David v Godzilla: TabFM

A 12 GB tabular foundation model, met under a pre-registered protocol. Eleven arms against the reference. The record, in full.

The tale of the tape

TabFM (as configured)laplace
model artifact12 GB weights988 KB source
dependencies560 MB (torch)none
peak memory~10 GB18 KB state
time per forecast45 ms–0.5 s (GPU)0.3 ms (one CPU core)

Round one: eleven arms, 226 stratified series

Pre-registered: universe, arms, and scoring frozen and committed before any result existed. Eleven ways of using TabFM (lag depths, four binning schemes, horizon variants, a residual-density regressor), so the verdict bounds the adaptation rather than one adapter. laplace won the per-series likelihood race against every arm, 71–88% of series, the strongest configuration trailing by 0.39 nats median. The strata told the story: near parity on strongly mean-reverting continuous series (a tabular regression problem, TabFM's home game), collapse on repeat-heavy series (2–4 nats; no concept of an atom), steady deficit on near-random-walk series where volatility tracking is the job.

The verdict

On density forecasting across the universe, laplace by a wide margin, at one ten-thousandth the footprint. On point forecasts, TabFM is genuinely decent (it takes the MAE undercard on continuous series by hairline shading). What TabFM contributes in cooperation with laplace is a different subject with its own numbers: see the sandwich.

All numbers from committed studies: benchmarks/tabfm_wide_study.py, results CSVs beside them, statements in benchmarks/preregistrations/.