A new AI framework called TabSieve makes a model show its work before it guesses a missing value in a spreadsheet.
Most tabular prediction tools either look at one row in isolation or lean on an LLM prompted with a chunk of the table, and both approaches have problems: the first ignores useful patterns sitting elsewhere in the data, and the second gets confused by noisy or irrelevant rows. TabSieve splits the job into two explicit steps instead. Given a table and a query row, it first picks a small set of informative rows as evidence, then predicts the missing value using only that evidence. The researchers trained it on a synthetic dataset of 40,000 reasoning examples built from 331 real tables, plus a reinforcement learning recipe called TAB-GRPO that rewards good evidence selection and correct predictions separately. On a held-out set of 75 classification and 52 regression tables, it beat the next-best baseline by 2.92% and 4.45% respectively.
The real news here isn't the accuracy bump, it's the auditability. Anyone who has tried to explain a model's spreadsheet prediction to a skeptical analyst knows the pain of a black box. Making the evidence-selection step explicit means you can actually check which rows drove a prediction, which matters more for adoption than a couple of percentage points on a benchmark.
That said, this is one paper's benchmark numbers on curated tables, not a production system, so treat the gains as a promising lead rather than a settled result.