Meituan's SKILL0 vs Skill1: Two Routes for Agent Skill Learning

> SKILL0: arXiv:2604.02268 | Skill1: arXiv:2605.06130 | Meituan / Zhejiang University / USTC

---

Opening: One Question, Two Answers

In April-May 2026, Meituan published two papers on agent skills just one month apart, addressing the same fundamental question: how should an agent learn, use, and evolve its skills?

The answers are diametrically opposed:

• SKILL0: Internalize skills into model parameters during training; achieve zero inference-time dependency.
• Skill1: Treat the skill library itself as the core competency; co-evolve selection, utilization, and distillation.

Both use 7B models. Both reach SOTA on ALFWorld (89.8% vs 97.5%). Their philosophies are entirely different.

---

SKILL0: From "Reading the Manual" to "Muscle Memory"

SKILL0's core diagnosis: inference-time skill retrieval suffers from three fundamental defects:

1. Retrieval noise: retrieved skills can be irrelevant or even misleading 2. Token overhead: skill content accumulates across multi-turn interactions 3. Capability parasitism: the model merely *executes* skills but never truly *learns* them

The analogy is human skill acquisition: novice reads the manual (explicit rules) → practitioner completes tasks from experience (internalized ability) → expert reacts automatically. Existing skill-augmented methods trap agents at stage one—re-reading driving-school notes on every trip.

SKILL0's solution: provide skill context during training, use RL to gradually internalize it into parameters, and remove skills entirely at inference.

Technical Implementation: ICRL + Dynamic Curriculum

In-Context Reinforcement Learning (ICRL): skills live in the context during training rollouts and are completely removed at inference. The RL objective directly drives the transition "from context-dependent to autonomous action."

Dynamic curriculum learning:

• Each skill file is associated with a validation subtask
• Periodically compare "with-skill" vs "without-skill" validation accuracy; the gap is the helpfulness score \(\Delta k\)
• \(\Delta k > 0\) keeps the skill; \(\Delta k \le 0\) removes it
• Skill budget follows a linear decay schedule down to zero

Context rendering: interaction history and skills are rendered as compact RGB images and compressed by a vision encoder. The model controls its own compression ratio \(c_t\), outputting action and compression ratio jointly each step. The reward combines task success and compression efficiency.

Key Results

| Metric | SKILL0 (7B) | SkillRL | AgentOCR | |---|---|---|---| | ALFWorld | 89.8% | 89.9% | ~80% | | Tokens/step | 0.38k | 2.21k | — | | Skill dependency at inference | ❌ none | ✅ yes | — |

On ALFWorld, SKILL0 nearly matches SkillRL (89.8% vs 89.9%) while reducing per-step tokens from 2.21k to 0.38k—roughly 5x savings. This validates the internalization route: skills can serve as training scaffolding and then be removed while preserving capability.

---

Skill1: A Unified Solution to the Three-Body Problem

Skill1's diagnosis: the skill library is not a burden but the agent's "external brain." The real problem is whether the three stages—selection, utilization, distillation—can co-evolve.

Problems in existing methods:

• Selection is unmanaged: RetroAgent doesn't optimize selection; SkillRL freezes the selection mechanism
• Reward signals are split: different stages use different reward sources, causing conflicts
• Capabilities operate in silos: good selection but poor use, or good use but no new skills generated

Skill1's solution: one policy model, one task-outcome signal, three stages sharing one optimization.

Technical Implementation: Credit Assignment from a Single Signal

Skill selection: the policy generates a query to retrieve candidate skills, then reranks to pick the best. Both query and rerank scores are produced by the policy and are directly optimizable.

Skill utilization: once a skill is selected, the policy interacts with the environment over multiple turns under that skill's guidance.

Skill distillation: after rollouts, the policy extracts reusable skills from trajectories (policy description + scenario description); only successful trajectories may be added to the library.

Credit assignment (key innovation):

• Utilization reward = task outcome \(r(\tau)\) (direct signal)
• Selection reward = low-frequency trend of \(r(\tau)\) (per-skill EMA score, reflecting long-term utility)
• Distillation reward = high-frequency deviation of \(r(\tau)\) from library trend (whether current experience exceeds library frontier)

All three capabilities derive from the same task-outcome signal. Low-frequency trends guide selection; high-frequency changes guide distillation. No auxiliary models, no extra reward sources.

Key Results

| Metric | Skill1 | RetroAgent | SkillRL | GiGPO | |---|---|---|---|---| | ALFWorld avg | 97.5% | 94.9% | 89.9% | 90.8% | | WebShop | 82.9% | — | 72.7% | — | | 6 subtask ranking | 5 first places | — | — | — |

Skill1 reaches 97.5% average success on ALFWorld, 2.6 points above RetroAgent, and is also best on WebShop.

Ablations: Library Is the Foundation, Distillation Is the Furnishing, Selection Is the Navigation

| Removed component | ALFWorld drop | Key finding | |---|---|---| | Entire library | -16.6 (→80.9%) | Heat/Pick2 drop 28+ points; multi-step procedural tasks depend most on skills | | Distillation | -5.1 (→92.4%) | Without distillation only raw trajectories are stored; library becomes noisy | | Selection | -5.7 (→91.8%) | Wrong selection wastes any utilization effort | | \(\lambda_1=0\) (no selection signal) | -3.5 | A single signal is insufficient; decomposition is needed | | \(\lambda_1=\lambda_2=0\) | -7.3 (→90.2%) | Worse than removing each alone; signals reinforce each other |

Training dynamics confirm the co-evolution: selection precision converges to 0.95 by step 20; utilization and distillation reach 0.8 by step 60. Good selection comes first, then good utilization, then good distillation—the capability chain activates in sequence.

---

Side-by-Side Analysis of the Two Routes

| Dimension | SKILL0 (Internalize) | Skill1 (Evolve) | |---|---|---| | Core idea | Skills are scaffolding; remove once learned | Skills are an external brain; refine forever | | At inference | Zero skill, zero retrieval | Continuous skill-library retrieval | | Training | ICRL + dynamic curriculum annealing | GRPO + single-signal credit assignment | | Token efficiency | Extreme (0.38k/step) | Grows with library | | ALFWorld score | 89.8% (7B) | 97.5% (7B) | | Best fit | High-frequency repetitive tasks, edge deployment | Complex variable tasks, cross-episode reuse | | Relation to Microsoft SkillOpt | SkillOpt's trained best_skill.md can be internalized by SKILL0 | SkillOpt is a training tool; Skill1 is a deployment framework |

Key insight: the two are not exclusive but complementary. SKILL0 validates a major proposition—skills can be parametrically absorbed, and procedural knowledge, like declarative knowledge, can be internalized through training. Skill1 shows the ceiling of the library framework when skills must be continuously updated and reused across tasks.

---

Knowledge vs Skills: A Divergence in the Internalization Path

A precise observation: "knowledge internalization relies on SFT, skill internalization relies on RFT."

• Knowledge (facts, concepts, text comprehension) → internalized via SFT / pretraining
• Skills (procedural operations, tool use, multi-step strategies) → require RL trial-and-error feedback

SKILL0 demonstrates via ICRL that the procedural nature of skills can indeed be migrated from context to parameters through reinforcement learning—mirroring the human path of deliberate practice toward automation. The cost is that the skill becomes "dead knowledge"—the model has learned it but no longer updates it, and cannot distill new skills for novel situations the way Skill1 can.

---

Industry Mapping

SKILL0 suits:

• On-device agents (tokens are expensive; inference must be lean)
• Fixed-procedure automation (customer service, form filling)
• Security-sensitive scenarios (where skill content should not leak through context)

Skill1 suits:

• Exploratory agents (research, data analysis—continually accumulate new strategies)
• Multi-tenant platforms (one library serves many users/tasks)
• Human-collaborative systems (skills are interpretable, auditable, shareable)

---

References

• SKILL0: Lu et al., "SKILL0: In-Context Agentic Reinforcement Learning for Skill Internalization", arXiv:2604.02268, 2026
• Skill1: Shi et al., "Skill1: Unified Evolution of Skill-Augmented Agents via Reinforcement Learning", arXiv:2605.06130, 2026
• Code: github.com/ZJU-REAL/SkillZero (SKILL0)
• Related: SkillRL (Xia et al., 2026), RetroAgent (Zhang et al., 2026), GiGPO (Feng et al., 2025)

> The bigger significance of these two papers is that they push agent-skill research from "how to write a good skill" to the meta-question of "how skills should interact with the model." SKILL0 proves skills can be digested; Skill1 proves skill libraries can self-evolve. A natural next step: run Skill1 first to grow a library, then run SKILL0 to internalize its essence—broad first, then deep.