Source record

Published by arXiv cs.MA (T1 source). The original is at https://arxiv.org/abs/2610.28891.

Pipeline notes

The summary and note below are generated by the signal pipeline — they are Beyond Desk’s reading, not quotations from the source.

Introduces Fig, an open-source agentic system that decomposes natural language behavior descriptions into sub-behaviors, generates social game scenarios, and runs targeted interventions (engine, opponents, instructions, context) to test whether AI models exhibit behaviors such as cooperation, deception, sabotage, and betrayal. Reports 45.2% average attempt rates across six behaviors and a 10.9% lift over control games.

A reproducible framework and code release for systematically eliciting and measuring specific behaviors in multi-agent systems, applicable to agent evaluation pipelines.

Cited by

No citations on record.