acceptodds
Withdrawn from ICLR 2027

Sufficient to Act: What a Representation Must Reveal When Its Consumer Optimises

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Representation learning usually defines sufficiency relative to prediction, but many systems consume representations by solving an optimisation problem and returning an action. We ask what information a representation must preserve when its consumer optimises. For a fixed downstream objective, query sufficiency and action sufficiency can separate sharply: there are input families for which answering all relevant queries requires bits while recovering the optimal action requires only . For constraint-generated minimum-cost hitting-set decisions, we then characterise universal-cost action sufficiency exactly: two inputs induce the same complete family of optimal actions under all strictly positive costs if and only if their hypergraphs have the same antichain of inclusion-minimal generators. Its union gives the exact action support, identifying candidates that can be removed with zero decision loss. We also show that pseudonymisation does not remove structure-determined nuisances and give exact conditions for quotienting the remaining. support while preserving the complete Argmin set. In a real-code study motivated by post-quantum cryptography (PQC) migration, the default model yields exact hypergraphs for 30 repositories; the action support retains a median 4.7% of graph vertices and the admissible quotient 3.5%, while pseudonymised full topology remains re-identifiable at about chance. On controlled synthetic instances, a graph-aware selector recovers the exact support on 37.0% of matched-budget cases but preserves the exact Argmin on 89.8% of sampled instance-cost pairs, showing that empirical objective preservation is weaker than universal action sufficiency.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
decision-sufficient representations, action sufficiency, representation learning, graph compression, combinatorial optimization, privacy, post-quantum cryptography

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_b9gMOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper asks what a released representation must preserve when its consumer is an optimiser rather than a predictor. In the proposed "two-room" setting, a private party holds a typed graph G*. The graph generates intervention constraints, a hypergraph H(G*). The private party must release a representation Z before an external party supplies per-node costs c and returns the complete set of minimum-cost hitting sets A(H, c).

Critical strengths
  1. A useful constraint-side framing. Prior decision-sufficiency work (Bennouna et al., 2025; Ye et al., 2026) treats the feasible set as known and the cost as uncertain. Here the private input generates the feasible set and the objective arrives after release. The three-layer decomposition G* -> H(G*) -> A(H, c) cleanly separates domain-specific constraint extraction from the decision-level statement, and so lets one result apply to several remediation domains (App. A.2).

  2. Transparent and checkable reporting. The paper provides per-repository statistics (Table 9) and the list of excluded repositories (App. D.2). It explicitly retracts an earlier coarsest-release claim (App. C.4) and clearly separates sampled-cost preservation from universal sufficiency: Table 3 reports 37.0% exact support against 89.8% sampled-cost Argmin, a useful caution for decision-focused representation learning.

Critical weaknesses
  1. The main theorem is classical, and its "must reveal" conclusion depends on an unstated extreme assumption.
  • Thm 2 follows directly from clutter/blocker duality. Under strictly positive costs, every optimum is a minimal transversal. Every minimal transversal is the unique optimum of an epsilon-cost vector, so c -> A(H, c) recovers Tr(H), and b(b(C)) = C (Edmonds & Fulkerson, 1970; Cornuéjols, 2001) then returns min H. Part (iii) is V(b(C)) = V(C), equivalently "the relevant variables of a monotone Boolean function are those in its prime implicates". This literature is not cited, while the necessity direction is claimed as the contribution (App. A.3, l. 813–815).
  • The path-generated minimum-cost blocking problem is itself established in attack-graph hardening (Sheyner et al., 2002; Noel et al., 2003).
  1. The two-room model is not internally consistent, and the experiments do not test the recommended release.
Questions for authors
  1. Thm 2(ii) follows from epsilon-exposure of minimal transversals plus b(b(C)) = C, and (iii) follows from V(b(C)) = V(C). Which statement in the paper is not covered by this derivation? Can you characterise the minimal sufficient statistic for bounded-ratio or category-constant cost classes? (For any class C, the set of minimal transversals that are optimal for some c in C is a natural candidate.)
  2. Does Z_min carry real identifiers or pseudonyms? How does the public room assign per-node costs to pseudonymous nodes? If costs are a function of Z, is the universal-cost premise still needed, and why is the hitting set not solved locally?
Reviewer_eHdAOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies what information a released representation must preserve so that a downstream optimizer can recover minimum-cost intervention plans after receiving a cost vector. It models these decisions as minimum-cost hitting sets and characterizes the constraints and candidate intervention sites needed to preserve all optimal plans for every strictly positive cost vector. The empirical evaluation uses a PQC-motivated study of real codebases and controlled synthetic experiments.

Critical strengths

The paper develops a coherent formulation and precise guarantees for preserving downstream decisions. The distinction between answering queries and retaining information needed for optimization is meaningful, and the characterization gives a systematic account of which constraints and candidates can be removed. The real-code and synthetic studies represent substantial effort, and the paper explicitly acknowledges the scope of its application model and privacy claims.

Critical weaknesses

My main concern is the depth of the contribution. The central combinatorial arguments—removing inclusion-redundant constraints and distinguishing different feasible families through appropriate costs—are elementary. The definitions, terminology, and flowchart introduce considerable complexity around these observations, while the additional insight remains limited. The release-before-costs setting explains the universal quantifier, but the practical need to preserve the complete set of optimal plans for every strictly positive cost vector remains insufficiently justified.

The connection to PQC migration is also not sufficiently convincing. The static import-graph and atomic path-blocking model leaves unclear how the selected interventions correspond to actual migration decisions involving coordinated edits and compatibility requirements. Consequently, the real-code and synthetic experiments provide limited evidence of practical value. Although the paper explicitly limits its privacy claims and measures re-identification, the early security and privacy framing is stronger than the demonstrated application evidence supports.

Overall, this is a serious piece of work, but the theoretical contribution is thin, and the application evidence does not sufficiently compensate for it. My recommendation is 1 — Reject.

Questions for authors

What is the main conceptual or technical insight of the action-equivalence characterization beyond inclusion-based redundancy removal and cost-based separation? In the motivating application, why is preserving the complete Argmin set for every strictly positive cost vector necessary?

Could the authors give a concrete PQC migration example in which actionable nodes and path-blocking constraints correspond to valid interventions, and explain when this abstraction remains useful despite coordinated changes and compatibility requirements? What practical decision benefit is established by the reported representation reductions?

Reviewer_YegfOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

Summary: I believe responsible reviewers cannot offer a valid summary to this work. Because objectively, the current submission is not ready for the formal venues. I can only guess, maybe the work is about: how to reveal less private information but still maintain a certain "action" (in the sense of optimal action maybe).

Critical strengths

Strengths: I have substantial questions about this paper. Given its current status, with the best of my ability, I withhold comments to this work's strength.

Critical weaknesses

Weakness: I can only list a few points with my best effort during reading.

  • What is the research gist here ? For me, the theory, the (maybe) privacy, the (maybe) cryptography, what is the connection ?
  • Can the authors explain what's the logical connection between the main theorem and the experiment ?
  • The paper has poor writing quality that it is impossible for readers to follow. The main cause is the broken logic. Too many problems to list, for example, some notations / symbols were used without clear definitions, like ln.220.
  • After reading, I cannot even understand the connection between the entire experiment and the paper's core research spine.
  • Also, why the hitting set ? I understand how it MAYBE work in the experiment, and also its appearance in the early description. but why you need it ? Also, is theorem 2 a new result or rephrasing existing results? Why does a theoretical proof need any "computational check" at around ln.1013 -- 1014 ?
Questions for authors

I already included my questions in the Weakness section.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.