MUSE: An Interactive Meta-Agent for Understanding and Steering LLM-powered Data Science Systems

Details

Wei-Hao Chen and four colleagues at UMass Amherst and collaborating institutions present MUSE, accepted at UIST 2026 (a top HCI venue, November 2-5, Detroit). MUSE sits as a meta-agent layer on top of existing LLM-powered data science agents, addressing a practical gap: users can invoke agentic workflows in natural language but struggle to diagnose failures or steer them when outputs are wrong. MUSE does three things: dynamically restructures low-level execution traces into multiple semantic levels (from high-level pipeline overview to low-level implementation details); lets users reference specific workflow steps for grounded questions, feedback, and revisions without manually searching execution logs; and implements mixed-initiative steering — proactively surfacing suspicious steps for inspection and scaffolding the repair process. A between-subjects study (n=15) demonstrates improved task efficiency and increased user confidence in understanding and steering agentic workflows. The meta-agent approach is agent-agnostic: MUSE wraps any underlying agentic data science system without modifying its internals, making it deployable on top of existing tools like Copilot for data analysis or similar.