Putting Task Expertise Into Rl

A query can return the right answer for the wrong reason, and training that rewards only the answer reinforces the flaw. In a research write-up, the correct query averages each order's total price, while a flawed one averages item prices. They match only because each order in that database has one item; with multi-item orders the flawed query would be wrong.

The researchers added VeriEQL, a bounded-verification solver that tests whether two queries are equivalent beyond one database, and reduced rewards when it refuted a match. In a pilot run, 32.8% of positive rewards went to non-equivalent queries. With cleaned data and a second reward rule, single-sample accuracy was 91.37% on their corrected Arcwise-Plat-SQL test set. That measures the whole recipe, and bounded checking is not a guarantee.