- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper asks how process knowledge can be reused across equivalent terminal outcomes when execution depends on order and the current state. Its carrier framework uses terminal correspondence to map independently obtained process relations into the target task, then updates proposal rankings as actions are executed. SymBuild combines the process and state ranks using and submits candidates to a fixed verifier in that order. The paper gives a non-identifiability result and a conditional completion guarantee, and evaluates the approach on CAD disassembly, synthetic Mini-Programs, exact-fill packing, and GRN target removal, with an additional next-action prediction study on IKEA assembly videos.
Critical strengths
-
The distinction between terminal correspondence and process knowledge is useful. Knowing which entities correspond does not tell us which action orders are valid; Theorem 1 makes that ambiguity explicit. The framework gives these sources of information separate roles and uses the transferred relations to guide proposals, with acceptance left to the verifier. This is a sensible starting point for reusing experience when terminal equivalence provides an incomplete description of execution.
-
The matched comparisons provide credible evidence that refreshing proposals helps. The length-OOD Mini-Program AUC rises from 42.92 to 64.67, a gain of 21.75 percentage points with a paired 95% CI of [18.33, 24.92] (Table 1). There are also gains in CAD and packing, and refresh remains useful with other aggregators and planning interfaces. Sharing the first action helps locate the effect in subsequent decisions. The separate development data, paired uncertainty estimates, and analysis of scoring overhead are also welcome. These choices make the empirical study a substantial part of the paper's value.
Critical weaknesses
-
My main reservation is the size of the algorithmic and theoretical advance. As the authors acknowledge, rank meet is the highest-rank rule described by Ho et al. [1]. That work also gives the corresponding rank bound for two strict input rankings. The union identity follows directly from taking the minimum. Theorem 2 adds a per-state worst-case characterization, but any ordering that queries each action in before leaving that set achieves the same bound of . Completion further requires accepted actions to preserve a complete suffix. The argument appears sound under these assumptions, but the guarantee is limited and does not establish optimal query cost over a full execution.
The paper's own comparisons also give little reason to prefer rank meet in several settings. On size-OOD packing, the product rule reaches 47.92 AUC with refresh and 41.39 without it, compared with 34.03 for SymBuild (Table 14(b)). On GRN OOD20, Combined refresh has a mean query score of 3.873 versus 7.608 for rank meet (Table 5). I agree that the framework can accommodate different aggregators. These results nevertheless put more weight on what the framework itself adds beyond familiar state-dependent heuristic updates, since rank meet has neither a strong novelty claim nor a consistent empirical advantage.
-
I found it difficult to separate the contributions of terminal transport and the different state updates in the experiments. In CAD, Static caches the full initial fused order, while SymBuild recomputes blocker scores and both ranks. In Mini-Programs, both ranks are refreshed by SymBuild and frozen by Static. In packing, both variants update the process rank, and only the residual-rank update differs. Updating ranks induced by a fixed transported structure is reasonable, but these comparisons do not all isolate the same operation. In particular, they leave open how much of the CAD and Mini-Program gain comes from residual refresh once the process side is updated in both policies.
The Mini-Program state-only control already uses the transported observed graph and slightly exceeds SymBuild on both splits (Table 2(a)); it therefore cannot serve as a baseline without transport. The mismatch intervention changes only the learned precedence matrix while keeping the observed expert aligned. The CAD alignment experiment is useful positive evidence, but GRN uses identity correspondence and mainly tests refresh. I would like the paper to make clearer which results establish the benefit of terminal transport when the corresponding transformation fails to preserve execution dynamics. A directed process can still be equivariant, so the presence of precedence constraints alone does not settle this point.
I lean toward weak rejection. The refresh effect is supported by a substantial set of experiments, but I am not yet persuaded that the paper establishes enough beyond known rank fusion and updating heuristics after state changes. A clearer answer to the first two questions below could change my assessment.
References
[1] T. K. Ho, J. J. Hull, and S. N. Srihari. Decision Combination in Multiple Classifier Systems. IEEE Transactions on Pattern Analysis and Machine Intelligence, 1994.
Questions for authors
-
How much of the CAD and Mini-Program improvement remains when process scores and ranks are updated in both policies, with only the residual information held fixed in Static? Please distinguish re-ranking cached scores on the remaining candidates from recomputing the underlying scores. This would help identify which update is responsible for the gains.
-
Could you walk through one evaluated example where a nonidentity terminal correspondence does not extend to a symmetry of the transition system, explaining which process knowledge is transferred and why it remains useful? For CAD, how does this relate to the yaw-lift and verifier direction mapping in Appendix D.2? This clarification would help me assess the paper's central setting.
-
In packing, each verifier call uses branch-and-bound to check that a complete solution still exists. How does total solver work or runtime, including proposal scoring, compare with directly solving the same anchored instance? The GRN timing analysis is helpful, but the practical advantage of reducing calls to the exact packing solver remains unclear.