Multimodal large language models (MLLMs) can improve their video spatial reasoning by spending extra compute at test time, using techniques such as planning, tool use, iterative perception, and reflection. However, a new arXiv paper argues that applying such computation indiscriminately is wasteful: it may redo steps that have already been resolved and can inject noise into the process.

The paper introduces ReTAC, or Residual Test-Time Agentic Computation, as a way to address this problem. Rather than repeating the full agentic loop, ReTAC focuses computation on the residual—the parts of the reasoning task that remain unsolved after earlier steps. This is meant to preserve the benefits of test-time computation while reducing redundancy and improving reliability.

The authors frame this as a correction to the common approach of simply scaling up test-time compute. The abstract does not provide experimental details, so the reported gains are not yet available from the source. Still, the core idea is clear: make agentic reasoning at test time more selective, not just more extensive.