{"id":"23b3cb19-b264-400b-b8ba-836c4f7926b2","arxiv_id":"2607.22987","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"Adversarial inverse reinforcement learning over consecutive vibration samples yields label-free early fault warnings, but the headline robustness claim depends on a threshold fitted to the test set.","lead":"The paper applies adversarial inverse reinforcement learning to industrial machinery fault detection, treating each next sensor reading as a \"proxy action\" and scoring whether transitions look like healthy dynamics. It claims AIRL is the only method with stable, non-saturated post-detection behavior across three run-to-failure datasets.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"AIRL's 'only method with non-saturated PDC' claim relies on a test-derived Otsu threshold; under the paper's own recommended max−2SE threshold, IMS PDC collapses from 90.15% to a single flagged file.","rationale":"The reader's weakest assumption identifies the same load-bearing flaw: PDC is computed with a test-derived Otsu threshold. The paper's own text in §5.2 explicitly states that Otsu partitions test scores globally, and §5.4 recommends max−2SE for actual deployment. Under max−2SE, Table 5 shows only one IMS detection across the entire test set, so the advertised 90.15% PDC is not a property of the learned reward under a usable threshold. This is not a disagreement about consensus; it is an internal inconsistency and a deployment-feasibility failure. The abstract's 'only method' claim therefore collapses as stated. The paper still contains a defensible conceptual critique of contextual-bandit MFD and a reproducible experimental protocol, but the headline comparative claim needs to be re-evaluated with thresholds computed solely from training data. I agree with the reader's REJECT verdict; no adjustment is needed.","tokens_in":14798,"tokens_out":5752,"duration_ms":53210,"concrete_test":"Recompute Table 4 PDC for AIRL using the max−2SE threshold recommended for deployment in §5.4, with τ = max(α_healthy) − 2·SE(α_healthy) computed from healthy training scores only. Count the number of test files after the GT onset (IMS: after mid-Nov; HUMS: after Day 24 #264; XJTU-SY: after ~500 min) that exceed τ. If IMS yields 1 file (as Table 5 suggests), the PDC drops from 90.15% to ~1/N_post-GT. Additionally, implement Otsu as described in §3.4 on healthy training scores only (not test scores) and recompute all Table 4 entries; if the entries change substantially, the paper's internal inconsistency (§3.4 vs §5.2) is material. This check directly settles whether the 'only method' claim survives under a deployable threshold.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—AIRL is the only method with meaningful, non-saturated PDC across all three datasets—rests entirely on Table 4, which the paper states uses Otsu's adaptive threshold 'which partitions test scores globally to maximize between-class variance' (§5.2). This threshold is not available in a label-free deployment: it requires observing the entire test-score distribution, including future failures. Section 3.4 claims all thresholds are 'computed once from healthy training scores and fixed at test time,' but §5.2 explicitly contradicts this for the PDC evaluation. The paper's own deployment recommendation is max−2SE (§5.4), and under max−2SE, Table 5 shows AIRL flags exactly one IMS test file (Nov 17) across the whole 8,940-file test set. Post-onset IMS files (after mid-Nov) number in the thousands; one positive implies PDC ≈ 0.03%, not 90.15%. The 90.15% figure is thus an artifact of the Otsu threshold's hindsight use of test data. The same concern applies to the abstract's 'only method' claim, which is not supported under any threshold that can be computed from training data alone. This is not a minor evaluation nuance; it invalidates the headline comparative result.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an adversarial inverse reinforcement learning (AIRL) framework for label-free machinery fault detection, treating raw vibration scalar transitions as state-action pairs and recovering a health reward from healthy training trajectories. It evaluates AIRL on three run-to-failure datasets (HUMS2023, IMS, XJTU-SY) and compares against one-class, reconstruction, and contextual-bandit baselines. The headline claim is that AIRL is the only method achieving meaningful, non-saturated post-detection consistency (PDC) across all three datasets, supported by Table 4 which reports PDC values of 79.09%, 90.15%, and 68.38% under an Otsu threshold. The paper also provides a threshold sensitivity analysis, recommends max−2SE for deployment, and includes an ablation on the discount factor.","tokens_in":15090,"tokens_out":4316,"duration_ms":40895,"significance":"If the headline claim were valid, the paper would offer a practically interesting demonstration of IRL for prognostics: a compact (644-parameter) model that requires no fault labels, with a systematic comparison to bandit-style RL baselines and a useful critique of the static CB formulation. The threshold sensitivity analysis and the explicit discussion of saturation failure are also valuable contributions. However, the central PDC result is not supported by the paper's own deployment recommendations: Table 5 shows that under the recommended max−2SE threshold, AIRL flags exactly one IMS test file across the entire 8,940-file test set, so the claimed 90.15% PDC on IMS is an artifact of a test-derived Otsu threshold, not a property of the learned reward. This invalidates the abstract's 'only method' claim and the main comparative conclusion.","major_comments":[{"comment":"The PDC values in Table 4 are computed using Otsu's threshold 'which partitions test scores globally' (Sec. 5.2). This contradicts Sec. 3.4, which states that 'all thresholds are computed once from healthy training scores and fixed at test time.' A global Otsu threshold requires observing the full test score distribution, including future failures, and is therefore unavailable in a label-free deployment. The central PDC claim, and the abstract's assertion that AIRL is the only method with non-saturated PDC, rest on this hindsight threshold and are not reproducible under a training-only calibration.","section":"Sec. 5.2 vs. Sec. 3.4"},{"comment":"Under the paper's own recommended deployment threshold (max−2SE, Sec. 5.4), Table 5 shows that on IMS AIRL flags exactly one anomalous file (Nov 17) across the entire 8,940-file test set. The IMS post-onset period includes thousands of files; one positive implies PDC of roughly 0.03%, not the 90.15% reported in Table 4. The claimed cross-dataset PDC consistency therefore collapses under a threshold that can actually be used in practice. The text in Sec. 5.2 acknowledging the two-threshold framework does not resolve this: Table 4 is not a deployable evaluation.","section":"Sec. 5.3, Table 5"},{"comment":"The Limitations paragraph states that 'AIRL remains the only method maintaining non-saturated FAR across all three datasets.' This is contradicted by Table 4, where VAE has FAR 89.1%, 50.0%, 46.8% and TCAE has FAR 41.8%, 88.3%, 86.3% on HUMS, IMS, and XJTU-SY respectively — all below the paper's own 90% saturation ceiling. The claim is therefore false under the authors' definition. This is not a minor wording issue; it is part of the concluding summary of the method's advantage.","section":"Sec. 6, Limitations"},{"comment":"Table 5 shows that on IMS, all max-based thresholds (Max, Max−SE, Max−2SE) produce exactly one anomalous file, reflecting a compressed score range. Appendix A(iii) claims the sequential model 'maintained dense, stable detection across the full threshold hierarchy.' These statements are mutually inconsistent: one point cannot be described as dense or stable. Additionally, the IMS FAR of 88% under Otsu (Table 4) is only two percentage points below the paper's own saturation ceiling, making the 'non-saturated' designation fragile and operationally questionable.","section":"Sec. 5.3 vs. Appendix A(iii)"}],"minor_comments":[{"comment":"Sec. 4.1 reports the XJTU-SY test split as 'Remaining 18,295' files, but Sec. 5.2 states that CTQN is evaluated on a 16,295-file subset because the last 2,000 test files are withheld as training data. The discrepancy should be reconciled or clarified in the experimental setup.","section":"Sec. 4.1 vs. Sec. 5.2"},{"comment":"The notation σ is used both for the logistic sigmoid in Eq. (3) and for the standard deviation of the Gaussian policy in Sec. 3.2. This is confusing; consider using a distinct symbol such as sigmoid or logistic.","section":"Sec. 3.3, Eq. (3)"},{"comment":"The reference to AAMAS 2026 in Sec. 1 and Ref. [26] should be checked for publication status and exact venue details to avoid citing a non-existent or future volume.","section":"General"}],"recommendation":"reject","confidential_remarks":"The central comparative claim — that AIRL is the only method with non-saturated PDC across all three datasets — cannot be restored by a local fix. Under the paper's own recommended deployable threshold, the IMS PDC collapses to a single detection, and the 'only method maintaining non-saturated FAR' statement is contradicted by the paper's own Table 4 (VAE and TCAE both have all three FARs below 90%). The methodological proposal and the critique of CB formulations may have merit, but the evaluation and the abstract-level conclusions would require a fundamental rework rather than a revision. Given the discrepancy between Sec. 3.4 and Sec. 5.2, I would also ask the authors to provide the full anomaly-score distributions and reproduce the PDC analysis under a training-only threshold before any resubmission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing you should know: the paper's main advertised result—that AIRL is the only method with non-saturated post-detection consistency across all three benchmarks—does not hold under the threshold the authors themselves recommend for deployment. Table 4's PDC numbers use Otsu's threshold computed from the full test-score distribution; Section 3.4 says all thresholds are computed from healthy training scores. Those two statements can't both be true. Under their own max-2SE threshold, Table 5 shows AIRL flags exactly one IMS file in 8,940, so the 90.15% PDC becomes effectively zero. The abstract's 'only method' claim is therefore unsupported.\n\nWhat's genuinely good: the paper is the first to apply state-only AIRL to machinery fault detection, and the critique of prior 'RL-based' MFD as static contextual bandits is well-argued and useful. The saturation-aware PDC idea is a real methodological contribution, even if this execution botches it. The ablation on gamma is sensible, and they ship code and data. The 'scope of the RL claim' section is refreshingly honest.\n\nThe soft spots are load-bearing. Besides the threshold contradiction, the CTQN baseline is trained on labeled faulty data from the test period, which makes the 'label-free' comparison unfair. The XJTU-SY 'precursor window' is defined post hoc, and similar early detections by baselines are called false alarms. The IMS FAR of 88% is close to the saturation ceiling; calling it 'non-saturated' because it is below the 90% line is a thin distinction.\n\nWho should read it: people working on RL for prognostics, and anyone who wants a case study in how threshold choice can manufacture a comparative result. It deserves a serious referee, but the authors need to redo the evaluation with thresholds computed only from training data, or retract the 'only method' claim. I'd be willing to engage with a revised version.","headline":"A useful conceptual critique of contextual-bandit MFD, but the headline PDC claim collapses under the paper's own recommended deployment threshold.","tokens_in":15622,"tokens_out":3784,"would_cite":false,"duration_ms":32424,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that machinery fault detection can be cast as adversarial inverse reinforcement learning over observed state transitions—using the next sensor reading as a proxy action—so that a 'health' reward learned from healthy data a","keywords":["machinery fault detection","adversarial inverse reinforcement learning","run-to-failure prognostics","label-free anomaly detection","state-only imitation","post-detection consistency","temporal credit assignment","industrial sensor monitoring"],"falsifier":"On the IMS bearing dataset, compute post-onset detection consistency for AIRL using the Max−2SE threshold calibrated strictly from healthy training scores (the paper's deployment recommendation) instead of a threshold fitted to the full test set. If the number of post-onset flagged files collapses from thousands to one, the claim of consistent non-saturated tracking across all three datasets fails.","tokens_in":14607,"feed_emoji":"⚙️","tokens_out":6043,"duration_ms":56433,"temperature":0.7,"pith_summary":"The paper claims that treating machinery fault detection as an adversarial inverse reinforcement learning problem—learning a reward that scores how plausibly each sensor reading follows from the previous one, with the next observation serving as a proxy action—lets a model detect incipient faults from healthy-only training data. Across three run-to-failure benchmarks (a helicopter gearbox and two bearing datasets), the authors report that this method is the only one whose post-detection consistency remains non-saturated on all three datasets, avoiding the two failure modes that dominate baselines: constant false-alarm saturation or complete detection collapse. If correct, this provides a label-free prognostic path that requires no fault examples, no manual reward design, and only a few hundred parameters at deployment.","feed_headline":"Label-free reward learning flags faults on all three test rigs","feed_subtitle":"Needs no fault labels; baselines either cry wolf constantly or miss the onset entirely.","key_machinery":"The load-bearing mechanism is the AIRL discriminator, which separates expert (healthy) transitions from policy-generated ones using the identity D(s,a,s′) = exp(r(s,a)+γV(s′)−V(s)) divided by exp(r(s,a)+γV(s′)−V(s)) plus π(a|s), where the 'action' a is the next sensor observation x_{t+1} (the state-only imitation trick), r is the learned reward, V a value function, π a stochastic Gaussian policy, and γ=0.99. The discriminator's confidence is converted into an anomaly score 1−D, and thresholds are calibrated on healthy training scores. All three networks are two-layer MLPs with 64 hidden units and a total of 644 parameters, so scoring a test file amounts to one forward pass over a single tran","core_discovery":"On the paper's own terms, the discovery is that the temporal structure of degradation is the information that matters, and that structure is accessible without labels. By learning a reward over consecutive observation pairs from healthy data alone—rather than classifying individual segments as normal or faulty—AIRL produces an anomaly score that behaves differently from both reconstruction-error methods and contextual-bandit classifiers. The paper argues that baselines fail in two characteristic ways: reconstruction and one-class methods saturate, flagging nearly every file so that high detection rates are artifacts of false-alarm persistence, while the bandit formulation collapses to a stat","pith_inferences":["The paper's own deployment recommendation (Max−2SE from healthy scores) is not the threshold used for its headline consistency claim (a global adaptive threshold fitted to the full test distribution); reconciling this gap is necessary before the headline result transfers to practice.","Because the proxy action is deterministic (the next observation), the learned reward is effectively a transition-density estimate; comparing AIRL to a simple density model over consecutive pairs would isolate what the adversarial training adds.","A natural transfer test the paper does not run: train on healthy windows of one bearing run and score a different bearing run of the same type under varying load; the claimed invariance of the reward to dynamics predicts such transfer would work, and it is directly testable.","The PDC metric is vulnerable to a detector that triggers early and never switches off; the saturation flag (FAR>90%) is a heuristic patch, and a per-file alert-rate normalization or a precision-recall-style summary might be a more principled fix."],"forward_implications":["Label-free deployment: a prognostic system could be trained on healthy-only vibration data and score new files at runtime without fault labels or manual reward engineering.","The contextual-bandit critique generalizes: 'RL-based' fault detectors that ignore state transitions and discounting are methodologically equivalent to supervised classifiers and cannot be expected to track gradual degradation.","Saturation-aware evaluation: post-detection consistency must be read together with false-alarm rate; AUROC computed against a rigid onset boundary is structurally inappropriate for run-to-failure data.","Edge compatibility: the 644-parameter model and constant per-file inference cost make the approach compatible with controllers that cannot buffer full windows."],"fun_headline_variants":["AIRL learns fault reward from transitions, no labels","Label-free AIRL: only consistent fault detection across 3 rigs","From healthy data: AIRL recovers health reward, detects faults","Inverse RL for fault detection: no labels, no saturation"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The headline cross-dataset consistency claim assumes that a global threshold fitted to the entire test-score distribution is a valid measure of post-detection reliability; under the paper's own deployment-valid threshold (Max−2SE computed from healthy training scores alone), AIRL flags only a single IMS file across the whole test set, so the 'only method with non-saturated PDC on all three datasets' claim does not survive the paper's recommended deployment protocol.","fun_headline_variants_meta":{"raw":{"variants":["AIRL learns fault reward from transitions, no labels","Label-free AIRL: only consistent fault detection across 3 rigs","From healthy data: AIRL recovers health reward, detects faults","Inverse RL for fault detection: no labels, no saturation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001391,"raw_usage":{"total_tokens":5452,"prompt_tokens":718,"completion_tokens":4734,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":462,"completion_tokens_details":{"reasoning_tokens":4662}},"tokens_in":462,"tokens_out":4734,"duration_ms":32180,"temperature":1.0,"reasoning_tokens":4662,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T03:55:56.222643+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On the IMS bearing dataset, compute post-onset detection consistency for AIRL using the Max−2SE threshold calibrated strictly from healthy training scores (the paper's deployment recommendation) instead of a threshold fitted to the full test set. If the number of post-onset flagged files collapses from thousands to one, the claim of consistent non-saturated tracking across all three datasets fails.","supporting_citations":[],"review_version":1}