📗 ArXiv: Essential Reading for LLMs

📗 ArXiv: Top AI Papers - Essential Reading

🧠 A concise guide to foundational and breakthrough AI papers that shaped the modern era of Large Language Models (LLMs).

---

🏗️ 1. Foundational Architectures

🔹 Attention Is All You Need

Vaswani et al., 2017

Introduced the Transformer — a model that looks at all words at once using self-attention, replacing slower step-by-step RNNs.
Why it matters: Every major LLM (BERT, GPT, etc.) builds upon this idea.

---

🔹 BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding

Devlin et al., 2018

Taught models to understand context both ways (left-to-right and right-to-left).
Why it matters: Revolutionized NLP by enabling fine-tuning for almost any text task.

---

🔹 GPT: Improving Language Understanding by Generative Pre-Training

Radford et al., 2018

Used unidirectional generative training — predicting the next word — to build scalable general-purpose language models.
Why it matters: Set the stage for GPT-2, GPT-3, and ChatGPT.

---

⚙️ 2. Model Adaptation & Efficiency

🔹 LoRA: Low-Rank Adaptation of Large Language Models

Hu et al., 2021

Fine-tunes large models cheaply by freezing most weights and learning small low-rank updates.
Why it matters: Enables efficient adaptation of huge models on modest hardware.

---

🔹 Retentive Network: RetNet — A Successor to Transformer

Sun et al., 2023

Replaces attention with retention, improving speed and long-sequence handling.
Why it matters: A step toward faster and memory-efficient Transformer alternatives.

---

🧩 3. Reasoning & Prompting

🔹 Chain-of-Thought Prompting Elicits Reasoning in Large Language Models

Wei et al., 2022

Shows that prompting models to “think step by step” improves reasoning and math performance.
Why it matters: Basis for today’s reasoning-enhanced prompts and tool-using LLMs.

---

🔹 The Illusion of Thinking

Explores how LLMs can appear to reason while really pattern-matching statistical structures.
Why it matters: Reminds us to critically assess “intelligence” in AI outputs.
(Note: this paper is a meta-discussion of reasoning illusion; see current research on interpretability & cognitive mirroring.)

---

🔹 Distilling the Knowledge in a Neural Network

Hinton et al., 2015

Compresses large “teacher” models into smaller “students” while preserving knowledge.
Why it matters: Key for mobile, embedded, and efficient deployment of LLMs.

---

🤝 4. Reinforcement & Alignment

🔹 RLHF: Learning to Summarize with Human Feedback

Stiennon et al., 2020

Uses human ratings to guide model training through reinforcement learning.
Why it matters: Core principle behind ChatGPT alignment and safe responses.

---

🔹 Expanding RL with Verifiable Rewards Across Diverse Domains

Explores broad reinforcement learning setups where rewards are automatically validated.
Why it matters: Pushes RLHF beyond text into general AI decision systems.
(See emerging research in “Verifiable RL” and cross-domain generalization.)

---

🧭 Summary — How to Read This List

Phase	Focus	Papers
🧱 Foundation	Core architecture & training	1 – 3
⚙️ Adaptation	Efficient fine-tuning & inference	4 – 5
🧩 Reasoning	Prompting & interpretability	6 – 8
🤝 Alignment	Human feedback & reinforcement	9 – 10

---

🪄 Beginner Roadmap

1. Start with Transformers — understand self-attention (Attention Is All You Need).
2. Move to pre-training (BERT, GPT) to learn language model foundations.
3. Learn adaptation tricks (LoRA, Distillation) to handle large models practically.
4. Explore reasoning (Chain-of-Thought) and awareness (Illusion of Thinking).
5. Finish with alignment (RLHF, Verifiable RL) — how AI learns to follow humans.

🏁 Each paper contributes a vital piece — from the birth of Transformers to alignment and reasoning. Together, they tell the story of modern AI.

---

✍️ Curated by Aishwarya Srinivasan
📎 Post compiled & validated by OpenAI GPT-5