Exploratory Study on Multi-User Program Synthesis: A Multi-Wizard Approach
summary
Proposes two study designs: a between-subjects lab with 24 participants (pairs vs individuals, each doing one task with and one without a synthesizer) and a field study observing 4–5 teams (20–25 participants) for one week on ongoing projects. Claims group-created specifications encode more edge cases; pilot regex tasks show classical synthesizers return lower-quality, slower candidate programs. Plans to compare LLM-based synthesizers for prompt flexibility versus traditional synthesis and to analyze incidents that lead to success or failure.
AI-written summary · about
chapters (16)
- 0:00Multi-user program synthesis intro
- 2:01Limitations of single-user tools
- 3:08Research goals and design
- 5:13Student study between-subjects
- 7:05Generalizing to professional teams
- 8:28Key questions and future work
- 10:18Hackathon tasks and conditions
- 11:22Data collection and analysis
- 13:08Expected collaborative phenomena
- 14:55LLMs and code review
- 16:50Teamwork configurations
- 19:26Community input on tools
- 20:35LLMs vs classical synthesizers
- 22:17Handling team disagreements
- 23:45Synthesis with vs without teams
- 24:28Closing remarks and Q&A
related talks
09mProgram Synthesis for Musicians: A Usability Testbed for Temporal Logic Specifications
54mA World Without Programmers?
28mScooPy: Enhancing Program Synthesis with Nested Example Specifications
41m"Do What I Mean": The History of AI and Program Synthesis
53mSemantics-based program manipulation in the age of LLMs
18mGrounded Copilot: How Programmers Interact with Code-Generating Models
01h 16mAnthropic Workshop: Build Agents That Run for Hours
16mMulti-modal Program Inference: a Marriage of Pre-trained Language Models and Component-based Synthes