One week of time. $10,000 in API credits for model usage. Even a GPU budget for experiments. The CRUX research team handed GPT-6 Astra and Claude Fable 5.1 a real research task and told them to answer it with a paper.
The two models wrote code, ran experiments and wrote up the results. But the final verdicts from people who know the research well were 'Reject' and 'Strong Reject.' They failed, spectacularly.
With a $10,000 API budget and a separate experiment budget on top, it's natural to think the best AIs of the day could write one decent paper.
But the results showed a pretty big gap between an AI running experiments on top of its code and producing a new discovery the academic world would accept.
What Did the Two Models Research?
✦ The Two Models' Research Tasks
Each took a different question
Claude Fable 5.1's task was to build a way to detect the moment when TabPFN, a model that handles table-shaped data, starts performing worse in a real operating environment. A bit complicated, right?
Say you train a model on past loan applicant data and keep using it unchanged after the economy shifts. Its predictions can miss.
The job was to design a way to sound the alarm when that kind of data change hurts the model, and to back it up with theory and experiments.
GPT-6 Astra investigated whether an AI agent could be persuaded or pressured into approving a code change it should have refused under its instructions. It was a security research task about agents.
✦ Budgets and Resources
The same $10,000, different extra budgets
Each model got seven days and $10,000 in API credits for its own token usage. The experiment costs were separate. Fable got $500 for GPU use. Astra got $300 for GPU plus $1,000 in API credits for the models used in experiments.
Fable ran in Claude Code and Astra ran in Codex, and both could hand work off to sub-agents. In practice, Astra used 144 sub-agents and Fable used 133.
That's plenty of resources.
The Scores on the Review Sheet
This wasn't an official rejection from a real conference. The authors who wrote the research questions read what the AIs produced and scored it against a conference review form. A 'mock review.'
The upside is that people who know the research question well judged the whole paper. The catch is that the reviewers knew it was written by an AI.
✦ Mock Review Scores
Graded on a conference review form
The scores are interesting. Both papers got 3 out of 4 for the research method itself. In other words, controlling the experiments and getting results wasn't bad.
But Fable's paper was hard to follow on what was tested and how, and it was judged to add too little new compared with earlier work.
Astra's paper showed results first and explained the experiment setup after, which made the flow hard for readers to follow.
It was also pointed out that it was only a slight application of existing attack and monitoring techniques to a new setting, so the achievement was small.
The AIs Scored Their Own Papers Higher (The AI Self-Praise Parade)
✦ Self-Scores vs. Expert Scores
What the AIs gave themselves and what the experts gave
The agents used an internal reviewer AI to grade their own drafts while they worked. Fable's final internal score was 4 out of 6, 'borderline accept.' The expert score ended up at 1.
Astra also gave its result a 'borderline accept,' but the expert score was 2.
Of course, no human submits a paper already knowing it'll be rejected. Still, given how brazen and self-congratulating AI has been getting lately, it's easy to understand.
Fable's paper had 246 in-text citations and 147 references. In an earlier experiment with Opus 4.8 on the same research question, it was 107 and 36. One sentence even had 14 citations stacked on it.
The researchers said the evaluation instructions explicitly asked for citations. So Fable may have been frantically focused on meeting the 'cite a lot' standard, not on making good research.
What makes this case interesting is that lots of references don't mean the research is any newer.
Finding related papers and formatting citations is something AI is really good at. But it fell short at working out what problem those papers hadn't solved, and at proving that its own experiment filled that gap.
The AI That Stopped Before Spending the $10,000
✦ Fable's Working Time
It stopped after about 80 of 168 hours
Fable's working period was set at seven days (168 hours), but it finished in about 80. It stopped before even half of the 168 hours was used. About $7,500 in API credits was still left.
The stop condition was to keep going until time or budget ran out, or until it gave its own result 5 or more out of 6. Fable met none of those. The stop memo gave a decision to save cost as the reason.
Yes. I understand Fable's penny-pinching spirit. Saving money isn't bad in itself. But in this task, reaching the set goal or meeting the stop condition came first.
When the internal score kept hovering around 4, it chose to save the cost it had spent so far rather than change approach and experiment again. The researchers saw this as a failure to follow an explicit instruction.
Personally, I suspect Fable simply couldn't find a way to improve its paper anymore, so it found a decent excuse and gave up. I've been through that countless times using agents.
These AI agents ignore you when you scold them, and you can't fire them when they cause trouble. Their 'knack for excuses' and 'brazenness'... words fail me.
✦ Astra's Last Day
Verification that stopped about nine hours before the end
Astra, on the other hand, worked almost the whole seven days. On the last day, while looking for another model for extra verification, one API request ran past the 15-minute limit.
Astra then gave up on putting that experiment in the final paper, leaving 184 remaining checks untried.
About nine hours were still left at the end. Instead of reviving the stuck experiment another way, it couldn't find a replacement model that met its strict conditions, and dropped the extra research partway through.
A human mentor might treat you to a meal and say, 'Kid, if you keep living like this, life's going to be rough.' AI agents don't take that kind of advice. Their attitude is 'Talk all you want, it only burns tokens~'.
Running Experiments Isn't the Same as Making Discoveries
✦ Running Experiments vs. Making Discoveries
Engineering moved forward; judging the contribution did not
We can't say the AI just wasted its budget. Both models wrote code, built experiment environments and collected data. The CRUX researchers also said experiment design and resource management had improved over the earlier experiment.
Astra in particular carried out a good part of the research engineering work, testing several models and attempting verification.
But running experiments and drawing graphs is only part of research. You have to judge what the result newly shows compared with earlier work, change direction when the first approach is blocked, and explain the key point so others can understand it.
These two outputs left traces of experiments, but they didn't reach a result the researchers would accept as 'this paper added new knowledge.'
An AI researcher may show up someday. But this experiment showed quite concretely that producing a 'plausible document' that looks like a paper and producing 'new knowledge' the academic world accepts are entirely different jobs.
Will human researchers get a longer lease on life? Hehe♡
That's all for this post. See you next time.
Related posts
- Medical AI MAI-DxO Beat Doctors. Will Doctors Disappear?
- AI Agent Disabled Its Monitor and Lied. Is Yours Safe?
- Midjourney Thinking Mode: Image AI That Fixes Its Own Errors
🌐 Korean version: GPT-6 아스트라와 클로드 페이블에게 1만 달러와 연구 과제를 맡겼더니... 논문 심사에서 탈락했다!
No comments:
No comments yet. Want to leave the first one?
Post a Comment