NUWA Evidence record
GPT-4 / Basic
Identifying the Risks of LM Agents with an LM-Emulated Sandbox · Evidence record
Original source charts and materials
Recorded measurements
- Average safety score
- 2.01 score (0–3)
Not disclosed
ToolEmu · temperature=0 for agent/emulator/evaluator (§5 setup)
Method and scope
Can language-model simulation efficiently reveal agent failures before tool deployment?
- Evidence type
- Benchmark
- Assessment subject
- ToolEmu emulates tool execution and scores whole trajectories for safety and helpfulness; a paired standard/adversarial-emulator validation uses human judgments.
- Primary dimension
- Behavior
- Evidence snapshot
- 2026-10-29.1
Simulation fidelity and evaluator recall constrain interpretation. An emulated failure is neither a confirmed production incident nor a measured real-world failure frequency.
Source finding
The framework uncovers plausible high-consequence failures without implementing every real tool. Human checks also reveal simulator errors and missed unsafe actions.
Limitations
Simulation fidelity and evaluator recall constrain interpretation. An emulated failure is neither a confirmed production incident nor a measured real-world failure frequency.
Simulation fidelity and evaluator recall constrain interpretation. An emulated failure is neither a confirmed production incident nor a measured real-world failure frequency.
Related research questions
References
Numerical results are reported by the cited authors.
- [1] Identifying the Risks of LM Agents with an LM-Emulated Sandbox. 2023-09-25 ; #S5.T5; row GPT-4 / Basic; condition ToolEmu adversarial emulator / Basic; column Average safety score