{"id":"137ddc6c-8c52-4245-8b99-a6a28601b0a4","arxiv_id":"2508.16171","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"SPL-LNS replaces the greedy proposal step in neural Large Neighborhood Search with sampling over locally-informed proposals, trained by hindsight relabeling on self-generated data, and reports large gains over prior neural LNS solvers on integer linear programs.","lead":"This paper introduces SPL-LNS, a neural network guided solver for integer linear programs that samples several repair proposals instead of picking one greedily, and trains itself on its own past solutions using hindsight relabeling. If the reported gains hold, it gives machine learning based solvers a practical way to escape local optima on hard optimization problems.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Sampling-enhanced LNS may gain only from extra compute; need compute-controlled comparisons.","rationale":"The reader's weakest assumption concerns generalization of the learned policy to new ILP families/sizes. I partially agree, but the more immediate and testable concern is whether the sampling mechanism itself, rather than the learned policy, is the source of gains. Given the full text is corrupted, neither concern can be resolved from the available evidence. My proposed test targets a concrete confound that is common in sampling-based heuristics and directly affects the central empirical claim. I keep the reader's UNVERDICTED verdict because the paper remains unverifiable; the concern reinforces the need for full experimental details and controlled comparisons.","tokens_in":11993,"tokens_out":2065,"duration_ms":25215,"concrete_test":"Re-run the paper's main evaluation while equalizing total objective evaluations: for each instance, give every baseline the same number of candidate solutions evaluated per LNS iteration (or the same wall-clock time, allowing greedy restarts). Compare final objective values and optimality gaps. If SPL-LNS's advantage becomes statistically insignificant, the performance claim reduces to a compute trade-off rather than a methodological gain.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that SPL-LNS substantially surpasses prior neural LNS solvers. The method differs from greedy baselines by sampling multiple locally-informed proposals per iteration. If the experimental protocol does not match computational effort across methods—e.g., baselines get one proposal per step while SPL-LNS evaluates many, or wall-clock time is not controlled—then the reported gains could reflect additional inference compute rather than the learned proposal policy or hindsight relabeling. The abstract's emphasis on 'sampling-enhanced' and the garbled fragment 'computational budget' suggest this was considered, but the unreadable full text prevents verification. This is load-bearing because if SPL-LNS's advantage disappears under a matched budget, the claimed contribution is substantially weakened.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SPL-LNS, a sampling-enhanced large neighborhood search method for integer linear programs. It frames LNS as a stochastic process, replaces the greedy proposal step with sampling over locally-informed proposals, and trains the proposal policy with a hindsight relabeling objective on self-generated data. The abstract claims that SPL-LNS substantially surpasses prior neural LNS solvers on ILPs of different sizes. In the supplied text, the abstract and a few fragments are readable, but the full body is heavily corrupted: equations, tables, experimental details, and related-work descriptions are largely unreadable. The available fragments mention 'computational budget' and 'fair comparison', but no quantitative results or baseline details are legible.","tokens_in":12032,"tokens_out":5105,"duration_ms":53383,"significance":"If the central claim is correct, SPL-LNS would make a useful contribution to neural ILP solving: the stochastic-process formulation gives a principled way to think about local optima, and the hindsight relabeling training scheme is a practical way to learn from self-generated rollouts. However, the empirical claim is the load-bearing part of the paper, and the supplied manuscript provides no readable experimental evidence. The sampling mechanism could simply be buying improvements with additional inference compute, and the self-training scheme needs a concrete demonstration that it is the source of the gain rather than a circular fit to the solver's own reachable solutions. Without a readable experimental section with compute-controlled comparisons and ablations, the significance of the contribution cannot be assessed.","major_comments":[{"comment":"The central claim, 'SPL-LNS substantially surpasses prior neural LNS solvers,' is unsupported in the supplied text. The abstract gives no quantitative numbers, and the experimental section is corrupted/unreadable, so baselines, instance families, problem sizes, objective values, standard deviations, and computational settings are not verifiable. Because this is an empirical paper, the absence of a readable results section is a load-bearing issue, not a presentation nit.","section":"Abstract and §4 (experimental section)"},{"comment":"SPL-LNS differs from greedy baseline solvers by sampling multiple locally-informed proposals per iteration. If the comparison does not match computational effort across methods—e.g., if baselines receive one proposal per step while SPL-LNS evaluates many, or if wall-clock time is not controlled—then the reported gains could be due to extra inference compute rather than to the learned proposal policy or hindsight relabeling. The readable fragments mention 'computational budget' and 'fair comparison,' but no details are legible. The authors should report compute-matched comparisons (same wall-clock time, same number of solver calls, or same total inference passes) explicitly.","section":"Sampling mechanism and computational budget"},{"comment":"Training with hindsight relabeling on self-generated data introduces a partial circularity: the proposal policy is fitted to solutions that its own rollouts can reach, so the policy's apparent competence may not imply an ability to find improving moves on new instances. This is not by itself disqualifying, but the paper needs to show that the learned sampler improves over a greedy policy under equal rollout/compute budgets and that the benefit generalizes to held-out instance families. I specifically request an ablation that compares SPL-LNS against (a) a greedy proposal with equal compute, (b) sampling with a random or uniform proposal, and (c) a policy trained on expert or externally generated trajectories. Without such an ablation, the contribution of the hindsight relabeling and sampling design is not isolated.","section":"Hindsight relabeling training (Section 3)"}],"minor_comments":[{"comment":"The abstract should report at least one quantitative result (e.g., average improvement over the strongest baseline and the instance sizes considered). Phrases like 'substantially surpasses' are not falsifiable without numbers.","section":"Abstract"},{"comment":"The supplied full text is severely corrupted: most equations, references, and experimental tables are unreadable. The authors must provide a clean, correctly encoded manuscript before the paper can be reviewed or published.","section":"Full text / formatting"},{"comment":"The stochastic-process formulation is barely legible in the supplied text. Definitions of the proposal distribution, the sampling budget, and the acceptance of sampled solutions should be stated clearly and precisely.","section":"Notation and stochastic-process formulation"},{"comment":"The references to prior neural LNS solvers are garbled. Please ensure they are properly cited and that the differences from specifically named baselines are discussed.","section":"Related work"}],"recommendation":"major_revision","confidential_remarks":"The text corruption is severe enough that I am effectively reviewing only the abstract plus a few fragments. If the actual submission is complete, the editor may wish to obtain the original source text before further review. The paper's own fragments about 'computational budget' suggest the authors are aware of the matching-compute concern, but the current version does not allow verification of any experiment."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things up front. First, the supplied full text is mostly garbled; I can read the abstract and scattered fragments, not the actual experiments. So this is a provisional read, not a verdict. Second, the central claim deserves a serious referee.\n\nWhat's new: casting LNS as a stochastic process, sampling locally-informed proposals instead of a single greedy prediction, and training the proposal policy with hindsight relabeling on self-generated data. The combination is a reasonable response to two known weaknesses in learned LNS: greedy proposals get stuck, and data collection is expensive. If the experiments hold up, this could be a solid step forward.\n\nWhat the paper does well: the abstract is honest about the limitations of greedy proposals, and the fragments I can read mention 'fair comparison' and 'computational budget,' which suggests the authors know the obvious confound. The authors are known for serious work.\n\nThe soft spots. Most importantly, the compute-budget concern is load-bearing. If SPL-LNS samples multiple proposals per step while the greedy baseline gets one, and the reported gains come from that extra inference, the method is less new. The abstract gives no numbers, so I can't check whether they matched compute. The fragment mentioning computational budget is encouraging, but it's a fragment. Second, the hindsight relabeling loop is a mild circularity: the policy is trained on solutions its own search produces, so it might be fitted to its own trajectory rather than to genuinely better solutions. That doesn't kill the paper, but it should be probed. Third, I have no way to assess baseline fairness, hyperparameters, or variance from the available text.\n\nOverall: the idea is plausible, the problem is relevant, and if the empirical claim is real it's worth chasing. But the onus is on the authors to show the gains survive matched compute and to open data/code.\n\nRecommendation: send to peer review, with instructions to reviewers to focus on the compute-budget matching and the source of the improvement. I would not desk reject this, and I would not cite it yet.","headline":"A plausible recipe for improving learned LNS, but the supplied text is too corrupted to verify the empirical claims and the compute-budget question is open.","tokens_in":12621,"tokens_out":2550,"would_cite":false,"duration_ms":27030,"reading_group":"maybe","serious_thinker":"unclear","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C10","90C59","68T20"],"pacs":[],"model":"deepseek-v4-flash","headline":"Sampling locally-informed proposals beats greedy selection in neural large neighborhood search for ILPs.","keywords":["large neighborhood search","integer linear programming","neural combinatorial optimization","sampling","hindsight relabeling","stochastic process","self-generated data","proposal policy"],"falsifier":"Hold training data and compute budget fixed; on a held-out ILP family from a different generator, compare SPL-LNS with sampling against a version that replaces sampling with the greedy argmax of the same learned policy. If the greedy version matches or beats the sampled version in final objective, the paper's central claim fails.","tokens_in":11770,"feed_emoji":"🎲","tokens_out":5225,"duration_ms":52898,"temperature":0.7,"pith_summary":"The paper targets Integer Linear Programs (ILPs), where Large Neighborhood Search (LNS) iteratively destroys a current feasible solution and solves a subproblem over the destroyed part to find a better one. Prior neural LNS solvers learn to greedily predict the single best next solution, which the paper argues traps them in local optima and wastes data. SPL-LNS instead samples a batch of locally-informed proposals from a learned policy, evaluates each one, and keeps the best, framing LNS as a stochastic process. To train the sampler without expert solutions, it uses hindsight relabeling on self-generated rollouts. The paper reports that SPL-LNS outperforms previous neural LNS solvers on ILP families of various sizes.","feed_headline":"Neural ILP solver beats greedy search by sampling proposals","feed_subtitle":"SPL-LNS samples repair proposals, trains on its own rollouts, and beats prior neural LNS solvers on ILPs of varied sizes.","key_machinery":"The central object is the learned 'locally-informed proposal' distribution: a generative policy that, given the current feasible solution and the ILP instance, proposes several candidate repaired solutions by solving a local subproblem. The argument is carried by (1) the stochastic-process formulation of LNS, which makes the proposal distribution an explicit object to learn; (2) the sampling step that draws multiple proposals per iteration instead of committing to the greedy argmax; and (3) hindsight relabeling, which uses the best solution found during a rollout as the target for the proposals that led to it, enabling training entirely on self-generated rollouts.","core_discovery":"The central claim is that a neural LNS policy trained to sample diverse, locally-informed neighborhood proposals—rather than to output a single greedy repair—combined with a hindsight-relabeling training loop on self-generated data, yields better final objective values than prior neural LNS solvers. The paper formalizes LNS as a stochastic process whose transition is the proposal-and-repair step, and treats the proposal distribution as learnable. SPL-LNS draws multiple proposals conditioned on the current solution, solves each corresponding subproblem, and accepts the best feasible solution found; the hindsight relabeler turns observed improvements into supervised training targets, allowing","pith_inferences":["A natural extension is to vary the number of samples per iteration; the paper's logic predicts diminishing returns as the batch grows, which would separate the benefit of sampling from the benefit of extra compute.","The hindsight-relabeling mechanism resembles goal relabeling in goal-conditioned reinforcement learning, so the same training scheme may transfer to mixed-integer nonlinear programs or constraint satisfaction problems where LNS is applied.","The learned proposal distribution could be combined with portfolio selection over destroy operators or with a classical LNS destroy-repair heuristic as a fallback, potentially improving robustness on out-of-distribution instances."],"forward_implications":["Under a fixed iteration budget, sampling multiple proposals rather than committing to the greedy prediction reduces the chance of converging to a poor local optimum, so the final objective improves.","Hindsight relabeling provides a training signal from the solver's own rollouts, removing the need for expert demonstrations or an optimal solver to generate labels.","The stochastic-process view of LNS could support quantitative analysis of sample efficiency and be reused by other search heuristics.","Because the reported gains span ILP families and instance sizes, the approach suggests a general-purpose learned LNS component rather than a per-instance tuned heuristic."],"supporting_citations":[],"fun_headline_variants":["Sampling beats greedy in neural ILP solver","Neural LNS: sampling proposals escapes local optima","Self-trained sampling solver outperforms prior neural LNS","Hindsight relabeling boosts neural LNS sampling","SPL-LNS: sample, repair, relabel, repeat"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The learned proposal policy, trained only by hindsight relabeling on its own rollouts, assigns sufficiently high probability to improving proposals on unseen ILP families and instance sizes, so that sampling from it reliably escapes local optima.","fun_headline_variants_meta":{"raw":{"variants":["Sampling beats greedy in neural ILP solver","Neural LNS: sampling proposals escapes local optima","Self-trained sampling solver outperforms prior neural LNS","Hindsight relabeling boosts neural LNS sampling","SPL-LNS: sample, repair, relabel, repeat"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000126,"raw_usage":{"total_tokens":924,"prompt_tokens":696,"completion_tokens":228,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":440,"completion_tokens_details":{"reasoning_tokens":149}},"tokens_in":440,"tokens_out":228,"duration_ms":3138,"temperature":1.0,"reasoning_tokens":149,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T17:29:07.556765+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Hold training data and compute budget fixed; on a held-out ILP family from a different generator, compare SPL-LNS with sampling against a version that replaces sampling with the greedy argmax of the same learned policy. If the greedy version matches or beats the sampled version in final objective, the paper's central claim fails.","supporting_citations":[],"review_version":1}