# Research papers every LLM engineer must read
**作者**: Amit Shekhar
**日期**: 2026-04-07T11:32:46.000Z
**来源**: [https://x.com/amitiitbhu/status/2045390052755411114](https://x.com/amitiitbhu/status/2045390052755411114)
---

## 1. Attention Is All You Need
The paper that introduced the Transformer. We will learn how self-attention replaces recurrence, and why this one architecture powers every modern LLM.
I decoded this paper, you can read these articles:
## 2. BERT
We will learn masked language modeling and why bidirectional context made BERT the default for understanding and classification tasks.
## 3. GPT-3: Language Models are Few-Shot Learners
We will learn how a 175B-parameter decoder-only model unlocks in-context learning, picking up new tasks from just a few examples in the prompt.
## 4. Scaling Laws for Neural Language Models
We will learn how loss decreases predictably with compute, data, and parameters, so we can plan model size and training before burning a single GPU hour.
## 5. Chinchilla
We will learn that most large models were undertrained, and that roughly 20 tokens per parameter is compute-optimal. At a fixed compute budget, a smaller model trained on more tokens beats a bigger one trained on less.
## 6. InstructGPT
The paper behind the methods that power ChatGPT. We will learn how RLHF turns a raw text predictor into a helpful assistant through supervised fine-tuning, reward modeling, and PPO.
## 7. Chain-of-Thought Prompting
We will learn how asking a model to think step by step dramatically improves reasoning on math, logic, and multi-step problems.
## 8. Retrieval-Augmented Generation
We will learn how combining a retriever with a generator lets models answer using fresh, factual documents, without retraining.
## 9. LoRA: Low-Rank Adaptation
We will learn how LoRA cuts trainable parameters by 10,000x by training tiny rank-decomposition matrices instead of the full weights. This is the foundation that QLoRA later used to fine-tune 70B models on a single GPU.
## 10. LLaMA
We will learn how a well-trained 13B model can outperform GPT-3 on most benchmarks, and how open weights reshaped the entire research landscape.
## 11. FlashAttention
We will learn how IO-aware attention cuts memory usage and speeds up training, without changing the math.
I decoded this paper, you can read the article:
## 12. DPO: Direct Preference Optimization
We will learn how to align a model on preference data directly, without a reward model and without reinforcement learning.
## 相关链接
- [Amit Shekhar](https://x.com/amitiitbhu)
- [@amitiitbhu](https://x.com/amitiitbhu)
- [4.4K](https://x.com/amitiitbhu/status/2045390052755411114/analytics)
- [Apr 7](https://x.com/amitiitbhu/status/2041479290580287543)
- [66K](https://x.com/amitiitbhu/status/2041479290580287543/analytics)
- [Apr 3](https://x.com/amitiitbhu/status/2040027305297526811)
- [192K](https://x.com/amitiitbhu/status/2040027305297526811/analytics)
- [Apr 5](https://x.com/amitiitbhu/status/2040751418576101454)
- [71K](https://x.com/amitiitbhu/status/2040751418576101454/analytics)
- [Apr 13](https://x.com/amitiitbhu/status/2043652464709427600)
- [18K](https://x.com/amitiitbhu/status/2043652464709427600/analytics)
- [Apr 11](https://x.com/amitiitbhu/status/2042925644112359442)
- [19K](https://x.com/amitiitbhu/status/2042925644112359442/analytics)
- [Upgrade to Premium](https://x.com/i/premium_sign_up)
- [2:32 PM · Apr 18, 2026](https://x.com/amitiitbhu/status/2045390052755411114)
- [4,414 Views](https://x.com/amitiitbhu/status/2045390052755411114/analytics)
---
*导出时间: 2026/4/18 18:32:36*