Dude: A Dual-Detection Multi-Agent System
for Paper-Code Discrepancy Detection

⭐️ Automatically Detecting Paper-Code Inconsistencies ⭐️

| ✅ Multi-LLM Platform Support | ⏯️ Support Task Resuming | 🌎 Offline & Online Compatibility |

Dude is a multi-agent system designed to automatically detect discrepancies between academic papers and their corresponding code implementations. It adopts a dual-detection architecture in which multiple specialized agents collaborate through granularity-aligned negotiation and a two-stage salience-filtering mechanism to identify inconsistencies, missing implementations, and deviations from the descriptions in the paper.

• Multi-Agent Architecture: Specialized agents for dual-detection modes (paper-oriented, code-oriented).
• Multi-LLM Platform Support: Configurations for Claude, Deepseek, GPT, and Kimi models.
• Support resuming interrupted tasks: Structured three-stage pipeline for systematic discrepancy detection that supports interrupted task recovery.
• Offline & online paper compatibility: Handles both local papers and online papers via URL.

• [2026.08.28] 🚩 [Features Update] We split the Paper Claimer Agent into two specialized agents: Paper Claimer and Paper Searcher. The former extracts claims from research papers, while the latter searches for external evidence to support those claims. This change helps prevent performance degradation caused by overly long contexts in a single agent.
• [2026.08.28] 🚩 [Features Update] We split the Code Verifier Agent into two specialized agents: Code Mapper and Code Verifier. The former maps claims to the codespace, while the latter verifies the paper-code discrepancy. This change helps prevent performance degradation caused by overly long contexts in a single agent.
• [2026.08.21] 🎉 Dude is now open source! The codebase includes the full implementation of the framework, configurations for mainstream agent platforms (ChatGPT, Claude Code, and OpenCode) as well as the prompts used by the system.
• [2026.08.20] 🎉 Dude has been accepted as a Main Conference paper in EMNLP 2026! See you in Budapest 🇭🇺!

Dude/
├── agents/                             # Multi-LLM agent configurations
│   ├── Claude/                         # Claude model agents
│   │   ├── code_claimer.md             # Extracts claims from code
│   │   ├── code_mapper.md              # Maps research claims to the codebase
│   │   ├── code_negotiator.md          # Negotiates code claim validity
│   │   ├── code_verifier.md            # Verifies code claims
│   │   ├── paper_claimer.md            # Extracts claims from papers
│   │   ├── paper_searcher.md           # Searches for external evidence 
│   │   ├── paper_negotiator.md         # Negotiates paper claim validity
│   │   └── paper_verifier.md           # Verifies paper claims
│   ├── Deepseek/                       # Deepseek model agents (same structure)
│   ├── GPT/                            # GPT model agents (.toml)
│   └── Kimi/                           # Kimi model agents
│
├── orchstrator-prompt/                  # Orchestrator workflow prompts
│   ├── orchestrator_main_prompt.md             # Initialization Prompt
│   ├── orchestrator_initialization_prompt.md   # Stage 1: Initialize analysis
│   ├── orchestrator_negotiation_prompt.md      # Stage 2: Resolution negotiation
│   └── orchestrator_code-summary_prompt.md     # Stage 3: Generate summary
│
├── info.md                              # Configuration metadata
├── README.md                            # This file

• A supported agent platform: Install and authenticate at least one of ChatGPT, Claude Code, or OpenCode before starting the workflow.

• jq: Required for processing and manipulating JSON files. Install via:

  # macOS
  brew install jq
  
  # Linux
  sudo apt-get install jq

For a new installation, clone the repository and enter its root directory:

git clone https://github.com/VinnyLiu0817/Dude.git
cd Dude

Dude stores platform-specific agent definitions under agents/. Copy all the agent files (.md or .toml) that matches the platform you plan to use into that platform's agents directory. The commands below install the agents globally for the current user:

The platform agent directory should be installed by default if you successfully install the agent platform. Otherwise, you can manully creat the directory.

# Codex
mkdir -p ~/.codex/agents

# Claude Code
mkdir -p ~/.claude/agents

# OpenCode
mkdir -p ~/.config/opencode/agents

If you want the agents to be available only in a particular project, copy them into that project's platform directory instead (for example, <project-root>/.claude/agents/ or <project-root>/.opencode/agents/).

Edit info.md to configure:

paper_path = "<path or url to paper PDF>"
code_path = "<path to code repository>"
Maximum_round for negotiation: 2

Choose your preferred LLM provider by selecting agents from:

• Claude: agents/Claude/
• Deepseek: agents/Deepseek/
• GPT: agents/GPT/ (TOML format)
• Kimi: agents/Kimi/

Each provider folder contains pre-configured agent prompts for all eight roles (4 code agents+ 4 paper agents).

To initialize the workflow, user can either paste the contents of ''<orchestrator_main_prompt.md>'' into Codex, Claude Code, or OpenCode, or have the agent read this markdown file directly.

MinerU uploads may fail during peak hours. If you encounter upload failures:

• Change paper_path to paper_url in info.md
• Provide a direct URL to the paper instead of a local file path
• Example:

  paper_url = "https://arxiv.org/pdf/XXXX.XXXXX.pdf"

When using Codex as your agent framework, ensure stable agent activation by configuring your agent information:

• Add your agent credentials to ~/.codex/config.toml
• This guarantees stable and reliable agent activation
• Without proper configuration, agents may fail to initialize consistently