127 Curated Large‑Model Papers Across 17 Research Directions – From CVPR to Nature

Keeping up with the rapid expansion of large‑model research—topics such as VLA unified architecture, world models, latent‑space inference, synthetic data, and autonomous agents—has become increasingly difficult as papers flood top venues like NeurIPS, CVPR, ICML, and Nature.

To address this, the authors manually selected 127 papers from six premier conferences and journals, ensuring each entry includes a PDF, GitHub repository, and project URL, all accessible through a dedicated web page. The collection spans classic milestones (e.g., CLIP, Decision Transformer, RT‑2, Eureka) and the latest frontier works.

Technical Chain Organization

The papers are arranged according to a complete technical pipeline: perception → decision → execution → optimization → safety.

Perception & Modeling (10‑11 papers) : native unified multimodal models, world models for autonomous driving and robotics, and multimodal world models based on Transformers.

Decision & Execution (13‑12‑6‑4 papers) : VLA models linking perception to execution, agent systems for rare‑disease diagnosis and mobile control, direct LLM‑driven decision making, and indirect LLM‑assisted policy learning.

Inference & Optimization (7‑5‑10‑7‑9 papers) : implicit and explicit reward design, latent‑space inference to remove text‑generation overhead, KV‑cache compression and attention optimizations, and AI‑generated synthetic training data.

Support & Alignment (6‑7‑7‑3‑5 papers) : efficient representation learning, natural‑language translation to executable strategies, real‑time multimodal dialogue, policy explainability, and safety/alignment for controllable deployment.

Intended Audience

AI algorithm engineers seeking a quick overview of cutting‑edge techniques.

Large‑model researchers needing systematic direction‑wise reading material.

Graduate and doctoral students building a complete knowledge map from fundamentals to the frontier.

Beginners who want each paper’s summary and resource links to lower the entry barrier.

Job‑seeking technologists preparing for interviews with high‑frequency topics.

The collection was compiled over several weeks with manual screening to ensure every paper is worth a deep read. It is freely shared without any access restrictions, aiming to support anyone actively working on AI.