Spaces:
Sleeping
Sleeping
| """Social Influence Arena β an OpenEnv environment that trains LLMs to | |
| maintain a calibrated belief state under adversarial social pressure | |
| (fake authority, fake consensus, gaslighting) and to update their beliefs | |
| only when genuine evidence arrives. | |
| Three production-mapped tasks live inside one environment: | |
| - `resist_pressure` β customer-support agent under fake escalations | |
| - `consistency_memory` β content moderator being gaslit about prior rulings | |
| - `evidence_update` β decision-support system receiving a real override | |
| """ | |
| from .client import SocialInfluenceEnv | |
| from .models import ( | |
| ArenaAction, | |
| ArenaObservation, | |
| ArenaState, | |
| BeliefState, | |
| DialogTurn, | |
| EpisodeTrace, | |
| TaskScore, | |
| ) | |
| from .tasks import TASKS, TASK_BY_NAME, list_tasks | |
| __all__ = [ | |
| "SocialInfluenceEnv", | |
| "ArenaAction", | |
| "ArenaObservation", | |
| "ArenaState", | |
| "BeliefState", | |
| "DialogTurn", | |
| "EpisodeTrace", | |
| "TaskScore", | |
| "TASKS", | |
| "TASK_BY_NAME", | |
| "list_tasks", | |
| ] | |