Foundational Canon

Landmark AI Papers

10 breakthrough publications that defined modern machine learning, attention, and RLHF.

architecture

mHC: Manifold-Constrained Hyper-Connections

DeepSeek Research

Replaces the residual connection paradigm (x + f(x)) used in every major transformer for the past decade. Projects residual paths onto Birkhoff Polytope manifold — neither explode nor collapse at any depth.

Read on arXiv →
architecture

Less is More: Recursive Reasoning with Tiny Networks

Jolicoeur-Martineau et al. (Samsung SAIL)

A 7M parameter model beats DeepSeek-R1, Gemini 2.5 Pro, and o3-mini on ARC-AGI benchmarks. Proves recursive reasoning can compensate for massive parameter counts.

Read on arXiv →
architecture

Holographic Transformers

Various

Physics-inspired attention using wave interference — builds holographic memory with cosines, sines, and phase modulation instead of standard dot-product attention.

Read on arXiv →
reasoning

Recursive Self-Aggregation for Deep Thinking

RSA Team

Test-time scaling via evolutionary algorithms — combines parallel and sequential scaling. Enables Qwen3-4B to compete with DeepSeek-R1 on AIME and LiveCodeBench.

Read on arXiv →
training

Agentic Context Engineering (ACE)

Stanford & SambaNova

Makes fine-tuning potentially irrelevant — evolves the context instead of weights. Delta updates and grow-and-refine mechanisms. 59.5% on AppWorld (+10.6pp).

Read on arXiv →
training

Verbalized Sampling

Stanford

Recovers 66.8% of base model diversity hidden by RLHF alignment. Single prompting technique: ask LLMs to generate multiple responses with probability estimates. 2x diversity, zero safety loss.

Read on arXiv →
training

GEPA: Reflective Prompt Evolution Outperforms RL

Various (ICLR 2026 Oral)

Prompt optimization via evolutionary search outperforms GRPO by 10-20% while using 35x fewer rollouts. Also beats MIPROv2 by 10%+.

Read on arXiv →
efficiency

Drag-and-Drop LLMs: Zero-Shot Prompt-to-Weights

Liang et al. (NUS, UT Austin, Oxford)

Eliminates per-task fine-tuning — maps task prompts directly to LoRA weights via hyper-convolutional decoder. 12,000x lower overhead than full fine-tuning.

Read on arXiv →
reasoning

Categories of Inference-Time Scaling

Sebastian Raschka

Overview of inference-scaling approaches — recursive language models, chain-of-thought, tree search, and their trade-offs for improved reasoning.

Read on arXiv →
rag

PageIndex: RAG Without Vector Databases (98.7% accuracy)

Akshay Kalane

RAG approach that achieves 98.7% on FinanceBench without vector DBs, embeddings, chunking, or similarity search. Challenges the assumption that RAG = vector DB.

Read on arXiv →