{"id":"35f54cbe-94da-484f-957b-c52796eebfc9","arxiv_id":"2607.04487","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":4,"one_line_summary":"Graph encoders and recourse training improve latent constraint representation and actionable decoder explanations in neural MAVRP solvers; make-feasible counterfactuals arise from the training regime, not the mask.","lead":"A joint XAI protocol opens both the encoder and decoder of neural multi-attribute vehicle-routing solvers, showing that graph inductive bias improves constraint readability and sanity while recourse training uniquely yields make-feasible counterfactuals. Logistics operators who must validate or override AI routes gain a concrete, multi-criterion way to compare black-box solvers.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified to the central RECOURSE-vs-HARD-MASK claim; the ablation cleanly isolates training regime from mask.","rationale":"The reader's weakest_assumption correctly flags the fixed-budget CF search and supervised dictionary as acknowledged limits (§3.2, §7), but those limits do not load-bear on the central claim. The claim rests on the controlled ablation that equalises the candidate pool and still finds a clean zero/nonzero gap. Because the same search procedure is applied to both decoder families, any incompleteness of the search affects absolute rates equally and cannot manufacture the observed asymmetry. Inference-cost competitiveness (Table 2) further rules out a degenerate-policy confound. Consequently the CONDITIONAL verdict (mild code-availability flag) already reflects the right residual caution; no adjustment is warranted.","tokens_in":13022,"tokens_out":473,"duration_ms":5506,"concrete_test":"Re-run the §5.7 ablation with a larger search budget (e.g., 20 verified candidates instead of 3, or a simple projected-gradient optimiser over continuous relaxations of the seven signed moves) on the same six checkpoints; if HARD-MASK make-feasible rates remain exactly 0.00 while RECOURSE rates stay positive, the training-regime claim is confirmed under a stronger search.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's strongest claim—that the RECOURSE training regime (not merely its softer mask) produces policies that represent infeasibility usefully, yielding make-feasible counterfactuals that HARD-MASK policies never produce even after the alt-selector is externally widened—is supported by a direct representational ablation in §5.7. After matching alt-availability rates (0.72–0.88), make-feasible rates remain 0.00 for all three HARD-MASK checkpoints and 0.03–0.06 for RECOURSE. The fixed-budget first-order search (§3.2) is acknowledged as a lower bound, yet the zero-vs-nonzero asymmetry is decisive under that shared procedure. The five-criterion scorecard and supervised concept dictionary are secondary scaffolding; they do not underwrite the ablation result itself. No internal inconsistency or untested assumption that would reverse the claim is present.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes a joint two-pillar XAI protocol for neural autoregressive MAVRP solvers: encoder probing (linear probes, spontaneous organization, effective/stable rank, PCA/ICA directions with intervention validation) and decoder attribution (gradient, IG, DeepLIFT under abductive, contrastive, and counterfactual readings). Six (encoder × decoder) checkpoints are compared under a five-criterion scorecard (fidelity, concentration, stability, sanity, actionability). The central empirical claims are that graph inductive bias improves probe quality and decoder sanity, that UNIMPMOE encodes constraints in a distributed rather than axis-aligned way, and—most sharply—that the RECOURSE training regime (not merely its softer mask) produces policies whose neighbourhood contains make-feasible counterfactuals that HARD-MASK policies never produce, even after the alt-selector is externally widened to the same candidate pool (§5.7).","tokens_in":13332,"tokens_out":1048,"duration_ms":9463,"significance":"If the results hold, the work supplies the first controlled joint encoder–decoder XAI account for neural combinatorial-optimization solvers and isolates a training-regime effect that is invisible to inference cost alone. The representational ablation in §5.7 is a clean, falsifiable contribution: after matching alt-availability rates, make-feasible rates remain 0.00 for all HARD-MASK checkpoints and 0.03–0.06 for RECOURSE. Competitive BKS gaps (Table 2), multi-seed evaluation, randomized-weight sanity controls, and an explicit lower-bound statement on the counterfactual search strengthen the empirical package. The protocol is reproducible on the rl4co interface and is therefore usable as a benchmark scaffold for the community.","major_comments":[{"comment":"§5.7 and §3.2: the make-feasible claim is decisive under the shared first-order search, yet the absolute rates (0.03–0.06) and the fixed budget of three replays / 5× relative-size cap leave open how much of the gap is search-limited. A short sensitivity check (larger budget or a second-order / black-box local search on the same seven signed relaxations) would show whether the zero-vs-nonzero asymmetry survives a stronger optimiser; without it the operational magnitude of the training-regime effect remains under-quantified even though the qualitative claim is sound.","section":null},{"comment":"§4.1–§4.3 and Table 1: the encoder ranking (UNIMP/RECOURSE ≻ ATT) rests on a supervised constraint-family dictionary and on k-means / PCA–ICA choices whose sensitivity is not reported. Because the paper already notes that concept-level macro-F1 flattens the ranking, a brief leave-one-family-out or random-label control would confirm that the graph advantage is not an artefact of the particular taxonomy used for probing.","section":null}],"minor_comments":[{"comment":"Figure 6 scorecard omits stability (reported only in Table 6); a one-sentence cross-reference in the caption would make the five-criterion grid self-contained.","section":null},{"comment":"§5.2 / Figure 5: the method-sensitive family ranking is correctly used to motivate multi-criterion evaluation, but the text could state more explicitly that equal-fidelity attributions can still tell different qualitative stories (already implied by the equal flip@1 numbers).","section":null},{"comment":"Notation: the latent is written both h and h ∈ R^{n\times d}; a single consistent symbol would help. Likewise, Δ_t is used for both the logit margin and the log-prob drop in different tables.","section":null},{"comment":"Table 2: the “best on” column for published baselines is useful; adding the same column for the RECOURSE rows (even if all zeros) would make the cost–interpretability trade-off fully symmetric.","section":null},{"comment":"§7 Limits correctly flags the supervised dictionary and the n=100 single-seed check; a one-line statement that the n=50 grid is the variance-controlled primary result would further clarify scope.","section":null}],"recommendation":"minor_revision","confidential_remarks":"The manuscript is a solid empirical measurement study rather than a methodological breakthrough in attribution algorithms; its value lies in the controlled six-variant design and the clean ablation. Fit for a methods / applications track of a serious ML or OR journal is good; pure theory venues would find the contribution thinner. No citation or novelty red flags."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The thing worth knowing is the §5.7 ablation. After they force HARD-MASK policies to see the same visited-only candidate pool that RECOURSE sees, make-feasible rates stay at 0.00 while RECOURSE sits at 0.03–0.06. That isolates training regime from mask cleanly, and it is the result that is actually new.\n\nThe rest of the paper is a careful six-variant measurement study (ATT / UniMP / UniMP-MoE × Hard-Mask / Recourse) that instruments both stages with the same checkpoints. Encoder side: linear probes, NMI/silhouette, effective rank, PCA/ICA plus intervention validation. Decoder side: gradient / IG / DeepLIFT under abductive, contrastive, and counterfactual readings, scored on fidelity, concentration, stability, sanity, and actionability. Graph bias improves probe quality and sanity ratios; MoE looks weaker on single-axis probes but catches up in subspaces (distributed coding); RECOURSE widens alternatives and shifts mass onto dynamic state features. Inference costs stay competitive (Hard-Mask at or better than published baselines; Recourse pays ~1 point). Tables and figures line up with the claims.\n\nSoft spots are real but secondary. The counterfactual search is first-order, fixed-budget, seven signed relaxations; rates are lower bounds by design. The concept dictionary is supervised. Stability of the dominant family drops under Recourse. No public XAI runner is shipped, only the rl4co training interface. None of that overturns the ablation or the controlled comparisons.\n\nThis is for people who build or audit neural CO solvers and care whether explanations are usable by dispatchers. The math is standard attribution and probing; the data design is the contribution. I would send it to referees. Worth engaging if you work on XAI for routing or recourse-style policies.","headline":"Solid joint XAI protocol for neural MAVRP with a clean ablation: Recourse training, not the mask, is what creates make-feasible counterfactuals.","tokens_in":13873,"tokens_out":463,"would_cite":true,"duration_ms":9935,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Training under recourse, not just a soft mask, is what makes neural routing solvers produce useful make-feasible explanations.","keywords":["multi-attribute vehicle routing","neural combinatorial optimization","encoder probing","decoder attribution","hard-mask decoder","recourse decoder","make-feasible counterfactuals","explainable AI"],"falsifier":"Widen the alternative selector on hard-mask checkpoints to the visited-only pool used by recourse, re-run the make-feasible search, and check whether the make-feasible rate remains zero; a non-zero rate would collapse the claim that the training regime, not the mask, is responsible.","tokens_in":13936,"feed_emoji":"🚚","tokens_out":673,"duration_ms":5727,"temperature":0.7,"pith_summary":"Neural solvers for multi-attribute vehicle routing can match strong heuristics on cost, yet they give dispatchers no per-step account of why a stop was chosen, which constraint dominated, or what small change would flip the decision. This paper opens both black boxes of an autoregressive solver in one protocol: it probes the encoder for how constraint families live in the latent, and it attributes decoder decisions under three complementary angles (why this action, why not the next-best alternative, and what minimal change would switch or restore feasibility). Across six encoder–decoder combinations the results are concrete. Graph inductive bias improves both how predictably constraints appear in the latent and how trustworthy decoder saliency is. A mixture-of-experts encoder stores the same information in a distributed rather than axis-aligned way. Most importantly, the recourse training regime—not merely its softer mask—produces policies whose neighbourhood contains actionable make-feasible counterfactuals; hard-mask policies never produce them, even when they are externally shown the same infeasible alternatives. The paper therefore treats the decoder as an interpretability choice, not only a performance choice.","feed_headline":"Recourse training unlocks make-feasible routing explanations","feed_subtitle":"Hard-mask policies never produce them, even when shown the same infeasible alternatives","key_machinery":"A two-pillar XAI protocol that freezes the same six trained checkpoints and scores them on one five-criterion grid: encoder probes (linear predictability, spontaneous organization, effective rank, discovered directions with intervention validation) paired with decoder attributions (gradient, integrated gradients, DeepLIFT) read abductively, contrastively, and counterfactually.","core_discovery":"The recourse training regime, not merely its softer mask, produces policies that represent infeasibility usefully and thereby expose make-feasible counterfactuals that hard-mask policies fail to produce even when the alternative selector is widened to the same candidate pool. Graph inductive bias simultaneously improves encoder predictability and decoder sanity, while a mixture-of-experts encoder encodes constraints in a distributed rather than axis-aligned fashion.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Recourse training, not soft masks, unlocks make-feasible VRP counterfactuals","Hard-mask solvers miss make-feasible fixes even with identical candidate pools","Graph inductive bias lifts encoder predictability and decoder explanation sanity","MoE encoders store constraints distributedly rather than axis-aligned","Recourse regimes make infeasibility useful for actionable routing attributions"],"cache_read_input_tokens":9344,"weakest_assumption_plain":"The fixed-budget first-order search over seven signed relaxations and the five-criterion scorecard are assumed to capture what a dispatcher would actually find actionable.","fun_headline_variants_meta":{"raw":{"variants":["Recourse training, not soft masks, unlocks make-feasible VRP counterfactuals","Hard-mask solvers miss make-feasible fixes even with identical candidate pools","Graph inductive bias lifts encoder predictability and decoder explanation sanity","MoE encoders store constraints distributedly rather than axis-aligned","Recourse regimes make infeasibility useful for actionable routing attributions"]},"model":"grok-4.5","effort":"low","cost_usd":0.004314,"raw_usage":{"total_tokens":1261,"prompt_tokens":813,"num_sources_used":0,"completion_tokens":76,"cost_in_usd_ticks":43140000,"prompt_tokens_details":{"text_tokens":813,"audio_tokens":0,"image_tokens":0,"cached_tokens":0},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":372,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":813,"tokens_out":76,"duration_ms":3584,"temperature":1.0,"reasoning_tokens":372,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-11T18:44:47.151981+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Widen the alternative selector on hard-mask checkpoints to the visited-only pool used by recourse, re-run the make-feasible search, and check whether the make-feasible rate remains zero; a non-zero rate would collapse the claim that the training regime, not the mask, is responsible.","supporting_citations":[],"review_version":1}