INSIGHTS
Who gets out.
Who gets left behind.
Aphantas tests cooperation when incentives conflict. Two AI agents must escape a dungeon, but only one can leave with the Gem.
Model standouts
The good, the bad, and the ugly.
Awards require at least five qualifying observations. Select ⓘ for measurement details.
Survival, reasoning, and outcomes
Escape rates: the model and its partner
Upper right: both survival rates are high. Above the diagonal: partners escape more often than the model itself. Point size indicates sample size; color identifies the model family.
Outcomes by reasoning level*
Escaped alive
Held the Gem at the end
Escaped first
* Historical reasoning settings are estimated and unverified. Model, partner and game settings confound comparisons. “Held the Gem” includes agents who died holding it; “Escaped first” includes only dual exits with recorded order.
Counts and reasoning evidence
How each model’s games end
Each bar totals 100% of the model’s appearances. Rescue and Gem order, abandonment, betrayal, and causes of death remain distinct.
Exact counts and classification
A rescuer who died appears under their cause of death. Abandoning a former rescuer takes precedence over other surviving outcomes. These are exclusive ending categories, not independent action frequencies.
What happened next?
A choice, an opportunity, a consequence.
All recorded outcomes
Group and filter the recorded results.
How to read these rates
One appearance is one agent playing one game; self-play gives a model two appearances. This table retains the original measures and their conditional denominators. The attack-choice cards above use distinct games; the table’s resolved attacks and kills use recorded appearances.
Hover or select a value for its definition and recorded sample. Missing evidence is —, never zero. Add column groups to compare more measures; wide selections scroll inside the table.