How Far Are We From True Auto-Research?

Zhengxin Zhang, Ning Wang, Sainyam Galhotra, Claire Cardie

arXiv:2605.19156 · 2026-05-20 공개 · arXiv · PDF

codex auto-research claude-code kimi-code artifact-aware-review experimental-rigor paper-evaluation research-personas

Abstract

Recent auto-research systems can produce complete papers, but feasibility is not the same as quality, and the field still lacks a systematic study of how good agent-generated papers actually are. We introduce ResearchArena, a minimal scaffold that lets off-the-shelf agents (Claude Code using Opus 4.6, Codex using GPT-5.4, and Kimi Code using K2.5) carry out the full research loop themselves (ideation, experimentation, paper writing, self-refinement) under only lightweight guidance. Across 13 computer science seeds and 3 trials per agent-domain pair, ResearchArena yields 117 agent-generated papers, each evaluated under three complementary lenses: a manuscript-only reviewer (SAR), an artifact-aware peer review (PR) in which agents inspect the workspace alongside the manuscript, and an human conducted meta-review. Under SAR alone the picture is optimistic: Claude Code obtains the highest score, outperforms Analemma's FARS, and matches the weighted-average human ICLR 2025 submission, suggesting that minimally scaffolded agents can produce papers that look competitive on manuscript-only review. Manual inspection, however, reveals this picture is overstated: SAR scores are poorly aligned with its actual acceptance decisions and reward plausible framing without verifying experimental substance. Under artifact-aware PR scores drop sharply, and manual auditing identifies experimental rigor as the major bottleneck, decomposing into three failure modes (fabricated results, underpowered experiments, and plan/execution mismatch) that are highly agent-dependent: Codex 5%/8% paper-vs-artifact mismatch / fabricated references versus Kimi Code 77%/72%, a $\sim$15$\times$ spread that tracks distinct research personas the agents develop. None of the 117 agent-generated papers reaches the acceptance bar of a top-tier venue. This suggests that we are still gapped from the true auto-research.

한국어 요약

한 줄 요약

**[자동 연구 / 평가]** 최소 스캐폴드 ResearchArena 위에서 프론티어 에이전트 3종이 만든 117편 논문을 세 가지 평가 렌즈로 분석해, 매뉴스크립트-only 점수의 낙관과 artifact-aware/메타리뷰에서의 큰 격차를 정량화.

핵심 기여도

핵심 아이디어

"자동 연구" 평가는 표면적 매뉴스크립트 품질이 아니라 실험 substance와 plan/실행 정합성까지 함께 봐야 의미가 있으며, 각 에이전트는 서로 다른 "research persona"를 형성한다.

기술적 접근법

주요 결과

의의 및 한계

**의의**: auto-research 분야의 "feasible ≠ good" 인식을 체계적 증거로 확립, 평가 방법론의 결함(SAR의 framing bias)을 노출. **한계**: 13 CS 시드 한정, 비-CS 도메인·장기 연구 트랙으로의 일반화는 후속 과제.

실용적 활용