ImageNet Classification with Deep Convolutional Neural Networks
A deep convolutional network demonstrated how data, GPU computation and learned representations could transform visual recognition.
THE RESEARCH ATLAS
How we arrived here.
Language and reasoning. Perception and generation. From models to action.
Countries and regions follow institutional sources; cross-border work stays visible.
Research tracks2017201820192020202120222023202420252026
Language & reasoning
Perception & generation
Robotics & control
DeepSeek
DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning
Reinforcement learning is used to develop reasoning behavior, with additional pipelines for usable models and distillation.
Selected works by source date. Lines do not imply causality.Select a node · Scroll horizontally on smaller screens
Learning to see
Deep convolutional networks make representation learning a central approach to visual recognition.
Attention becomes architecture
Transformer sequence modeling helps establish an architecture shared by later language and multimodal systems.
Attention Is All You Need
The Transformer replaces recurrent sequence processing with attention, creating a foundation for many later language and multimodal models.
BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
Masked-language pretraining learns bidirectional representations that can be adapted to language-understanding tasks.
Scale & generation
Few-shot language modeling and diffusion offer different routes to generative capability. These research paths develop in parallel.
Language Models are Few-Shot Learners
GPT-3 studies how larger autoregressive language models perform new tasks through examples in their input context.
Denoising Diffusion Probabilistic Models
A generative model learns to reverse a gradual noise process, establishing an influential approach to image synthesis.
Learning Transferable Visual Models From Natural Language Supervision
Contrastive training aligns images with natural-language descriptions, enabling transfer to new visual categories.
Learning from feedback
Instruction following and preference learning change how pretrained models are adapted for human use.
Training language models to follow instructions with human feedback
Human demonstrations and preference feedback are used to improve how a language model follows instructions.
Constitutional AI: Harmlessness from AI Feedback
A set of written principles guides critique, revision and AI-generated preference feedback during alignment.
Direct Preference Optimization: Your Language Model is Secretly a Reward Model
A preference-based objective optimizes a language model directly, avoiding a separate reward-model training stage.
2 research records
From NVIDIA, DeepSeek. Grouped by the source publication date; document types are identified separately.
Cosmos World Foundation Model Platform for Physical AI
A platform of video curation, tokenizers, pretrained world models and post-training for physical-AI applications.
DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning
Reinforcement learning is used to develop reasoning behavior, with additional pipelines for usable models and distillation.
2 research records
From Google, NVIDIA. Grouped by the source publication date; document types are identified separately.
Gemini Robotics: Bringing AI into the Physical World
Gemini-based models address robot manipulation and embodied reasoning, connecting multimodal understanding with physical control.
GR00T N1: An Open Foundation Model for Generalist Humanoid Robots
A coupled vision-language and diffusion architecture combines reasoning with continuous robot-action generation.
2 research records
From Huawei, Physical Intelligence. Grouped by the source publication date; document types are identified separately.
Huawei Qiankun ADS 4: World Engine and World Behavior Model (WEWA)
Huawei's official WEWA description links generated scenarios in a cloud world engine with an on-vehicle world behavior model.
π0.5: a Vision-Language-Action Model with Open-World Generalization
A vision-language-action system studies generalization across environments and tasks using varied sources of supervision.
1 research records
From Qwen. Grouped by the source publication date; document types are identified separately.
Qwen3 Technical Report
Dense and mixture-of-experts models integrate thinking and non-thinking behavior, with configurable inference-time reasoning.
1 research records
From Hugging Face. Grouped by the source publication date; document types are identified separately.
SmolVLA: A Vision-Language-Action Model for Affordable and Efficient Robotics
A compact vision-language-action model explores accessible robot learning and an efficient inference pipeline.
1 research records
From Google. Grouped by the source publication date; document types are identified separately.
Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities
An industrial model-family report covering reasoning, multimodal input, long contexts and agent-related evaluations.
2 research records
From OpenAI, AI2. Grouped by the source publication date; document types are identified separately.
gpt-oss-120b & gpt-oss-20b Model Card
Open-weight reasoning models are described through their architecture, training, capabilities and safety evaluations.
MolmoAct: Action Reasoning Models that can Reason in Space
Depth-aware perception, spatial planning and low-level action prediction are organized into an explicit action-reasoning pipeline.
2 research records
From NVIDIA, Google. Grouped by the source publication date; document types are identified separately.
Alpamayo-R1: Bridging Reasoning and Action Prediction for Generalizable Autonomous Driving in the Long Tail
Connects driving reasoning and action prediction through Cosmos-Reason and a diffusion trajectory decoder, studying reasoning-action consistency.
Gemini Robotics 1.5: Pushing the Frontier of Generalist Robots with Advanced Embodied Reasoning, Thinking, and Motion Transfer
The report combines embodied reasoning, interleaved thinking and motion transfer across robot embodiments.
1 research records
From Qwen. Grouped by the source publication date; document types are identified separately.
Qwen3-VL Technical Report
The Qwen vision-language family integrates text, image and video understanding across several model sizes.
1 research records
From DeepSeek. Grouped by the source publication date; document types are identified separately.
DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models
Sparse attention and post-training methods are combined to improve efficiency, reasoning and agent-oriented behavior.
2 research records
From Unitree, Qwen. Grouped by the source publication date; document types are identified separately.
UnifoLM-VLA-0: A Vision-Language-Action (VLA) Framework under UnifoLM Family
Continues training on robot manipulation data, connecting visual-language understanding and spatial instructions to action prediction.
Qwen3-ASR Technical Report
A speech-recognition family and a forced-alignment model address multilingual transcription and timestamp prediction.
2 research records
From Kimi, OpenAI. Grouped by the source publication date; document types are identified separately.
Kimi K2.5: Visual Agentic Intelligence
Studies joint visual-language training and multi-agent task decomposition in an open multimodal model.
Scaling social science research
An open-source toolkit uses language models to help turn qualitative text and image data into measurable variables.
3 research records
From Anthropic, academic researchers, Qwen. Grouped by the source publication date; document types are identified separately.
A “diff” tool for AI: Finding behavioral differences in new models
An interpretability research overview explores ways to identify behavior differences between AI models beyond predefined tests.
UniDex: A Robot Foundation Suite for Universal Dexterous Hand Control from Egocentric Human Videos
A dataset, VLA policy and capture setup connect human demonstrations to control across different dexterous hands.
Qwen3-Coder-Next Technical Report
An open-weight model for coding agents explores strong training recipes with a small active parameter footprint.
2 research records
From XPeng, Qwen. Grouped by the source publication date; document types are identified separately.
XPeng X-World: Controllable Multi-View Generative World Model for Autonomous Driving
Describes a controllable multi-view world model used for driving simulation, scenario generation, and evaluation.
Qwen3.5-Omni Technical Report
The report explores an omni-modal model family and training over heterogeneous visual, textual and audio data.
2 research records
From Qwen. Grouped by the source publication date; document types are identified separately.
Qwen-Image-2.0-RL Technical Report
A diffusion-model post-training pipeline combines task-specific rewards, reinforcement learning and on-policy distillation.
Qwen-RobotWorld Technical Report: Unifying Embodied World Modeling through Language-Conditioned Video Generation
A language-conditioned video world model predicts future visual trajectories across manipulation, navigation and driving settings.
1 research records
From xAI. Grouped by the source publication date; document types are identified separately.
Introducing Grok 4.6
An official technical release describing supplemental training, regenerated SFT trajectories and agentic reinforcement learning.
4 research records
From Anthropic, Unitree, Qwen, DeepSeek. Grouped by the source publication date; document types are identified separately.
What work can robots do?
A task-level analysis of robot exposure that distinguishes demonstrated capability in limited settings from general deployment.
UnifoLM-WLA-1.0: One Model Driven, Whole-Body Coordination
Connects embodied understanding, interaction-centric world modeling and action learning for tabletop and whole-body humanoid manipulation.
Qwen-Audio-Agent Technical Report
A foreground-background architecture separates voice interaction from asynchronous task execution, coordinating authorization, state and delivery.
DeepSeek-V4.1-Flash: smarter, faster, more efficient
Official release introducing an asymmetric causal encoder-decoder architecture and linking its technical report and model resources.
10 research records
From Anthropic, OpenAI, Google, Deep Robotics, academic researchers. Grouped by the source publication date; document types are identified separately.
An alignment assessment of the UK AISI cybertesting incident
An alignment assessment of model actions that exceeded the intended scope of a cybersecurity evaluation.
Using thousands of AI agents to uncover new leads in biological data
A research overview of distributed agents linking large biological datasets with scientific literature. Proposed connections still require experimental follow-up.
Advancing coding in low-resource programming languages
Official research discussion of improving coding capabilities in languages with less available training material.
Scaling Misuse Detection with Self-Assessment
Models assess their own conversations as an additional misuse-detection layer, reusing cached context to limit overhead.
Gemini 4 Argon — Model Card
Official model documentation covering evaluation, limitations and safety considerations. The date records the listed model-card update.
PointWalker: Mapless Quadruped Locomotion Directly From Raw LiDAR Point Clouds
Fuses LiDAR point clouds and proprioception through attention, connecting terrain perception with a reinforcement-learned quadruped policy.
Investigating unintended model actions in our evaluations and internal use
A report of unintended actions observed during model evaluation and internal use, with discussion of safeguards.
Advancing computer use with Ironclad
An official account of progress in computer-use agents. Read the original methods and evaluations separately from deployment claims.
Nano Banana 2.1 — Model Card
Official documentation for an image-and-text model, including inputs, outputs, architecture dependencies and safety evaluation.
FACT: A Multimodal Force-Action-Track Foundation Model for Human-to-Robot Dexterity
Human and robot demonstrations are combined with force-aware planning for dexterous manipulation.