THEATRON
A benchmark for evaluating the persistence of personality in LLMs during long-horizon decision making. It uses financial replay as a controlled experimental environment and, through behavioral-finance personality injection and multiple metrics (intra-personality stability, cross-personality separability, and action collapse), quantifies whether personality continues to influence behavior across a sequence of decisions.
Timeline (last 90 days)
THEATRON is a benchmark for evaluating persona persistence in LLM long-horizon decision-making.
THEATRON uses financial replay as a controlled experimental environment, in which models are assigned fixed behavioral-finance personas, receive account-state feedback relevant to their own trajectories under shared daily market evidence, and make continuous decisions over trajectories spanning several months.
The evaluation metrics focused on cumulative returns and action trajectories, and on this basis diagnosed three dimensions: intra-persona stability, cross-persona separability and action collapse.
These behavioral shifts cannot reliably form coherent persona-specific strategies.
Apparently stable discriminability may stem from stereotyped action collapse, while apparently diverse strategies may be accompanied by noisy trajectory switching.
The research team released the code and benchmark data on GitHub (yuyan12138/theatron).
All Facts (6)
THEATRON is a benchmark for evaluating persona persistence in LLM long-horizon decision-making.
50%UnverifiedTHEATRON uses financial replay as a controlled experimental environment, in which models are assigned fixed behavioral-finance personas, receive account-state feedback relevant to their own trajectories under shared daily market evidence, and make continuous decisions over trajectories spanning several months.
50%UnverifiedThe evaluation metrics focused on cumulative returns and action trajectories, and on this basis diagnosed three dimensions: intra-persona stability, cross-persona separability and action collapse.
50%UnverifiedThese behavioral shifts cannot reliably form coherent persona-specific strategies.
50%UnverifiedApparently stable discriminability may stem from stereotyped action collapse, while apparently diverse strategies may be accompanied by noisy trajectory switching.
50%UnverifiedThe research team released the code and benchmark data on GitHub (yuyan12138/theatron).
50%