Apple RLTL;DR achieves 12-13% Pass@1 where GRPO stalls at 0%, extracting 17-token lessons from failed rollouts and training the task-to-insight mapping into weights.

SPLASH switches LLM attention parallelism layout live mid-request with under 0.51% overhead, improving throughput 1.3-1.73x over fixed-layout deployments on B200 GPUs.

ContextRender tracks tool-result reuse in a persistent execution graph, matching full-history performance with a 6K budget and cutting inference cost by 10.2-32.2%.

1.5-SPSA trains OPT-30B on 60 GB instead of 600 GB, using probe-space preconditioning to beat backpropagation accuracy on SST-2 in 70 steps.

UW and Meta’s Context Language Models treat context as an editable file, achieving 11.4% higher accuracy and 21.5% fewer FLOPs on deep-research tasks.

Six research systems from 2026 treat the agent harness as a separately optimizable artifact alongside model weights, cutting inference cost by up to 98.6% and raising agent success rates.

Six systems that decide what long-horizon agents are allowed to forget: ACON, Self-GC, StateComp, Forget Reasoning, PaMER, and DRSR — with exact numbers and design tradeoffs.

From separate GPU pools to different model checkpoints and custom silicon, five systems show prefill and decode splitting into distinct computational products.

Five systems let clients verify that a remote LLM ran the claimed model on their prompt, without re-running inference or taking the provider at their word.