flâneur

Directed graphs from papers with OpenAI o3 | Notes

kosti.bearblog.dev · 596 words · saved by 1 readers

I've been having fun asking O3 to read papers, extract hypotheses and results, and then construct "causal"1 diagrams, i.e., directed acyclic graphs that contain the relationships of different variables. Strong associations are connected with black lines, weak or negative associations with red lines. Here's an example from "World Models and Consistent Mistakes in LLMs". Kind of useful, although one would have to read the paper to understand what is going on. Asked o4-mini to write a mini-paper on Newton's gravitational law, specifically the dependence on 1 / 𝑟 2 , using fictitious experiments. It produced such a mini-paper with the table Then asked in a separate chat for a causal diagram based on the mini-paper: What happens if we only give to o3 the table without any explanation? Without a lot to go on, o3 focusses on the functional relationships, but the graph looks correct given the table. With ChatGPT agent, we can even scale this across papers, getting perhaps some nice summaries

30 Jul, 2025 I've been having fun asking o3 to read papers, extract hypotheses and results, and then construct "causal"1 diagrams, i.e., directed acyclic graphs that contain the relationships of different variables. Strong associations are connected with black lines, weak or negative associations with red lines. Here's an example from "World Models and Consistent Mistakes in LLMs". Prompt: Read this paper and then do a causal diagram that captures the hypotheses tested in it. Add nodes for hypotheses and results. You may use dark lines for hypotheses that have sufficient evidence and…

saved by

related reading