Danny Xie

Evolutionary Optimization of Instruction-Tuned Large Language Model Inference Parameters for Spanish Text Simplification

Large Language Models (LLM) have become valuable tools for automatic text simplification, yet their output quality is highly sensitive to inference-time parameters such as temperature, repetition penalty, top-p, and top-k. These parameters are typically set heuristically rather than systematically optimized. In this work, we apply Covariance Matrix Adaptation Evolution Strategy (CMA-ES) to identify high-performing inference configurations for an instruction-tuned Large Language Model (LLM) fine-tuned for Spanish Text Simplification (TS) on the Financial Education Corpus IN SpAnish (FEINA) benchmark. Using System output Against References and against the Input sentence (SARI) as the optimization objective and Sentence Bidirectional Encoder Representations from Transformers (SBERT) as a monitoring metric, CMA-ES converges in 9 of 25 generations, improving test-set SARI from 36.57 (default) to 38.87 (+6.3%) while increasing SBERT from 0.85 to 0.88. Grid search analysis reveals that repetition penalty and temperature are the dominant factors influencing simplification quality, while top-𝑘 and maximum tokens have negligible effects. The strong positive correlation between SARI and SBERT across all 2,500 evaluated configurations provides empirical support for the single-objective approach within the studied setting. Our framework offers a repli- cable methodology for inference parameter optimization applicable to Spanish and other under-resourced languages in Latin American Natural Language Processing (NLP) contexts.

PineSORT: A Simple Online Real-time Tracking Framework for Drone Videos in Agriculture

We introduce PineSORT, a novel Multiple Object Tracking (MOT) system for drone-based agricultural monitoring, specifically tracking pineapples for yield estimation. Our approach tackles key challenges such as repetitive patterns, similar object appearances, low frame rates, and drone motion effects. PineSORT enhances the tracking accuracy with motion direction cost, camera motion compensation, a three-stage association strategy, and overlap management. To handle large displacements, we propose an ORB-based camera compensation technique that significantly improves the Association Accuracy (AssA). Evaluated via 5-fold cross-validation against BoTSORT and AgriSORT, PineSORT achieves statistically significant gains in our Identity-Switch Penalized IDF1 (ISP-IDF1) metric, along with gains in IDF1 (Identity F1 Score), HOTA (Higher Order Tracking Accuracy) and AssA. These results confirm its effectiveness in tracking low-FPS drone footage, making it a valuable tool for precision agriculture.

From Semantic Reasoning to Geometric Precision: Optimizing Multimodal Foundation Models via Custom Prompting and Parameter-Efficient Fine-Tuning for Crop Counting

Precision agriculture increasingly relies on drone-based monitoring; however, traditional deep learning models often require extensive labeled datasets to maintain accuracy. This paper presents a comparative study of multimodal foundation models against state-of-the-art detectors, including YOLOv11 and RT-DETR, for crop detection under limited data samples. We evaluate the Segment Anything Model (SAM) 3 and the Qwen~3 and 3.5 series using Low-Rank Adaption (LoRA), a parameter-efficient fine-tuning technique on a specialized pineapple plantation dataset. Our results show that while foundation models match or exceed classical detectors with only 5 and 10 training images, a significant ``localization gap'' exists in Large Language Models' reasoning capabilities. Specifically, the Qwen ``Thinking'' (or reasoning-based) variants introduce spatial noise that degrades coordinate precision. Conversely, SAM-based pipelines coupled with RT-DETR offer the most robust few-shot performance. These findings demonstrate that while multimodal models provide a scalable alternative to exhaustive manual labeling, bridging the gap between semantic reasoning and geometric precision remains critical for industrial-grade autonomous farming.

Squeeze Every Bit of Insight: Leveraging Few-shot Models with a Compact Support Set for Domain Transfer in Object Detection from Pineapple Fields

Object detection (OD) typically demands large annotated datasets and substantial computational resources. To address these challenges, we propose a novel two-stage pipeline that integrates Visual Foundation Models (VFMs) for object proposal generation with few-shot learning models enhanced by Mahalanobis distance-based classification. Our approach improves upon traditional Euclidean-based methods by incorporating data covariance through support and context prototypes. We specifically focus on scenarios with only just a few annotated images, reflecting real-world limitations where large-scale labeling is not feasible. Validated on pineapple detection from drone imagery, our method outperforms state-of-the-art (SOTA) few-shot models using minimal labeled data. Extensive experiments show that FastSAM, when combined with a Mahalanobis distance variant that applies singular value decomposition (SVD) and diagonal loading for regularization, achieves the highest mean average precision (mAP), offering a practical and effective tool for crop monitoring and management.

Simple Object Detection Framework without Training

This research introduces a simple framework for Object Detection (OD) based on few-shot methods and Visual Foundation Models (VFM). The framework comprises of three core modules: (i) object proposal, (ii) embedding creation, and (iii) object classification. We evaluated six distinct VFMs to generate the object proposals. We compared the performances of four feature extractors to optimize the object representation, including convolutional neural networks and transformer-based models. Furthermore, we investigated four few-shot methods for classifying objects using minimal labeled data. Our framework provides a scalable and cost-effective solution, specifically applied to OD for pineapple localization in the drone imagery of large pineapple fields, where labeled data are scarce and expensive.

Peeling Back the Pixels: Stabilizing Latent Diffusion for Data-Efficient Generation of Pineapple Fields

While Latent Diffusion Models (LDMs) excel at synthetic data generation, they suffer from severe training instability on the small, homogeneous UAV datasets typical of precision agriculture. In this work, we investigate these architectural bottlenecks by training a Denoising Diffusion Probabilistic Model (DDPM) from scratch on constrained pineapple plantation imagery. Our systematic comparison reveals that continuous latent spaces beta-VAE preserve complex crop structures, whereas discrete architectures (VQ-VAE) fail under scarce data conditions. Additionally, we demonstrate that replacing destabilizing global attention mechanisms with a purely convolutional U-Net backbone ensures smooth convergence in repetitive visual domains. Ultimately, combining this robust architecture with DDIM sampling yields a stable, highly efficient framework for low-data agricultural image synthesis. Code and models will be available at https://github.com/imagine-laboratory/Latent-Diffusion.

Learning Compact Representations of Agricultural Fields: A Study of Variational Autoencoders Variants for Aerial Drone Imagery

The integration of aerial drone imagery into precision agriculture enables large-scale, high-resolution monitoring of crop conditions but also introduces challenges due to the dimensionality and variability of visual data. Variational Autoencoders (VAEs) and their variants provide a promising framework for learning compact latent representations that preserve meaningful crop features while reducing computational complexity. In this work, we investigate the suitability of VAE-based architectures for analyzing aerial imagery of pineapple fields. Using a dataset of approximately 5000 drone-acquired images, we evaluate reconstruction fidelity, robustness to noise, and the structure of latent spaces across different VAE variants, including VQVAE and β-VAE. Our analysis demonstrates that VAEs not only capture fine-grained plant features but also enable latentspace clustering that separates foreground (pineapple plants) from background (soil, vegetation, and other elements). These findings highlight the potential of VAE-derived representations for supporting downstream tasks such as plant counting, yield estimation, and stress detection. By releasing both dataset and code at https://github.com/imagine-laboratory/VAE, we aim to establish a reproducible benchmark for generative models in agricultural monitoring, advancing data-driven approaches for sustainable pineapple cultivation.

Tracking Through Words: A Novel Framework for Data Association in Tracking-by-Detection using Large Language Models

Multiple Object Tracking is a fundamental problem in computer vision that involves tracking multiple objects over time. This is typically addressed using a tracking-by-detection paradigm, which involves detecting objects in each frame and associating them across time to construct coherent trajectories. Established methods, such as Simple Online and Realtime Tracking (SORT) and DeepSORT, leverage Kalman filters, appearance features, and Intersection over Union-based cost metrics, operating within well-defined algorithmic frameworks. In this work, we focused on tracking pineapples in drone-captured videos of large-scale pineapple plantations, exploring the novel use of Large Language Models (LLMs) as reasoning engines within the MOT pipeline, particularly for data association via prompt engineering. We investigated whether prompt-engineered LLMs can effectively perform tasks such as data association using SORT-like mechanisms. Our preliminary results suggest that, despite the inherent limitations in numerical precision and multi-step logic, LLMs exhibit promising capabilities in interpreting geometric relationships and reasoning about object identity across frames. These findings open new avenues for hybrid human-AI reasoning systems in vision-based tracking applications.

Evaluation of the Influence of Multispectral Imaging for Object Detection in Pineapple Crops

Normally, most studies related to Object Detection focus only on RGB images. However, this research explores the feasibility of utilizing multispectral drone images, incorporating RGB channels with near-infrared, and red-edge channels, to perform Object Detection (OD) using drone images of pineapple crops. There are two main challenges when dealing with multi-spectral images. The first challenge is related to the alignment of the images when dealing with different cameras. Multispectral image alignment corrects for camera position and exposure time differences. We use SIFT and ORB for feature-based exposure matching after initial phase alignment. The second challenge is how to incorporate the extra channels into the RGB images, also known as channel fusion. Here, we studied two fusion techniques: early and late fusion. These techniques offer a comprehensive perspective on the potential of multispectral data to enhance object detection accuracy, although the anticipated leap in performance compared to conventional RGB imagery faced challenges. Finally, this research proves that using the correct alignment images process, considering the Vegetation Indexes, and also using the early fusion technique can assist in getting better results in order to improve the precision agriculture techniques.

Artificial Intelligence in STEM Education: Interactive Hands-on Environment using Open Source Electronic Platforms

This article describes an interactive methodology to teach Artificial Intelligence (AI) through the contructivism philosophy of learning by doing, using, open source electronic platforms, like Arduino, Snap Circuits, Raspberry Pi and Circuit Playground, with an interactive hands-on approach Workshops. These are provided to high school and non-engineering students by (previously trained) engineering students volunteers. The methodology proposed is designed to highlight, in different learning activities, key concepts about Artificial Intelligence (AI). AI abstractes the human intelligence processes through algorithms and computer systems, taking advantage of the amount of data generated nowadays to create innovative, effective, efficient, accurate and at low cost solutions, applied in different fields. The main purpose is to motivate the participants to explode its creativity, improving their innovation skills to provide solutions for XXI century problems, better quality of life, health, among others. A survey will be conducted for the students to find insights about effectiveness of the proposed methodology to better acquire knowledge about AI knowledge. We encourage instructors to use similar interactive hands-on methodologies and to include AI concepts with STEM activities into general education courses. Other concerns of AI, is about the fairness of these algorithms, the inclusion and diversity is a key player in how these systems are built, and it can have consequences as the person perspective when building it The idea of the need for diversity and inclusion of the AI field.

The Women's Antenna: An experience of community technology construction led by Cabécar women of Costa Rica

The article delves into the experiences of three remarkable women leaders, Mina A., Mina X., and Mina S., within Costa Rica's Cabécar indigenous community in Alto Pacuare. Referred to as "Mina," these women play integral roles in decision-making despite their remote locations. Their initiation of the Association of Cabécars Women of Alto Pacuare led to impactful ventures, including the establishment of the "Casa de las Mujeres" (House of Women). Nonetheless, they express apprehensions about the potential internet-induced impact on their age-old culture.Participating in a hackathon organized by Sulá Batsú Cooperative, these leaders proposed a platform aimed at preserving ancestral wisdom. Bolstered by institutional backing, they developed a community network merging natural metaphors to bridge the digital divide.This innovative "walkie talkie" style network seeks to fuse Cabécars women's traditional insights with modern technology. The initiative encompasses technical training, infrastructure development, and content creation. Its accomplishments encompass the establishment of a radio network, intranet, and the Okamasuei platform. These tools facilitate communication, knowledge dissemination, and education, fostering connectivity between Cabécars women and younger generations. The project underscores the leaders' commitment to transparency, cultural respect, and women's leadership.Reflecting on lessons learned, the article underscores the importance of adapting to community rhythms, appreciating cultural contexts, and engaging the youth. It concludes that the project's triumph significantly contributes to gender equality and social transformation. Moving forward, the endeavor aims to integrate younger generations and expand knowledge-sharing to encompass other communities. It aspires to shape public policies and advocate for community networks as viable alternatives for digital connectivity, emphasizing the intersection of tradition and innovation.

IEEE en el TEC Contribuyendo con el avance de la ciencia y la tecnología para el beneficio de la humanidad: Sea voluntario, una forma de cambiar el mundo

The article delves into the experiences of three remarkable women leaders, Mina A., Mina X., and Mina S., within Costa Rica's Cabécar indigenous community in Alto Pacuare. Referred to as "Mina," these women play integral roles in decision-making despite their remote locations. Their initiation of the Association of Cabécars Women of Alto Pacuare led to impactful ventures, including the establishment of the "Casa de las Mujeres" (House of Women). Nonetheless, they express apprehensions about the potential internet-induced impact on their age-old culture.Participating in a hackathon organized by Sulá Batsú Cooperative, these leaders proposed a platform aimed at preserving ancestral wisdom. Bolstered by institutional backing, they developed a community network merging natural metaphors to bridge the digital divide.This innovative "walkie talkie" style network seeks to fuse Cabécars women's traditional insights with modern technology. The initiative encompasses technical training, infrastructure development, and content creation. Its accomplishments encompass the establishment of a radio network, intranet, and the Okamasuei platform. These tools facilitate communication, knowledge dissemination, and education, fostering connectivity between Cabécars women and younger generations. The project underscores the leaders' commitment to transparency, cultural respect, and women's leadership.Reflecting on lessons learned, the article underscores the importance of adapting to community rhythms, appreciating cultural contexts, and engaging the youth. It concludes that the project's triumph significantly contributes to gender equality and social transformation. Moving forward, the endeavor aims to integrate younger generations and expand knowledge-sharing to encompass other communities. It aspires to shape public policies and advocate for community networks as viable alternatives for digital connectivity, emphasizing the intersection of tradition and innovation.