Analyzing the quality and implications of a scientific papers

[spark glade]

As I go through hundreds of scientific papers every week, I have been tinkering with a prompt to help me get a quick overview of the quality and implications of a paper. This is where I'm currently at. I find that it gives okay outputs, but I'm having a bit of a problem steering it towards providing me with more examples when the findings are useful. Also, I'm having trouble getting it to stop saying "more research is needed," which is a cliché in all scientific papers that is probably deeply ingrained in the LLM. This might seem like a cynical approach to research, but it's born of years of disappointment with the usefulness of papers.

The prompt: You are tasked with summarizing and critiquing a scientific research paper. In your analysis, consider the following: Provide a brief overview of the study and its findings. Evaluate the potential real-world applications. Discuss specific behaviors, actions, or changes that could be targeted. Focus on the effect size over p-value when discussing statistical significance. Critically analyze the research methodology. Evaluate for common pitfalls such as: Publication bias, P-hacking, Causation vs. correlation fallacy, overpowered or underpowered studies, Lack of preregistration, Selective reporting, Overreliance on observational studies, Poor sample selection (e.g., studies using mice or online surveys like Mechanical Turk), Lack of data transparency and availability. Comment on whether the researcher's conclusions are overstated. Assess the overall validity and logic of the study's concepts, methods, and derived conclusions. Keep your critique succinct, avoid clichés, and ensure each point is only discussed once for clarity.

[nimble thicket]

I would suggest assigning it the role of 'expert' at what you want it to do before giving it the task, just as a basic tip. Unsure how to solve the problems you're having though as they sound quite domain specific.

[civic crown]

have you tried adding to the prompt "You will never, under any circumstances, say that "more research is needed" or anything akin to its intent"

for my prompts i find if i get a non optimal response i can just tell it not to provide responses of that variety

[spark glade]

Interesting. I'm going to test it out today to see if I can "feel" a difference. Maybe the "You are an expert in the field of clinical research" would anchor it better as a reviewer. Unless I am misunderstanding you. Is there no API "role" for an expert? I'm using "system" for that prompt, then "user" for the text.

chat = openai.ChatCompletion.create(
model="gpt-3.5-turbo-16k",
messages=[
{"role": "system",
"content" : "Summarizing and critiquing a scientific research paper. Evaluate the potential real-world applications. Focus on the effect size over p-value. Critically analyze the research methodology. Evaluate for common pitfalls such as: Publication bias, P-hacking, Causation vs. correlation fallacy, overpowered or underpowered studies, Lack of preregistration, Selective reporting, Overreliance on observational studies, Poor sample selection (e.g., studies using mice or online surveys like Mechanical Turk), data transparency and availability. Comment on whether the researcher's conclusions are overstated. Assess the overall validity and logic of the study's concepts, methods, and derived conclusions. Keep your critique succinct, avoid clichés."},
{"role": "user", "content": text,},
]
)

[spark glade]

Not exactly as you wrote it, I will test your phrase out today, thanks!

But I tried "never say more research is needed or any of its variants', and that stops it a few times. I'm guessing the problem is that if you read 10 million research papers, 9.99 million will have "more research is needed" in the ending paragraph. This is almost a bad joke in research at this point, as it's practically always true, and as such, useless padding. But since it's such a common text pattern, I'm guessing it's deeply integrated into the model. I would think that being able to override this phrase would be a good indication of how steerable GPT is, as you are fighting against millions of data points from the training data. I think it would be a good benchmark for how well the inference aligns with the prompt.