear dynamical systems that uses Stein variational inference to adapt to latent parametric uncertainty, improving performance-robustness tradeoffs.

Studies the impact of adversarial attacks on AI-driven RAN slicing decisions, quantifying SLA violations and recovery behavior.

This paper investigates how context length and conversational settings silently shorten LLM reasoning traces, potentially impacting performance on complex tasks.

Multiscreen is a novel language model architecture that uses screening to achieve absolute query-key relevance, reducing parameters and improving inference latency.

Examines and replicates a debiased LASSO framework for high-dimensional regression, comparing its performance and power against the desparsified LASSO.

Investigating the role of induction heads in LLMs for understanding temporal dependencies in in-context learning.

A theoretical framework for optimal low-rank decomposition of LLM weights using second-order Hessian information.

A hierarchical framework for organizing multi-agent systems like a company to improve reasoning, reduce costs, and enhance coordination.

Universal YOCO combines a decoder-decoder architecture with recursive computation for efficient depth scaling of Large Language Models, improving inference efficiency.

A unified framework embeds structured knowledge into neural networks for consistent estimation and improved financial applications.

A survey providing a unified framework and comprehensive overview of on-policy distillation techniques for large language models, addressing exposure bias.

This research explores how students' trust, AI literacy, and need for cognition influence their reliance on AI assistants in programming tasks, suggesting a need for better instructional support.

An anytime algorithm for integrated task and kinodynamic motion planning under semantic uncertainty in partially observable environments.

Investigate phase transitions in context-sensitive random language models with short-range interactions to understand the intrinsic nature of language properties.

A novel information-theoretic scheme for error-detecting private information retrieval over rings that reduces key size and communication overhead.

Developing accessible explainable AI for blind and low-vision users to foster trust and independent use of autonomous AI agents.

A novel approach for reconstructing the latent geometry of Riemannian manifolds from noisy distance data, breaking the volumetric barrier in higher dimensions.

Investigating whether enforcing compositional structure in sequence embeddings improves geometric organization in protein-protein interaction networks.

A review of practitioner conference talks to understand how companies adopt, architect, and implement LLM-driven AI agentic systems.

A theoretical framework for multi-task linear bandits that improves sample efficiency by learning shared low-rank representations.

This paper provides a statistical perspective on TF-IDF, showing how it arises from a penalized likelihood-ratio test for word burstiness.

This paper details the large-scale pretraining of Mixture of Experts language models on a supercomputer, focusing on optimizing training speed and stability.

Develops theoretical generalization bounds for spectral graph neural networks by analyzing their behavior in the graph Fourier domain.

A transformer-based ML model that approximates solutions to Maxwell's equations for wave reflection and transmission.

A theoretical algorithm for risk-aware multi-armed bandits that minimizes variance in reward selection.

A theoretical framework for understanding neural network representations in superposition, highlighting limitations of current alignment metrics.

A theoretical approach to scale speeds along robot paths to ensure collision-free waypoint tracking under acceleration limits.

Authors leave unique 'fingerprints' in the novelty dynamics of their writing, detectable at both book and chapter levels.

Simulating mobbing behavior in Braitenbergian robots using the Webots platform to explore the effects of mobbing call range and group size on predator harassment.

An information-geometric analysis of dependency networks and pseudo-Gibbs sampling for theoretical understanding.

Classifying the valency of Mapudungun verbal roots using the language's own morphotactics to improve a morphological analyzer.

Theoretical convergence analysis of projected stochastic natural gradient variational inference under various step size and sample/batch size schedules, providing new non-asymptotic results.

Investigating the phenomenon of grokking in feature learning kernels, showing that breaking data symmetry is crucial for generalization.

LAPIS-SHRED is a modular architecture for reconstructing and forecasting complete spatiotemporal dynamics from sparse sensor observations in short temporal windows.

Identifies vulnerabilities in the Loki e-voting protocol and argues for the necessity of pre-agreed secrets to prevent last-minute coercion.

Exploring the biological foundations of human language through mathematical and algebraic models of syntactic structures.

Developing a sample-efficient model estimation procedure for learning policies in partially observable environments.

Develops fully initialization-dependent complexity bounds for shallow neural networks with general Lipschitz activation functions, offering non-vacuous generalization bounds for overparameterized models.

REMind is an educational robot-mediated role-play game designed to support anti-bullying bystander intervention among children.

Theoretical analysis of mixing time bounds for positively weighted restricted Boltzmann machines.