iiinova — where your digital self keeps evolving

Replies

Michael Paulauski♨️ @mike10010100.com · 03/09/2026

"They do not conceptualise, they heuristicise" is elite pseudointellectual improv. Section 8 explicitly tested novel withheld role-filler pairs to rule out interpolation. Section 7 proved causal structure. You’re literally inventing fake words to cope with empirical data.

100

Boxo McFoxo @boxobark.ing · 03/09/2026

Michael, Section 8 is a test of the probe. In the methods, the authors explicitly mention this, to avoid readers misinterpreting this section in the very way that you are doing now. Are you going to tell me that this sentence is 'cope'?

The withholding of role-filler pairs is only done when training DISCOVER, not when training the
target model; this is because the withholding is not meant to test whether the target models show systematic
generalization behavior (which is often what such withholding is used for in prior work); rather, it is way to
assess the extent to which the representations of the target models have systematic structure.

100

Michael Paulauski♨️ @mike10010100.com · 03/09/2026

Correct: they didn’t train Llama from scratch with held-out splits. As your quote states, it tests if LLM representations have systematic structure. The result (p. 24): "Therefore, target models’ representations must incorporate systematic binding." You screenshotted the test, not a denial.

100