Research & policy
Academic papers, official research, regulatory material, patents, and standards are grouped together with their evidence labels intact.
arXiv cs.CL
Nov 12, 2026
MAPE evolves agent memory programs with an explicit serving-cost objective
The MAPE preprint evolves complete memory programs through a shared lifecycle interface and an iterative diagnosis-and-proposal process. On LoCoMo and LongMemEval, the authors report absolute accuracy gains of up to 12 percentage points over manually designed baselines at 35% of their serving cost. Ablations identify retaining raw conversations as retrievable evidence and composing evidence according to query intent as useful patterns. These results describe the authors’ tested systems and serving-cost assumptions.
- The authors evaluate evolved memory programs on LoCoMo and LongMemEval.
Why it mattersAgent memory design affects evidence retention and operating cost. Replication with matched models, datasets and cost accounting would test transfer to research and financial workflows.
arXiv cs.AI
Nov 12, 2026
Decision-model study separates acting ability from teaching value
A preprint compares decision models and general language models as actors and retrospective judges, then uses JEV to teach a Qwen2.5-1.5B policy. The authors report success increasing from 72.8% to 82.9% on ALFWorld and from 56.8% to 71.4% on WebShop against GRPO. Their experiments suggest that a model with weak action selection can still provide useful learning signals after seeing a complete trajectory. The findings retain the tested environments and policy-training setup.
- The preprint evaluates decision models as actors, judges and policy teachers.
Why it mattersEvaluation and teaching roles may have different model requirements. Agent-shopping teams can test whether retrospective action judgments improve held-out task completion and spending control.
arXiv cs.LG
Nov 12, 2026
Metal Sandbox measures scheduling gains for local GPU coding agents
Metal Sandbox combines isolated agent workspaces, native command execution and command-level scheduling across interconnected Macs. Controlled experiments on four M4 Mac mini nodes report command-throughput gains of 1.68–2.05 times under concentrated demand. A replay of sixteen measured traces reports 18.8% lower mean tool waiting and 22.3% lower batch elapsed time, assuming unchanged phase durations. A separate sustained-throughput estimate assumes continuously replenished sessions. The authors release the infrastructure and Harbor integration.
- The authors evaluate command scheduling on four M4 Mac mini nodes.
Why it mattersLocal agent economics depend on scheduling as well as kernel speed. Measured batch results and trace-based estimates supply distinct inputs for capacity planning.
Federal Reserve
Nov 13, 2026
Fed stability report documents AI financing exposure and surveyed risk concerns
The November Financial Stability Report says advanced-computing debt issuance exceeded $680 billion through October 2026, about 31% of nonfinancial investment-grade issuance. It estimates outstanding related public debt above $1.1 trillion and describes complex data-center obligations and financial-sector links. In a September–October survey of 23 market contacts, 87% cited AI among salient risks, compared with 50% in the spring survey of 20 contacts. Survey responses express participant concerns. The report generally uses information available through October 28.
- The official report includes a data-center leverage analysis and a survey of 23 market contacts.
Why it mattersThe report connects AI infrastructure investment with debt markets and potential loss transmission. Borrower concentration, contingent obligations and refinancing exposure provide concrete monitoring questions.