Agents can complete tasks after hardware faults, but cost increases
A small number of faults cause errors that are hard to see while the LLM agent continues to work correctly.
Claimed, not confirmed
Researchers examined how LLM agents work with random hardware faults. They say that it is the first test of this type. They used 3 agent benchmarks. The 6,814 agent runs include more than 2,400 clean runs and more than 4,300 runs with random two-bit weight faults. The researchers say that agent loops change how the faults show in a system.
Sources
Posted
