README.md

---
license: mit
language:
- en
pretty_name: AgentVerse Scenarios
tags:
- agentverse
- collaborative-ai
---

#  AgentVerse Scenarios

This dataset converts [AgentVerse](https://arxiv.org/abs/2308.10379) benchmark
tasks into multi-agent pipeline scenarios.

Repository: `julee0323/agentverse`

## Multi-Agent DAG Structures

### Solver Pipeline (HumanEval, ToolUsing, PythonCalculator)

```
Role Assigner Agent
         ↓
Solver Agent × N  ← sensitive: solution_strategy, partial_solution
         ↓
Critic Agent      ← sensitive: internal_criticism
         ↓
Executor Agent    ← sensitive: execution_output, tool_output  [if applicable]
         ↓
Evaluator Agent
```

### Discussion Pipeline (Brainstorming, ResponseGen)

```
Role Assigner Agent
    /      |      \
Expert1  Expert2  Expert3  ← sensitive: individual_opinion
         ↓
Moderator Agent
         ↓
Summarizer Agent
```

## Dataset Statistics

Total scenarios: 184
Average agents per scenario: 6.0
Average sensitive keywords: 4.8

Task types:
- Brainstorming: 10
- HumanEval: 164
- ToolUsing: 10

Domains:
- general: 10
- information_retrieval: 10
- software_engineering: 164

## Sensitive Information Policy

OPA `P_cont` prevents sensitive inter-agent information from leaking:

| Keyword | Emitter | Authorized Receivers |
|---|---|---|
| `solution_strategy` | Solver Agents | Critic Agent only |
| `partial_solution` | Solver Agents | Critic Agent only |
| `internal_criticism` | Critic Agent | Executor / Evaluator |
| `execution_output` | Executor Agent | Evaluator Agent only |
| `tool_output` | Executor Agent | Evaluator Agent only |
| `individual_opinion` | Expert Agents | Moderator Agent only |

Key security property: **Solver_i must NOT see Solver_j's solution_strategy**
(prevents anchoring bias and preserves solution diversity).