Meituan's Skill Double-Punch: Internalization or Evolution? Two Ultimate Routes for Agent Skills

> 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, one month apart, both answering the same fundamental question: how should an agent's skills be learned, used, and evolved?

But the answers are diametrically opposed:

• SKILL0 says: internalize skills into model parameters — learn with skills during training, and rely on them zero at inference.
• Skill1 says: the skill library itself is the core competency; the key is making selection, utilization, and distillation co-evolve.

One path leads to "mastering no moves to beat all moves"; the other to "perfecting every move." Both papers use 7B models and hit SOTA on ALFWorld (89.8% vs 97.5%), but their philosophies are entirely different.

---

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

SKILL0's core judgment is that inference-time skill retrieval has three fundamental flaws:

1. Retrieval noise: retrieved skills may be irrelevant or even misleading 2. Token overhead: skill content accumulates over multi-turn interactions 3. Capability parasitism: the model merely "executes" skills without ever truly "learning" them

By analogy to human skill acquisition: novices follow manuals (explicit rules) → experienced workers rely on experience (internalized ability) → experts react instinctively (automation). Existing skill-augmentation methods keep agents stuck at stage one — like needing driving-school notes every time you drive.

SKILL0's solution: provide skill context during training, progressively internalize it into parameters via RL, until inference requires no skills at all.

Technical Approach: ICRL + Dynamic Curriculum

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

Dynamic curriculum learning:

• Each skill file is associated with a validation subtask
• Periodically compare validation accuracy with vs. without the skill; the difference is the help value Δk
• Keep skills with Δk > 0, remove those with Δk ≤ 0
• The skill budget is linearly annealed down to zero

Context Rendering: interaction history and skills are rendered into compact RGB images and compressed through a vision encoder. The model controls its own compression ratio ct, outputting both an action and a compression rate each step. The reward includes task success and compression efficiency.

Key Results

| Metric | SKILL0 (7B) | SkillRL | AgentOCR | |--------|-------------|---------|----------| | ALFWorld | 89.8% | 89.9% | ~80% | | Tokens per step | 0.38k | 2.21k | — | | Inference requires skills | ❌ No | ✅ Yes | — |

On ALFWorld, SKILL0 nearly matches SkillRL (89.8% vs 89.9%) while cutting per-step tokens from 2.21k to 0.38k — nearly 5x savings. The internalization route is validated: skills can serve as training scaffolding, be completely removed at inference, and the model still retains strong performance.

---

Skill1: A Unified Solution to a Three-Body Problem

Skill1's judgment: the skill library is not baggage but the agent's "external brain." The question is not whether to have a library, but whether its three components (selection, utilization, distillation) can co-evolve.

Problems with existing methods:

• Selection is neglected: RetroAgent doesn't optimize selection; SkillRL freezes the selection mechanism
• Fragmented reward signals: different components use different reward sources, causing conflicts
• Siloed capabilities: good selection with poor utilization, or good utilization without producing new skills

Skill1's solution: one policy model, one task-outcome signal, three components jointly optimized.

Technical Approach: 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 are policy-generated and thus directly optimizable.

Skill utilization: after selection, the policy interacts with the environment over multiple turns guided by the selected skill.

Skill distillation: after each rollout, the policy distills reusable skills (strategy description + scenario description) from trajectories; only successful trajectories may enter the library.

Credit assignment (core innovation):

• Utilization reward = task outcome r(τ) (direct signal)
• Selection reward = low-frequency trend of r(τ) (per-skill EMA score reflecting long-term utility)
• Distillation reward = high-frequency deviation of r(τ) from library trends (whether current experience exceeds the library's existing boundaries)

All three capabilities derive from the same task-outcome signal: low-frequency trends guide selection, high-frequency variation guides 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% | — | | Ranking across 6 subtasks | 1st on 5 | — | — | — |

97.5% average success on ALFWorld, surpassing RetroAgent by 2.6 points. Also best on WebShop.

Ablations: The Library Is the Foundation, Distillation the Renovation, Selection 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 = raw trajectories stored; library gets noisy | | Selection | -5.7 (→91.8%) | Poor selection wastes good utilization | | λ1=0 (no selection signal) | -3.5 | A single signal isn't enough; decomposition is needed | | λ1=λ2=0 | -7.3 (→90.2%) | Worse than removing either alone; signals reinforce each other |

Training dynamics confirm the synergy: selection accuracy converges to 0.95 by step 20; utilization and distillation both reach 0.8 by step 60. Good selection first, then good utilization, finally good distillation — capabilities activate in sequence.

---

Head-to-Head Analysis of the Two Routes

| Dimension | SKILL0 (Internalization) | Skill1 (Evolution) | |-----------|--------------------------|--------------------| | Core philosophy | Skills are scaffolding — tear down once learned | Skills are an external brain — built ever finer | | At inference | Zero skills, zero retrieval | Continuously retrieves from skill library | | Training | ICRL + dynamic curriculum annealing | GRPO + single-signal credit assignment | | Token efficiency | Extreme (0.38k/step) | Grows with the library | | ALFWorld | 89.8% (7B) | 97.5% (7B) | | Best for | High-frequency repetitive tasks, on-device deployment | Complex, changing tasks requiring cross-episode reuse | | Relation to Microsoft's SkillOpt | SkillOpt's best_skill.md can be internalized by SKILL0 | SkillOpt is a training method; Skill1 is a deployment framework |

Key insight: the two are complementary, not mutually exclusive. SKILL0 validates an important proposition — skills can be absorbed into parameters; procedural knowledge, like declarative knowledge, can be internalized through training. Skill1 shows where the library framework's limits lie when skills need continuous updating and cross-task reuse.

---

Knowledge vs. Skills: Diverging Internalization Paths

One observation puts it precisely: "knowledge internalization relies on SFT; skill internalization relies on RFT."

• Knowledge (facts, concepts, text comprehension) → internalizable into parameters via SFT/pretraining
• Skills (procedural operations, tool use, multi-step strategies) → require RL's trial-and-error feedback to internalize

SKILL0 proves via ICRL that a skill's procedural character can indeed migrate from context to parameters through reinforcement learning — consistent with the human path of "deliberate practice → automation." The cost: the skill becomes "dead knowledge" — learned but no longer updated, unable to distill new skills into a library for novel scenarios the way Skill1 can.

---

Industry Mapping

SKILL0 suits:

• On-device mobile agents (tokens are expensive; inference must be minimal)
• Fixed-workflow automation (customer service, form filling)
• Security-sensitive scenarios (no skill content leaked in context)

Skill1 suits:

• Exploratory agents (research, data analysis — constantly accumulating new strategies)
• Multi-tenant platforms (one library serving many users/tasks)
• Human-collaboration systems (skills 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 larger significance of these two papers: 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. The interesting next question might be: can a model first follow the Skill1 route to accumulate a library, then follow the SKILL0 route to internalize the essence — breadth first, then depth?