AGI-Papers
A curated, actively updated archive of research paper reviews toward artificial general intelligence, organized around agents, architecture, training, evaluation, RAG and knowledge, and on-device AI.
AGI-Papers collects in-depth reviews of AI research papers, mostly written first for the author's social-media audience and then archived here, sometimes including pre-release insights or early notes before wider publication. Entries are organized into themes including autonomous agents and planning frameworks, LLM architecture innovations such as Transformers, Mamba and mixture-of-experts, pre-training data and scaling laws, post-training methods such as RLHF, DPO and GRPO, evaluation benchmarks and methodology, retrieval-augmented generation and knowledge graphs, on-device and edge AI, and industry trends.
Recent reviews cover topics such as a multi-agent architecture for parallel long-context reading, training small coding agents with reinforcement learning instead of large-model distillation, why frontier models' multi-turn conversations degrade after several turns, and surveys of self-improving agents and agent lifespan effects on long-running tasks.