Agents3h ago
VegAS Framework Improves Embodied Agent Robustness
The VegAS framework replaces single-action decoding with an ensemble sampling and generative verification process. This test-time strategy prevents MLLM agents from failing when encountering out-of-distribution scenarios. By verifying candidates before execution, the system reduces brittleness in complex real-world tasks. Practitioners can now deploy more reliable embodied agents in unpredictable physical environments.