PersonaCodeBench: Evaluating Coding Agents in Persona-driven Human-Interactive Sessions
Summary
- Researchers introduce PersonaCodeBench, a benchmark for evaluating coding agents in persona-driven, multi-turn interactive sessions.
- The benchmark turns user characteristics like expertise, preferences, and constraints into verifiable obligations.
- A controller governs user simulation and requirement disclosure, while separate verifiers assess native correctness and persona fulfillment.
- Evaluation of major frontier agents on PersonaCodeBench establishes a comprehensive leaderboard for personalized requirement fulfillment.
Why it matters
As coding agents become more interactive, evaluating how well they adapt to user preferences, expertise, and constraints is critical for real-world deployment.