AI Research Paper Analyzer
First-principles analysis, methodology audits, fallacy identification, and proposed improvements for the most influential AI papers from 2016–2026. Latest first.
AI Safety · Political Persuasion
July 2025 / Published 2026The Levers of Political Persuasion with Conversational AI
Three large-scale experiments (N=76,977) across 19 LLMs and 707 UK political issues. Post-training (+51%) and prompting (+27%) drive more persuasion than model scale. More persuasive AI systematically produces less accurate information.
Hackenburg, Tappin, Hewitt et al. · UK AI Security Institute, Oxford, LSE, Stanford, MIT · arXiv:2507.13919
Large-scale and methodologically ambitious, but undermined by UK-only sampling, demand characteristics in crowdsourced setting, and circular AI fact-checking.
Reasoning · Reinforcement Learning · Open Source
January 2025DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning
DeepSeek-R1 achieves reasoning performance comparable to OpenAI-o1 using RL without supervised fine-tuning — open-sources six distilled models from 1.5B to 70B. 5,176+ citations. Triggered a ~$600B NVIDIA market cap loss on release day.
DeepSeek-AI · China · arXiv:2501.12948
Technically rigorous, fully reproducible with open-source weights, and independently validated. Minor concerns about benchmark saturation and RL reward signal opacity.
Foundation Models · Open Source · Scaling
July 2024The Llama 3 Herd of Models
Meta's Llama 3 — trained on 15T+ tokens with instruction tuning and RLHF — matches or exceeds GPT-4 on many benchmarks while remaining open-source. The most cited AI paper of 2024 and the most influential open-source model release since GPT-2.
Meta AI · Aaron Grattafiori et al. · arXiv:2407.21783
Comprehensive technical report with strong engineering documentation and genuinely reproducible results through open weights. Constrained by self-reported benchmarks and limited safety evaluation transparency.
Foundation Models · Multimodal · RLHF
March 2023GPT-4 Technical Report
OpenAI's GPT-4 achieves human-level performance on professional and academic benchmarks — top 10% on the bar exam, top 11% on SAT — and demonstrates emergent multimodal capabilities. Defines the modern era of frontier AI deployment.
OpenAI · arXiv:2303.08774
Groundbreaking capability demonstration severely undermined by deliberate opacity on architecture, training data, compute, and safety methodology — arguably the least scientifically transparent high-impact paper of the decade.
Alignment · RLHF · Instruction Following
March 2022Training Language Models to Follow Instructions with Human Feedback (InstructGPT)
A 1.3B InstructGPT model outperforms a 175B GPT-3 model on human preference — demonstrating that alignment with human intent via RLHF dramatically outweighs raw scale. The foundational paper of the modern AI alignment era.
Long Ouyang, Jeff Wu, Xu Jiang et al. · OpenAI · arXiv:2203.02155
Rigorous experimental design with strong ablations and honest reporting of limitations. The foundational RLHF methodology paper that shaped every subsequent aligned LLM. Limited by labeler demographic narrowness.
Protein Folding · Structural Biology · Nobel Prize 2024
July 2021Highly Accurate Protein Structure Prediction with AlphaFold
DeepMind's AlphaFold2 solves the 50-year protein folding problem — predicting 3D protein structure from amino acid sequence with atomic accuracy. Awarded the 2024 Nobel Prize in Chemistry. Predicted structures for 200M+ proteins across virtually every known species.
John Jumper, Richard Evans et al. · DeepMind · Nature 596, 583–589 (2021)
Arguably the most impactful application of deep learning to a hard scientific problem in history. Peer-reviewed in Nature, independently verified at scale, and awarded the Nobel Prize. Assessed at 10 with the single note that its transformer architecture's theoretical justification for biological reasoning remains incompletely understood.
Scaling · Few-Shot Learning · Foundation Models
May 2020Language Models are Few-Shot Learners (GPT-3)
GPT-3's 175B parameters enable in-context learning from just a few examples without gradient updates. Established scaling laws empirically, launched the foundation model paradigm, and made LLMs commercially viable. 28,000+ citations.
Tom Brown, Benjamin Mann et al. · OpenAI · arXiv:2005.14165
Exceptional empirical rigor with comprehensive evaluation across 30+ benchmarks and honest reporting of failure modes. Primary limitation is underdocumented training data and closed-access deployment.
Pre-Training · Transfer Learning · NLP
October 2018 / NAACL 2019BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
BERT's bidirectional masked language modeling achieves SoTA on 11 NLP benchmarks simultaneously — demonstrating that pre-training on large corpora transfers to every downstream NLP task. Made pre-training the dominant NLP paradigm. 40,000+ citations.
Jacob Devlin, Ming-Wei Chang, Kenton Lee, Kristina Toutanova · Google AI · arXiv:1810.04805
Methodologically near-flawless: reproducible, open-sourced, rigorously ablated, and the most broadly impactful NLP paper of its decade.
Pre-Training · Transfer Learning · Generative Models
June 2018Improving Language Understanding by Generative Pre-Training (GPT-1)
OpenAI's GPT-1 demonstrates that unsupervised causal language model pre-training on large text followed by task-specific fine-tuning achieves strong NLP performance — the first proof-of-concept for the GPT paradigm that would reshape AI.
Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever · OpenAI · OpenAI Blog
Foundational paper introducing the pre-train + fine-tune paradigm with rigorous multi-task evaluation. The fundamental insight was simultaneously explored by BERT from a different architectural direction.
Transformers · Attention · Architecture
June 2017Attention Is All You Need
The transformer architecture replaces recurrence and convolution with self-attention — enabling parallelizable training at scale. The foundation of every major AI system: GPT, BERT, DALL-E, AlphaFold, Whisper, and virtually all modern AI. Most cited AI paper of all time: 110,000+ citations.
Ashish Vaswani, Noam Shazeer, Niki Parmar et al. · Google Brain & Google Research · arXiv:1706.03762
One of the most impactful scientific papers in computing history — rigorous, reproducible, and universally adopted. The universal architecture of AI.
Reinforcement Learning · Game Playing · Search
January 2016Mastering the Game of Go with Deep Neural Networks and Tree Search (AlphaGo)
DeepMind's AlphaGo defeats European Go champion Fan Hui 5-0 — the first computer program to defeat a professional Go player without handicap. Demonstrates that deep RL combined with Monte Carlo tree search can master games of superhuman complexity. Nature cover paper.
David Silver, Aja Huang, Chris J. Maddison et al. · DeepMind · Nature 529, 484–489 (2016)
Scientifically exemplary: peer-reviewed in Nature, fully described methodology, independently verifiable results through public matches, and honest reporting of computational requirements. The gold standard for high-stakes AI capability papers.