Reasoning, software and computer use.
Official sourceAll research hubs
What can intelligence do?
Models & intelligence
Reasoning, multimodality, model architectures and the evidence behind capability claims.
12 topic maps Generated briefs under source review
Research briefs
01
Frontier Reasoning Models & Test-Time Compute
System 2 deliberate reasoning, process reward models, and inference-time search scaling
Evidence review pending
02Mixture-of-Experts (MoE) & Multi-Head Latent Attention
Sparse activation scaling, auxiliary-loss-free routing, and memory bandwidth optimization
Evidence review pending
03Context Engineering & Long-Context Architecture
Million-token context windows, needle-in-a-haystack retrieval, prompt caching, and attention sink dynamics
Evidence review pending
04Reinforcement Learning from Verifiable Rewards (RLVR)
Deterministic reward oracles, formal theorem verification, and self-directed policy optimization
Evidence review pending
05Multimodal Reasoning & Vision-Language-Action Models
Native multimodality, spatial intelligence, 4D world representations, and cross-modal attention
Evidence review pending
06Post-Training Distillation & Speculative Decoding
Knowledge transfer from reasoning teachers, speculative token drafting, and edge model optimization
Evidence review pending
07Diffusion Transformers & Generative Neural Video
DiT architectures, flow matching, spatio-temporal attention, and cinematic video synthesis
Evidence review pending
08Neural Audio Synthesis & Conversational Speech Models
Neural audio codecs, zero-shot voice cloning, expressive prosody modeling, and full-duplex conversational audio
Evidence review pending
09Sparse Attention, State Space Models & Linear Transformers
Mamba 2, RWKV, FlashAttention-3, sub-quadratic attention, and hybrid state-space transformer backbones
Evidence review pending
10Representation Engineering & Mechanistic Interpretability
Sparse Autoencoders (SAEs), concept steering vectors, dictionary learning, and circuit analysis
Evidence review pending
11Adversarial Robustness & Jailbreak Defense Architectures
Automated red-teaming, prompt injection defenses, multi-turn jailbreak mitigation, and robust alignment
Evidence review pending
12Mathematical Theorem Proving & Formal Verification AI
Lean 4, Isabelle provers, AlphaProof, AlphaGeometry, and neuro-symbolic automated deduction
Evidence review pending
Inside Models & intelligence
The labs. Selected models.
Selected model families, checked against provider documentation between 13 September 2026 and 9 October 2026. Specifications are provider-reported; local experiments remain separately dated.
Long-horizon agents, coding and reasoning.
Official source
Multimodal systems and agentic workflows.
Official sourcexAI
Configurable reasoning and tool use.
Official source