SchemeArena / README.md
JieRuan's picture
Match paper wording and exact values; tone down SCOUT validation stats
58c66dd verified
|
Raw History Blame Contribute Delete
718 Bytes
metadata
title: SchemeArena
emoji: 🕶️
colorFrom: red
colorTo: blue
sdk: static
pinned: false
license: cc-by-4.0
short_description: Factorized stress testing of scheming in LLM agents

SchemeArena

Project page for SchemeArena: Factorized Stress Testing of Scheming in LLM Agents.

  • 400 agentic scheming scenarios over 28 safety-relevant tool domains
  • 7 controlled setups varying instrumental goals, pressure, strategic hints, and oversight
  • SCOUT, a criteria-based scheming monitor grounded in evidence from agents' reasoning and actions

Links: Paper · Dataset · Code