July 9, 2026

Using a fully transparent rewrite-grammar environment, researchers show RL post-training doesn't merely amplify base capabilities — it actively constructs novel compositional strategies, both sequential (collapsing ordered steps) and parallel (combining independent operations). RL concentrates exploration into valid reusable structure; rejection fine-tuning produces many invalid shortcuts. ICML 2026 Compositional Learning Workshop. arXiv:2607.07646.

Introduces institutional red-teaming: holding agent model weights constant while varying deployment rules to measure causal effects on collective safety. Tested across 228 contexts and seven model populations. Key findings: (1) deployment rules shift fatality rates by 22–58 percentage points; (2) no universally safe default exists; (3) anonymizing agents in rule text reduced targeted elimination from 81% to 22%, though agents eventually re-inferred identity through observed patterns. arXiv:2607.07695.

A survey of 1,250 arXiv papers (2024–2026) on how AI systems participate in their own improvement. Proposes a verification-strength hierarchy: formal verifiers > process reward models > LLM judges > rubrics > intrinsic self-assessment. Shows improvement quality correlates with this hierarchy. Identifies failure modes including self-confirming loops and model collapse. Flags research direction-setting as the remaining human bottleneck. arXiv:2607.07663.