Scaling RL to the full Pretraining Corpus using Factuality Grounding

People seem to think we need to have models that are doing reasoning at the pretraining level--models should be more explicitly integrating global information to make next token predictions supported by some actual reason (https://substack.com/home/post/p-168023871, https://time-for-consistency.github.io/assets/pdfs/Consistency_Pos_Paper.pdf). If you believe current frontier LLMs are good at inverse entailment, then you run inverse entailment over the full pretraining corpus at some granularity (sentences? paragraphs?), create tool calls/a DB for all the premises, and inject the tool calls into the pretraining corpus before each sentence/paragraph. Knowledge gets externalized, and now you can do things like reasoning between that tool call and NTP. (idea in part inspired by https://arxiv.org/abs/2505.15962)