August 9, 2026

Proposes a Manage-Execute-Audit (MEA) loop that separates task-state tracking from execution, using a manager to hold verified task state, a fresh-context executor per subtask, and a read-only auditor to check environment state before each new round. Reported gains include Qwen 3.7-Plus rising from 51.8% to 80.7% on WeaveBench and Claude Opus 4.7 rising from 20.0% to 34.3% on an OSWorld 2.0 subset.

Introduces Recursive Synthetic Terminal Tasks (RST), a recursive verified-synthesis framework that automatically generates over 37,000 verifiable long-horizon terminal-agent tasks, keeping instructions, environments, solutions, and verifiers consistent at roughly $0.05 per task. Task difficulty compounds across rounds, with median reference-solution length growing from 67 to 374 lines and DeepSeek-V4-Pro pass@4 dropping from 90% to 2.5% over 15 rounds.

Identifies a 'privilege illusion' failure mode in on-policy self-distillation, where a student model learns behavior dependent on training-time privileged information it cannot access at inference. Proposes Dual-Path and Dual-Source Anchoring to align reference and rollout behavior bidirectionally, improving over prior on-policy self-distillation by roughly 2-2.8 points across Qwen3 models from 4B to 32B.