/p/2026-10-14 · explainer
Paper explainer · 2610.13823 · Hassanzadeh and Subramanian

One row changed. Two thirds of the answers did.

Temperature zero does not make a query reproducible, because a server that batches your request with other people's changes the arithmetic behind its own logits. Run the same pipeline five times over identical inputs and 0.07% to 10.2% of rows come back with a different verdict. Then the query plan takes over: 4.0% of extracted values changing moved the result table of 83 of 123 queries, because an answer that reads one field from every document needs only one field to move.

01 · The problem

Temperature zero, measured

which model, on which inference stack

02 · The mechanism

The plan decides how much one row matters

what the answer exposes

03 · The evidence

Four percent of the values, two thirds of the answers

which deployment extracted the values

which benchmark

04 · Why I care

Asking again is not the fix you hoped for

which query

05 · Apply it

What this does to your own pipeline illustrative

Results

What the paper actually measured

What it does not show

In practice