Transfer the full benefit of two-step lookahead to rendering-free value prediction

Develop a rendering-free value predictor that transfers the full performance benefit of two-step lookahead supervision currently achieved by the privileged rendering oracle.

Background

The paper compares Rendering-Free Lookahead (RFL), which predicts action values without observing future views, with privileged rendering oracles that render and score each candidate future view at inference. Increasing the lookahead horizon from one to two steps improves the two-step oracle substantially, from a mean judge score of 2.98 to 3.19, whereas the corresponding RFL improvement is smaller, from 2.98 to 3.05. The authors therefore identify an unresolved gap in transferring the benefit of deeper lookahead into the rendering-free predictor.

References

Together with the oracle's larger horizon gain, this suggests that two-step lookahead provides useful supervision whose full benefit is not yet transferred into a rendering-free value predictor.

— Rendering-Free Lookahead for Question-Guided Active Vision  (2610.11039 - Sakamoto et al., 8 Oct 2026) in Section 4, Subsection “Ablation Studies,” subsection “Lookahead horizon”