ResearchArena Benchmarks AI Research Paper Quality

A recent study has unveiled ResearchArena, a streamlined framework that allows off-the-shelf AI agents to independently navigate the entire research process—covering ideation, experimentation, writing, and self-improvement—with minimal oversight. The platform underwent testing with 13 computer science seed topics, yielding 117 papers from three trials per agent-domain combination. The agents assessed were Claude Code with Opus 4.6, Codex with GPT-5.4, and Kimi Code with K2.5. Evaluations were conducted through three perspectives: a manuscript-only review (SAR), an artifact-aware peer review (PR), and a human-led meta-review. Claude Code excelled under SAR, surpassing Analemma's FARS and equaling human-written works. This research underscores the variability in quality among automated research outputs and the need for systematic evaluation. The findings set a foundation for future advancements in automated research.