SkillZip Pro: Execution-Aware Dynamic Compression of Progressively Loaded Skills for Self-Evolving Agents
Xiaofan Bai, Chao Liu, Hongqiang Lin, Di Wu, Mingli Song, Xuan Jin, Xipeng Cao, Yuhong Li
cs.AI
Quick verdict
Claims evaluation-free skill-bundle compression cutting bundles 38% without quality loss; strong engineering, but evidence rests on small, private evaluations.
Summary
SkillZip Pro is an evaluation-free compressor for production agent skills, which are directory bundles loaded progressively (root at activation; references, schemas, scripts and subskills on demand). Unlike flat prompt compressors, it models the bundle as a resource graph and jointly optimizes content placement across four cost layers — catalog, activation, deployment and path-weighted runtime — under hard coverage, routing-preservation and byte-identity constraints, backed by a witness hierarchy (literal containment, deterministic coverage, frozen-model entailment) and an independent disk-level audit that falls back to the verbatim bundle on failure. Two orthogonal axes cover deployment needs: One-Shot vs Continual (reusing state across evolution patches) and Persistent (rewrites the shipped bundle) vs Transient (builds a per-run execution view). On BFCL-v4, LiveMathematicianBench and SpreadsheetBench bundles and real self-evolved libraries, it reports the largest faithful savings (+23.1% shipped, +17.7% per-run on real libraries at routing 1.000), and continual mode with repacking uses about one third of the compressor calls of full rebuilds. On a private Chinese content-moderation skill it removes 38.1% of deployed tokens and 10.4% of paired per-run tokens, with accuracy 89% against uncompressed session baselines of 92%/91%. Main limitations: small held-out sets (102 pooled tasks), tiny direct-call rollouts (n=3–8), no significance testing, a handicapped SkillReducer baseline, and headline quality claims resting on one private production skill.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● 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
- Compressing progressively loaded skill bundles across files, with activation-scoped placement, yields larger faithful savings than root-only or flat compression while preserving routing: on real evolved libraries +23.1% shipped and +17.7% per-run at routing 1.000, versus routing collapse to 0.000 for root-only and flat baselines (Table IX).
- On a production Chinese content-moderation skill, witnessed compression removes 38.1% of deployed bundle tokens and 10.4% of pooled paired per-run tokens while accuracy (89.00%) stays within the observed variation of uncompressed session baselines (92.00%/91.00%) (Table XXIII).
- Unwitnessed aggressive compression is unsafe: 71.4–75.8% deployment savings cost 18–26 accuracy points, almost entirely via additional false positives from deleted exemption rules.
- Continual compression with triggered repacking matches full rebuilds (drift −0.083) at roughly one third of the compressor calls (3.31 vs 10.00 per round) with +48.1% shipped saving (Table X).
- Self-evolution is intrinsically redundant — 55%±3% of each round's added content repeats earlier text — so compression should be enabled from the first evolution round (40.8% cumulative stream saving vs 18.5% when delayed to round 13).
Strengths
Correct formalization of a real deployment problem
The four-layer cost model (catalog/activation/deployment/path-weighted, Eq. 3–7) and Propositions IV.1–IV.2 capture placement trade-offs that flat compressors ignore; Table IX confirms the predicted failure empirically: Expert Progressive gains +25.0% shipped but regresses the always-loaded layer by −2.9%, while SkillZip Pro improves every layer at routing 1.000.
Safety stack is engineered and demonstrated, not asserted
Witness hierarchy (W1 literal containment ≻ W2 coverage ≻ W3 frozen-model entailment), byte-locked executables, contiguous verbatim interface contracts, and a fresh-process disk audit; real fallbacks are shown: one of three real libraries is republished verbatim after failing the never-inflate check (Sec. VI-M), and the multi-entry audit catches a renamed public subskill (independence 0.500→1.000, Tables XVII–XVIII).
Unusual honesty about failure modes
Reproducibility apparatus exceeds norms
Non-English production evidence exposes a real limitation and a fix
Weaknesses
No statistical uncertainty quantification
Small, partly synthetic evaluation base
Key competitor tested handicapped
Retention metric underpins baseline disqualification but is fuzzy
Externally unverifiable headline
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
SkillZip Pro (Alibaba Group; Zhejiang University; arXiv 2608.30785) targets a real gap in agent infrastructure: production skills are directories loaded progressively (root at activation; references, schemas, scripts, subskills on demand), so flat-text prompt compression mis-prices cost and can break loading. The system compresses the whole bundle with two pillars — activation-aware cross-file compression (host-entailment pruning, activation-scoped sharing, guarded capsules) and routing preservation (locked routing table, per-line reference preservation, independent disk audit) — across two orthogonal axes: One-Shot/Continual and Persistent/Transient (Sec. V, Table I). Central claims: (1) bundle-aware placement beats root-only and flat compression at equal fidelity; (2) witnessed removal enables 38.1% deployment and 10.4% paired per-run savings on a production Chinese moderation skill without quality loss; (3) unwitnessed compression at 71.4–75.8% loses 18–26 accuracy points; (4) continual mode with repacking achieves rebuild-grade results at ~1/3 the compressor calls.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 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 SkillZip Pro?
A compression tool for LLM agent skills stored as directories: a root file loaded at activation plus references, scripts and subskills loaded on demand. It shrinks the whole bundle across files while keeping every required file reachable, and it decides what to remove without task rollouts, using typed contracts, deletion witnesses and a disk-level audit that falls back to the original bundle when a check fails.
How much does SkillZip Pro compress?
On the authors' benchmark and real evolved-library experiments it saves roughly 12–25% depending on workload and library size (e.g., +23.1% shipped tokens on real evolved libraries). On a production Chinese content-moderation skill it removed 38.1% of deployed tokens and 10.4% of measured per-run tokens. Unprotected compression could reach 71–76% savings but lost 18–26 accuracy points.
Does compressing an agent skill hurt accuracy?
With the paper's witness and audit safeguards, measured accuracy on a 100-task production set was 89% versus uncompressed same-session baselines of 92% and 91% — within observed session variation but nominally 2–3 points lower, and no statistical test is reported. Aggressive compression without safeguards dropped accuracy by 18–26 points, mainly through more false positives.
Is code or data available for this paper?
The abstract links a GitHub repository (yutou520131/SkillZip-Pro). The production content-moderation skill, its evaluation tasks and the industrial harness are not public, so the headline production numbers cannot be independently verified from the paper alone; the benchmark bundles derive from public benchmarks (BFCL-v4, LiveMathematicianBench, SpreadsheetBench).
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.
- The provided text is a sampled excerpt with explicit [...sampling gap...] markers; several tables (e.g., Tables II–VI, XI–XIV) and some section text were skipped, so judgments about those parts rely on in-text cross-references.
- The bibliography was not visible in the provided excerpt; citations [1]–[32] appear only in-text.
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.