- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
SpatialMind treats pointing with a frozen VLM as candidate selection and observation acquisition. A registry of dependent zoom procedures produces candidate points; a learned graph model combines a boosted-tree prior, prefill features, and typed messages to score their correctness. Adaptive SpatialMind adds a cost-conditioned acquisition and stopping policy. Experiments span three backbones and five benchmarks, with image-grouped cross-fitting and held-out-benchmark evaluation. The paper reports accuracy and call-efficiency gains and motivates the method through Bayesian scoring and policy-class arguments. The main unresolved issue is whether graph correction itself explains the improvement over a matched selector.
Critical strengths
-
A useful formulation of dependent visual observations. The registry deliberately generates different evidence through self-centered and anchor-centered zoom chains. Separating candidate coverage from selection clarifies why obtaining more correct candidates need not improve the returned answer. Reusing the answer model over partial histories gives a practical connection between selection and adaptive acquisition, rather than treating the two as unrelated procedures.
-
Credible system-level gains under a careful evaluation protocol. Image-grouped cross-fitting, nested training of learned components, held-out-benchmark evaluation with overlap removal, clustered resampling, and multiplicity adjustment are substantial strengths. Table 3 reports 70.71% accuracy for SpatialMind versus 65.74% for the plain call and 69.23% for registry voting; the 1.48-point margin over voting has a 95% interval of [0.88, 2.08]. Adaptive SpatialMind reaches 69.05% with 2.93 backbone calls versus 7.91 for full acquisition. These results support the utility of the overall system under the stated call-count measure.
Critical weaknesses
-
The learned baseline does not isolate the observation-graph contribution. The internal prior in Appendix D.1 uses 36 features, including procedure success rates and retrieval-conditioned rates, and is trained over partial histories with its own fitting choices. The boosted-tree baseline in Appendix E instead uses 22 features, default classifier settings, and full histories. SpatialMind additionally receives prefill states. Therefore, the 70.71% versus 69.08% comparison changes the prior, information available, and relational model together. The zero-correction internal prior is an especially direct missing control, since Proposition 1 explicitly states that it is contained in the model class. A same-feature scorer without dependency relations would then distinguish the benefit of those relations from the richer inputs. This attribution matters because the observation dependency graph is the central methodological claim.
-
The Bayesian interpretation exceeds what is established for the actual objective. Proposition 1 characterizes the population optimum of binary cross-entropy, while Equation 12 combines BCE with a normalized pairwise ranking loss. Appendix B proves a useful two-node ordering property, but ranking consistency is not calibrated posterior estimation: the pairwise optimum depends on joint label probabilities. The combined objective therefore does not inherit the stated posterior interpretation without further conditions. Likewise, exact messages in an idealized Markov chain and optimality over all history-dependent policies do not establish exact inference or optimality for the implemented network and approximate policy iteration. These results remain useful motivation, but the abstract's claim that the method learns Bayes' rule should be qualified accordingly.
Questions for authors
-
What accuracy does the exact internal boosted prior achieve with the graph correction disabled, under identical folds, retrieval features, partial-history training, and fitting choices? With prefill inputs also held fixed, how does a scorer without parent/child relations perform? These controls would most directly affect my recommendation.
-
Which result applies to posterior calibration under the complete Equation 12 objective, rather than to its BCE component or two-node ordering alone? Please either establish the needed conditions or narrow the Bayesian claim. A BCE-only comparison and held-out calibration measurements would be informative if posterior estimation remains a central claim.
-
At comparable mean call cost, how does the acquisition policy compare with a training-selected fixed schedule using the same answer model? The paper already computes fixed-schedule values during policy training; reporting this control would clarify how much benefit comes from adapting acquisition to each query.