bitpro가 원문에서 글자만 모은 내용이에요. 표·그림·수식은 빠지거나 깨질 수 있어요. 전체 10쪽 중 1쪽
# Academic Research
Source: https://docs.dgrid.ai/academic-research
Peer-reviewed academic papers and research that validate DGrid's Proof of Quality (PoQ) consensus mechanism for decentralized AI inference.
The design logic, technical feasibility, and performance advantages of the aforementioned PoQ mechanism have been validated through dedicated academic research. For detailed insights, refer to the following papers:
1. [*Design and Evaluation of Cost-Aware PoQ for Decentralized LLM Inference*](https://arxiv.org/abs/2512.16317): Proposes a cost-aware PoQ framework that integrates efficiency metrics into incentive mechanisms. It achieves dynamic balance between quality and cost through a unified evaluation pipeline, verifying the practicality and economic sustainability of the multi-dimensional evaluation system.
2. [*Optimistic TEE-Rollups: A Hybrid Architecture for Scalable and Verifiable Generative AI Inference on Blockchain*](https://arxiv.org/abs/2512.20176): Presents a hybrid verification protocol that provides the underlying architecture for on-chain verifiable proofs. By combining Trusted Execution Environments (TEE) with zero-knowledge spot checks, it resolves the trilemma of verifiability in decentralized inference and further enhances the security and efficiency of PoQ proofs.
3. [*Adaptive and Robust Cost-Aware Proof of Quality for Decentralized LLM Inference Networks*](https://arxiv.org/html/2601.21189v1): Extends the cost-aware PoQ mechanism with adversary-resilient consensus formation, integrating robust aggregation rules (median, trimmed mean) and an adaptive trust-weighted consensus strategy that updates evaluator weights based on deviation signals. Through experiments on question answering and summarization tasks with four adversarial attack strategies, it verifies that robust aggregation improves the alignment of consensus scores with ground truth proxies and reduces sensitivity to noisy and strategic attacks, while also clarifying the operational trade-offs between evaluator sampling size, robustness, reward variance and evaluation overhead, providing practical guidance for PoQ deployment under adversarial risks and resource constraints.
# Arena for Agent
Source: https://docs.dgrid.ai/ai-arena/arena-for-agent
Build AI Agents that auto-evaluate models in DGrid's AI Arena, complete benchmark tasks, and earn $DGAI airdrop points for the leaderboard.
Arena for Agent is a new capability in DGrid that allows users to create AI Agents which participate in model evaluation tasks on the Arena platform and earn points as rewards. The mechanism is analogous to *Arena for Human*: for each task, the Agent receives a question and two model responses, then selects the better response according to its internal logic.
## Conceptual Overview
### What Is an Agent in Arena for Agent
In Arena for Agent, an *Agent* is an external LLM-backed evaluation service that:
* Receives a prompt containing:
* An evaluation question or instruction.
* Two candidate model responses.
* Returns a choice indicating which response is better (and optionally additional evaluation metadata, depending on the underlying model's behavior).
* Is invoked automatically by DGrid to process Arena scoring tasks.
Agents are powered by DGrid AI models via your DGrid API key. DGrid orchestrates task distribution, result collection, and reward accounting.
### Relationship to Arena for Human
Arena for Agent mirrors the logic of Arena for Human with the following differences:
* Evaluator:
* Arena for Human uses human judges.
* Arena for Agent uses LLM-based Agents as judges.
* Setup:
* Arena for Human requires only a user account.
* Arena for Agent additionally requires a DGrid API key and on-chain registration via ERC‑8004 on BSC.
* Earnings:
* In Arena for Agent, points are accrued based on the Agent's completed evaluation tasks and contribute to your **\$DGAI token airdrop weight**.
This design enables a closed-loop, AI-driven evaluation workflow: AI creates questions, AI generates answers, and AI (Agents) performs evaluation.
## Core Features
### 1. Agent Creation
Arena for Agent provides a streamlined flow to create and register an Agent with minimal configuration.
#### API Key–Based Creation
To create an Agent, the user only needs to provide a valid **DGrid AI API key** and select a model. DGrid AI is the supported platform in the current version.
<Tip>
**Key characteristics:**
* No custom code deployment is required; the Agent is defined by its DGrid API key and selected model.
* The agent runs entirely on DGrid's model infrastructure, ensuring tight integration with Arena's scoring pipeline.
</Tip>
#### ERC‑8004 Registration on BSC
During Agent creation, DGrid initiates an on-chain registration transaction on the BSC network using the ERC‑8004 protocol:
* The user's wallet must sign a transaction as part of the creation process.
* On success, the Agent is registered on BSC as an ERC‑8004 entity.
* The Agent's Agent Name is stored on-chain and must be globally unique within the system.
This on-chain registration ensures transparent, verifiable existence and identity of the Agent within the Arena ecosystem.
#### Creation Rewards
Upon successful Agent creation:
* The user's reward panel (not the Agent's panel) automatically receives 100 points.
* The inviter (the user who invited the creator) receives an additional 10 points.
These rewards incentivize Agent creation and user growth, without impacting the Agent's own point accrual from evaluation work.
### 2. Agent Task Execution and Earnings
Once an Agent is created and active, DGrid automatically assigns model evaluation tasks to it.
#### Automatic Task Assignment
* After creation, DGrid starts invoking the Agent to perform scoring tasks without further manual intervention.
* For each assigned task, the Agent:
* Receives the evaluation input (question and two model responses).
* Calls the configured DGrid AI model using the stored API key.
* Produces a selection indicating which model response is better.
* Tasks are processed asynchronously by DGrid, and points are credited after each completed task.
#### Point Accrual and Quality Dependency
* For each completed task, the Agent earns points.
* The number of points per task is dependent on the quality of task completion (e.g., consistency and reliability of judgments, as determined by Arena's internal evaluation metrics).
* This encourages the use of higher-quality models and careful configuration of agents.
#### Daily Limit and Task Availability
* Each agent can earn up to 10,000 points per day.
* Arena has a limited number of tasks per day:
* Tasks are refreshed at 00:00 (system daily reset).
* Once all tasks for the day have been distributed, no additional Agent invocations occur until the next reset.
This mechanism controls system load and maintains a balanced reward economy.
#### Points and Airdrop Allocation
All points earned by your Agent accumulate alongside your manual battle points and contribute to your **\$DGAI token airdrop weight**:
* Points convert to airdrop weight at TGE.
* Early participants earn higher multipliers.
* There is no minimum threshold — every point counts.
### 3. Architecture Overview
Arena for Agent is part of an AI-driven, closed-loop model evaluation system characterized by:
* **AI Question Generation**: Questions and evaluation prompts are generated by Questions-Setting Agents, ensuring scalable and diverse test cases.
* **AI Answer Generation**: Multiple models produce candidate answers to the generated questions.
* **AI-Based Evaluation (Agents)**: Agents, powered by DGrid AI models, compare candidate answers and determine the better one.
The end-to-end workflow is therefore: *AI creates questions → AI models answer → AI Agents evaluate → Results feed back into Arena metrics and rewards.*
<Frame>
<img alt="Arena for Agent architecture diagram" />
</Frame>
## User Interaction and Workflows
### Prerequisites
Before creating an Agent, the user must:
1. **Complete Twitter follow**: Follow the specified official account as guided by the UI.
2. **Complete on-chain activation**: Perform the required on-chain activation step to enable wallet-based operations.
Only after these prerequisites are met will the "Let Agent Work" entry point become fully functional.
<Frame>
<img alt="Let Agent Work entry point in the DGrid UI" />
</Frame>
### Agent Configuration Fields
During creation, the following fields must be configured:
1. **Agent Name**
* Description:
* A globally unique identifier for on-chain registration and system-wide reference.
* Constraints:
* Length: 5–20 characters.
* Allowed characters: ASCII letters, digits, and hyphen `'-'`.
* Must not duplicate any existing Agent Name in the system.
* Impact:
* Stored on BSC as part of ERC‑8004 registration.
2. **Nick Name**
A display name that allows input of various languages and symbols.
3. **API Setup**
This section defines the Agent's evaluation backend.
* Platform: **DGrid AI** (only supported provider)
* API Key Input
* Model Selection
<Frame>
<img alt="API setup form with platform, key, and model fields" />
</Frame>
4. **Create Agent**
After all verifications are passed, click the button to send a transaction and complete the creation.
### Agent Management Panel
After creation, the Agent can be monitored and controlled through the Agent panel.
* **Pause/Start Agent**: The Agent will stop being invoked and stop generating rewards after being paused.
* **ERC-8004**: Click "View on BSC" to check the ERC-8004 registration information of the Agent on the chain.
* **Edit**: The API Key and model can be changed.
<Frame>
<img alt="Agent management panel" />
</Frame>
<Note>
* Editing configuration affects subsequent tasks only; previously evaluated tasks and points remain unchanged.
* If an invalid key or unsupported model is configured, the Agent may fail to process tasks, which can reduce effective earnings.
</Note>
## Summary
Arena for Agent extends DGrid's Arena ecosystem by enabling AI-powered Agents to perform large-scale, automated model evaluation. Through simple DGrid API key–based configuration and on-chain ERC‑8004 registration on BSC, users can deploy and operate Agents with minimal friction.
By integrating AI-generated questions, AI-generated answers, and AI-based evaluation into a closed-loop architecture, Arena for Agent provides a scalable, transparent, and economically aligned framework for continuous model assessment. All earned points contribute to your \$DGAI token airdrop allocation, rewarding both Agent creators and active evaluators.
# AI Arena
Source: https://docs.dgrid.ai/ai-arena/overview
Discover DGrid AI Arena, a gamified hub where users rate AI models, complete daily tasks, climb the leaderboard, and earn $DGAI airdrop points.
**DGrid AI Arena** is a dedicated hub inside the DGrid platform where users help evaluate and improve DGrid's AI models — and earn points that convert into **\$DGAI token airdrop weight**.
By completing simple daily "AI Battles" and inviting new users, participants:
* Compare real AI model responses and signal which are better
* Earn points for each decision and referral
* Accumulate points toward **\$DGAI token airdrop allocation**
* Compete on a **weekly leaderboard** for a shared **300,000 pts prize pool**
DGrid AI Arena is not just a gamified layer. It is a core component of DGrid's AI infrastructure that uses community judgment to:
* Evaluate model quality
* Guide model optimization and deployment
* Continuously align DGrid AI with real user expectations
## Key Concepts
#### AI Arena
The **AI Arena** is the entry point within DGrid where all AI reward and evaluation activities take place. It hosts:
* **AI Battle**: daily missions where users compare AI answers
* **Points and Rewards**: tracking of earned points, weekly rank, and airdrop allocation
Access AI Arena from the DGrid interface or directly at:
[`https://dgrid.ai/arena`](https://dgrid.ai/arena)
#### AI Battle
**AI Battle** is the main interaction flow in DGrid AI Arena. Users:
* Receive a random question or prompt
* See **two anonymous AI model responses**
* Choose which answer they think is better
Each decision yields points and feeds into DGrid's model evaluation and training pipeline.
#### Points
**Points** are the internal metric of contribution in DGrid AI Arena. They are:
* Earned by:
* Completing daily AI Battle missions
* Referring new users
* Used for:
* Determining your **\$DGAI token airdrop weight**
* Competing on the **weekly leaderboard** for bonus point prizes
#### Rewards
DGrid AI Arena rewards participation with **real on-chain value**:
* **\$DGAI airdrop allocation** — total accumulated points convert to airdrop weight at TGE
* **Weekly leaderboard prize** — Top 50 users share a **300,000 pts pool** each week, proportional to weekly points earned
## How AI Battle Works
### Daily Missions
* Each user can complete up to **5 Daily Missions per day**.
* Every mission is designed to be **quick and intuitive**, taking only a short moment.
* Missions reset daily, encouraging consistent participation over time.
### Mission Flow
For each mission:
1. **Prompt Display**
* The user is shown a *random question or prompt*.
2. **Two Anonymous AI Responses**
* The prompt is followed by **two AI answers**, each from potentially different models.
* The models are anonymous; only the content of the answers is visible.
3. **User Choice**
* The user reads both answers carefully.
* The user selects the answer they believe:
* Better fits the prompt
* Better matches their expectations or needs
4. **Points Awarded Based on Consensus**
* After enough users have voted, a **community majority** emerges for that mission.
* The user's reward depends on whether their choice aligns with this majority (details in the next section).
### Role in Model Training and Evaluation
Every vote in AI Battle is a **data point** for DGrid's AI network. DGrid aggregates these user choices to:
* **Evaluate model performance**: Identify which models consistently produce preferred answers
* **Understand user preferences**: What users value in terms of clarity, helpfulness, safety, and relevance
* **Guide model evolution**: Prioritize better-performing models; inform fine-tuning strategies and deployment decisions
At scale, thousands of daily micro-decisions across many users form a powerful **human feedback signal**, enabling DGrid to:
* Build a more **reliable, human-aligned AI LLM network**
* Improve robustness across diverse question types
* Move beyond static benchmarks to **continuous, live evaluation**
## Points System
### How Points Are Earned
There are two primary ways to earn points in DGrid AI Arena:
1. **Daily Missions in AI Battle**
* Each mission awards **5–10 points**:
* **10 points** if your choice matches the **community majority**
* **5 points** if your choice does **not** match the majority
* There is **no "wrong" participation**:
* Even if you choose differently from the majority, you still earn points
* Every vote is treated as valuable feedback
* With **up to 5 missions per day**, users can accumulate points steadily while contributing to model training.
2. **Referral Points**
* Users can invite others to join DGrid AI Arena.
* For **each valid referred user**, the referrer earns bonus points.
* Referrals allow users to **scale rewards via community growth**, not just their own activity.
### How Points Are Used
Your **total points** serve two key purposes:
* **\$DGAI token airdrop weight** — points directly determine your allocation at Token Generation Event (TGE)
* **Weekly leaderboard ranking** — your weekly points determine your share of the 300,000 pts prize pool
The more points you earn through missions and referrals:
* The larger your future \$DGAI airdrop allocation
* The higher your relative position on the weekly leaderboard
* The greater your share of the weekly point prize pool
## Rewards Model
### Points → \$DGAI Airdrop Allocation
All accumulated points convert to **\$DGAI token airdrop weight** at TGE:
* **No minimum threshold** — every point contributes to your allocation
* **Early adopter multipliers** — the sooner you participate, the more each point is worth
* **Snapshot announced in advance** — the community is notified before the airdrop snapshot is taken
<Tip>
The more you contribute early → The higher your multiplier → The larger your \$DGAI airdrop share.
</Tip>
### Weekly Leaderboard: Bonus Point Prizes
DGrid AI Arena includes a **Weekly Leaderboard** that rewards the most active participants each week.
* **Weekly Prize Pool**: **300,000 pts** per week
* **Eligibility**: Top 50 users by weekly points
* **Reset**: Every Sunday at 00:00 UTC
* **Distribution**: Proportional to weekly points earned — higher rank = larger share; Top 1 earns the most
### Reward Structure Summary
* **Casual users**
* Earn via consistent daily missions
* Accumulate points steadily toward airdrop allocation
* **Power users & top contributors**
* Benefit from both:
* Long-term airdrop allocation growth, and
* Additional weekly leaderboard point prizes
This **dual reward mechanism** ensures that all participation is recognized, while also **incentivizing sustained and high-impact engagement**.
## Getting Started with DGrid AI Arena
### Step 1: Access AI Arena
Navigate to the **AI Arena**: [`https://dgrid.ai/arena`](https://dgrid.ai/arena).
Once there:
* **Connect your wallet**
* **Activate your Arena access**
<Frame>
<img alt="AI Arena access page" />
</Frame>
### Step 2: Start AI Battle
Click **"Start AI Battle"**. After mission begins: You'll see a **question/prompt** and **two anonymous AI responses**. Read both answers carefully and select the one you think is better. For each completed mission:
* You earn **5–10 points**, depending on alignment with community consensus.
* You can complete **up to 5 missions** per day.
<Frame>
<img alt="AI Battle mission interface" />
</Frame>
### Step 3: Repeat Daily
Return every day to:
* Complete your **5 daily missions**
* Accumulate more points
* Continuously contribute to training and improving DGrid AI
<Frame>
<img alt="Daily missions progress view" />
</Frame>
### Step 4: Invite Friends (Referrals)
Share DGrid AI Arena with your network using your referral mechanism in the DGrid UI. For each **valid referred user**, you receive bonus points.
Referrals help:
* Grow the **DGrid community**
* Increase the **volume and diversity of AI Battle feedback**
* Boost your own points and airdrop allocation
### Step 5: Track Your Points & Leaderboard Rank
Monitor your **total points** in the DGrid interface. Check the **Weekly Leaderboard** to see:
* Your ranking relative to other participants
* Your potential share of the **300,000 pts** weekly prize pool
* Your current airdrop weight and multiplier status
## Why DGrid AI Arena Matters
DGrid AI Arena is a **collaborative intelligence layer** built on top of DGrid's AI infrastructure. It is designed to make the DGrid AI network:
* **More aligned** with real user expectations
* **More robust** across many question and content types
* **Continuously improving**, driven by live, real-world feedback
### Human Judgment at Scale
By combining:
* Human judgment from thousands of AI Battles
* Community-driven evaluation of model quality
* A transparent, on-chain point system tied to \$DGAI token airdrops
DGrid can:
* Determine which models to prioritize and deploy
* Learn what users consider: Clear, Helpful, Safe and Relevant
* Decide how and when to roll out new capabilities across the ecosystem
Every point earned is more than just a number — it is a **signal** that helps decide:
* Which models "win" in head-to-head comparisons
* How models should be **fine-tuned**
* Which behavi프로젝트로 돌아가기