Microsoft Research's Resource2Skill distills tutorial videos, code, articles and artifacts into a hierarchical skill library, gaining 11.9 points across seven domains.

Purdue's NPO claims rich rollout feedback can replace GEPA-style search — but its 'beats GEPA with fewer rollouts' claim hides an unresolved confound.

vLLM claims 14.6-106x cost savings for agentic serving, a number with a caveat. The real engineering: KV cache paging, DCP/PCP/DEP parallelism, and three failure cases.

TypeSafe AI's Jev claims 193.6x faster and 444.6x cheaper LLM classification with no hallucinations, but independent tests found only about 5x-25x.

Dream-RSI replays discovery history to cheaply test exploration policies, but its 'World Model' framing overstates what a pure replay mechanism can actually do.

Do more in-context examples always help? Many-Shot CoT-ICL finds negative scaling, a similarity-retrieval trap, and order sensitivity on reasoning tasks.

How RecursiveMAS swaps text between AI agents for continuous vectors, making multi-agent systems differentiable end-to-end with just 0.31% trainable parameters.