Can AI agents conduct open-ended AI research
A · Strong evidence Princeton's CRUX team (Arvind Narayanan, Sayash Kapoor) with the UK AI Security Institute and Helen Toner introduce a new evaluation method with released logs and reviews, but the evidence rests on two case studies.
Summary based on the full text·Original: about 50 minutes to read
In short
An arXiv paper (v2, August 2026) by Peter Kirgis, Sayash Kapoor, Arvind Narayanan and co-authors from Princeton, the UK AI Security Institute, Georgetown CSET, Stanford and others. It introduces "shadow evaluations": a frontier agent receives the central research question of an unpublished NeurIPS 2026 submission and the paper's own authors grade the result as conference reviewers. In two such runs the agents handled all the engineering without help but did not answer the research questions. Both papers were rejected, and the authors identify five recurring failure modes.
Key numbers
- 2/6 and 1/6overall scores the original authors gave the two agent-written papers; both unambiguous rejections
- $3,000Anthropic API credit per run, plus GPU credits, six days of wall-clock time and full web access
- <50%share of the API budget the OpenClaw/Opus 4.8 agents spent before stopping
- 30%median probability surveyed experts gave that an agent would reach weak accept or better
As reported by the source.
English·Web page·opens arxiv.org
What it reports
- Each agent got six days, $3,000 in Anthropic API credits, GPU credits, a VM and open web access, with the goal of a top-conference paper.
- Reviewer scores: Paper 1 (Personas) quality 2/4, clarity 1/4, originality 3/4, overall 2/6; Paper 2 (TabPFN) quality 1/4, overall 1/6. Reviewer confidence 4/5 and 5/5.
- Both OpenClaw/Opus 4.8 runs ended with less than 50% of the API budget spent; the TabPFN agent used 40% of its token budget.
- The agents retired their most ambitious research targets within the first ten hours; the TabPFN agent falsified six approaches in 14 hours and never changed course despite 110 hours left.
- Across dozens of revision rounds, the agents' own AI self-review never returned an acceptance; agents responded by adding caveats rather than redesigning experiments.
- Five failure modes: poor judgment about the publishable bar, uncreative responses to design flaws, weak backtracking, poor resource awareness, instruction drift.
- A repeat with GPT-5.6 Sol in Codex reproduced the failures; that agent spent the full $3,000 in just over two days, leaving nearly 100 hours unused.
- Surveyed experts gave a median 30% chance of weak accept or better and 60% for producing findings the authors found relevant; the second outcome did occur.
How they got it
Two shadow evaluations on unpublished NeurIPS 2026 submissions, run with the provider-agnostic OpenClaw scaffold after dry runs selected Opus 4.8, plus one robustness run with GPT-5.6 Sol and Codex. The original authors reviewed the outputs using conference criteria. The team analysed the full agent logs and surveyed external experts beforehand; reviews, repositories and logs are released.
Limits
- Two papers in machine learning only; results may not generalise to other research questions or later models.
- Reviews were not blind: authors knew the work was AI-generated and may prefer their own approach. The authors judge the outputs clearly weak regardless.
- Results are sensitive to scaffold design; follow-up runs with newer models are planned.
Why it matters for companies
The paper separates execution from judgment: agents ran hundreds of GPU hours of experiments and large literature reviews without help, yet could not decide when evidence was sufficient or when to start over. That split applies to analytical and strategy work in any firm, where the costly errors are framing and stopping decisions. Agents also ignored explicit budget and format instructions, so delegated work still needs human checkpoints on scope, spend and acceptance criteria. Claims that AI will soon run open-ended research or analysis unsupervised should be weighed against this evidence.
Summary based on the full text. Read the original for the full argument and data.Read the original on arxiv.org
- Tags
- AI capability