{"id":"3b5913ff-9b47-4aca-8728-5d80a44b2404","arxiv_id":"2504.15062","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"OPO learns data acquisition decisions end-to-end through a differentiable optimise-predict-optimise chain, and on a drone reconnaissance shortest path benchmark it lowers relative decision regret from 0.469 (random search) to 0.240 after tuning.","lead":"This paper proposes OPO, a pipeline that decides which data to collect by using the quality of the final optimization decision as the objective instead of proxy scores such as coverage. On a drone-routing version of the Warcraft shortest-path problem, the learned acquisition strategy beats random search and lowers decision regret.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No proxy-objective baseline (coverage or mutual information) is tested, so the central claim that decision-focused DA improves on existing heuristics is not supported by the experiments.","rationale":"The central mechanism of OPO is internally coherent: the chain rule in Eq. (9) is valid, the counterfactual gradient is computable under Assumption 3.1, and the linear surrogate is fully expressive for binary acquisition decisions (every feasible binary vector is an extreme point of the convex hull of the feasible set, so a supporting linear objective exists). The reader's concern about the convex hull boundary condition is therefore less serious than stated. The genuinely load-bearing gap is empirical: the paper's motivating contrast is with proxy objectives like coverage, yet no coverage or mutual-information baseline appears in Section 4. The only comparison is random search over pi, which is not a standard DA heuristic. Since the headline contribution is a practical method, the absence of this comparison leaves the central claim unsupported. A separate weakness is the validation-set tuning of the pi learning rate for the best initialization; the reported 17% improvement uses best-of-N selection and could be mostly noise. Both issues would be settled by the proposed baseline comparison with multiple seeds. I keep the reader's CONDITIONAL verdict, since the method is plausible and the math is sound, but the empirical evidence needs strengthening.","tokens_in":8835,"tokens_out":25045,"duration_ms":224867,"concrete_test":"Implement a coverage heuristic (e.g., select up to 36 tiles maximizing spatial spread under the orienteering constraint) and a mutual-information or entropy baseline on the same Warcraft shortest-path setup, evaluate their test-set regret using the same pretrained masked ViT, and report mean plus/minus standard deviation over multiple random route seeds. If the coverage or MI baseline's regret is within one standard deviation of OPO's best reported 0.240 (or its mean), the claim that decision-focused acquisition outperforms proxy-objective heuristics is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper motivates OPO by arguing that data acquisition is usually solved heuristically for proxy objectives such as coverage, which risks objective misalignment. The experiments, however, only compare learned OPO to random search over pi and to fixed-pi baselines on the drone reconnaissance task (Section 4.3). There is no coverage-based or information-based acquisition baseline. Random search is not a proxy-objective heuristic, so the result that OPO beats random search does not establish that decision-focused acquisition is better than the coverage heuristics the paper critiques. The reported 17% relative-regret improvement over the best non-learnt method is also based on best-of-N selection and validation-set tuning of the pi learning rate (Tables 1-2), which may overstate the advantage. The mathematical mechanism is coherent: for binary s, every feasible solution is an extreme point of conv({s: A(s)<=0}), so a linear surrogate can in principle represent any optimal DA strategy; but this representational point does not substitute for an empirical comparison against the proxy baselines the paper is meant to improve upon.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes OPO (optimise-predict-optimise), a differentiable pipeline for making constrained data-acquisition decisions in contextual stochastic optimization. A binary acquisition vector s is chosen by solving a linear surrogate max ⟨π,s⟩ subject to A(s)≤0; the selected contextual variables are passed through a masked ViT that predicts the uncertain coefficients θ; the downstream optimization problem is solved; and a PFL or DFL loss is backpropagated through the two optimization layers. The experiments concern a drone-reconnaissance variant of the Warcraft shortest-path problem, where an orienteering constraint limits the number of visited tiles. Ablations vary fixed vs. learned π, pretrained vs. fine-tuned vs. jointly trained predictors, and PFL vs. DFL losses, and the paper reports that warm-starting plus fine-tuning with a learned π gives the best results, with a claimed 17% relative-regret improvement over the best fixed-π baseline.","tokens_in":9046,"tokens_out":11196,"duration_ms":106328,"significance":"The idea of optimizing data-acquisition decisions against downstream decision loss is natural and potentially useful, and the paper's decomposition into a linear surrogate, a masked predictor, and a differentiable solver is clean. Strengths include a clear statement of the full-observation assumption, released code, and a sensible ablation design. The linear-surrogate justification is actually stronger than the paper claims: every binary feasible solution is an extreme point of its convex hull, so a linear objective can represent any optimal deterministic DA strategy. The main weaknesses are empirical: no coverage or mutual-information baseline is tested, and the headline relative-regret numbers are not transparently defined. If those are fixed, the paper would be a solid contribution to decision-focused learning and data acquisition.","major_comments":[{"comment":"The central quantitative claim, a 17% relative-regret reduction from 0.289 to 0.240, is not verifiable from the paper. 'Relative regret' is never defined, and the values 0.469, 0.289, 0.260, and 0.240 do not appear in any table. The table columns labeled lossDFL contain values around 38–43 while the perfect-information objective is 29.64, suggesting that the tables report absolute objective values rather than the regret defined in Eq. (7). Please define the metric, state the selection protocol, and report both the raw objective and the regret so the claim can be checked.","section":"Sec. 4.3, Tables 1–2"},{"comment":"The paper motivates OPO by arguing that DA is typically solved by proxy objectives such as coverage or mutual information, but the experiments include no such baseline. Random search over π is an uninformative control, not a proxy-objective heuristic. Therefore the paper does not demonstrate its claimed advantage over existing heuristics, only over random search. I recommend adding at least a coverage-oriented orienteering baseline (e.g., reward proportional to geographic spread or a hand-crafted visiting pattern) and, if possible, a mutual-information or informative-path-planning baseline to the drone task.","section":"Secs. 1, 4.3, 5"},{"comment":"The Full Observation Assumption is load-bearing because the counterfactual gradient in Eq. (10) requires fully observed z and θ to compute ∂I/∂s. When historical data are themselves only partially observed, the proposed training loop cannot be applied. This assumption narrows the practical scope of the paper considerably, and the text should either discuss relaxations or explicitly restrict the claims to the full-observation setting.","section":"Assumption 3.1, Eq. (10)"}],"minor_comments":[{"comment":"Since the feasible solutions are binary, every feasible s is an extreme point of the convex hull of the feasible set; the boundary-condition justification for the linear surrogate is therefore automatically satisfied and should be stated as a fact rather than as a belief.","section":"Sec. 2.1"},{"comment":"The relationship between the MIN and BEST columns is unclear; for example, in the PFL-training lossDFL FIXED-FT row BEST (39.66) is worse than MIN (37.40) under a minimization objective. Explain how the best initialization is selected and why BEST can exceed the minimum.","section":"Table 1 caption"},{"comment":"The π learning rate is tuned on the best-performing validation initialization only, so the improvement in Table 2 may confound initialization choice with learning-rate choice; report the sweep over multiple initializations or explicitly acknowledge this in the limitations.","section":"Sec. 4.2, Table 2"},{"comment":"The notation ∂I/∂s followed by a vector of token gradients is dimensionally ambiguous; define the Jacobian explicitly.","section":"Eq. (10)"},{"comment":"The manuscript contains unfinished formatting artifacts ('Draft A PREPRINT', an unresolved 'available here' link for the code) and a few grammatical slips; these should be corrected before resubmission.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a promising but early preprint. The strongest path to acceptance is to add a coverage baseline, define the regret metric transparently, and tighten the novelty claims. I am not requesting a new theory; the linear-surrogate argument is acceptable once the automatic vertex fact is stated. Please also verify the claim of being 'the first decision-focused model for data acquisition' against the active learning and informative path planning literature."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague — quick take: OPO is a real extension, not a repackaging. It treats data acquisition as a trainable optimise-predict-optimise loop, replaces proxy coverage objectives with a learned linear surrogate on the acquisition variables, and differentiates through a masked tokenizer and a downstream solver. The masked-input differentiable tokenization (diag(s) t(z) + (1-s)⊗Λ) is a neat piece of engineering, and having two non-sequential implicit layers is genuinely new as far as I know. The authors are also honest about their assumptions: Section 2.1 admits the convex-hull boundary condition is an assumption, and Assumption 3.1 states the full-observation requirement up front. Code is linked and the math is coherent. Credit where due.\n\nSoft spots. The stress-test note is right: there is no coverage or mutual-information baseline. The paper motivates OPO by saying DA is usually solved for proxies like coverage, then only compares against random search over pi and fixed-pi. So the empirical claim 'decision-focused DA beats what people actually do' is unsupported. Random search is a useful control, but it is not the incumbent heuristic. That should be fixable in a revision. Second, the headline 17% relative-regret improvement is selected via best-of-N validation init and learning-rate tuning on the best validation init (Tables 1-2). Not fraudulent, but the reported advantage is optimistic. Third, the linear surrogate justification is representational rather than constructive: any binary feasible point is an extreme point of the convex hull, so a linear objective can in principle encode any optimal strategy, but that does not tell you gradient descent will find that pi or that the surrogate will transfer. Minor: 'first decision-focused model for data acquisition' and 'first instance of non-sequential optimisation layers' are strong claims; the first depends on literature coverage that is thinner than I'd like, and the second is a novelty claim best softened.\n\nBottom line: the method is sensible and the architecture is a genuine contribution. The missing baseline is a real gap but not a fatal one; the assumptions are stated and the mechanism is sound enough to deserve referee time. I'd send this out with a request for a proxy-baseline comparison, a less cherry-picked number, and toned-down claims.","headline":"A useful and mostly sound extension of decision-focused learning to data acquisition, but the experiments need a proxy-objective baseline before the central claim is supported.","tokens_in":9523,"tokens_out":1719,"would_cite":true,"duration_ms":15891,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C27","90C15","90C35","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that data acquisition decisions should be optimised for the downstream decisions they enable, and that OPO, a fully differentiable optimise-predict-optimise model with a learnable linear surrogate for acquisition value…","keywords":["decision-focused learning","data acquisition","differentiable optimization","contextual stochastic optimization","surrogate linear objective","orienteering problem","masked input prediction","shortest path"],"falsifier":"Run the OPO training loop on a dataset that is only partially observed, so that ground-truth context for unobserved variables is not available, and check whether the gradient computation can be replaced or the method reverts to random-search performance; if it cannot, the Full Observation Assumption is load-bearing. Separately, construct a data acquisition problem whose optimal solution lies in the interior of the convex hull of the feasible acquisition set; if the linear surrogate $\\langle \\pi, s\\rangle$ cannot recover that optimum, the paper's boundary assumption fails.","tokens_in":8647,"feed_emoji":"📷","tokens_out":7510,"duration_ms":63551,"temperature":0.7,"pith_summary":"OPO makes a simple but underused claim: the point of collecting data is to make better downstream decisions, so the data acquisition problem itself should be optimised for decision quality. The paper models acquisition as a constrained binary choice, replaces the unknown value of each acquisition plan with a learnable linear surrogate $\\langle \\pi, s\\rangle$, and trains the whole pipeline—acquisition, masked prediction, and downstream optimisation—end-to-end by differentiating through two optimisation layers. On a drone reconnaissance task where a limited-range drone chooses which tiles of an aerial image to photograph before a shortest-path problem is solved, the learned acquisition strategy outperforms random search and fixed heuristics, with relative regret falling from 0.469 to 0.240 after tuning. If correct, this gives a general recipe for aligning costly data collection with what a decision-maker actually cares about.","feed_headline":"Learning which data to collect cuts decision regret by half","feed_subtitle":"An end-to-end drone-reconnaissance model tunes data acquisition to the downstream shortest-path objective, beating random search.","key_machinery":"The machinery is a surrogate linear objective parameterised by vector $\\pi$ over binary acquisition decisions $s$, constrained by known constraints $A(s) \\le 0$; solving it gives $s^*(\\pi)$. Around this, the paper builds a fully differentiable chain: a masked tokenization layer $I(z,s;\\Lambda)=\\operatorname{diag}(s)t(z)+(1-s)\\otimes \\Lambda$ replaces unobserved image segments with a learnable mask token $\\Lambda$, a masked-input vision transformer predicts travel costs, and a shortest-path solver produces the decision. Two non-sequential differentiable-optimisation layers carry gradients: blackbox differentiation through the acquisition solver and a perturbed Fenchel-Young loss through the downstream shortest path. The counterfactual gradient $\\partial I/\\partial s$ is what connects a change in acquisition choice to a change in prediction, and it is what requires the Full Observation Assumption.","core_discovery":"The central claim is that data acquisition decisions can be treated as part of an optimise-predict-optimise chain and learned end-to-end, rather than fixed by coverage heuristics. Concretely, the paper posits an unknown value function $p(s)$ over binary acquisition decisions, and proposes to approximate its optimal solution by maximising a linear surrogate $\\langle \\pi, s\\rangle$ subject to known acquisition constraints; $\\pi$ is then learned so that the entire system minimises downstream decision loss. On the drone reconnaissance problem, the best configuration combines warm-starting the surrogate parameter $\\pi$ from a random search, then fine-tuning the prediction model and $\\pi$ together with the decision loss, which the paper reports as the first decision-focused model for data acquisition. The experiments show the learned acquisition strategy beats random search and fixed strategies in downstream objective value, with the best tuned model reducing relative regret by about 17% compared with the best non-learned strategy.","pith_inferences":["An extension the paper leaves open is reading $\\pi$ as a shadow price: if the acquisition budget is relaxed, the learned score could rank which additional sensors or tiles to buy, though the paper only frames $\\pi$ as relative value within a fixed budget.","Because OPO treats acquisition as constrained token selection, the same architecture could plausibly be adapted to language tasks where deciding which tokens or documents to read before answering is itself the acquisition decision, an extension the authors only conjecture.","The observed sensitivity of $\\pi$ optimisation to the learning rate points to a testable improvement: schedule the $\\pi$ learning rate dynamically after fine-tuning begins to stabilise the two-level optimisation.","The main scaling bottleneck is the two heuristic solver calls per batch; replacing the combinatorial acquisition layer with a solver-free differentiable surrogate would let OPO scale to larger orienteering-style problems, a direction the paper explicitly identifies as needed."],"forward_implications":["If OPO is correct, any constrained data acquisition problem with a differentiable downstream predictor can in principle be aligned with decision quality, not just with coverage or prediction error.","Warm-starting $\\pi$ via cheap random search and then fine-tuning end-to-end is a practical recipe that outperforms training from scratch or fixing the acquisition strategy.","The learned $\\pi$ values carry meaning as relative values of acquiring each piece of context, which could be used to rank data sources before a full acquisition budget is fixed.","PFL and DFL objectives give different behaviour under learned acquisition, so choosing the training loss must be guided by the downstream task structure, not by convention.","The experimental protocol gives a concrete benchmark: on Warcraft shortest path with orienteering constraints, a learned DA strategy reaches relative regret 0.240, a number future methods can be compared against."],"supporting_citations":[{"why":"Supplies the blackbox differentiation method that gives gradients through the combinatorial acquisition solver.","marker":"Pogančić et al. (2020)"},{"why":"Supplies the perturbed Fenchel-Young loss used for the differentiable downstream shortest-path layer.","marker":"Berthet et al. (2020)"},{"why":"Supplies the masked image modeling encoder that OPO adapts as its prediction model.","marker":"Xie et al. (2022)"},{"why":"Provides the antecedent for learning linear surrogates of nonlinear optimisation problems with fixed constraints, which OPO extends to data acquisition.","marker":"Ferber et al. (2023)"},{"why":"Empirically motivates choosing PFYL over other differentiable-optimisation losses and documents PFL outperforming DFL on this shortest-path task.","marker":"Tang & Khalil (2024)"},{"why":"Provides the routing heuristics used to solve the orienteering subproblem in the reconnaissance experiments.","marker":"Furnon & Perron (2024)"}],"fun_headline_variants":["Learn which data to collect to cut decision regret","Decision-focused data acquisition beats random search","End-to-end data acquisition for downstream decision quality","Drone recon learns which images to capture","Differentiable data acquisition beats coverage heuristics"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The training set must contain fully observed contextual vectors and true parameters for every example, because the counterfactual gradient in Eq. (10) needs to say how the prediction would have changed if different data had been acquired; if the training labels are themselves only partially observed, the OPO loop cannot run.","fun_headline_variants_meta":{"raw":{"variants":["Learn which data to collect to cut decision regret","Decision-focused data acquisition beats random search","End-to-end data acquisition for downstream decision quality","Drone recon learns which images to capture","Differentiable data acquisition beats coverage heuristics"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001583,"raw_usage":{"total_tokens":6306,"prompt_tokens":926,"completion_tokens":5380,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":542,"completion_tokens_details":{"reasoning_tokens":5313}},"tokens_in":542,"tokens_out":5380,"duration_ms":34385,"temperature":1.0,"reasoning_tokens":5313,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:34:04.764441+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the OPO training loop on a dataset that is only partially observed, so that ground-truth context for unobserved variables is not available, and check whether the gradient computation can be replaced or the method reverts to random-search performance; if it cannot, the Full Observation Assumption is load-bearing. Separately, construct a data acquisition problem whose optimal solution lies in the interior of the convex hull of the feasible acquisition set; if the linear surrogate $\\langle \\pi, s\\rangle$ cannot recover that optimum, the paper's boundary assumption fails.","supporting_citations":[{"cited_title":"Learning with differentiable pertubed optimizers","cited_arxiv_id":null,"evidence_quote":"Supplies the perturbed Fenchel-Young loss used for the differentiable downstream shortest-path layer."},{"cited_title":"M., Huang, T., Zha, D., Schubert, M., Steiner, B., Dilkina, B., and Tian, Y","cited_arxiv_id":null,"evidence_quote":"Provides the antecedent for learning linear surrogates of nonlinear optimisation problems with fixed constraints, which OPO extends to data acquisition."},{"cited_title":"and Khalil, E","cited_arxiv_id":null,"evidence_quote":"Empirically motivates choosing PFYL over other differentiable-optimisation losses and documents PFL outperforming DFL on this shortest-path task."},{"cited_title":"and Perron, L","cited_arxiv_id":null,"evidence_quote":"Provides the routing heuristics used to solve the orienteering subproblem in the reconnaissance experiments."}],"review_version":1}