Copy Initial State Baseline
A control baseline strategy for evaluating a model's prediction capability in which the initial state is output directly as the prediction. It is used to determine whether a model genuinely performs state inference rather than simply copying.
Timeline (last 90 days)
On long-history tasks, some models can capture a considerable number of correct changes, clearly exceeding the baseline of directly copying the initial state.
The copy-initial-state baseline can achieve a not-low score on cell-level accuracy when, after multi-step operations, the grid often changes only locally and its overall layout remains similar to the initial state.
A model's performance must clearly surpass the copy-initial-state baseline to prove that it is genuinely doing state inference rather than static copying.
All Facts (3)
On long-history tasks, some models can capture a considerable number of correct changes, clearly exceeding the baseline of directly copying the initial state.
50%UnverifiedThe copy-initial-state baseline can achieve a not-low score on cell-level accuracy when, after multi-step operations, the grid often changes only locally and its overall layout remains similar to the initial state.
50%UnverifiedA model's performance must clearly surpass the copy-initial-state baseline to prove that it is genuinely doing state inference rather than static copying.
50%