acceptodds
Withdrawn from ICLR 2027

CutWorld: A Latent Dynamics Model for Unified Branch-and-Cut Control

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Branch-and-Cut solvers repeatedly choose a branching variable, a cutting plane, and an open search node. These choices reshape an exponentially large search tree, but standard learned branching rules score candidates only from the current relaxation. We present , a latent decision-dynamics model that plans over candidate branching actions without solving their child linear programs. A bipartite GATv2 encoder maps the current relaxation to a global latent and per-variable embeddings. A causal Transformer then predicts the latent successor and its variable embeddings conditioned on a candidate and branch direction, allowing the policy to be re-applied to imagined states. We combine the resulting rollout value with the predicted dual-bound progress and cost-to-go, and use the shared representation to order nodes and select the root Gomory cut. Because the learned components only select among valid branches, globally valid cuts, and open-node orderings, they do not alter feasibility or bound tests. On fixed Set Cover benchmarks generated with one seed, the neural configurations complete all 10 medium instances (, 300 s limit) and all 5 hard instances (, 600 s limit) in the Python/HiGHS harness, whereas the classical rules reach their time limits. A depth-one rollout reduces the medium-tier shifted-geometric-mean tree-node count by relative to imitation (3,197 versus 5,733); at hard scale, combining rollout and root cuts reaches 2,603 nodes versus 4,633 for imitation. We report completion, tree-node counts, wall-clock time, LP time, and cuts, and explicitly separate completed-tree measurements from timeout lower bounds. The benchmark is intended as a controlled feasibility study; its small, single-seed evaluation does not establish statistical superiority over industrial MIP solvers.

Primary area:
optimization
Keywords:
World Model, optimization, Integer Programming

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_u5bBOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes CutWorld, a latent dynamics model for learned control of branch-and-cut (B&C) search in binary mixed-integer programs. A bipartite GATv2 encoder maps the current LP relaxation to a global latent state and per-variable embeddings. A causal Transformer predicts a successor latent and successor variable embeddings conditioned on a candidate variable and branch direction. The policy can therefore be re-applied to imagined states during a short latent rollout. The system also includes value, reward, cost-to-go, integrality, and root Gomory-cut selection heads; the learned components select admissible actions, while LP feasibility, bound, integrality, and pruning remain solver-certified.

The evaluation uses synthetic binary Set Cover instances in a custom Python/HiGHS B&C harness. The reported benchmark has 10 medium instances of size 500×1000500\times1000 with a 300-second limit and 5 hard instances of size 1000×20001000\times2000 with a 600-second limit, all generated with one seed. The paper reports that neural configurations complete all listed instances, while classical controls time out. Relative to imitation, a three-step rollout reduces the medium completed-tree shifted geometric mean from 5,733 to 3,197 nodes, and the combined neural configuration reports 2,603 nodes on the hard tier. The paper explicitly states that these results are descriptive, do not establish superiority over industrial solvers, and lack independently verified optima, multi-seed intervals, and a tuned SCIP baseline.

The central empirical quantity is the shifted geometric mean

SGMδ(y)=exp⁡(1n∑i=1nlog⁡(yi+δ))−δ\mathrm{SGM}_{\delta}(y) = \exp\left(\frac{1}{n}\sum_{i=1}^{n}\log(y_i + \delta)\right) - \delta

with shifts δN=10\delta_N=10 for nodes and δT=1\delta_T=1 for time. For timeout runs, the reported node count is only a partial-tree lower bound, whereas neural rows use completed-tree counts. The reported incumbent gap is relative to the best feasible incumbent observed in the same budget, not to an independently certified optimum.

Critical strengths
  1. A coherent model-based formulation for B&C. The transition of both the global latent and the per-variable embeddings directly addresses the need to reapply a branching policy after an imagined action. This is a technically meaningful extension of myopic learned branching.

  2. Explicit safety and evaluation caveats. The paper keeps feasibility, pruning, and bound logic outside the learned model, gives a conditional correctness argument, distinguishes timeout lower bounds from completed trees, and reports the rollout wall-time trade-off. These choices make the limitations visible and reduce the risk of interpreting a heuristic as a correctness certificate.

Critical weaknesses
  1. The experimental design does not support the main performance conclusion. The benchmark has only 10 medium and 5 hard instances, uses one generator seed, and reports no confidence intervals, paired tests, or independently verified optima. Moreover, classical rows are partial trees at timeout while neural rows are completed trees; a smaller partial tree is not comparable to a completed-tree count. The rollout-versus-imitation comparison is at least within completed neural rows, but the comparison with timed-out classical controls and the broader superiority claim remain unsupported. The relative incumbent gap is also defined against the best incumbent found within the same budget, so it is not an optimality gap. In addition, several neural rows are marked complete but still report different incumbent-gap values; under the paper's stated exactness conditions, completed runs on the same instance should return the same optimum, so the aggregation and completion semantics need an audit. This combination makes the reported 44.2% node reduction and the 2,603-node hard result descriptive rather than evidence of a reliable advantage.

  2. Baseline and configuration attribution are incomplete and internally ambiguous. The main tables do not include a tuned production-quality SCIP baseline, even though the supplementary code contains a separate SCIP harness. The supplied tools/benchmark_solvers.py configures neural_best_d3 with lookahead_depth=3, mf_blend_alpha=0.20, and size_weight=1.0, while Appendix A.5 reports that adding the size penalty increases completed-tree nodes (3,197 to 4,320 on medium and 2,603 to 4,376 on hard). The paper does not make clear which exact configuration produced the headline 2,603 result or whether the reported ablations were selected after observing the test benchmark. The archive contains source/configuration files but no checkpoint, trajectory data, or raw result JSON, which prevents independent reproduction of the tables.

Questions for authors
  1. Data separation: What are the exact instance IDs, generator seeds, and file hashes for the 800/100/100 training/validation/test split, and for the 10 medium plus 5 hard benchmark instances? Can you demonstrate that no benchmark instance, trajectory, or near-duplicate appears in training, validation, or hyperparameter selection?
  2. Generation protocol: Were the reported benchmark instances generated independently from the 1,000 training instances, or are they a subset/re-generation of the same seed stream? Please provide the exact command and generator implementation used for both datasets.
  3. Split reproducibility: The supplied train.py calls split_files(...) without passing the command-line --seed, so the file split uses the helper’s default seed. Is this intentional, and which split seed actually produced the reported checkpoint and validation metrics?
  4. Equal solver settings: Were LP tolerances, presolve, primal heuristics, cut limits, node selection, warm starts, and termination checks identical for every method? Please provide a machine-readable configuration for each table row and explain any method-specific differences.
  5. Censored comparisons: Can you report common-budget anytime curves for every method, including incumbent value, primal-dual gap, open-node count, and time? How would the conclusions change when timeout lower bounds are not compared with completed-tree counts?
  6. Optimality verification: For every claimed completion, was the returned objective independently checked by SCIP, HiGHS, or another exact/independent solver? Why do completed neural configurations show different incumbent-gap values on the same benchmark if they all satisfy the stated exactness conditions? What numerical tolerances and certificate checks were used to rule out false “optimal” statuses?
  7. Statistical power: Can you rerun the benchmark over multiple independent generator seeds and report per-instance paired differences with confidence intervals or an appropriate paired test? Please include the number of failures and not only aggregate SGM values.
  8. Out-of-distribution generalization: Does the method retain its advantage on other MILP families, densities, coefficient distributions, and size ratios, or is the result specific to fixed-density Set Cover? Please provide at least one family and size regime not used during training or model selection.
  9. Dynamics validity: What are the one-step and open-loop K=1,2,3K=1,2,3 prediction errors for the latent, per-variable embeddings, reward, and value heads? More importantly, how often does rollout rank candidates correctly when compared with actually solving their child LPs?
  10. Causal rollout ablation: With the same checkpoint, candidate budget, node ordering, cut setting, and MF blend, what is the isolated effect of rollout depth? Please reconcile the code’s neural_best_d3 setting (size_weight=1.0) with Appendix A.5, where the size penalty appears harmful.
  11. Cost-to-go targets: How are subtree-size and cost-to-go labels constructed for non-DFS SCIP visitation? Are all future nodes used to form a target that would be unavailable at inference, and how do you prevent this from becoming a target-construction or trajectory-order leakage issue?
  12. Branch directions: Are both branch directions explicitly represented in every training trajectory and in every rollout? Please report direction-label coverage, the handling of the root direction, and whether any supplied trajectories use the code’s zero-direction fallback.
  13. Cut correctness: After converting tableau cuts to the original variables, what exact validity and numerical checks are applied? Please report the number of generated, rejected, and committed cuts, the cut budget, and a comparison with standard HiGHS/SCIP separation under the same budget.
  14. Runtime and hardware: What are the hardware, precision, batch, and CPU/GPU settings for the reported wall times? Please publish per-node encoding, policy, rollout, cut, LP, peak-memory, and total-time distributions, since the rollout reduces nodes but is slower than MF-blend.
  15. Reproducibility artifacts: Can you release the exact checkpoint checkpoints/highs retrain/phase4 final.pt, the trajectory or instance files (or immutable hashes), raw benchmark JSON, environment versions, and the command lines that generated Tables 1–5? The supplied archive appears to contain source and configs but not these artifacts.
Reviewer_LPxeOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

CutWorld learns action-conditioned latent dynamics for short-horizon lookahead in branching-variable selection, with shared representations and prediction heads additionally used for node ordering and root Gomory-cut selection. Evaluation uses 15 Set Cover instances in a Python/HiGHS research harness. Some configurations report fewer search nodes, although rollout is slower than a simpler policy blend.

Critical strengths
  • Approximating expensive branching lookahead with learned dynamics is a meaningful direction. Predicting variable-level successor representations allows the branching policy to be reapplied in imagined states and is a technically motivated part of the design.
Critical weaknesses
  • The closest prior work is neither discussed nor evaluated, leaving the central contribution insufficiently established. PlanB&B (Strang et al., AAAI 2026) already learns B&B dynamics and uses planning over imagined states to improve branching. The manuscript neither cites nor compares with this directly relevant work, and provides no explanation for its absence. Consequently, the claim of “a first realization of latent decision dynamics inside B&C” is insufficiently supported. Differences in transition architecture, training objectives, and additional control modules may provide contributions, but their distinction and value are not established.
  • The central results contain an unresolved inconsistency, and the method-to-experiment description lacks clarity. Tables 1–2 report 100% completion for all neural methods alongside differing nonzero incumbent gaps. Under the stated queue-exhaustion definition, exactness conditions, and same-instance best-feasible reference, valid completed searches should return an optimal objective and approximately zero relative gap. The absence of independently precomputed optima does not resolve this inconsistency. Until clarified, the completed-tree and efficiency comparisons cannot be reliably interpreted. Moreover, branching rollout is the technical core, but the additional node/cut components lack a clear account of their training, deployment, and use across experimental configurations. Comparing cuts against no cuts also does not establish the value of the proposed cut-selection rule itself.
Questions for authors
  • Completion and gaps: How can exhausted searches satisfying the stated exactness conditions have nonzero incumbent gaps? Please provide per-instance final objectives, certified bounds, termination reasons, and gap inputs to clarify whether this is a reporting issue or a failure of those conditions.
  • Distinction from PlanB&B: What are the key additions over PlanB&B in state transitions, policy reapplication, and planning objectives? Which specific additions are supported by the existing experiments?
  • Node/cut implementation and motivation: Please clarify how these components are trained and used at inference, which experimental configurations enable them, and how the cut-scoring parameters are obtained. Why are both components introduced alongside branching rollout, and can their individual contributions be distinguished from the existing results?
Reviewer_Z9wQOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces CutWorld, a latent decision-dynamics framework for Branch-and-Cut (B&C) control. The authors model B&C as a sequential decision process and learn (i) a GATv2 bipartite encoder that maps the relaxation graph to a latent state and per-variable embeddings, (ii) a causal Transformer dynamics model that predicts latent evolution under candidate actions, and (iii) policy, value, cost-to-go, and integrality heads. A per-variable transition head enables the policy to be re-applied in imagined states, and a search-aware rollout objective combines predicted dual-bound progress, value bootstrap, and cost-to-go to score candidate branching variables. The framework unifies branching, open-node ordering, and root Gomory-cut selection under a shared latent representation. The authors provide a correctness argument for admissible learned control and report a controlled Set Cover study with component ablations and explicit timeout semantics.

Critical strengths
  1. Unlike most learning-to-cut work that isolates cut selection, CutWorld places branching, node ordering, and root cut selection under one shared latent representation and dynamics model.
  2. Per-variable transition head is a genuine technical contribution. Eq. (12) addresses a real limitation of graph-level latent rollout: without per-variable predictions, the policy cannot re-score candidate variables in imagined states.
Critical weaknesses

W1. Experimental scale and statistical power are insufficient for the stated claims

The benchmark contains only 10 medium and 5 hard instances from a single generator seed. The authors acknowledge in Section 6 that SGM differences are descriptive, yet the abstract and conclusion use phrases such as "substantially reduces the completed-tree node count" and "general-purpose world model." With 15 instances and one seed, no paired test, no confidence interval, and no multi-seed replication, these claims are over-stated. The paper should either expand the evaluation (multiple seeds, larger held-out sets, paired tests) or substantially soften its conclusions.

W2. Benchmark breadth is too narrow for a "general-purpose" claim

Synthetic Set Cover is an acceptable starting point (Gasse et al. and SORREL use it), but the scale is far below community standards: 15 instances vs. the ~100-instance test sets typical of the Gasse family, and no MIPLIB 2017 subset. The paper only evaluates one problem family. Whether the per-variable transition and root cut selection generalize to Combinatorial Auction, Facility Location, or Maximum Independent Set is completely unknown. At minimum, a MIPLIB 2017 subset should be included to delineate applicability, even if results are not favorable.

W3. Wall-clock cost and inference amortization are not analyzed

Rollout reduces tree size but increases wall time. This is the central engineering trade-off, yet the paper does not report a breakdown of encoding, policy, rollout, and LP re-solve time per node, nor does it characterize the tree-size regime in which rollout becomes worthwhile. The performance profiles in Figure 3 are informative but do not resolve this question. Without an amortization analysis, the practical value of latent planning remains unclear.

W4. Cut-selection contribution is not isolated

The "Policy + root cuts" vs. "Policy" comparison does not report cut count, cut quality (efficacy/violation), LP re-solve time, or the gain of attention scoring (Eq. 21) over simple heuristics. The paper's own framing—"root cut selection rather than a learned general-purpose separator"—is honest, but the cut-selection component remains under-evaluated relative to its prominence in the framework.

W5. Proposition 1 needs tightening

The proposition omits a key condition: whether both branch children are created and retained, with direction dtd_t affecting only processing order. If only one child is kept, the proposition does not hold. The phrase "preserves the set of optimal integer solutions returned" is also imprecise—B&C returns one optimal solution, not a set. The proposition should state that no optimal solution is removed and that optimality is preserved under the same termination conditions. Proposition 1 and Appendix Proposition 2 are essentially the same result and should be merged.

Questions for authors
  1. Branch direction semantics. Does the action at=(vt,dt)a_t=(v_t,d_t) always create both children, with dtd_t only determining exploration order? If so, please state this explicitly in Proposition 1. If not, how is completeness preserved?
  2. Relation to GCNN. How does the "Policy (imitation)" row relate to the Gasse et al. (2019) GCNN? Can the authors provide a controlled comparison on the same instances and harness?
  3. SCIP baseline. Can the authors report either (a) a SCIP baseline with heuristics/separators disabled to isolate branching, or (b) a tuned SCIP baseline? Without this, how should the 0% vs. 100% completion contrast be interpreted?
  4. Cut-selection ablation. What is the gain of attention-based root cut selection over random selection and highest-violation selection on the same certified pool? What are the cut counts, LP re-solve times, and efficacy statistics?
  5. Rollout depth. Can the authors report K=1,2,3,5K=1,2,3,5 and prediction error as a function of KK, to substantiate the "shallow imagination" claim and identify where long-horizon dynamics break down?
  6. Inference cost breakdown. What are the per-node encoding, policy, rollout, and LP re-solve times? Under what tree-size or parallelism regime does rollout become wall-clock competitive?

Comments on the paper

1 comment

Dear authors and reviewers,

The author-reviewer discussion period closes on November 18 (AoE). Please engage in response periods so that we have our issues discussed thoroughly.

AC

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

Related papers

Discussion (0)

Sign in to comment.