Deconstructing AI Prose Quirks
The subtle 'quirks' in AI-generated prose, often a result of 'overfitting,' are increasingly being mimicked by human writers, blurring the lines of textual authenticity. This challenge tasks you with building a sophisticated multi-agent system using Gemini 3 Pro and Langroid to perform linguistic forensics, accurately identifying AI-generated content, human-mimicking-AI content, and purely human writing. Your system will employ A2A protocol for collaborative analysis, leveraging Gemini 3 Pro's Deep Think mode for profound semantic and stylistic insights. Agents will adapt their thinking budgets based on the complexity of the text, providing detailed reports on characteristic AI linguistic patterns and suggesting 'de-AI-ification' strategies to restore human-like prose, effectively becoming a 'prose purity' guardian.
What you are building
The core problem, expected build, and operating context for this challenge.
Develop a Langroid multi-agent system with Gemini 3 Pro Deep Think mode and OpenAI o3, using A2A protocol and adaptive thinking budgets for linguistic forensics to analyze AI prose quirks.
How work is evaluated
Evaluation will assess the system's ability to accurately classify text origins and provide insightful linguistic analysis. It will also examine the efficiency of A2A communication and adaptive budget utilization.
Challenge datasets
Public datasets, plus the private datasets you have access to.
What you should walk away with
Utilize Gemini 2.5 Pro's Deep Think mode for granular linguistic pattern recognition.
Design and implement A2A communication protocols for agent collaboration.
Develop techniques to differentiate between various text origins (AI, human-mimicking-AI, human).
Implement a multi-agent system using Langroid for collaborative text analysis.
Propose strategies for 'de-AI-ifying' text back to natural human prose.
Reference links and supporting material
Sample data for 2 tasks
Challenge setup
Evaluation will assess the system's ability to accurately classify text origins and provide insightful linguistic analysis. It will also examine the efficiency of A2A communication and adaptive budget utilization.
Challenge input
{'texts': [{'id': '1', 'content': 'sample text 1'}, {'id': '2', 'content': 'sample text 2'}...]}
Agent execution
No agent tool is listed for this challenge.
Evaluated output
{'analyses': [{'id': '1', 'origin': 'AI-generated', 'confidence': 0.9, 'quirks': ['formulaic phrasing', 'lexical redundancy']}, ...]}
- Verify that agents communicate effectively using the A2A protocol, exchanging analysis results and insights.
- Confirm that the system intelligently adjusts reasoning budgets based on text complexity, using Deep Think...
- Classification Accuracy target: 0.9
- Python execution harness
Summary shows the first task, up to three tools, and up to three evidence items. Open the technical recipe to view all configured items.
View technical recipe
Configured tools
No tool records are attached.
Evaluation contract
- Verify that agents communicate effectively using the A2A protocol, exchanging analysis results and insights.
- Confirm that the system intelligently adjusts reasoning budgets based on text complexity, using Deep Think only when necessary.
Metrics
- Classification Accuracy · Target 0.9
- De-AI-ification Score · Target 4.5
Workflow tasks
- Task 1
Input: {'texts': [{'id': '1', 'content': 'sample text 1'}, {'id': '2', 'content': 'sample text 2'}...]}
Output: {'analyses': [{'id': '1', 'origin': 'AI-generated', 'confidence': 0.9, 'quirks': ['formulaic phrasing', 'lexical redundancy']}, ...]}
Execution harness: Python
- Task 2
Input: {'text': 'AI-generated text'}
Output: {'original_text': '...', 'suggested_rewrite': '...', 'recommendations': ['use simpler vocabulary', 'add personal anecdote']}
Execution harness: Python
Recipe freeze marker
No recipe freeze time is recorded.