AI Research
Beyond Transformers: The Next Wave of AI Architecture Research in 2026
Explore 2026 AI architecture research: state space models, Mamba, hybrid attention, and what long-context efficiency means for enterprise AI costs.
Articles by Tomás Herrera
NLP & LLMs
Retrieval-Augmented Generation at Scale: What the 2026 Benchmarks Actually Tell Us
A critical read on 2026 RAG benchmarks: what leaderboard scores really measure, where they mislead, and how to build evaluation that predicts production.
AI Safety & Alignment
Beyond the Red Team: Building a Practical AI Safety Evaluation Stack for Enterprise Agents
A practical, layered architecture for turning one-off red team campaigns into a repeatable AI safety evaluation stack that ships with your enterprise agents.
AI Research
Beyond Benchmarks: How 2026 Reasoning Models Are Reshaping Enterprise Decision-Making
2026 reasoning models don't have to be benchmark winners to earn a place in your enterprise. Here's how to deploy them for real decision quality, cost control,
Reinforcement Learning
RL-Driven Cost Optimization: How Reinforcement Learning Is Cutting Token Spend in Production Agent Workflows
Reward shaping, RLHF distillation, and plan caching are cutting agent token spend in production. Here's how RL turns token waste into a learnable, compounding objective.
AI Research
Reasoning Models on the Front Line: A 2026 Benchmark of Multi-Step Inference and Tool-Use Agents
A practical 2026 benchmark scorecard for reasoning models on multi-step inference and tool-use agents — accuracy, latency, cost, and reliability.
AI Research
The Reasoning Frontier: What Comes After Chain-of-Thought in 2026
From chain-of-thought prompting to test-time compute and RLVR: how reasoning models work, why verifiers are the new moat, and where LLM reasoning research goes in 2026.
Reinforcement Learning
RL at the Frontier of Reasoning: How 2026 Reinforcement Learning Trains Agents to Think
Key Takeaways
NLP & LLMs
Long-Context LLMs in 2026: How Model Architectures Scale to Millions of Tokens
In 2026 the million-token context window is an infrastructure problem, not a spec sheet feature. Sparse attention, Mixture-of-Experts, KV-cache optimization, positional encodings, and hybrid architectures determine whether long-context models actually scale in production.
Reinforcement Learning
Beyond Chat: How Reinforcement Learning Is Training the Next Generation of Agentic AI Systems
How reinforcement learning trains agentic AI — from verifiable rewards and GRPO to tool use, reward hacking, guardrails, and cost. A practitioner's pl
AI Research
Benchmarking the Reasoners: How 2026 AI Research Is Reinventing How We Measure Reasoning Agents
Why classic AI benchmarks stopped measuring reasoning — and how trajectory, contamination-resistant, and governance-aware evaluation is redefining 2026 agent quality.
NLP & LLMs
Reasoning Models in 2026: How Test-Time Compute Is Reshaping LLM Capabilities
All three new illustrations address concepts that are explained verbally but benefit from visual representation, consistent with the article's existing illustration strategy.
AI Safety & Alignment
Technical Review & Expert Q&A Addendum
Technical accuracy: High. No factual errors found; two clarifications needed (blast radius vs. envelope; toxicity vs. refusal drift) — both addressed in the Q&A.
AI Research
Test-Time Compute and the Reasoning Revolution: What Scaling Inference Unlocks in 2026
Test-time compute is scaling reasoning at inference. What the 2026 reasoning revolution unlocks for LLMs and agents.
Reinforcement Learning
RLHF 2.0: How RLAIF and Constitutional AI Are Reshaping LLM Alignment in 2026
Human feedback built today's frontier models — but it can't scale. This guide explores how RLAIF, Constitutional AI, DPO, KTO, and RLVR are reshaping LLM alignment in 2026, cutting cost while preserving safety.
AI Safety & Alignment
Guardrails in the Wild: A Practical Field Guide to Deploying LLM Safety Systems That Don't Kill Your UX
Layer your controls, keep the happy path fast, and measure false positives. A practical field guide to shipping LLM safety systems that don't break UX.
AI Safety & Alignment
AI Alignment in 2026: New Approaches to Making Large Language Models Safer and More Reliable
A practical guide to the 2026 LLM alignment landscape — from RLHF and Constitutional AI to mechanistic interpretability, enterprise security defenses, and EU AI Act compliance.
Reinforcement Learning
RLHF to GRPO and Beyond: How Reinforcement Learning Is Transforming LLM Training Pipelines
A comprehensive guide to RLHF, GRPO, DPO, and the reinforcement learning techniques reshaping how large language models are aligned and trained.
AI Safety & Alignment
Constitutional AI and Beyond: How Safety Frameworks Are Shaping Model Development in 2026
Constitutional AI and Beyond: How Safety Frameworks Are Shaping Model Development in 2026
AI Research
Reasoning in Large Language Models: Chain-of-Thought, Tree-of-Thought, and What's Next
Chain-of-thought prompting changed how LLMs reason. This deep-dive covers CoT, ToT, GoT, OpenAI o1/o3, process reward models, and what's next for LLM reasoning.
Reinforcement Learning
RLHF vs GRPO in 2026: A Complete Guide to LLM Alignment Methods
Large language models do not naturally know what humans want them to do. GPT-3 demonstrated powerful in-context abilities but routinely produced outputs that were unhelpful, harmful, or simply...
AI Safety & Alignment
Mechanistic Interpretability: Opening the Black Box of Neural Networks in 2026
For most of deep learning's history, neural networks have been black boxes: you train them, they work, and you have only a rough intuition about why. Behavioral testing — probing inputs and outputs — tells you what a model does. It never tells you ho
NLP & LLMs
Sparse Mixture of Experts: How MoE Architectures Are Reducing LLM Inference Costs by 60%
MoE architectures cut LLM inference costs by up to 60% through sparse activation. Here's how they work and when they actually save you money.