PEARL turns the first divergence between two same-history branches with different outcomes into a reusable decision memory, strengthens only the negative advantages of recurring low-reward decisions, and reaches 97.9% mean success on ALFWorld at 4B, 2.7 points above the strongest baseline.
Letting a cloud model write the program while a small on-device model answers bounded questions about user data that appear only at run time improved a code-only agent's success rate on AppWorld in all six backbone-model and test-partition combinations, by up to 14.9 percentage points, with user data staying on the device.
This work systematically studies how agentic post-training interacts with MoE expert selection, observing that off-the-shelf MoE routers already show operation-grouped specialization (e.g., READ, UPDATE) that standard RL perturbs; the authors propose a hierarchical routing control framework that aligns turn-level expert selection with operation labels via mutual information, regularizes token-level adjacent consistency with a gap threshold, and adds entropy gating for stability, improving success rates by over 10 points across PPO, GRPO, LOOP, and GiGPO on AppWorld and AutomationBench while also improving inference efficiency.
ControlScope compares three nested revision permissions from the same public execution state—continuing the current program (Keep), editing only the next tool call's data arguments (Arg), and replacing the unfinished workflow (Full)—evaluating one-time and repeated reviews across filesystem tasks, ALFWorld, and AppWorld, and finds that Full completes 15–16 of 20 filesystem tasks versus 13 for Keep across two source programs and three reasoning-reviewer draws, 10–13 versus 13 across four fast draws, scores 85/86/87 on 87 ALFWorld tasks and 134/134/127 on 134 tasks, shows small net differences on 585 AppWorld V1 official-test task instances, while frozen replays expose viable agent-written replacements interrupted by later revision.
The work introduces Harness-Zero: an agent-as-harness setup in which a harnessing agent reviews and minimally rewrites the student agent's proposals at its response boundary, translating guidance from an evolved harness (tools, middleware, skills, memory) into executable supervision inside the target harness's action space, followed by SFT on the reviewed trajectories; across SpreadsheetBench Verified, AppWorld, and USPTO retrosynthesis, training-free agent-as-harness averages 81.1% over six benchmark-model settings versus 78.1% for code-as-harness, while the distilled Qwen3.5-9B under a minimal target harness alone raises the macro average from 23.3% to 44.3%, surpassing the 41.7% of the base model with the evolved harness still attached, and recovers 82.
You’re all caught up