{"id":"582b9b57-57c2-4230-bdcd-22f91e934015","arxiv_id":"2608.07398","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Inverse optimization of a mixed-integer production planning model on 50 training plans reveals that Dow planners weight avoiding understock and stable cycle lengths most heavily.","lead":"Using historical production schedules, the authors infer what objective Dow's production planners were implicitly optimizing, and find that avoiding inventory shortages and keeping cycle lengths steady dominate their choices. The work shows a way to convert tacit human expertise into an interpretable model, which could make decision-support tools in industry more accurate and trusted.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Omitted objectives (acknowledged in §2.3) can bias the learned weights; without a sensitivity check, the claimed dominance of understock and cycle-length objectives is not secured.","rationale":"I read the paper as an applied inverse-optimization case study whose value rests on two things: accurate prediction of expert decisions and an interpretable explanation of those decisions. The predictive part is supported by test-set RMSE improvements over a one-parameter inventory-cost baseline and by the transparent use of real Dow data. The interpretive part is the central claim: the learned alpha values reveal true planner priorities. For that claim to hold, the objective model must contain the true objective or at least not omit terms that are correlated with the included ones. The authors themselves flag incompleteness in §2.3. Since they do not report scaling factors, lambda, or any sensitivity analysis, the reader cannot distinguish a true priority signal from omitted-variable bias. The expert interview, while valuable, is not an independent test because the same expert supplied the candidate objectives and was asked to react to post-hoc hypotheses. I therefore agree with the reader's weakest assumption. A simulation-based sensitivity check is the single most direct way to test whether an omitted correlated objective can manufacture the observed dominance of alpha_sl and alpha_cl. If the method is robust in simulation, the concern is weakened; if not, the paper's central interpretive claim is conditional on an unverified assumption. No change to the reader's CONDITIONAL verdict is needed.","tokens_in":15002,"tokens_out":11665,"duration_ms":99125,"concrete_test":"Run a controlled simulation that mirrors the Dow case: choose a known true objective that includes the seven terms plus one held-out business-specific term correlated with C_sl (e.g., a penalty on late delivery to high-priority customers), generate 50 feasible 200-period plans by solving the forward MILP (P), and then estimate the seven-term model from these plans with the paper's suboptimality-loss method. If the estimated alpha_sl or alpha_cl is materially inflated relative to the true value, or if the dominance ranking changes when the held-out term is added to the estimation set, then the acknowledged omitted objectives are a plausible explanation for the paper's headline finding. Repeating the exercise with the actual confidential objectives, if they can be anonymized, would settle the issue directly.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the estimated weights alpha describe the planners' true priorities, not merely that they fit historical plans. That requires the seven-term weighted objective in (P) to be a correct specification of the planners' objective. Section 2.3 explicitly states that 'there are some additional business-specific objectives that we could have incorporated; however, for confidentiality reasons, we only consider the more generally applicable objectives.' The specification is therefore known to be incomplete. If any omitted objective is correlated with an included term—for example, a confidential customer-priority or product-margin objective that also leads to higher end-of-campaign inventory—the affected alpha will absorb that omitted priority. The reported dominance of alpha_sl and alpha_cl would then be an artifact of the chosen objective set, not a faithful description of planner preferences. The expert-validation loop in §4.3 does not remove this concern: the same expert helped enumerate the candidate objectives in §2.3, and the hypotheses were generated after seeing the learned weights, so the confirmations are not independent evidence against omitted-variable bias. No sensitivity analysis with alternative objective sets, no data release, and no identifiability analysis is provided. This is a load-bearing external-validity gap in the central interpretive claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper applies data-driven inverse optimization to infer the objective function of expert production planners at a Dow chemical plant. The forward problem is a mixed-integer linear program whose constraints encode campaign, cycle, and inventory rules, and whose objective is a weighted sum of seven hypothesized cost terms (inventory holding, four inventory-range penalties, cycle-length consistency, and production-gap reduction). Weights are learned from 50 historical production plans by minimizing a suboptimality loss with L1 regularization using a cutting-plane method. The authors compare one-, five-, and seven-parameter models, report test-set RMSE improvements, and then extend the model to time- and product-dependent weights. They interpret the estimated weights as revealing that avoiding inventory shortages and maintaining consistent cycle lengths dominate planner decisions, and they report expert-planner confirmation of several hypotheses and rejection of two.","tokens_in":15233,"tokens_out":5822,"duration_ms":55429,"significance":"If the central interpretation holds, this is a valuable demonstration that inverse optimization can extract interpretable, data-efficient models from real industrial planning data with rich discrete constraints. The strengths are the genuine industrial case study, the transparent expert-validation protocol that includes falsified hypotheses, the explicit treatment of planner suboptimality via suboptimality loss, and the data efficiency achieved with only 50 training plans. These features make the paper a useful empirical contribution to the inverse-optimization and apprenticeship-learning literature. However, the central claim that the learned weights faithfully describe the planners' true priorities depends on the completeness of the hypothesized objective set and on several modeling choices that are not fully disclosed or stress-tested.","major_comments":[{"comment":"The central interpretive claim that the estimated α values reveal the planners' true priorities is load-bearing, but Section 2.3 explicitly states that additional business-specific objectives were excluded for confidentiality. If any omitted objective is correlated with an included cost term, the corresponding α absorbs that omitted priority, so the reported dominance of α_sl and α_cl could be an artifact of the incomplete candidate set. Because the same expert helped enumerate the candidate objectives in §2.3 and later validated hypotheses in §4.3, the confirmation is not independent evidence against this omitted-variable bias. Please add sensitivity analyses with alternative or expanded objective sets, or at minimum state clearly that the learned weights are conditional on the chosen specification and cannot be interpreted as unbiased causal priorities.","section":"§2.3 and §4.3"},{"comment":"The interpretation of α magnitudes relies on the scaling factors ρ_k, described only as 'calculated beforehand,' and on the regularization parameter λ in (IOP-SL), whose value and selection procedure are not reported. Different choices of ρ_k change the relative scale of the cost terms and therefore the comparability of α values across objectives; different λ values induce different shrinkage and sparsity, so the observation α_ic=0 in Figure 7 could be a regularization artifact rather than a planner preference. Please report how ρ_k is computed, the value or tuning method for λ, and a sensitivity analysis with λ=0 and a range of ρ choices to show that the ranking of weights is stable.","section":"§2.3, problem (P), and §3.2"},{"comment":"The suboptimality-loss formulation (IOP-SL) can admit multiple optimal cost vectors, and the paper does not analyze identifiability of α or report variation of the estimates across different algorithmic restarts or initializations. The five training-testing splits provide variance across data subsets, but not evidence that the specific estimated α is unique. Please add an identifiability analysis, report the range of optimal α vectors obtained from different master-problem starts, or discuss the non-uniqueness and its implications for the weight-ranking claims.","section":"§3.2 and §3.4"},{"comment":"The expert-validation loop is presented as confirming practical validity, but the same expert who helped elicit the candidate objectives in §2.3 later confirmed or refuted hypotheses generated from the learned weights. The two false hypotheses (Hypotheses 3 and 6) show that the validation has some discriminative power, but the 'True' conclusions are still vulnerable to confirmation bias because the expert is not independent of the model-building process. Please solicit feedback from a second planner not involved in the study, or explicitly reframe the expert interviews as illustrative and hypothesis-generating rather than confirmatory evidence for the weight interpretation.","section":"§4.3"}],"minor_comments":[{"comment":"The RMSE comparisons are presented only graphically; please report the numerical mean and standard deviation over the five splits so that readers can assess whether the improvements of model-7 over model-5 are substantial relative to the variability.","section":"§4.2 and Figures 5, 8, 11"},{"comment":"The feasible set C for the cost vector c is not specified. Please state explicitly whether nonnegativity of α is enforced, how the trivial all-zero solution is excluded, and what bounds, if any, are imposed on the weights.","section":"§3.2"},{"comment":"Model-7-t&p estimates 7×4×8 = 224 weights from only 50 training plans, and the paper does not report the regularization strength or parameter counts for this model. Please add a discussion of overfitting risk and the effective number of parameters.","section":"§4.4"},{"comment":"There are several typos in this section, including 'The is also true' in the paragraph introducing model-7-t&p and 'deviating form the targeted cycle length' in Hypothesis 9; these should be corrected.","section":"§4.4"},{"comment":"The explanation for the moderate value of α_g is speculative ('we do not know exactly why these plant downtimes occurred'). Consider labeling that explanation as a hypothesis requiring future investigation rather than a conclusion.","section":"§4.3, Hypothesis 6 discussion"},{"comment":"The statement that results are 'normalized if necessary' is vague; please specify the normalization applied to the data and confirm that the RMSE values reported in inventory-profile units are indeed in units of the batch size β in all figures.","section":"§4.1 and §4.2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's main contribution is empirical, and the expert validation is critical to the interpretation. Because one co-author is affiliated with Dow and may be the interviewed expert, the independence of the validation is questionable; the authors should disclose whether the validating planner is a co-author. The absence of data and code, while understandable for confidentiality, weakens reproducibility. The methodological novelty relative to Moghaddass and Terekhov (2021) and Holani et al. (2026) appears limited, so the paper should be judged primarily on the strength of the case study and the robustness of its interpretations."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know before you decide whether to spend time on this. It is a real industrial application of inverse optimization—70 historical production plans from a Dow plant, a rich MILP with campaign sequencing and inventory constraints, and an unusually honest expert-validation section. And the central interpretive claim, that the learned weights reveal what the planners actually care about, is less secure than the abstract makes it sound.\n\nWhat is new: as far as I can tell, this is the first application of data-driven inverse optimization to real production planning data with a non-trivial MILP forward model. The methodological core is the suboptimality-loss framework of Moghaddass and Terekhov (2021); the time- and product-dependent extensions are simple but appropriate. The paper does several things well. It reports predictive RMSE improvements over a one-parameter baseline across five random splits, it shows convergence of the cutting-plane algorithm, and it lists its hypotheses and explicitly marks which ones were rejected by the expert (Hypotheses 3 and 6). That transparency is genuinely good.\n\nThe soft spots are real, and the main one is exactly what the stress-test note says. Section 2.3 admits that business-specific objectives were omitted for confidentiality. If an omitted objective correlates with an included term—for example, a confidential customer-priority objective that also leads to higher end-of-campaign inventory—the affected weights absorb that priority, and the 'understock and cycle consistency dominate' conclusion becomes partly an artifact of the chosen objective set. The expert interviews do not resolve this: the same expert helped enumerate the candidate objectives, and the hypotheses were generated after seeing the learned weights, so the confirmations are not independent evidence. A sensitivity analysis with a different or expanded objective set, or an identifiability check on the weights, would have materially strengthened the interpretation. The absence of code and data is also a limitation; the numbers cannot be reproduced.\n\nNone of this kills the paper as an application demonstration. It kills only the strong version of the interpretive claim. The prediction gains and the honest reporting carry the paper. Who is this for: researchers in inverse optimization who want a realistic industrial testbed, and practitioners who want to see whether IO can work on messy, real-world problems. It deserves a serious referee. The referee should press for a discussion of omitted-variable bias, ideally a robustness experiment with an alternative objective set, and either code/data release or an explicit statement of why confidentiality prevents it. An equal-weight baseline would be nice but is not essential.\n\nRecommendation: send it to review. I would be happy to see it appear after moderate revision, with the interpretive claims scaled back to match the evidence.","headline":"Credible industrial inverse-optimization study with honest validation; central preference claims need a sensitivity check before being taken literally.","tokens_in":15732,"tokens_out":3162,"would_cite":true,"duration_ms":26918,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C11","90B30","90C90"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that expert planners' hidden objectives can be recovered from historical production plans, and that in a real industrial case the recovered weights make avoiding inventory shortages and keeping cycle lengths consistent…","keywords":["inverse optimization","production planning","mixed-integer linear programming","apprenticeship learning","human-in-the-loop decision-making","objective inference","industrial case study","suboptimality loss"],"falsifier":"Re-estimate the weights after appending a plausible additional objective term, such as a penalty for unmet service levels or a term encoding demand-forecast bias; if $\\alpha_{\\mathrm{sl}}$ and $\\alpha_{\\mathrm{cl}}$ no longer dominate, the claimed priority ranking is an artifact of the hypothesized objective set.","tokens_in":14806,"feed_emoji":"🏭","tokens_out":7922,"duration_ms":63416,"temperature":0.7,"pith_summary":"The paper tries to establish that the implicit objectives of human production planners can be learned from historical plans instead of being elicited directly. It models the planning problem as a mixed-integer linear program with known constraints and an unknown weighted-sum objective, and infers the seven objective weights from 50 past production plans using suboptimality-loss inverse optimization. The learned weights indicate that avoiding inventory shortages and maintaining consistent cycle lengths dominate planners' decisions, while inventory holding cost itself contributes nothing. If correct, this makes tacit expertise explicit enough to build a decision-support model that predicts planner behavior and earns planner trust.","feed_headline":"Hidden goals of expert planners decoded from 50 production plans","feed_subtitle":"Learned weights show avoiding understock and consistent cycle lengths dominate planner decisions.","key_machinery":"The load-bearing mechanism is the suboptimality-loss formulation of inverse mixed-integer linear optimization. For each observed plan, it introduces a nonnegative slack $\\epsilon_i$ measuring how much better a feasible plan could be under candidate weights, then minimizes the sum of these slacks plus an $\\ell^1$ penalty on the weights, avoiding the mixed-integer bilevel structure that would otherwise make the inverse problem intractable. The resulting semi-infinite linear program is solved by a cutting-plane algorithm that alternates between a linear master problem and forward MILP cut-generating problems. The output is the vector of seven objective coefficients $\\alpha_k$ in the weighted-sum objective of the forward production planning model.","core_discovery":"The central claim is that expert planners' historical production plans encode a recoverable objective function, and that the recovered weights are both interpretable and practically valid. Across five training-test splits of a seven-term model, the estimated weights show $\\alpha_{\\mathrm{sl}}$ and $\\alpha_{\\mathrm{cl}}$ are largest, $\\alpha_{\\mathrm{ic}}$ is effectively zero, and understock penalties outweigh overstock penalties ($\\alpha_{\\mathrm{sl}} \\gg \\alpha_{\\mathrm{su}}$, $\\alpha_{\\mathrm{el}} \\gg \\alpha_{\\mathrm{eu}}$). Time-dependent weights decline for $\\alpha_{\\mathrm{sl}}$ and rise for $\\alpha_{\\mathrm{cl}}$ across the horizon, and product-dependent weights vary across the eight products. An expert planner interview corroborated the main hypotheses, including the primacy of avoiding understock and keeping cycle lengths consistent, while invalidating two derived hypotheses about start-versus-end inventory attention and deliberate downtime avoidance.","pith_inferences":["If any confidential omitted objective is correlated with an included term, the learned dominance of understock avoidance and cycle consistency could be partly an artifact of the chosen objective set; a robustness study adding plausible omitted terms would reveal how stable the ranking is.","The same suboptimality-loss pipeline should transfer to other rolling-horizon planning settings where constraints are well specified but human objectives are tacit, such as logistics scheduling or hospital resource planning.","A stronger validity test would be to elicit pairwise preferences from planners independently and check whether the learned weights reproduce the rank order implied by those preferences."],"forward_implications":["Adding the learned objective terms yields predicted inventory profiles that match actual planner decisions more closely than an inventory-cost-only baseline.","The recovered weights give a quantitative priority ordering: understock penalties dominate overstock penalties, and cycle-length consistency and gap avoidance are strong drivers.","Time- and product-dependent extensions show planners focus more on early periods and on certain products, and these extensions improve predictive accuracy.","An interpretable model whose objectives match planner reasoning is more likely to be adopted as a trusted decision-support tool."],"supporting_citations":[{"why":"Supplies the suboptimality-minimization inverse integer optimization formulation on which problem (IOP-SL) is built.","marker":"Moghaddass and Terekhov (2021)"},{"why":"Provides the cutting-plane algorithm for inverse mixed-integer linear programs that underlies the solution method.","marker":"Wang (2009)"},{"why":"Gives polyhedral and trust-region insights for inverse mixed-integer optimization, supporting the theoretical treatment.","marker":"Bodur et al. (2022)"},{"why":"Establishes inverse optimization with noisy data, justifying the framework's tolerance for imperfect human decisions.","marker":"Aswani et al. (2018)"},{"why":"Frames apprenticeship learning as recovery of an expert's objective, the conceptual basis for interpreting learned weights.","marker":"Abbeel and Ng (2004)"},{"why":"Introduces the suboptimality-loss idea for inverse convex optimization that the integer extension adapts.","marker":"Keshavarz et al. (2011)"}],"fun_headline_variants":["Inverse optimization exposes planners' true objectives","Planners' hidden objectives: understock and cycle consistency","Inverse optimization: decode expert planners' real goals","From actions to objectives: inverse learning on production plans"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The learned weights are trustworthy only if the planner's true objective is exactly a weighted sum of the seven hypothesized cost terms; the paper itself notes that further business-specific objectives were excluded for confidentiality, and an omitted objective correlated with an included term could bias the weights.","fun_headline_variants_meta":{"raw":{"variants":["Inverse optimization exposes planners' true objectives","Planners' hidden objectives: understock and cycle consistency","Inverse optimization: decode expert planners' real goals","From actions to objectives: inverse learning on production plans"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000577,"raw_usage":{"total_tokens":2719,"prompt_tokens":937,"completion_tokens":1782,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":553,"completion_tokens_details":{"reasoning_tokens":1720}},"tokens_in":553,"tokens_out":1782,"duration_ms":13203,"temperature":1.0,"reasoning_tokens":1720,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T14:27:02.931673+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-estimate the weights after appending a plausible additional objective term, such as a penalty for unmet service levels or a term encoding demand-forecast bias; if $\\alpha_{\\mathrm{sl}}$ and $\\alpha_{\\mathrm{cl}}$ no longer dominate, the claimed priority ranking is an artifact of the hypothesized objective set.","supporting_citations":[],"review_version":2}