QUIT: A Human-in-the-Loop Platform for AI Research Automation

Large language models (LLMs) can help researchers search the literature, generate ideas, write code, and draft papers. However, in current practice, these capabilities are often used through fragmented interactions hidden inside long chat histories, making the resulting research process difficult to reproduce, audit, or resume after failure. We present QUIT, a human-in-the-loop platform for AI research automation that organizes the research process into four stages: Query, Understand, Implement, and Tell. Rather than using conversational memory as the primary interface between stages, QUIT externalizes intermediate decisions and outputs into structured artifacts. This design avoids the token cost and contextual instability of repeatedly carrying long interaction histories, while making the research process easier to trace, inspect, and reproduce. Moreover, QUIT supports stage-level invocation and repair, allowing researchers to locally inspect intermediate artifacts, identify failure points, and intervene with timely corrections.