Why choose this project topic?
A study of keeping a document-grounded assistant within its authorised task when sources contain misleading instructions gives this topic a concrete purpose beyond simply adding an AI model. Comparing instruction violations, grounded answers and refusal overreach helps a student demonstrate capability, inspect failure cases and explain where human review or a simpler method remains necessary.
How effectively can prompt injection resistance in local qa support keeping a document-grounded assistant within its authorised task when sources contain misleading instructions, evaluated using instruction violations, grounded answers and refusal overreach?
Choose a feasible licensed or synthetic dataset and available compute budget, then agree a narrow evaluation for keeping a document-grounded assistant within its authorised task when sources contain misleading instructions with your supervisor.
Proposed project objectives
- 01Define the task, evidence and evaluation assumptions for keeping a document-grounded assistant within its authorised task when sources contain misleading instructions.
- 02Create harmless adversarial passages in a local fictional corpus, compare instruction separation strategies and score task adherence and answer grounding.
- 03Measure instruction violations, grounded answers and refusal overreach against a stated baseline and analyse failure cases.
A suggested research approach
Create harmless adversarial passages in a local fictional corpus, compare instruction separation strategies and score task adherence and answer grounding. Check dataset permissions and keep evaluation examples separate from model development. Record model versions, prompts or training settings, then compare instruction violations, grounded answers and refusal overreach using repeatable runs and retain incorrect outputs for a transparent limitations discussion.
What you will need
- Fictional document corpus
- Harmless adversarial test prompts
- Local evaluation harness
Keep your project scope clear
Testing must stay within owned systems and cannot establish complete protection from future attacks. Generated project writing does not include a trained or deployed AI application.
Artificial Intelligence project chapter outline
Use this outline as a starting point. You can edit the chapter titles to match your department’s format during setup.
- Chapter 1Introduction
- Chapter 2Literature Review
- Chapter 3System Analysis and Design
- Chapter 4System Implementation and Testing
- Chapter 5Summary, Conclusion and Recommendations
Turn this topic into your own final year project.
Your title, department, research question and outline are ready. Add your institution, personalise the details and continue to your project workspace.
Generate the Complete Project Generation uses your word balance. Review the draft and supply your own verified research findings.