Why is zeroing attention weights after softmax the wrong way to do causal masking?

Sreedath Panat calls it naive masking: take softmax, set the upper triangle to 0 and divide each row by its new sum. The rows sum to 1 and look identical in form to a correct mask, but softmax divides by the sum of e raised to every score in the row, and those scores include the dot products of the current query with future keys. Information from tokens that would not exist at inference is already baked into every remaining weight.

• Panat says the naive approach may work to some extent, but it still lets attention scores look into the future.
• In the notebook, the naive and correct masks cannot be told apart just by looking at the resulting numbers.

▸In their words

“So the moment you calculate attention weights from raw attention scores using softmax, some information from the future is already baked in.” Sreedath Panat · at 20:26 — open in the transcript →

Source: our own transcription of the episode. The full conversation is on The attention family (2026-09-24). Read the whole episode summary →

▸More from this episode