AUSO: From skill internalization to action-level use¶
Fidelity: core-mechanism mini-suite. Exercises skill internalization, action-distribution JSD, bounded utilization and reuse phase switching.
Paper information¶
| Field | Details |
|---|---|
| Paper | Paper or original source |
| Institution | University of Science and Technology of China |
| First public release | 2026-08-21 (arXiv v1) |
| Author code | Yes: Action-Skill. |
| Local adapter or method key | auso |
| Local reproduction code | src/auto_research/agent_research/latest_20260824.py |
Paper method and evidence¶
External skill retrieval costs context, but complete internalization loses task-specific choice; trajectory-level success cannot tell which action benefits. A skill-conditioned teacher first internalizes reusable skills, then the system explores with outcome feedback. Action-level JSD between with-skill and without-skill policies yields bounded GRPO advantage weights. Bounds prevent flipping the original trajectory advantage's sign.
flowchart LR
A[Skill-conditioned teacher] --> B[Internalization]
B[Internalization] --> C[Outcome-driven exploration]
C[Outcome-driven exploration] --> D[Action-level JSD]
D[Action-level JSD] --> E[Bounded advantage adjustment]Original paper figure¶
Figure 2 shows the main stages. From the original paper; copyright remains with the authors.
Core equation¶
The bound is \(1-\beta\le w_t\le1+\beta\).
Reported results¶
ALFWorld ID/OOD success is 94.3/67.9 vs 85.1/54.7 without action-level internalization. SearchQA overall is 47.5 vs SkillRL 47.1. The paper also tests WebShop; no industrial A/B.
Local reproduction¶
On matched PlanBench-mini (120 episodes), SkillRise vs AUSO joint success stays 1.0 and average cost goes 0.6550→0.5955 (−9.09%). AUSO creates 12 skills, reuses them 108 times and performs 108 action updates. Stable metrics · batch index. No ALFWorld/WebShop/SearchQA, LLM token rollouts or full GRPO; the deterministic mini-suite validates the lifecycle and reweighting control flow, not paper success rates.
Unified isolated L2.1 capability result¶
On held-out toolroute-l2.1-v1, 60 episodes per seed 42/43/44, joint success is 0.7556, plan step F1 0.7887, recovery 0.6667 and mean cost 4.5192. Cross-episode skill memory maintains success while reducing decision cost. Metrics · Protocol.