Building Neuro-Symbolic Reasoning Graphs from Documents, Seeking Ways to Train LLMs

Rough_Practice7631 · reddit · 2026-09-07

A Reddit developer shares an in-progress project that extracts facts and logical rules from documents and links them into reasoning graphs ending in final conclusions — a data structure meant to capture the core of human reasoning while stripping away redundant language.

The author's open question: could these graphs be used to train LLMs? One idea is linearizing the graphs into chain-of-thought samples, though that risks losing the richness encoded in the graph structure. The post solicits community input on training approaches.

Original post →

More from Research

Research channel →