AIAI agents and the trade-offs of optimising 3D reconstructions
The author of “Agentic Optimisation” says an agent in their AgentSTAR experiments could sometimes accept a lower score for a more plausible 3D reconstruction.
TLDR
An account of AgentSTAR experiments describes an AI agent trying to match a reconstructed object and its motion to target silhouettes. Flat 2D slices could satisfy that scoring objective without producing a plausible 3D object, but the author says the agent could sometimes recognise an implausible result and revise it—even at the expense of its score. The author calls this “agentic optimisation”: the model stays fixed while revising the solution, rather than updating its own weights as in reinforcement learning. They argue that the model’s prior knowledge may help it work with imperfect objectives, but can also prevent it from reaching novel solutions it considers implausible.
AI agents and the trade-offs of optimising 3D reconstructions
The author of “Agentic Optimisation” says an agent in their AgentSTAR experiments could sometimes accept a lower score for a more plausible 3D reconstruction.