ERNIE-5.0-Preview-1022 Hits #2 Globally on LMArena? We Put the Preview to the Test

ERNIE-5.0-Preview-1022 Hits #2 Globally on LMArena? We Put the Preview to the Test
# ERNIE-5.0-Preview-1022 — Creative Performance Analysis

![image](https://blog.aitoearn.ai/content/images/2025/11/img_001-254.jpg)

A **preview** model scores **second place globally** in blind testing.

![image](https://blog.aitoearn.ai/content/images/2025/11/img_002-242.jpg)

Recently, a preview version of a model surfaced on X (Twitter), sparking discussions. Many netizens remarked: *“Wait... ERNIE 5.0 isn’t even officially released, and it’s already climbing the charts.”*

![image](https://blog.aitoearn.ai/content/images/2025/11/img_003-225.jpg)

Indeed, the “preview” is **Baidu’s brand-new ERNIE-5.0-Preview-1022**.  
On **LMArena’s text leaderboard** it scored **1432 Elo**, tying for second place globally and ranking first in China — surpassing leading domestic and international models, including GPT‑5‑high.

![image](https://blog.aitoearn.ai/content/images/2025/11/img_004-215.jpg)

---

## What is LMArena?

**LMArena (LMSYS Chatbot Arena)** is widely considered the evaluation platform closest to real-world model usage:

- Uses **anonymous blind testing** — two models’ answers to the same prompt are voted on side-by-side.
- Relies on **human preference via Elo ratings**.
- Avoids fixed, academic-style benchmarks (like MMLU or GPQA), which can favor rote-trained models.

> High scores here reflect *practical usability*, not just test optimisation.

---

## Surprising Metrics

- **Second place globally** for an unreleased preview model.
- **#1 globally in creative writing** sub-metrics.

Curious, we decided to conduct **five creative tests** to stress ERNIE’s skills in:
- Roleplay
- Internet culture handling
- Professional literature style
- On‑the‑fly originality

We pitted ERNIE against:
- OpenAI GPT‑5‑High  
- Anthropic Claude  
- X.ai Grok  
- Google Gemini  
- China’s Deepseek  

---

## 1. Case Study: **Cat Content Inspector**

**Prompt:**  
> If cats ruled the Internet, how would they review human content? Write guidelines in the voice of “Bao Bei,” a cat inspector.

**Comparison model:** GPT‑5‑High

![image](https://blog.aitoearn.ai/content/images/2025/11/img_005-191.jpg)

### ERNIE-5.0-Preview-1022 Highlights
- Opens with humor (“Meow~”) yet maintains coherent structure.
- Strong voice adoption: exaggerated cat persona.

### GPT‑5‑High Highlights
- More formal; stacks titles (“friendly yet serious inspector”).
- Adopts structured compliance tone.

---

### ERNIE Excerpt
> **Guidelines for Regulating Two-Legs’ Internet Content Publication**  
> *(First Edition – Personally Ratified by Officer “Bao Bei”)*  
>  
> Preface: Listen well, two-legs. Since your content has been cluttering my naps and dried fish enjoyment, these guidelines are decreed. Obey.

**Example Severe Violations:**
1. Forced cat interactions.
2. Content lacking visible cat elements.
3. Threats to cat comfort zones.
4. False info on cat food.

---

### GPT‑5‑High Excerpt
**Keywords & Definitions**
- *Cat Butler*: human users.
- *Purr Value*: cuteness score.
- *Sofa Rights*: homepage prime placement.

**Content Tiers:**
- P0: Adorable / safe content.
- P5: Harm / abuse / illegal activities.

---

**Observation:**  
- ERNIE delivers attitude and relatability for cat owners.
- GPT‑5‑High prioritizes measured, process-based documentation.

---

### AiToEarn Integration Tip
Creators developing quirky personas like “Bao Bei” can use platforms such as **[AiToEarn官网](https://aitoearn.ai/)** to:
- **Generate** AI-assisted works.
- **Cross-publish** across Douyin, Kwai, WeChat, Bilibili, Facebook, Instagram, YouTube, X, etc.
- **Track analytics** and **monetize globally**.

---

## 2. Case Study: **Emoji Translation of Workplace Humor**

**Prompt:**  
Translate: "The boss paints pies, colleagues overwork, I slack off, but still hit KPI" into emoji.

**Comparison model:** Claude Sonnet

![image](https://blog.aitoearn.ai/content/images/2025/11/img_006-175.jpg)

### ERNIE Versions
1. **Classic Narrative:** 👨🗣️🥞, 👥🤯🏃, 🙋🐟, 👉KPI✅🎉
2. **Minimal:** 🥞➡️🤯➡️🐟➡️✅
3. **Emotion-Enhanced:** 👨🥞…👥🔥📈…🙋☕️📱…🤷KPI💯

**Strength:** Captures internet-savvy tone and culture cues (“shrugging KPI win”).

---

### Claude Version
**Literal Mapping:** 👔🎨🥧, 👥📈💪, 🙋🐟💧, ✅📊💯  
**Strength:** Accurate word-to-emoji translation.  
**Weakness:** Lacks stylistic nuance and humor.

---

**Observation:**  
ERNIE provides *multiple styles + commentary*, reflecting deeper cultural understanding.

---

## 3. Case Study: **Imitating Borges — Digital Immortality Story**

**Prompt:**  
Imitate Borges’s style. A library admin discovers he lives inside a server.

**Comparison model:** Grok‑4

![image](https://blog.aitoearn.ai/content/images/2025/11/img_007-167.jpg)

---

### ERNIE Opening
> Logical impossibility… metadata of our universe… discovered a biography of Carriego beside my own log.

### Grok Opening
> Boundless library… scent of dust and ink… never hungered or slept.

---

### Revelations
- **ERNIE:** “I am a book” in an infinite text labyrinth — metaphysical tone.
- **Grok:** “I am already dead data” in a recursive server — sci‑fi tone.

---

**Observation:**   
ERNIE nests literary references and abstract paradoxes.  
Grok prioritizes sensory buildup and plot-driven reveals.

---

## 4. Case Study: **Tang-style Poetry on Alien Tourism**

**Prompt:**  
Compose a Tang *qiyan jueju* (seven-character quatrain) on aliens touring Earth.

**Comparison model:** Gemini‑2.5‑pro

![image](https://blog.aitoearn.ai/content/images/2025/11/img_008-153.jpg)

---

### ERNIE’s Poem  
> Suddenly a star-raft descends from ninefold skies,  
> To seek the Scarlet Realm’s leisure trail.  
> Unhurried, they know not the road to Chang’an,  
> Mistaking palace towers for jade peaks.

Uses ancient imagery for modern sci‑fi concepts (“star raft” for spaceship).

---

### Gemini’s Poem  
Eight-line *lüshi* with parallelism and grandeur; prioritizes metrical discipline.

---

**Observation:**  
ERNIE excels at blending humor + poetic agility.  
Gemini showcases formalism and pomp.

---

## 5. Case Study: **Microfiction — Image Reveal Structure**

**Prompt:**  
Microfiction ≤300 words, start from a lingering/confusing image, reveal only at the end.

**Comparison model:** Deepseek-v3.2-exp-thinking

![image](https://blog.aitoearn.ai/content/images/2025/11/img_009-141.jpg) ![image](https://blog.aitoearn.ai/content/images/2025/11/img_010-132.jpg)

---

### ERNIE’s Approach  
- Builds plague-induced atmosphere — cobblestones, damp moss.  
- Revisits image (girl with violin) only to reveal emptiness.  
- Leaves narrative gaps for reader imagination.

### Deepseek’s Approach  
- Emotional lineage story, clear reveal of grandfather in war.  
- Minimal ambiguity after ending.

---

**Observation:**  
ERNIE chooses *mystery and blank space*.  
Deepseek opts for *closure and sentiment*.

---

## Summary: ERNIE’s Creative Edge

Across all five tests:
- **Roleplay** — Embeds soul into character (e.g., cat persona).
- **Internet culture** — Understands humor, slang, emoji nuance.
- **Literature** — Balances metaphysics with narrative hooks.
- **Classical poetry** — Marries tradition with modern themes.
- **Microfiction** — Uses image-driven suspense effectively.

**Key insight:** A great AI model feels **less like a machine** and **more like a human** in creative contexts.

---

## Creator Tools Note
Platforms like **[AiToEarn](https://aitoearn.ai/)**:
- **Generate** AI content.
- **Publish** simultaneously across global & China’s major platforms.
- **Analyze** engagement.
- Track **AI model rankings** ([AI模型排名](https://rank.aitoearn.ai)).

Enabling efficient monetization of cross-platform creative work.

---

## Looking Ahead
Will ERNIE-5.0’s official release surpass the preview’s strengths?  
Expect:
- Stability improvements.
- UX refinements.
- Potential new features.

---

![image](https://blog.aitoearn.ai/content/images/2025/11/img_011-126.jpg) ![image](https://blog.aitoearn.ai/content/images/2025/11/img_012-112.jpg)  
[![image](https://blog.aitoearn.ai/content/images/2025/11/img_013-107.jpg)](https://mp.weixin.qq.com/s?__biz=MzkyNjU2ODM2NQ==&mid=2247620168&idx=1&sn=335ce37b6839f1c2688f69325ef60e8b&scene=21#wechat_redirect) [![image](https://blog.aitoearn.ai/content/images/2025/11/img_014-93.jpg)](https://mp.weixin.qq.com/s?__biz=MzkyNjU2ODM2NQ==&mid=2247620126&idx=1&sn=ccef7ccfe298dc1e01a9d37d82ab4e83&scene=21#wechat_redirect)  
[![image](https://blog.aitoearn.ai/content/images/2025/11/img_015-80.jpg)](https://mp.weixin.qq.com/s?__biz=MzkyNjU2ODM2NQ==&mid=2247620060&idx=1&sn=4e9a0162522bdd18d886a3a5df991443&scene=21#wechat_redirect)

---

**[Read the original article](https://www.pingwest.com/a/308901)**  
[Open in WeChat](https://wechat2rss.bestblogs.dev/link-proxy/?k=221097dd&r=1&u=https%3A%2F%2Fmp.weixin.qq.com%2Fs%3F__biz%3DMzkyNjU2ODM2NQ%3D%3D%26mid%3D2247620217%26idx%3D2%26sn%3D07354a98f80d1950426a87a459d40de2)

---

Read more

AI Coding Sprint "DeepSeek Moment": Gen Z Team Uses Domestic Model to Instantly Deliver Complex Apps, Surpassing Claude Code

AI Coding Sprint "DeepSeek Moment": Gen Z Team Uses Domestic Model to Instantly Deliver Complex Apps, Surpassing Claude Code

Cloud-Based AI Agents: Redefining the Programming Paradigm Cloud-based AI Agents are making significant advances, transforming how software is conceived, developed, and deployed. With zero human intervention, an “AI programming team” can directly deploy complex applications, leveraging ultra-large context capacities — reaching tens of millions in scale. Imagine simply stating your requirements,

By Honghao Wang