Six learned controllers scheduling compute for frozen large models, cutting LLM calls 33-65% or robot latency 34% without retraining the target model weights.

KV cache grows with every token. Seven 2026 architectures use fixed-size recurrent state instead. Analysis: GDN-2, Kimi Linear, Mamba-3, HOLA, CKDA, KDN.

Six MoE inference runtimes stream expert weights from SSD, reducing peak active memory to 3 GiB while running 35B models at 20+ tok/s on consumer hardware.

Prime Intellect’s Prime Sandboxes replace gVisor containers with full Linux microVMs, fixing silent reward signal errors in agentic RL training at scale.

HySparse2 cuts prefill FLOPs 5.02x at 1M tokens and reduces KV cache from 12 GB to 2.69 GB, using two-level KV sharing in a split Transformer for agents.

Google builds a hypernetwork that generates personalized LoRA adapters from user context via forward passes, requiring no gradient descent on the device.

Harness-Zero distills specialized agent harness behavior into model weights, raising macro success from 23.3% to 44.3% after the harness is removed.

AMD Hyperloom autonomously profiles and tunes inference stacks on AMD GPUs, delivering a median 1.73x gain across 16 workloads with no human in the loop.

RRSI from Google Cloud AI Research prevents harness overfitting via edit budgets and leakage screening, gaining 4.7 OOD points on 30% fewer policy tokens.