Must-Read AI Research Papers

Introduced the Transformer architecture - the foundation of every major AI model since. Replaced recurrent networks with self-attention, enabling full parallelisation during training. Without this paper, GPT, Claude, Gemini, and every LLM would not exist in their current form.

Read: The abstract, Section 3 (Model Architecture), and Figure 1. The maths is secondary; understanding the parallel attention mechanism is the key insight.

The paper that proved scaling works. GPT-3 (175B parameters) demonstrated that a sufficiently large language model could perform tasks from just a few examples in the prompt - without any fine-tuning. Established few-shot prompting as a paradigm. The scaling laws appendix (that bigger models trained on more data get predictably better) changed AI investment trajectories permanently.

Read: Sections 1–3 for the core argument. Skip the appendix unless you want the benchmark details.

Described RLHF (Reinforcement Learning from Human Feedback) - the technique that turned raw language models into helpful assistants. A fine-tuned 1.3B RLHF model was preferred by human evaluators over a raw 175B GPT-3. This paper explains why ChatGPT is usable while base GPT-3 is not.

Read: Section 2 (methodology) and the human evaluation results. The RLHF training diagram is the most useful single figure.

Demonstrated that including step-by-step reasoning examples in the prompt dramatically improves performance on arithmetic and logic tasks. The simplest intervention with one of the largest effects on model reasoning accuracy. Every "think step by step" prompt instruction traces back to this paper.

Showed that sampling multiple reasoning chains and taking the majority vote improves accuracy significantly over a single CoT sample. The theoretical justification for using sampling + voting instead of greedy decoding for hard reasoning tasks.

Empirically demonstrated that model performance degrades for facts placed in the middle of long contexts. Every practical recommendation about placing critical information at the start or end of context traces to this paper.

The original RAG paper. Introduced the pattern of combining a retrieval component (dense passage retrieval) with a sequence-to-sequence generator. Established the architecture that every modern RAG system follows, though implementations have evolved substantially.

Introduced the ReAct prompting pattern (Reason → Act → Observe → repeat) that became the standard for tool-using agents. Every major agent framework (LangChain, LlamaIndex) implements this loop. The clearest exposition of why interleaving reasoning with tool calls outperforms either alone.

Introduced Constitutional AI - using a set of principles to have the model critique and revise its own outputs for safety, rather than relying entirely on human feedback. Reduced the need for human labelling of harmful content. The core technique behind Claude's safety training.

Demonstrated that model performance scales predictably as a power law with model size, data size, and compute. The paper that justified the "just scale it" approach to AI and influenced trillion-dollar investment decisions in AI infrastructure.

Showed that reasoning capability could emerge from pure reinforcement learning on verifiable problems - without supervised fine-tuning on human-written reasoning chains. The model learned to produce chain-of-thought reasoning spontaneously. Significant because it matches o1-class reasoning at dramatically lower training cost.

The vLLM paper. Introduced PagedAttention - paged KV cache management that reduces GPU memory waste and enables dramatically higher serving throughput. If you run LLMs in production, this paper explains the core technique enabling cost-efficient serving.