Essential Papers - AI Interview Guide

Organized by topic. For each paper I've noted the key contribution and whether it's worth reading in full vs just the abstract and introduction.

Foundations

Attention Is All You Need (Vaswani et al., 2017) arxiv.org/abs/1706.03762 The original transformer paper. Multi-head attention, positional encoding, encoder-decoder architecture for translation. Historical importance is high; many implementation details have been superseded. Read: Abstract + Section 3 (Model Architecture). The rest is specific to translation.

BERT: Pre-training of Deep Bidirectional Transformers (Devlin et al., 2018) arxiv.org/abs/1810.04805 Masked language modeling + next sentence prediction for bidirectional encoders. Foundation for embedding models and classification. Encoder-only; predates GPT-style decoder models. Read: Abstract + Sections 3-4 for the pre-training approach.

Language Models are Few-Shot Learners (GPT-3) (Brown et al., 2020) arxiv.org/abs/2005.14165 Demonstrated that scale alone enables few-shot learning without fine-tuning. Introduced the prompt engineering paradigm. 175B parameter model. Read: Abstract + Section 2 (approach) + the few-shot results tables.

Attention and Efficiency

FlashAttention: Fast and Memory-Efficient Exact Attention (Dao et al., 2022) arxiv.org/abs/2205.14135 IO-aware attention implementation using CUDA tiling. Makes long-context practical. Every production LLM inference stack uses this. Read: Abstract + Section 3. The CUDA kernels are optional unless you're writing inference code.

GQA: Training Generalized Multi-Query Transformer Models (Ainslie et al., 2023) arxiv.org/abs/2305.13245 Grouped query attention reduces KV cache by 4-8x with minimal quality loss. Used in Llama 3, Mistral, most modern models. Read: Abstract + Section 2. Short and practical.

RoFormer: Enhanced Transformer with Rotary Position Embedding (Su et al., 2021) arxiv.org/abs/2104.09864 Rotary position embeddings encode relative position by rotating Q/K vectors. Better length generalization than absolute position encodings. Now standard. Read: Abstract + Section 3.2 (the RoPE formulation).

RAG and Retrieval

Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (Lewis et al., 2020) arxiv.org/abs/2005.11401 Original RAG paper. DPR retrieval + BART generation. Most production RAG systems use different components but the same architectural pattern. Read: Abstract + Sections 2-3 (model) + Section 4 (experiments).

Lost in the Middle: How Language Models Use Long Contexts (Liu et al., 2023) arxiv.org/abs/2307.03172 Demonstrates that LLMs recall information at beginning and end of long contexts better than the middle. The U-shaped recall curve. Essential for designing RAG context placement. Read: Abstract + Section 3 (experiments). The finding is simple; the graphs tell the story.

RAGAS: Automated Evaluation of Retrieval Augmented Generation (Es et al., 2023) arxiv.org/abs/2309.15217 Introduces four automatic metrics: faithfulness, answer relevancy, context precision, context recall. Now the standard RAG evaluation framework. Read: Abstract + Section 3 (metrics definitions). Essential for RAG practitioners.

Dense Passage Retrieval for Open-Domain Question Answering (Karpukhin et al., 2020) arxiv.org/abs/2004.04906 DPR: using separate BERT encoders for queries and documents. Foundation of dense retrieval. Shows dense retrieval beats BM25 on QA tasks. Read: Abstract + Sections 2-3. The training approach is the important part.

Fine-Tuning and Adaptation

LoRA: Low-Rank Adaptation of Large Language Models (Hu et al., 2022) arxiv.org/abs/2106.09685 Trains low-rank matrices alongside frozen base model weights. 0.5% of parameters, full fine-tuning quality. The standard for efficient fine-tuning. Read: Full paper. Short (8 pages) and practically essential.

QLoRA: Efficient Finetuning of Quantized LLMs (Dettmers et al., 2023) arxiv.org/abs/2305.14314 4-bit NF4 quantization + double quantization + LoRA enables fine-tuning 65B models on single GPUs. Practical guide to consumer-hardware fine-tuning. Read: Abstract + Sections 2-3 (approach) + Section 4 (results).

Direct Preference Optimization (Rafailov et al., 2023) arxiv.org/abs/2305.18290 Trains on preference pairs directly without a reward model. More stable than PPO-based RLHF. Increasingly used as RLHF replacement. Read: Abstract + Section 4 (method). The math is approachable.

RLHF: Learning to summarize with human feedback (Stiennon et al., 2020) arxiv.org/abs/2009.01325 Foundational RLHF paper on summarization. Reward model + PPO training. Pre-InstructGPT but demonstrates the approach. Read: Abstract + Sections 2-3.

InstructGPT: Training language models to follow instructions (Ouyang et al., 2022) arxiv.org/abs/2203.02155 Applies RLHF to GPT-3 to create GPT-3.5. Demonstrates that alignment via RLHF produces more helpful models even at smaller scale. Read: Abstract + Sections 2-3 (methods) + Section 4 (results).

Quantization

GPTQ: Accurate Post-Training Quantization for Generative Pre-trained Transformers (Frantar et al., 2022) arxiv.org/abs/2210.17323 Layer-by-layer quantization using Hessian second-order information. 3-4 bit quantization with minimal quality loss. Read: Abstract + Section 3 (method). Technical but accessible.

AWQ: Activation-aware Weight Quantization for LLM Compression (Lin et al., 2023) arxiv.org/abs/2306.00978 Identifies 1% of salient weights using activation magnitudes and protects them. Consistently better than GPTQ at the same bit width. Read: Abstract + Section 3.

Agents and Reasoning

ReAct: Synergizing Reasoning and Acting in Language Models (Yao et al., 2022) arxiv.org/abs/2210.03629 Reasoning + Acting loop: think about what to do, act, observe. The foundational agent pattern used everywhere. Read: Full paper. Short and essential for understanding agents.

Chain-of-Thought Prompting Elicits Reasoning in Large Language Models (Wei et al., 2022) arxiv.org/abs/2201.11903 Showing reasoning steps before answering dramatically improves performance on multi-step problems. The paper behind CoT prompting. Read: Abstract + Section 2 (method) + the few-shot examples in the appendix.

Tree of Thoughts: Deliberate Problem Solving with Large Language Models (Yao et al., 2023) arxiv.org/abs/2305.10601 Extends CoT to explore multiple reasoning paths in a tree structure. Better on hard planning problems. Read: Abstract + Section 2 (method). Useful context for understanding reasoning strategies.

Toolformer: Language Models Can Teach Themselves to Use Tools (Schick et al., 2023) arxiv.org/abs/2302.04761 Training LLMs to self-generate API calls. Pre-function-calling but foundational for understanding why tool use works. Read: Abstract + Section 3.

Evaluation

Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena (Zheng et al., 2023) arxiv.org/abs/2306.05685 Systematic evaluation of using GPT-4 as a judge. Position bias, verbosity bias, agreement with humans. Essential reading for anyone building LLM-based evaluation. Read: Full paper. Short and directly applicable.

Large Language Models are not Fair Evaluators (Wang et al., 2023) arxiv.org/abs/2305.17926 Documents position bias in LLM-as-judge: judges prefer the first presented option. Calibration techniques. Read: Abstract + Sections 3-4. Informs how to design robust LLM judges.

Inference and Scaling

Scaling Laws for Neural Language Models (Kaplan et al., 2020) arxiv.org/abs/2001.08361 Power law relationships between model size, data, compute, and loss. Foundation for understanding why bigger models are better. Read: Abstract + Section 4 (empirical scaling laws). The derivations are optional.

Efficient Large Scale Language Modeling with Mixtures of Experts (Artetxe et al., 2021) arxiv.org/abs/2112.10684 Mixture of experts for LLMs. Foundation for understanding Mixtral and similar models. Read: Abstract + Section 2.

vLLM: Efficient Memory Management for LLM Serving with PagedAttention (Kwon et al., 2023) arxiv.org/abs/2309.06180 PagedAttention: manages KV cache like OS virtual memory. 2-4x throughput improvement over standard serving. Read: Abstract + Sections 2-3. Essential for infrastructure engineers.

Small Language Models

Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone (Abdin et al., 2024) arxiv.org/abs/2404.14219 Microsoft's Phi-3-mini (3.8B) outperforms Llama 8B through heavy data curation. Demonstrates quality > quantity for training data. Read: Abstract + Section 2 (data) + benchmark results.

How to Read Papers Efficiently

For most papers, you don't need to read every section:

Abstract (2 min): What problem, what approach, what result?
Introduction (5 min): Context, motivation, main contributions
Method/Model section (10 min): How does it actually work?
Key results table or figure (5 min): Does it work?
Conclusion (2 min): What does it mean?

Skip: Related work (unless you're writing a paper), proofs, full experimental appendices.

For code-oriented engineers: many papers have companion GitHub repos with implementations that are faster to understand than the paper itself.