{"id":"6a9aae0a-8084-4438-b774-c0c3dac1513c","arxiv_id":"2607.16031","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A single pretrained tabular transformer reaches ~90% Macro F1 with 120 labels per contingency and matches a non-deployable oracle transfer model for unseen faults with 10 labels.","lead":"This paper tests a pretrained tabular foundation model (TabPFN) for pre-fault power-grid security assessment, reporting about 90% Macro F1 from only 120 labelled operating conditions per contingency on the IEEE 68-bus system. It also proposes a modified electrical-distance encoding that lets the same model match an oracle transfer-learning baseline for unseen contingencies using just 10 new labels.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Unseen-contingency claim rests on same-bus evaluation; truly novel fault locations are untested, and the few-shot oracle match is only partial.","rationale":"The sample-efficiency result (Macro F1 ~90% at 120 samples) is well supported by the experiments and is credible given TabPFN's strong prior; this part of the central claim is not the weakest link. The generalization claim, however, is the most load-bearing because it underpins the paper's promise of handling new contingencies without retraining. The reader's weakest assumption correctly identifies that the evaluation design conflates 'unseen contingency' with 'same-bus, different clearing line.' My analysis confirms this: EDC depends only on the faulted bus, so held-out contingencies at known buses are not truly novel from the model's perspective. Table IV shows the failure mode when label distributions diverge, and the aggregate metrics hide this. The few-shot result also has a secondary issue: the abstract claims a match with the oracle, but Table V shows the match is only on two of four metrics, and the sampling procedure for the 10 labels may be optimistic. These are concrete, addressable concerns rather than fatal flaws; they warrant a conditional verdict and the suggested test. Since the reader already arrived at CONDITIONAL, my independent stress-test agrees, so the verdict should be UNCHANGED.","tokens_in":16755,"tokens_out":6673,"duration_ms":66722,"concrete_test":"Perform a leave-one-bus-out evaluation: for each bus with contingencies, hold out all contingencies at that bus (e.g., buses 17 and 22), train the TFM with EDC/mod-EDC on all remaining buses, and evaluate zero-shot and few-shot performance on the held-out bus. Report Macro F1 and balanced accuracy separately for these truly novel locations. If performance drops significantly below the reported 0.88 average, the 'unseen contingency' generalization claim must be restricted to same-bus contingency variation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim has two parts: sample efficiency (credible) and generalization to unseen contingencies. The generalization experiments (Section IV-E) hold out one contingency at a time while keeping other contingencies at the same bus in the training context. Table IV shows that 20 of 22 contingencies share a faulted bus with at least one training contingency; only buses 17 and 22 have a single contingency. Since EDC is computed solely from the faulted bus (Eq. 6), a held-out contingency at a known bus has an identical EDC vector to same-bus training contingencies. The task therefore reduces to choosing among decision boundaries of electrically (and often topologically) similar faults, not to generalizing to a new location. The paper itself shows that when stable-label proportions differ among same-bus contingencies (e.g., bus 54 range 0.40, Macro F1 0.70), zero-shot EDC degrades sharply. The reported aggregate averages across all 22 contingencies are dominated by favorable same-bus cases, so the zero-shot 'unseen contingency' claim is not established for novel buses. The few-shot TFM+10 result is more robust because it injects 10 direct target labels, but the abstract overstates it: Table V shows TFM+10 matches JDA-best only on balanced accuracy and Macro F1, while specificity is substantially lower (0.84 vs 0.98) and precision is also lower (0.91 vs 0.94). Furthermore, the few-shot protocol may enforce a minimum of one sample per class (Section IV-A), which would make 10 random samples more informative than a truly random 10-sample TDS budget, especially for rare-class contingencies. The paper should either restrict the claim to 'unseen clearing lines at known buses' or provide evidence for truly novel locations.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes using TabPFN, a tabular foundation model, for pre-fault dynamic security assessment (DSA). The central claims are: (i) a single TFM, applied without retraining or hyperparameter tuning, achieves an average Macro F1 of about 90% with only 120 labelled operating conditions per contingency, roughly two orders of magnitude fewer than conventional databases; (ii) a single pooled TFM outperforms a multi-label DSA baseline across 22 contingencies; and (iii) for unseen contingencies, an electrical-distance-coordinate (EDC) encoding enables zero-shot generalization, and adding just 10 labelled samples of the new contingency with a modified EDC (mod-EDC) matches the best achievable transfer-learning oracle (JDA-best). The paper also characterizes when EDC fails and discusses computational costs.","tokens_in":17148,"tokens_out":6043,"duration_ms":60960,"significance":"The sample-efficiency result is significant and, if confirmed, directly addresses the TDS database-generation bottleneck in data-driven DSA. Using a fixed pretrained model (TabPFN) avoids circular training on the DSA labels, and the comparison against a non-deployable oracle is a useful reference point. The paper is also honest in reporting that EDC degrades when same-bus contingencies have very different stable-label proportions (Table IV). However, the generalization-to-unseen-contingencies claim, as currently stated, is not fully supported: the experiments mostly test held-out contingencies at buses that already appear in training context, and the few-shot 'matches JDA-best' claim is only true on two of the four reported metrics. The central idea is promising, but the evidence for the strongest conclusions needs to be either strengthened or re-scoped.","major_comments":[{"comment":"The zero-shot 'unseen contingency' experiment does not actually test generalization to unseen fault locations. Because EDC in Eq. (6) is computed only from the faulted bus, any held-out contingency at a bus that also appears among the 21 training contingencies has an EDC vector identical to same-bus training contingencies. The task then reduces to choosing among decision boundaries of electrically close—and often topologically identical—faults, not to extrapolating to a new location. Table IV shows precisely that when same-bus contingencies have different stable-label proportions (e.g., bus 54, range 0.40, Macro F1 0.70), zero-shot EDC degrades sharply. The aggregate averages across all 22 contingencies are dominated by these favorable same-bus cases; the truly unseen-bus cases (buses 17 and 22) are not reported separately. To support the advertised 'unseen contingency' generalization, t","section":"Section IV-E, Table IV, Eq. (6)"},{"comment":"The claim that 'using just 10 labelled samples of the new contingency with EDC encoding matches the best achievable transfer learning oracle model' is only partially correct. Table V shows TFM+10 equals JDA-best on balanced accuracy (0.88 vs 0.88) and Macro F1 (0.87 vs 0.84), but on specificity TFM+10 is 0.84 versus 0.98 for JDA-best, and on precision 0.91 versus 0.94. Since Section IV-B explicitly states that specificity is the most operationally critical metric for DSA, the current abstract and conclusion overstate the few-shot result by omitting this substantial gap. The claim should be qualified to the metrics on which parity actually holds.","section":"Table V and Abstract"},{"comment":"The few-shot protocol uses stratified sampling with a minimum of one sample per target class, as described in Section IV-A. This means the '10 additional labelled samples' in Section IV-E are not a random draw from the target distribution. For a contingency with only 1% insecure cases (e.g., contingency 5), the protocol guarantees an insecure example in the context, while a truly random 10-sample draw would provide one with probability roughly 0.1. Section IV-E states that samples are 'randomly sampled,' which is inconsistent with the stratified protocol. This constraint substantially changes the few-shot setting and may inflate the reported performance of TFM+10. Please clarify whether the minimum-1-per-class rule applies to the few-shot experiment and, if so, either report results under random sampling or qualify the 'just 10 labelled samples' claim.","section":"Section IV-A and IV-E"}],"minor_comments":[{"comment":"Figures 10 and 11 report means over four seeds without variance bars or confidence intervals. Given that the crossover between EDC and mod-EDC occurs at n=10, showing the spread across seeds would help assess whether this crossover is robust.","section":"Figures 10 and 11"},{"comment":"The definition Q_k = EDC_pre - EDC_post should clarify whether the same set of eight reference nodes is used for the post-fault network and whether the Z-bus matrix is recomputed after the line trip. This detail is needed to reproduce the mod-EDC encoding.","section":"Equation (8)"},{"comment":"The last paragraph of Section V says 'the proposed workflow positively answers the three questions posed in Section III-C,' but the three questions are posed at the beginning of Section IV, not in Section III-C.","section":"Section V"},{"comment":"The sentence 'TFM is the only model which correctly classifies up to 80% of all insecure cases on average when only 120 training samples are available' is slightly ambiguous; it should be explicit that this refers to average specificity across contingencies.","section":"Section IV-A"}],"recommendation":"major_revision","confidential_remarks":"The sample-efficiency result is solid and likely publishable, but the generalization claim is currently over-sold. The most important fix is to either add a leave-one-bus-out experiment or explicitly re-scope the 'unseen contingency' claims to 'unseen contingency at a known bus.' The few-shot claim also needs qualification on specificity/precision, and the sampling protocol should be clarified. No concerns about the citation pattern or novelty disclosure."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, what you should know: the core sample-efficiency result holds up. A single TabPFN with 120 labelled samples per contingency reaches roughly 90% Macro F1 on the 68-bus system, beating tuned XGBoost and the other baselines, with no retraining and no hyperparameter search. That is a real, practically relevant finding. The pooled multi-contingency setting also works well, and the robustness to severe class imbalance is well demonstrated. The mod-EDC encoding is a reasonable extension, and the paper honestly documents when EDC-based zero-shot generalization fails (Table IV), which is more than most papers do.\n\nThe soft spot is the unseen-contingency claim. Section IV-E holds out one contingency at a time while other contingencies at the same bus remain in the training context. Since EDC is computed from the faulted bus, a held-out fault at a known bus has an identical EDC vector to same-bus training faults. The task reduces to choosing among decision boundaries of electrically similar faults, not generalizing to a new location. Only two contingencies (buses 17 and 22) are truly novel bus locations, and the paper does not report them separately. The aggregate zero-shot numbers are dominated by same-bus cases, and the paper itself shows severe degradation when same-bus contingencies have different stable-label proportions (bus 54: range 0.40, Macro F1 0.70). So the 'generalizes to unseen contingencies' phrasing in the abstract overstates what is actually shown.\n\nThe few-shot TFM+10 result is more robust because it injects 10 direct target labels, but the 'matches the oracle' claim is only partial: Table V shows TFM+10 matches JDA-best on balanced accuracy and Macro F1, but specificity is substantially lower (0.84 vs 0.98) and precision is lower (0.91 vs 0.94). Also, the stratified sampling enforces at least one sample per class, which makes 10 samples more informative than a truly random 10-sample budget, especially for rare-class contingencies.\n\nOther issues are minor: the study is only on one 68-bus system, no code or data is provided, and there are no error bars (though the mean over four seeds is reported). These are all addressable.\n\nMy take: this deserves a serious referee. The sample-efficiency result is valuable and likely to hold. The generalization claim needs to be restricted to 'unseen clearing lines at known buses' or backed by a truly novel-bus experiment. I would engage with it and ask for the artifacts and a re-analysis of the unseen-bus cases.","headline":"A credible sample-efficiency result for TabPFN in pre-fault DSA, but the unseen-contingency claim is overstated: zero-shot evaluation mostly holds out faults at buses that remain in the training context.","tokens_in":17638,"tokens_out":2361,"would_cite":true,"duration_ms":23425,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A single tabular foundation model, used without retraining, can assess power-system dynamic security per contingency with roughly 120 labelled operating conditions and can adapt to an unseen fault with ten labelled samples, matching an impr","keywords":["dynamic security assessment","tabular foundation model","in-context learning","pre-fault DSA","electrical distance coordinates","contingency generalization","sample efficiency","power system stability"],"falsifier":"Hold out contingency 15, the only contingency at bus 17 with no same-bus partner, train on the other 21 contingencies with EDC encoding, and measure zero-shot Macro F1 on 12,000 samples. If the TFM does not clearly beat a label-proportion-based baseline, the 'unseen contingency as interpolation' story fails for the case the current experimental design avoids.","tokens_in":16706,"feed_emoji":"⚡","tokens_out":4457,"duration_ms":41644,"temperature":0.7,"pith_summary":"This paper argues that data-driven dynamic security assessment (DSA) of power grids does not need the large per-contingency labelled databases that current practice assumes. Using a tabular foundation model that classifies by in-context learning—reading labelled examples at inference time rather than updating weights—one model achieves about 90% Macro F1 across 22 contingencies with only 120 labelled operating points per contingency, about two orders of magnitude less data than baselines require. The same model, when given an electrical-distance-coordinate encoding of fault location, can generalise to a contingency it has never seen; adding just 10 labelled samples of that new contingency matches the best theoretically achievable transfer-learning oracle, which is not deployable in practice. If right, this would replace the workflow of training, tuning and maintaining a separate classifier for every credible fault, and make pre-fault security assessment more practical under limited simulation budgets.","feed_headline":"One grid model needs 120 labels per fault, not 12,000","feed_subtitle":"In-context learning lets a single tabular model replace per-contingency classifiers and adapt to new faults from 10 samples.","key_machinery":"The central object is the tabular foundation model operating by in-context learning: a transformer pretrained on many synthetic tabular datasets that, given labelled context rows and unlabelled query rows, outputs the Bayesian posterior predictive in one forward pass, with no parameter updates. Coupled to it is the scenario vector: EDC, an 8-dimensional electrical-distance coordinate of the faulted bus computed from the network impedance (Z-bus) matrix, and its extension mod-EDC, the concatenation of pre- and post-fault EDCs, which acts as a label-free proxy for how the clearing action reduces synchronising power. The EDC makes unseen contingencies continuous with training contingencies; mod","core_discovery":"The central discovery is that pre-fault DSA can be reframed as an in-context learning problem solvable by a single pretrained tabular foundation model. On the IEEE 68-bus test system with 22 three-phase bus-fault contingencies, the TFM—with no retraining and no hyperparameter tuning—matches or beats tuned gradient-boosting, MLP, decision-tree and SVM baselines at every training size from 12 to 12,000 samples per contingency, reaching about 90% average Macro F1 with 120 samples per contingency. For unseen contingencies, the paper shows that a continuous scenario vector matters: one-hot fault IDs fail because an unseen ID is orthogonal to training data, whereas electrical distance coordinates","pith_inferences":["The label-distribution dependence observed in the paper's same-bus table suggests a pre-screening rule the authors do not state: before trusting zero-shot EDC for a new contingency, estimate the new contingency's stable-label proportion; if it differs sharply from the nearest trained contingency, spend the 10 TDS samples.","The mod-EDC idea—using the change in electrical coordinates induced by a topology change as a label-free severity proxy—could be transferred to other security assessments (e.g., voltage instability or cascading failure) where post-contingency topology is known.","Because this class of tabular foundation model is pretrained once and reused across tasks, the same model instance could serve multiple operational tasks (pre-fault DSA, post-fault trajectory prediction, contingency screening) without separate training pipelines—a direction the paper mentions but leaves to future work.","A direct testable extension: repeat the leave-one-contingency-out study on a system with several buses having only one contingency, isolating the no-same-bus case that the current study leaves under-sampled; if zero-shot performance collapses there, the generalization claim needs qualification."],"forward_implications":["Operating utilities could shrink TDS database-generation budgets by roughly two orders of magnitude while keeping high accuracy, since 120 labels per contingency suffice.","A single model can replace per-contingency classifiers: the pooled TFM outperforms a multi-label baseline on balanced accuracy and Macro F1, and is robust for contingencies with very rare insecure cases.","For a new/unseen contingency, 10 labelled samples added to the context—without finetuning—give accuracy on par with an impractical oracle transfer-learning approach; this is an operationally realistic few-shot adaptation.","Zero-shot assessment of new contingencies becomes possible when the new fault's electrical-distance coordinate lies near training contingencies with similar stability-label proportions.","The approach's computational cost is dominated by a single forward pass, avoiding the scaling cost of per-contingency hyperparameter tuning."],"fun_headline_variants":["Single tabular model cuts DSA training labels to 120 per fault","One foundation model replaces per-contingency classifiers in DSA","120 labels per fault suffice for dynamic security assessment","In-context learning slashes DSA label needs to 120 per contingency","Ten samples adapt DSA model to new faults via electrical distance"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The claim about zero-shot generalization to unseen contingencies rests on the assumption that a held-out fault's dynamic behaviour is predictable from electrically close faults at the same bus—something the paper shows holds when label proportions are similar and fails when they differ (its Table IV).","fun_headline_variants_meta":{"raw":{"variants":["Single tabular model cuts DSA training labels to 120 per fault","One foundation model replaces per-contingency classifiers in DSA","120 labels per fault suffice for dynamic security assessment","In-context learning slashes DSA label needs to 120 per contingency","Ten samples adapt DSA model to new faults via electrical distance"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00077,"raw_usage":{"total_tokens":3293,"prompt_tokens":832,"completion_tokens":2461,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":576,"completion_tokens_details":{"reasoning_tokens":2373}},"tokens_in":576,"tokens_out":2461,"duration_ms":16007,"temperature":1.0,"reasoning_tokens":2373,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T21:32:36.047383+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Hold out contingency 15, the only contingency at bus 17 with no same-bus partner, train on the other 21 contingencies with EDC encoding, and measure zero-shot Macro F1 on 12,000 samples. If the TFM does not clearly beat a label-proportion-based baseline, the 'unseen contingency as interpolation' story fails for the case the current experimental design avoids.","supporting_citations":[],"review_version":1}