Technical Research & Publications

Bridging academic research with practical business applications in AI and analytics

TECHNICAL DOCUMENT TDM_DEV_GUIDE_v0.2

Core Architectures of Text Diffusion Models: A Developer's Guide

An analysis of recent research revealing two dominant architectural paradigms for text generation diffusion models: Discrete Diffusion and Continuous Diffusion. This guide outlines a core five-component architecture, providing a clear and consistent understanding of both approaches for developers and technical leaders.

Key Topics: Autoregressive vs. Non-Autoregressive models, Forward/Reverse processes, Denoising networks, Objective functions, Corruption schedules

READ FULL DOCUMENT →
EXPERIMENTAL VALIDATION ACCELERATING_DDM_20250813

Speculative Sampling and Cosine Schedules: Accelerating Discrete Diffusion Models

Comprehensive experimental validation combining Zhang (2025)'s theoretically optimal cosine corruption schedule with De Bortoli et al. (2025)'s speculative sampling acceleration. Our performance assessment on 6.5M parameter TDLM models demonstrates significant real-world speedups through algorithmic optimization rather than architectural changes.

Key Results: 2.41x maximum speedup (Speculative Linear), 40% model call reduction, 1.82x speedup with best quality (Cosine), Model-agnostic performance gains

VIEW VALIDATION STUDY →
EXPERIMENTAL VALIDATION TDLM_DOMINANCE_20250812

Immediate Diffusion Dominance at Tiny Scale in Hardware-Constrained Settings

Experimental validation revealing a surprising performance anomaly: discrete diffusion models immediately outperform autoregressive models from the first gradient steps at 6.54M parameter scale. Contradicts established research patterns showing AR initial superiority, suggesting novel scale-dependent dynamics that warrant investigation.

Key Results: 37.4% perplexity advantage, 4.3x better performance at step 2, Immediate dominance without crossover event, Consumer RTX 3070 Ti validation

VIEW VALIDATION STUDY →
EXPERIMENTAL VALIDATION ZHANG_COSINE_20250809

Zhang's Fisher-Rao Optimal Cosine Schedule: Experimental Validation

Practical validation of Zhang (2025)'s theoretical work proving cosine scheduling is Fisher-Rao optimal for masked discrete diffusion models. Our experimental testing on a 49.2M parameter TDLM demonstrates how optimal sampling trajectories minimize geometric path length and deliver measurable performance gains.

Key Results: 2.6x generation speedup, Fisher-Rao optimal step allocation, Smart computational budget optimization, No quality trade-offs

VIEW VALIDATION STUDY →
EXPERIMENTAL VALIDATION HRM_PATHFINDING_20250809

HRM Pathfinding Validation: Small-Scale Reasoning Emergence

Testing Wang et al. (2025)'s Hierarchical Reasoning Model on custom 30x30 city logistics pathfinding tasks. Despite scaling down to 2.1M parameters, the brain-inspired dual-module architecture demonstrates spatial reasoning capabilities, achieving 18% exact accuracy with sophisticated dual validation detecting true reasoning vs false positives.

Key Findings: 6.1-hour training breakthrough, 97.8% token vs 18.2% sequence accuracy gap, Adaptive computation evidence, Alternative optimal solution discovery

VIEW VALIDATION STUDY →

Core Services

Financial Modeling & Analysis

Comprehensive financial modeling solutions for strategic decision-making and business growth.

  • Integrated 3-Statement Models
  • Driver-Based Budgeting & Forecasting
  • DCF & Business Valuation
  • Scenario Planning & Sensitivity Analysis
  • Puerto Rico Private Company Valuation
  • Startup Financial Planning for Funding

Ideal for: CFOs, startup founders, investors evaluating PR opportunities, M&A transactions

Data & Business Intelligence

End-to-end BI solutions leveraging Microsoft Azure cloud platform for actionable insights.

  • Azure Data Architecture & Warehousing
  • ETL/ELT Pipeline Development
  • Power BI Dashboards & Reporting
  • Data Modeling & Governance
  • Custom Advanced Analytics
  • Predictive Modeling & Forecasting

Ideal for: Operations leaders, IT/BI heads, organizations with data silos, cloud migration projects

About Jorge A. Arroyo

xP&A Consultant Specializing in Financial Planning & Advanced Analytics

With over 10 years immersed in finance and analytics, I saw a critical need among local businesses: the need to bridge the gap between complex financial planning and the wealth of operational data often locked away in separate systems. SENS ADVISOR was born from an idea to address this: to empower organizations by integrating these domains through expert Extended Planning & Analysis (xP&A) and full-stack enterprise application development that transforms how businesses manage their financial and operational workflows.

I mainly focus on designing and deploying cloud-based xP&A and Business Intelligence systems using Microsoft Azure, Power BI, and advanced Excel techniques, while also architecting sophisticated full-stack business applications with enterprise-grade security, role-based access control, and real-time data synchronization.

For the past 4 years, I've pursued AI and LLM research as a purposeful hobby to stay current with rapidly evolving technology and to rigorously test what works versus what doesn't in real business implementations. This is how I maintain a ground-truth understanding of AI's current capabilities and limitations which allows me to provide clients with realistic assessments of AI opportunities, honest guidance on implementation risks, and evidence-based recommendations for efficiency gains.

Professional Journey

Work Experience

2019 - Present

xP&A Consultant

SENS ADVISOR

2018 - 2019

Financial Planning & Analysis Manager

Grupo Cabrera

2016 - 2018

Lead Financial Analyst for PROMESA

Public Corporation for Supervision and Insurance of Cooperatives

2015 - 2016

Executive Assistant to Comptroller

Government Development Bank for Puerto Rico

Education

2024

Advanced Valuation Certification

NYU Stern Executive Education

2023

MS, Big Data and Analytics

BarcelonaTech

2022

MS, Accountancy (Data Analytics)

University of Illinois Urbana-Champaign

2014

MPA, Government and Public Policy

University of Puerto Rico

2012

BA, History of the Americas

University of Puerto Rico

Get In Touch

Ready to transform your planning and analytics? Let's discuss your specific needs.

LinkedIn: Jorge A. Arroyo
Location: San Juan, Puerto Rico