View Proposal

Proposer
Gavin Abercrombie
Title
Safety for "Agentic" Language Model Workflows
Goal
Develop a framework to evaluate safety for language models deployed in multo-step workflows
Description
Language models are frequently employed in (semi-)autonomous, multi-step workflows (so-called "Agentic AI"). While safety, i.e. production of harm-free outputs has traditionally been evaluated with static single-run benchmarks, this paradigm requires dynamic evaluation throughout the task being carried out. This project will aim to: - Map the relevant safety issues and existing resources - Develop an evaluation framework - Evaluate a range of models and tasks
Resources
https://aclanthology.org/2026.findings-acl.727/
Background
Url
Difficulty Level
High
Ethical Approval
None
Number Of Students
1
Supervisor
Gavin Abercrombie
Keywords
llms, safety
Degrees
Master of Science in Artificial Intelligence
Master of Science in Data Science
Master of Science in Software Engineering
Back to List