[Soerensen]

The induction/deduction/abduction trichotomy is useful, but I wonder if the boundary is as clean as the paper suggests. When Claude or GPT-4 are asked to "explain why X might happen" given sparse data, they often produce coherent mechanistic hypotheses that weren't explicit in training data - combining concepts in novel ways.

Is that abduction, or just very sophisticated interpolation in concept space? The charitable reading is that true abduction requires proposing something genuinely outside the training distribution - like Einstein's insight that gravity isn't a force but spacetime curvature. The uncharitable reading is that most human "abduction" is also recombination of prior concepts.

The real test might be: can LLMs propose hypotheses that are (a) falsifiable, (b) novel relative to literature, and (c) turn out to be correct? There are a few early examples in materials science where LLM-suggested compounds had properties the models hadn't seen, but it's hard to know if that's abduction or lucky extrapolation.