Show HN: "data-to-paper" – autonomous stepwise LLM-driven research
github.com
github.com
Overall, because of the happy to serve alignment of commercial AI it’s more likely to praise us than be critical, which would mean that OTS models may not fit in to the reviews of methods and conclusions.
Quite an interesting situation, where academia becomes intermediated via LLMs.
I'm happy to see this directly stated. Is there any guidance for domain experts on the types of mistakes an LLM will make? The process will be different from vetting a university student's paper so they are unlikely to know what to look out for. How often will a domain expert reject generated papers? Given the large vetting burden, does this save any time versus doing the research the traditional way? I'm honestly wary domain experts won't be used, careful review won't be performed, and believable AI slop will spread in academic channels that aren't ready to weed out these flawed papers. We're relying pretty heavily on personal ethics here, right?
While it’s kind of clear form the context that it’s about type 2, I doubt a paper like this would pass a peer review without stating it explicitly, in particular with respect to the data set that could potentially include both. Rigor is essential in drawing scientific conclusions.
I guess this is a good example about the statistical nature of LLMs outputs (type 2 is the most common) and consequentially their limitations...
Biases in data will always exist, as this is the nature of our world. We need to think about them carefully and understand the challenges they introduce, especially when training large "foundational" models that encode a vast amount of data about the world. We should be particularly cautious when interpreting their outputs and when using them to draw any kind of scientific conclusions.
I think this is one of many reasons why we implemented the system with inherent human overseeing and strongly encourage people to provide input and feedback throughout the process.
A tool to smoothly catalyze “data to paper” or better yet “data to prevention or treatment” is what we need.
First of all, it's inevitable that LLMs will be/are used in this way and it's great to see development and discussion in the open! That's really important.
Secondly, this will absolutely destroy some areas of science even more than they have already been.
Why? First, science as all of humankind is always a balance between benevolent and malevolent actors. Science already battles data forgery, p-hacking and replication issues. Giving researchers access to tools like this will mean that some conventional quality assurance processes will fail hard. Double-blind peer review will no longer work when there are 10:1 or 100:1 AI generated to high-quality submissions.
Second, doing analysis and writing a paper is one bottleneck of science, but epistemologically, it's not the important one. There are innumerable ways to analyze extant data and it's completely moot to do any analysis in this way. Simmons, Nelson and Simonsohn / Gelman et al. etc have shown: Given a dataset, (1) the findings you can get are practically always from very negative effects to very positive effects, depending on the setup of the analysis. So having one analysis is pointless, especially without theory. (2) even when you give really good labs the same data and question, almost nobody will get the same result (many labs experiment).
What does this tell us? There are a few parts of science that are extremely important and without them science is not only low-impact, it even has a harmful effect by creating costs for pruning and distilling findings. The really important part are causal analyses, and they practically always involve data collection. That's why sciences with strong experimental traditions fare a bit better - when you need to run a costly experiment yourself in order to publish a paper, this creates a strong incentive to think things through and do high-impact research.
So yeah, we've seen this coming and it must create a big backlash that prevents this kind of research from being published, even if vetted humans.
Source: am a scientist, am a journal editor.
Sorry, I just had to. Hottest day of the year in the UK today and warm weather causes me to lose inhibition.
However, it is an open secret that this is already true, is the thing. Meaningful peer review is already confined to islands within a system that has devolved into generating content. The automation of the process doesn't represent a tipping point, and I don't think that the ethically disclosed production of 'research' by large language models is going to represent a significant part of the problem. The errors of the current system will be reduced to absurdity by the existent ethical norms.
"Why is it that only those who have never fought in a battle are so eager to be in one?"
With regard to the debate, I think it's good not to engage in too much black-and-white thinking. Science itself is a pretty muddy affair, and we still haven't grown beyond simplistic null hypothesis significance testing (NHST), even decades after its problematic implications became clear.
That's why it's so important to look at the macro implications: I.e. how does this shift costs? As another comment nicely put it, LLMs are empowering good science, but they are potentially empowering bad science at an order of magnitude more.
Seems to me to be scientific in the same manner ELIZA is therapeutic.
We will be happy to explain and even correct ourselves, if needed, if approached in a civil, respectful manner.
Do you belong to some popperian philosophy? It can't be feyerabendian, since his thinking put virtue as foundational for science. Do you agree with the large journal publishers, that the essence of science is to increase their profits?
Not sure why you think you've earned my respect, and it would be very hard for me to violate your rights since we communicate by text alone.
Your example paper has omitted non-English characters in the names of anyone who gets cited. Look especially at citation [5], where a lot of the authors have European characters in their names which get omitted.
There is also possibly a missing × or ⋅ in "1.81 10^5" on page 3.
Here is a small idea for taking it one step further in the future. Perhaps there could be an additional stage where once the initial data is analyzed and some candidate research ideas generate, a domain knowledge stage is incorporated. So Semantic Scholar API helps generate a set of reference papers currently, instead those papers could be downloaded in full, put into a local RAG, and then have agents read in detail each paper with the summary of the current data in context, effectively doing research, store it’s summaries and ideas in the same RAG, then combine all that context specific research into the material for the further development of the paper.
There is a link to awesome-agents and I’d be curious what their opinion is of various other agent frameworks, especially as I don’t think they actually used any.
For my proposed idea above I think txtai could provide a lot of the tools needed.
Thank you. I’m honored you found it useful.
> From there, you could continue clicking to trace the origins of each fact or idea.
Exactly! I think you would like automated semantic knowledge graph building example in txtai.
Imagine how much could be done when price/token drops by another few orders of magnitude! I can envision a world with millions of research agents doing automated research on many thousands of data sets simultaneously and then pooling their research together for human scientists to study, interpret and review.
txtai has some demos of automated semantic graph building that might be relevant.
I noticed you didn’t really use any existing agent frameworks, which I find very understandable as their value added can be questionable over DIY approaches. However txtai might fit better with your overall technology style and philosophy.
Has your team studied latest CoT, OPA, or research into Cognitive architectures?
this is obviously just my initial impression on a distracted Sunday but I’m very encouraged by your project and I will absolutely be following it and looking at your source code.
The detractors don’t understand LLMs and probably haven’t used them in the way you have and I have. They don’t understand that with CoT and OPA that they can be used to reason and think themselves.
I’ve used them for full automated script writing, performing the job of a software developer. I’ve also used them to create study guides and practice tests, and then grade those tests. When one implements first hand automated systems with agent frameworks using the APIs it gives a deeper understanding of their power over the basic chat usage most are familiar with.
The people arguing that your system can’t do real science are silly, as if the tedious process and logical thinking is something so complex and human that the LLMs can’t do it when used within a cognitive framework, of course they can!
Anyway I’m very exited by your project. I hope this summer to spend at least a week dedicated to setting it up and exploring potential integrations with txtai for use on private knowledge bases in addition to your public Scholarly published papers.
1. Traceability, transparency and verifiability. I think the key question for me is not only whether AI can accelerate science, but rather how we can use AI to accelerate science while at the same time enhancing key scientific values, like transparency, traceability and verifiability.
More and more these days when I read scientific papers, published either at high impact journals or at more specialized journals, I find it so hard, and sometimes even frustratingly impossible, to understand and check what exactly was done to analyze the raw data and get to the key results, what was the specific chain of analysis steps, what parameters where used, etc, etc. The data is often not there or is poorly annotated, the analysis is explained poorly, the code is missing or is impossible to track, etc etc. At all, it became practically impossible to repeat and check the analysis and the results of many peer reviewed publications.
Why are papers so hard to follow and trace? Because writing clear and fully traceable and transparent papers is very hard, and we don’t have powerful tools for doing this, and it requires the scientific process itself (or at least the data analysis part) to be done in an organized and fully traceable way.
Our data-to-paper approach is designed to provide ways to use AI powerfully, not only to speed up science (by a lot!), but also at the same time to use AI to enhance transparency, traceability and verifiability. Data-to-paper sets a standard for traceability and verifiability which imo exceeds the current level of human created manuscripts. In particular:
1. “Data-Chaining": by tracing information flow through the research steps, data-to-paper creates what we call “data-chained” manuscripts, where results, methodology and data are programmatically linked. See this video (https://youtu.be/mHd7VOj7Q-g). You can also try click-tracing results in this example ms: https://raw.githubusercontent.com/rkishony/data-to-paper-sup...
See more about this and more examples in our preprint: https://arxiv.org/abs/2404.17605
2. Human in the loop. We are looking at different ways to create a co-piloted environment where human scientists can direct and oversee the process. We currently have a co-pilot app that allows users to follow the process, to set and change prompts and to provide review comments at the end of each steps (https://youtu.be/Nt_460MmM8k). Will be great to get feedback (and help!) on ways in which this could be enhanced.
3. P-value hacking. Data-to-paper is designed to raise an hypothesis (autonomously, or by user input) and then go through the research steps to test the hypothesis. If the hypothesis test is negative, it is perfectly fine and suitable to write a negative-result manuscript. In fact, in one of the tests that we have done we gave it data of a peer reviewed publication that reports a positive and a negative result and data-to-paper created manuscripts that correctly report both of these results.
So data-to-paper on its own is not doing multiple hypothesis searches. In fact it can help you realize just how many hypotheses you have actually tested (something very hard for human research even when done honestly). Can people ask data-to-paper to create 1000 papers and then read them all and choose only the single one in which a positive result is found? Yes - people can always cheat and science is built on trust, but it is not going to be particularly easier than any other of the many ways available for people to cheat if they want.
4. Final note: LLMs are here and are here to stay and are already used extensively in science doing (sadly sometimes undisclosed: https://retractionwatch.com/papers-and-peer-reviews-with-evi...). The new models of ChatGPT5, ChatGPT6, ... will likely write a whole manuscript for you in just a single prompt. So the question is not whether AI will go into science (it already does), but rather how to do so and use AI in ways that fosters, not jeopardizes, accountability, transparency, verifiability and other important scientific values. This is what we are trying to do with data-to-paper. We hope our project stimulates further discussions on how to harness AI in science while preserving and enhancing key scientific values.
thanks for the honest and thoughtful discussion you are conducting here. Comments tend to be simplistic and it's great to see that you raise the bar by addressing criticism and questions in earnest!
That said, I think the fundamental problem of such tools is unsolvable: Out of all possible analytical designs, they create boring existing results at best, and wrong results (i.e. missing confounders, misunderstanding context ...) as the worst outcome. They also pollute science with harmful findings that lack meaning in the context of a field.
These issues have been well-known for about ten years and are explained excellently e.g in papers such as [1].
There is really one way to guard against bad science today, and that is true pre-registration. And that is something which LLMs fundamentally cannot do.
So while tools such as data-to-paper may be helpful, they can only be so in the context of pre-registered hypotheses where they follow a path pre-defined by humans before collecting data.
[1] http://www.stat.columbia.edu/~gelman/research/unpublished/p_...
I can’t but fully agree: pre-registered hypothesis is the only way to fully guard against bad science. This in essence is what the FDA is doing for clinical trials too. And btw lowering the traditional and outdated 0.05 cutoff is also critical imo.
Now, say we are in a utopian world where all science is pre-registered. Why can’t we imagine AI being part of the process that creates the hypotheses to be registered? And why can’t we imagine it also being part of the process that analyzes the data once it’s collected? And in fact, maybe it can even be part of the process that help collects the data itself?
To me, neither if we are in such a utopian world, nor in the far-from-utopian current scientific world, there is ultimately no fundamental tradeoff between using AI in science and adhering to fundamental scientific values. Our purpose with data-to-paper is to demonstrate and to provide tools to harness AI to speed up scientific discovery while enhancing the values of traceability and transparency and make our scientific output much more traceable and understandable and verifiable.
As of the question of novelty: indeed, research on existing public datasets which we have currently done cannot be too novel. Though scientists can also use data-to-paper with their own fascinating original data. It might help in some aspects of the analysis, certainly help them keep track of what they are doing and how to report it transparently. Ultimately I hope that such co-piloting deployment will allow us delegating more straight forward tasks to the AI and letting us human scientists to engage in higher level thinking and higher level conceptualization.
My concern is an ecological one within science, and your argument addresses the frontier of scientific methods.
I am sure both are compatible. One interesting question is what instruments are suitable to reduce negative externalities from bad actors. Pre-registration works, but is limited to few fields where the stakes are high. We will probably similarly see a staggered approach with more restrictive methods in some fields and less restrictive ones in others.
That said, there remain many problems to think about: E.g. what happens to meta-analyses if the majority of findings comes from the same mechanism? Will humans be able to resist the pull of easy AI suggestions and instead think hard where they should? Are there sensible mechanisms for enforcing transparency? Will these trends bring us back to a world in which trust was only based on prestige of known names?
Interesting times, certainly.
This doesn't seem correct to me at all. If new data is provided and the LLM is simply an advanced tool that applies known analysis techniques to the data, then why would they create “boring existing results”?
I don’t see why systems using an advanced methodology should not produce novel and new results when provided new data.
There is a lot of reactionary or even luddite responses to the direction we are headed with LLMs.
I assume you mean that LLMs can generate new insights in the sense of producing plausible results from new data or in the sense of producing plausible but previously unknown results from old data.
Both these things are definitely possible, but they are not necessarily (and in fact often not) good science.
Insights in science are not rare. There are trillions of plausible insights, and all can be backed by data. The real problem is the reverse: Finding a meaningful and useful finding in a sea of billion other ones.
LLMs learn from past data, and that means they will have more support for "boring", i.e. conventional hypotheses, which have precedent in training material. So I assume that while they can come up with novel hypotheses and results, these results will probably tend to conform to a (statistically defined) paradigm of past findings.
When they produce novel hypotheses or findings, it is unlikely that they will create genuinely meaningful AND true insights. Because if you randomly generate new ideas, almost all of them are wrong (see the papers I linked).
So in essence, LLMs should have a hard time doing real science, because real science is the complex task of finding unlikely, true, and interesting things.
I’d be surprised if you have used LLMs beyond the classic chat based linear interface that is commonly used and still have the opinions you do.
In my opinion, once you combine RAG and agent frameworks with raw observational input data they can absolutely do real reasoning, analysis, and create new insights that are meaningful and will be considered genuine new science. This project/group we are discussing have practically proven this with their replication examples. The reason this is possible is because the LLM is not just taught how to repeat information but it can actually reason and analyze at a human level and beyond when utilizing it’s capabilities within a well designed cognitive architecture using agents.
The whole paper is "we took an existing dataset and ran the simplest reasonable model (a logistics regression) on it". That's about 5-10 minutes in R (or Python, or SAS, or whatever else). It's a very well-understood process, and it's a good starting point to understand the data, but it can't be the only thing in your paper, this isn't the 80's anymore.
The overall style is verbose and flowery, typical of LLMs. Good research papers should be straightforward and to the point. There's also strange mixing of "we" and "I" throughout.
We learn in the introduction that interaction effects were tested. That's fine, I'd want to see it set up earlier why these interaction effects are posited to be interesting. It said earlier that "a comprehensive investigation considering a multitude of diabetes-influencing lifestyle factors concurrently in relation to obesity remains to befully considered", but quite frankly, I don't believe that. Diabetes is remarkably well-studied, especially in observational studies like this one, due to its prevalence. I haven't searched the literature but I really doubt that no similar analysis has been done. This is one of the hardest parts of a research paper, finding existing research and where its gaps are, and I don't think an LLM will be sufficiently capable of that any time soon.
There's a complete lack of EDA in the paper. I don't need much (the whole analysis of this paper could be part of the EDA for a proper paper), but some basic distributional statistics of the variables. How many responses in the dataset were diabetic? Is there a sex bias? What about age distribution? Are any values missing? These are really important for observational studies because if there's any issues they should be addressed in some way. As it is, it's basically saying "trust us, our data is perfect" which is a huge ask. It's really weird that a bunch of this is in the appendix (which is way too long to be included in the paper, would need to be supplementary materials, but that's fine) (and also it's poorly formatted) but not mentioned anywhere in the paper itself. When looking at the appendix, the main concern that I have is that only 14% of the dataset is diabetic. This means that models will be biased towards predicting non-diabetic (if you just predict non-diabetic all of the time, you're already 86% accurate!). It's not as big of an issue for logistic regression, or for observational modeling like this, but I would have preferred an adjustment related to this.
In the results, I'm disappointed by the over-reliance on p-values. This is something that the statistics field is trying to move away from, of a multitude of reasons, one of which is demonstrated quite nicely here: p-values are (almost) always miniscule with large n, and in this case n=253680 is very large. Standard errors and CIs have the same issue. The Z-value is the most useful measure of confidence here in my eyes. Effect sizes are typically the more interesting metric for such studies. On that note, I would have liked to see predictors normalized so that coefficients can be directly compared. BMI, for example, has a small coefficient, but that's likely just because it has a large range and variance.
It's claimed that the AIC shows improved fit for the second model, but the change is only ~0.5%, which isn't especially convincing. In fact, it could be much less, because we don't have enough significant figures to see how the rounding went down. p-value is basically meaningless as previously stated.
The methods section says almost nothing that isn't already stated at least once. I'd like to know something about the tools which were used in this section, which is completely lacking. I do want it highlight this quote: "Both models employed a method to adjust for all possible confounders inthe analysis." What??? All possible confounders? If you know what that means you know that that's BS. "A method"? What is your magic tool to remove all variance not reflected in the dataset, I need to know! I certainly don't see it reflected in the code.
The code itself seems fine, maybe a little over-complicated but that might be necessary for how it Interfaces with the LLM. The actual analysis is equivalent to 3 basic lines of R (read CSV, basic log reg with default parameters 1, basic log reg with default parameters 2).
This paper would probably get about a B+ in 261, but shouldn't pass a 400-level class. The analysis is very simple and unimpressive for a few reasons. For one, the questions asked of the dataset are very light. More interesting, for example, might have been to do variable selection on all interaction terms and find which are important. More models should have been compared. The dataset is also extremely simple and doesn't demand complex analysis. An experimental design, or messy data with errors and missing values, or something requiring multiple datasets, would be a more serious challenge. It's quite possible that one of the other papers addresses this though.
You suggested some directions for more complex analysis that could be done on this data - I would be so curious to see what you get if you could take the time to try out running data-to-paper as a co-pilot on your own - you can then give it directions and feedback on where to go - will be fascinating to see where you take it!
We also must look ahead: complexity and novelty will rapidly increase as ChatGPT5, ChatGPT6 etc are rolled in. The key with data-to-paper is to build a platform that harnesses these tools in a structured way that creates transparent and well-traceable papers. Your ability to read and understand and follow all the analysis in these manuscripts so quickly speaks to your talent of course, but also to the way these papers are structured. Talking from experience, it is much harder to review human-created papers at such speed and accuracy...
As for your comments on “it's certainly not close to something I could submit to a journal” - please kindly look at the examples where we show reproducing peer reviewed publications (published in a completely reasonable Q1 journal, PLOS One). See this original paper by Saint-Fleur et al: https://journals.plos.org/plosone/article?id=10.1371/journal...
and here are 10 different independent data-to-paper runs in which we gave it the raw data and the research goal of the original publication and asked it to do the analysis reach conclusions and write the paper: https://github.com/rkishony/data-to-paper-supplementary/tree... (look up the 10 manuscripts designated “manuscriptC1.pdf” - “manuscriptC10.pdf”)
See our own analysis of these manuscripts and reliability in our arxiv preprint: https://arxiv.org/abs/2404.17605
Note that the original paper was published after the training horizon of the LLM that we used and also that we have programmatically removed the original paper from the result of the literature search that data-to-paper does so that it cannot see it in the search.
Thanks so much again and good luck for the exam tomorrow!
Even if this thing works I wouldn’t call it “end-to-end scientific research”. IMHO the most challenging and interesting part of scientific research is coming up with a hypothesis and designing an experiment to test it. Data analysis and paper writing is just a small part of the end-to-end process.
> Towards this goal, data-to-paper systematically guides interacting LLM and rule-based agents through the conventional scientific path, from annotated data, through creating research hypotheses, conducting literature search, writing and debugging data analysis code, interpreting the results, and ultimately the step-by-step writing of a complete research paper.
Then it’s all just wrong, automated p-hacking. You’re supposed to start with the hypothesis, not generate it from the data you’re about to publish.
But I guess these days in many fields of science and in popular parlance "data" has become synonymous with "observation" and "writing papers" with "research", so.