Are there examples of the outputs the LLMs under test generated? I couldn't find any detailed ones in the paper or code.
The result here seems to be "Our Judge LLM gave another LLM a 21% grade for some code it generated", which is ... not qualitatively meaningful at all to me.
We evaluate several frontier models on PaperBench, finding that the best-performing tested agent, Claude 3.5 Sonnet (New) with open-source scaffolding, achieves an average replication score of 21.0%.
overall i REALLY like this paper and effort, but this part sounds like a bit of bullshit. they dont have the ability to implement retries and backoffs to deal with rate limits?
I've been developing a more elaborate variation on the "chat with a pdf" idea for my own use as a researcher. It's mostly designed for a historian's workflow but it works pretty well for science and engineering papers too. Currently Flash 2.0 is the default but you can select other models to use to analyze pdfs and other text through various "lenses" ranging from a simple summary to text highlighting to extracting organized data as a .csv file:
(Note: this is not at all a production ready app, it's just something I've been making for myself, though I'm also now sharing it with my students to see how they use it. If anyone reads this and is interested in collaborating, let me know).
I regularly paste papers into LLM interfaces but they all spit out generic non-helpful answers. Your app is the only one i've seen that actually helps me understand.
I'm building such tools at https://sugaku.net, right now there's chatting with a paper and browsing similar papers. Generally arXiv and other repositories want you to link to them and not embed their papers, which makes it hard to build inline reading tools, but it's on my roadmap to support that for uploaded papers. Would love to hear if you have some feature requests there
One feature could be that it automatically fetches the papers that it refers to and also feeds them through the llm. And maybe apply that recursively. This could give the AI a better overview of the related literature.
What were the human PhDs able to do after more than 48 hours of effort? Presumably given that these are top-level PhDs, the replication success rate would be close to 100%?
Depending on how well the exact algorithms, implementation details, and experimental design were documented, replication can easily take days, if not weeks. (Personally, I would start by filtering out papers that cannot be replicated by well-skilled researchers in a fixed amount of time and only give the replicatable ones to the agents.)
The point is to bootstrap self-improving AI. Once a measurement becomes a goal, model makers target saturating it.
There is a coefficient of intelligence replication ie: Model M with intelligence I_m, can reproduce a model N with intelligence I_n. When (I_n / I_m) > 1 we'll have a runaway intelligence explosion. There are of course several elements in the chain - akin to the Drake equation for intelligent machines - and their combined multiplicative effect determines the overall intelligence of the system. If f(paper) -> code is the weakest part of the chain, it makes sense to target that.
You don’t see the value of independent replication of findings?
The agent didn’t have access to the code, although they acknowledge it could theoretically be in the training set, even then the original code wouldn’t conform to the structure of the test.
There is a planet-wise eternal 100% safe AI solution that can be a billion dollar startup, too:
Put all the GPUs in cloud/s controlled by international scientists (now you can use your GPU on any device, can earn money by renting it when you don’t need it, nothing changes except you need to be online to us it, but we’ll have 5G and better worldwide. You can develop, sell or release free math-proven safe AI models in this cloud “AI App Store”, etc).
Because the main risk is an AI agent botnet - current GPUs are like nukes that are 100% unprotected - any hacker can make a virus with AI agent component just to steal money, this AI will be not aligned at all, will become a per perpetual and eventually autonomous botnet.
The result here seems to be "Our Judge LLM gave another LLM a 21% grade for some code it generated", which is ... not qualitatively meaningful at all to me.
reply