PRACTICE: From Experience to Expertise in Self-Evolving Embodied Agents
Ziyi Bai, Siqi Li, Tinglei Huang, Börje F. Karlsson
cs.LG
Quick verdict
Skill learner evolves a frozen executor's library, +9.7/+2.6 pts over experience baselines; solid design, but internal number conflicts and no variance tests.
Summary
The paper introduces PRACTICE, a framework in which a small trainable skill learner (Qwen3-VL-8B) maintains a persistent library of skill cards (action-pattern sequences plus usage specifications) for embodied executors whose parameters stay frozen. Instead of fixed prompt-engineered maintenance workflows, the learner emits structured ADD/REVISE/MERGE/REMOVE edits that are hierarchically consolidated. Training proceeds in stages: BPE-style pattern discovery seeds the initial library; oracle-grounded SFT (Stage 0) teaches generation, editing, and consolidation; failure-aware SFT (Stage 1) contrasts successful and failed same-task trajectories from eight executors; and Skill OPD (Stage 2) applies top-32 forward-KL distillation from a frozen Qwen3-VL-32B teacher on student-generated edit prefixes, with no reward or policy-gradient term. With a Qwen3-VL-32B executor, PRACTICE reaches 49.7% average success on EB-ALFRED and 58.3% on EB-Habitat, exceeding the strongest experience-augmented baselines (MemCompiler, 40.0 and 55.7) by 9.7 and 2.6 points; with a GPT-5.4 executor it reports 67.7/72.0. Stage-wise ablations attribute +1.6/+3.0/+4.4 points to Stages 0-2, and failure-aware learning beats a success-only control by 2.6 points. Main limitations: internal numerical inconsistencies (EB-Habitat paired gains of +8.7 in Appendix C.3 vs +21.0 implied by Table 1; 40.0% vs 49.7% in Appendix C.2), no seeds, significance tests, or variance estimates, and cross-protocol comparisons underlying state-of-the-art claims.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 3
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A trainable skill learner can maintain and refine a persistent skill library for frozen embodied executors, replacing fixed prompt-engineered update workflows.
- Under the same Qwen3-VL-32B executor, Practice exceeds the strongest experience-based baselines by 9.7 points on EB-ALFRED (49.7% vs 40.0%) and 2.6 points on EB-Habitat (58.3% vs 55.7%).
- Contrasting successful and failed same-task trajectories from heterogeneous executors yields failure-aware skills worth +2.6 average points over a success-only control (Table 4).
- Skill OPD, top-32 forward-KL distillation on student-generated edit prefixes without reward or policy-gradient terms, delivers the largest learned-stage gain (+4.4 points, Table 3).
- Practice-augmented GPT-5.4 reaches 67.7/72.0 on EB-ALFRED/EB-Habitat, claimed as state of the art on EB-Habitat among all methods.
Strengths
Learnable update policy replaces prompt-engineered curation
Library maintenance is cast as a trainable policy emitting structured ADD/REVISE/MERGE/REMOVE batch edits with hierarchical consolidation (Sec 3.1-3.2, Eq. 5-7, Algorithm 1), a genuine conceptual step beyond the fixed workflows of prior skill/memory systems surveyed in Sec 2.2.
Executor-controlled headline comparison checks out
Within the same-executor block of Table 1 (Qwen3-VL-32B), Practice's 49.7 (EB-ALFRED) and 58.3 (EB-Habitat) exceed the strongest experience baseline MemCompiler (40.0 and 55.7) by exactly the claimed 9.7 and 2.6 points.
Clean stage-wise ablation chain
Unusually detailed skill-level analysis
Reproduction-friendly training specification
Weaknesses
Internal numerical contradictions
No uncertainty quantification
Protocol-mixed state-of-the-art claims
Specification inconsistencies impede reimplementation
A same-executor competitor outperforms Practice in the paper's own Table 2
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
PRACTICE (arXiv 2608.30760) targets experience-augmented embodied agents: instead of post-training executors or maintaining skill/memory libraries via fixed prompt pipelines, it trains a compact skill learner (Qwen3-VL-8B) that reads executor trajectories plus the current skill library and emits structured batch edits (ADD/REVISE/MERGE/REMOVE), hierarchically consolidated into the next library version. The executor stays frozen, positioning the contribution as complementary to post-training. Central claims: (1) the learner-library loop constitutes a learnable update policy rather than a prompted heuristic (Sec 3.2, Eq. 7); (2) it beats the strongest experience-based baselines by 9.7 points on EB-ALFRED (49.7% vs MemCompiler 40.0%) and 2.6 on EB-Habitat (58.3% vs 55.7%) under the same Qwen3-VL-32B executor; (3) failure-aware supervision from same-task success/failure contrast is worth +2.6 over a success-only control (Table 4); (4) Practice-augmented GPT-5.4 sets SOTA on EB-Habitat (72.0) among all methods.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is PRACTICE in one sentence?
PRACTICE is a framework that trains a small 'skill learner' model (Qwen3-VL-8B) to automatically extract, add, revise, merge, and remove entries in a persistent skill library used by a frozen embodied agent, so the agent improves from its own interaction experience without retraining the agent itself.
What results does PRACTICE achieve?
With a Qwen3-VL-32B executor it reports 49.7% average task success on EB-ALFRED and 58.3% on EB-Habitat, exceeding the strongest prior experience-based baselines (MemCompiler) by 9.7 and 2.6 percentage points; with a GPT-5.4 executor it reports 67.7 and 72.0.
How is the skill learner trained?
In three stages: supervised fine-tuning on successful oracle trajectories (skill generation, editing, consolidation); failure-aware fine-tuning contrasting successful and failed trajectories from eight different executor models on the same tasks; and on-policy distillation (Skill OPD) matching a frozen Qwen3-VL-32B teacher's top-32 token distributions on the learner's own generated edits, using forward KL only.
Is code or data available for PRACTICE?
The paper points to a public project page (baai-agents.github.io/PRACTICE) as 'project resources', but the text does not explicitly confirm a code release, prompt templates, or edit schemas; training used proprietary teacher models (Qwen3.7-Max, GPT-5.4-class).
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-16 are image content; only captions and surrounding prose were extractable, so figure-level claims (e.g., Fig 8 bubble positions, CI widths) could not be independently verified.
- Cell alignment in Tables 1-2 was partially garbled by PDF extraction; all cited values were cross-checked against in-text averages (e.g., 49.7/58.3, +9.7/+2.6, stage sums in Table 3) before use.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.