Primary Literature

Landmark AI Research Papers

The seminal papers that defined modern Transformers, Parameter-Efficient Fine-Tuning (LoRA), ReAct Agent Loops, and Self-Reflective RAG.

Deep Learning & Transformers2017

Attention Is All You Need

Vaswani et al. (Google Brain & Google Research)

Replaced recurrent neural networks and convolutions entirely with multi-head self-attention mechanisms, founding the modern generative AI era.

Core Innovations:

•Introduced the Transformer encoder-decoder architecture.

•Replaced recurrent O(N) sequential computation with parallelizable O(1) step matrix operations.

•Proposed scaled dot-product attention and sinusoidal positional encodings.

Why it matters: Every major modern LLM (GPT-4, Claude, Gemini, LLaMA) is directly derived from the transformer architecture pioneered in this paper.

Read on arXiv

LLM Fine-Tuning & Efficiency2021

LoRA: Low-Rank Adaptation of Large Language Models

Hu et al. (Microsoft)

Freezes pre-trained model weights and injects trainable rank decomposition matrices into each transformer layer, reducing trainable parameters by 10,000x.

Core Innovations:

•Hypothesized that weight updates during adaptation have a low intrinsic dimension.

•Decomposed update matrix Delta W into W_0 + B * A where B and A have rank r << d.

•Enables training multi-gigabyte models on single consumer GPUs without inference latency overhead.

Why it matters: Made enterprise custom fine-tuning economically viable and democratized open-source LLM specialization.

Read on arXiv

Agentic AI2022

ReAct: Synergizing Reasoning and Acting in Language Models

Yao et al. (Princeton & Google Research)

Introduced the paradigm of interleaving reasoning traces ('thoughts') and task-specific actions ('tools') to guide autonomous agent execution.

Core Innovations:

•Showed that combining Chain-of-Thought reasoning with external tool observation eliminates hallucination and improves goal achievement.

•Outperformed pure reasoning and pure acting baselines across HotpotQA and ALFWorld.

•Established the foundation for modern agent frameworks like LangGraph and CrewAI.

Why it matters: The conceptual blueprint for all tool-using AI agents in production today.

Read on arXiv

RAG Systems2023

Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection

Asai et al. (University of Washington, Allen AI)

Trained an LLM to dynamically determine *when* to retrieve documents and critique its own generated responses using reflection tokens.

Core Innovations:

•Introduced reflection tokens: [Retrieve], [IsRel], [IsSup], [IsUse].

•Eliminates redundant retrieval when the model already possesses parametric knowledge.

•Enforces strict attribution checking against retrieved context before presenting answers.

Why it matters: Pioneered adaptive self-correcting RAG architectures, reducing hallucination by over 40% on open-domain QA.

Read on arXiv