{"id":"cd1553d4-126f-4687-bcc5-f60abe70f65a","arxiv_id":"2611.17038","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"DINGO-TGR uses neural posterior estimation plus importance sampling to accelerate multiparameter gravitational-wave GR tests by over an order of magnitude, reproducing bilby-tgr posteriors and finding both GW150914 and GW250114 consistent with GR.","lead":"A neural-network method runs multiparameter tests of Einstein's relativity on gravitational wave data about 200 times faster than standard Bayesian sampling, and matches the slow method's answers on two real black hole mergers. An apparent small deviation from GR in GW150914 turns out to be consistent with noise (p=11.5%), and the higher-SNR event GW250114 is cleanly consistent with GR.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"GW250114's 0.041% importance-sampling efficiency and ESS-filtered calibration leave the proposal-coverage assumption under-tested in the low-ESS regime; direct bilby-tgr agreement nonetheless supports the two-event claim.","rationale":"I read the central claim in good faith: DINGO-TGR with IS correction reproduces bilby-tgr/dynesty PCA posteriors and delivers an order-of-magnitude wall-clock speedup. The direct comparisons for GW150914 and GW250114 are strong, with eigenvector angles below 1.5 degrees and PC1 medians agreeing within the quoted intervals; the 1000-injection null test is honestly reported with p=11.5%, and the authors explicitly acknowledge nonstationary-noise and waveform-modeling limitations (b284, b326). The remaining load-bearing assumption is proposal coverage for the target posterior (b114). The reader identified the same assumption, and I partially agree: the low ESS for GW250114 and the ESS>100 filtering of calibration are genuine gaps. I do not fully adopt the reader's reading of the three KS p-values, because three values near 0.05 among 21 tested parameters are plausible sampling fluctuations under good calibration, whereas the structural gap is the absence of a realistic-noise coverage check at the low-ESS operating point. Because this concern is real but addressable, and because the direct event comparisons already provide substantial support for the central claim, the reader's CONDITIONAL verdict is appropriate; my stress test does not move it.","tokens_in":25640,"tokens_out":8762,"duration_ms":65671,"concrete_test":"Generate 20 GW250114-like injections in realistic Gaussian noise, run DINGO-IS at the event's 2.5-million-proposal operating point without imposing an ESS>100 filter, and compare each weighted PC1 posterior median and 90% interval with an independent bilby-tgr/dynesty run on the same data; if the DINGO-vs-bilby PC1 median differences exceed the quoted 90% interval widths in more than 2 of 20 cases, proposal coverage bias is real rather than a benign low-ESS artifact.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The method's central claim, that DINGO-TGR with importance-sampling correction reproduces bilby-tgr/dynesty posteriors in the PCA-compressed null test, is valid only if the neural proposal covers the 21-dimensional target posterior; the authors state this condition at b114. The supporting evidence is strongest where the proposal works well and weakest where it struggles. GW250114 has ESS 1037 from 2.5 million proposal samples, an efficiency of 0.041%, and with self-normalized weights a small number of large-weight samples can dominate the weighted mean and covariance that feed the PCA. The P-P calibration in Fig. 10b already shows KS p-values of 0.034 (phi12), 0.047 (phiJL), and 0.055 (tc), and all 500 calibration injections are described as having ESS>100. The 16+12 recovery-validation injections are zero-noise, so they do not exercise proposal coverage under realistic noise. This is not an inconsistency and the authors do not hide the coverage condition, but the central quantitative claim for GW250114 is certified least precisely in exactly the low-ESS regime where importance sampling is most sensitive to proposal mismatch. The direct event-level agreement with bilby-tgr is the main evidence that coverage held for these two data sets.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"The paper presents DINGO-TGR, a neural posterior estimation framework for multiparameter gravitational-wave tests of general relativity. It combines normalizing flows with importance sampling to jointly infer binary parameters and six PN phase deformations, then applies PCA to identify well-constrained linear combinations. The method is validated on zero-noise injections and applied to GW150914 and GW250114, showing agreement with conventional bilby-tgr analyses while reducing wall-clock time by orders of magnitude. A 1000-injection null test yields an empirical p-value of 11.5% for the GW150914 offset.","tokens_in":25871,"tokens_out":1457,"duration_ms":13592,"significance":"If the method holds, it offers a substantial computational speedup for multiparameter GR tests, enabling large-scale injection studies that are currently prohibitive. The combination of NPE with importance-sampling correction and PCA-based compression is a natural extension of existing work, but the validation on real events and the null test demonstrate practical utility. The paper is clearly written and the claims are appropriately qualified.","major_comments":[{"comment":"The importance-sampling efficiency for GW250114 is only 0.041% (ESS 1037 from 2.5M samples). The P-P calibration in Fig. 10b shows KS p-values of 0.034 (phi12), 0.047 (phiJL), and 0.055 (tc), indicating that the neural proposal does not fully capture these parameters. The central claim that DINGO-TGR reproduces bilby-tgr posteriors for real events rests on the proposal covering the target posterior, which is least well-certified for GW250114. The authors acknowledge the coverage condition (b114) and report the ESS, but the low efficiency and marginal calibration for several parameters mean that the agreement with bilby-tgr for GW250114 should be interpreted with caution. I recommend adding a discussion of how the low ESS might affect the weighted covariance estimate that feeds the PCA, and whether the agreement with bilby-tgr is robust to the choice of ESS threshold.","section":"§II.B, §III.B, Appendix A","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The BGR injection tests are all zero-noise, so they do not probe the proposal coverage under realistic noise conditions. The authors note this implicitly (the injections are described as zero-noise), but the validation of the method for real events would be strengthened by at least a few BGR injections with realistic noise. The current validation shows that the pipeline can recover injected deviations in the absence of noise, but does not test whether the importance-sampling correction works correctly when the proposal is imperfect and the likelihood is noisy. This is a limitation that should be stated more explicitly.","section":"§III.A, Table II, Fig. 3","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The 1000-injection null test uses a fixed GW150914-like configuration and fixed PSDs. The authors note that nonstationary noise and waveform uncertainties are not accounted for (b284). However, the test also uses the same trained NPE model for all injections, so the null distribution is conditional on the neural proposal's coverage holding for all 1000 noise realizations. The median ESS is ~894, which is reasonable, but the spread of ESS across injections is not reported. If some injections have very low ESS, their contribution to the null distribution may be unreliable. I suggest reporting the range of ESS values and checking whether the p-value changes when excluding low-ESS injections.","section":"§III.C, Fig. 7","authors_acknowledge":null,"authors_acknowledge_p":null}],"minor_comments":[{"comment":"The abstract states 'GW150914 exhibits a mild offset from GR in the leading PCA mode, whereas GW250114 is consistent with GR.' This is accurate, but it might be helpful to also state the p-value in the abstract for balance.","section":"Abstract","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The JSD threshold of 0.1 bits is adopted from Ref. [41]. It would be useful to briefly justify why this threshold is appropriate, or at least note that the results are not sensitive to small changes in the threshold.","section":"§II.C, Eq. (22)","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The lower panels of Figs. 4 and 5 compare PC1 medians and credible intervals with previous results. The figure captions could be more explicit about which previous result is shown (e.g., 'Kumar (arXiv:2511.16879)' for GW150914 and 'LVK (arXiv:2509.08099)' for GW250114).","section":"§III.B, Fig. 4 and Fig. 5","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The upper panel of Fig. 7 shows median PC1 vs median PC2 for the 1000 injections. It would be informative to also show the 90% credible interval widths or the ESS values for these injections, perhaps as a color scale.","section":"§III.C, Fig. 7","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The P-P plots show KS p-values for each parameter. For GW250114, several parameters have p-values below 0.05. The authors should comment on whether these are expected given the number of parameters tested (21), or whether they indicate a systematic issue.","section":"Appendix A, Fig. 10","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The paper uses 'bilby-tgr' and 'dynesty' as benchmarks. It would be helpful to state the version numbers of these codes used in the analysis.","section":"General","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The importance weights are defined as w_i = L(d|θ_i, H) π(θ_i|H) / q(θ_i|d). It might be worth noting that the proposal density q includes the synthetic phase sampling contribution, as described in Eq. (11).","section":"§II.B, Eq. (12)","authors_acknowledge":null,"authors_acknowledge_p":null}],"recommendation":"major_revision"},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"Hong,\n\nThis paper does what it says on the tin: it combines DINGO-IS neural posterior estimation with the TIGER multiparameter GR test framework, adds PCA compression, and demonstrates order-of-magnitude speedups on two real events plus a 1000-injection null test. The cross-validation against bilby-tgr/dynesty for both GW150914 and GW250114 is the strongest part of the paper — eigenvector angles below 0.4 degrees and PC1 medians agreeing within errors. That is real evidence that the method works where it matters.\n\nThe 11.5% p-value for the GW150914 offset is honestly presented and does not overclaim. The injection recovery tests (16 GR, 12 BGR) behave sensibly. The paper is clearly written and the limitations are stated, not hidden.\n\nThe soft spots are the ones the stress-test note flags. GW250114 has IS efficiency 0.041% (ESS 1037 from 2.5M samples), and the P-P calibration shows KS p-values of 0.034, 0.047, and 0.055 for phi12, phiJL, and tc. That means the proposal coverage is marginal for this event. The authors acknowledge the coverage condition at b114, and the direct agreement with bilby-tgr is the main evidence that it held, but the central quantitative claim is certified least precisely in the low-ESS regime. The circularity concern about PCA basis construction is real but minor — the paper uses injection-specific bases for the null test and a common basis for method comparison, which is standard practice.\n\nNo code release is a practical limitation for reproduction, but the paper is not claiming a code release.\n\nWho this is for: anyone doing multiparameter GR tests who wants to scale up injection studies. The method is ready to use if you have access to DINGO and the training infrastructure. The paper deserves a serious referee — the engineering is sound, the validation is adequate, and the claims are proportionate to the evidence.\n\nI would accept this for peer review. The GW250114 coverage issue needs a robustness check (e.g., repeat with more proposal samples or a different flow), but that is a revision point, not a rejection.\n\nBest,\n[your name]","headline":"DINGO-TGR is a solid engineering advance that makes multiparameter GR null tests practical at scale; the GW250114 coverage concerns are real but do not undermine the central claim.","tokens_in":26554,"tokens_out":573,"duration_ms":7230},"pith_extraction":{"msc":[],"pacs":["04.30.-w","04.80.Nn","07.05.Kf","95.55.Ym"],"model":"deepseek/deepseek-v4-flash","headline":"DINGO-TGR reproduces conventional multiparameter GR null-test posteriors while cutting inference time from hundreds of hours to under an hour.","keywords":["gravitational waves","tests of general relativity","neural posterior estimation","importance sampling","principal component analysis","post-Newtonian parametrization","TIGER","IMRPhenomXPHM"],"falsifier":"Run DINGO-TGR on a high-SNR event (e.g., GW250114) with a deliberately mismatched training prior that excludes the true binary parameters, and check whether the importance-sampling efficiency drops to near zero and the recovered PCA posteriors become biased relative to a conventional nested-sampling analysis; if the method still returns correct posteriors despite the mismatch, the proposal coverage assumption is not the weakest link.","tokens_in":28908,"feed_emoji":"🕳️","tokens_out":3001,"duration_ms":42648,"temperature":0.7,"pith_summary":"This paper demonstrates that a neural posterior estimation framework, DINGO-TGR, can perform multiparameter tests of general relativity on gravitational-wave signals in hours rather than weeks, by jointly inferring binary parameters and six post-Newtonian phase deformations, then using principal component analysis to isolate the two best-constrained combinations. The key result is that the framework reproduces the conventional Bayesian posteriors for both the first detected event GW150914 and the higher-signal-to-noise-ratio event GW250114, while cutting per-event inference time from roughly 200 hours to under an hour for GW150914 and to 4.4 hours for GW250114. A large injection campaign of 1000 GW150914-like signals shows that the mild offset from general relativity seen in GW150914's leading PCA mode is not statistically significant (upper-tail p = 11.5%), meaning noise fluctuations alone can explain it. The paper's central claim is that importance-sampling-corrected neural inference makes rapid, large-scale null tests of general relativity practical, enabling significance assessments that were previously too expensive.","feed_headline":"Neural pipeline cuts GR-test inference from weeks to minutes","feed_subtitle":"DINGO-TGR matches standard Bayesian GR-test results in 38 minutes and 4.4 hours for two real events.","key_machinery":"The central machinery is DINGO-TGR, which combines (i) a normalizing-flow neural posterior estimator trained on 3 million simulated BBH signals to produce a 20-dimensional proposal for the binary parameters and six PN deformations, (ii) synthetic phase sampling to augment the proposal to 21 dimensions, and (iii) importance-sampling reweighting to correct the proposal toward the true target posterior. After marginalizing over the 15 standard parameters, principal component analysis (PCA) diagonalizes the 6x6 covariance matrix of the PN deformation posteriors; only the leading and subleading modes (PC1 and PC2) pass the Jensen-Shannon divergence threshold of 0.1 bits and are retained for the GR null test. The importance-sampling correction is what makes the neural approximation reliable: it removes the flow's bias by weighting each sample by the ratio of the true likelihood times prior to the proposal density.","core_discovery":"The paper establishes that combining neural posterior estimation with importance sampling (DINGO-IS) can jointly infer 15 standard binary parameters and six post-Newtonian (PN) phase deformation parameters within the TIGER parametrization, using IMRPhenomXPHM waveforms that include spin precession and higher-order modes, and that the resulting posteriors agree with conventional bilby-tgr/dynesty analyses to within eigenvector angles below 1.5 degrees for the leading PCA mode. For GW150914, the 21-dimensional inference takes 38 minutes of wall-clock time (versus 177 hours for nested sampling); for GW250114 it takes 4.4 hours (versus 238 hours). The paper further shows that the observed offset of the leading PCA mode in GW150914 has an empirical upper-tail p-value of 11.5% when tested against 1000 GR injections with independent noise realizations, so the GR null hypothesis is not rejected at the 5% level.","pith_inferences":["The paper's reliance on zero-noise injections for validation means the proposal coverage under realistic noise is not directly tested; the 0.041% importance-sampling efficiency for GW250114 and the borderline KS p-values (0.034, 0.047, 0.055) for three parameters in the P-P calibration suggest that the flow may not fully capture the tails of the posterior under realistic noise, which could affect ","The PCA basis is constructed separately for each event and each injection, so the leading mode does not have a fixed physical interpretation across events; comparisons between events (e.g., GW150914 vs GW250114) are only meaningful if the basis is held fixed, which the paper does only for the real-event comparisons (using the bilby-tgr basis). The 1000-injection null test uses injection-specific b","The paper does not test sensitivity to resolvable nonzero PC2 displacements, as the authors acknowledge; future work should design injections that specifically probe the second PCA mode to establish whether the method can detect deviations that are orthogonal to the leading mode.","The restriction to the inspiral phase through 3.5PN order and the use of a single waveform model (IMRPhenomXPHM) means that waveform-modeling systematics are not quantified; a cross-validation with multiple waveform families (e.g., SEOBNRv4PHM, surrogate models) would be needed to assess whether the observed offsets or limits are robust against modeling uncertainties."],"forward_implications":["Multiparameter tests of general relativity can now be run on individual events in tens of minutes instead of hundreds of hours, making large-scale injection campaigns (thousands of simulations) computationally feasible for assessing statistical significance.","The leading PCA mode for GW150914 is not statistically significant (p = 11.5%), so the mild offset does not warrant a claim of new physics; noise fluctuations remain a plausible explanation.","For higher-SNR events like GW250114, the method yields tighter constraints that are fully consistent with GR, showing that the framework scales to louder signals without loss of calibration.","The amortized neural training means that once a model is trained for a given parameter region, it can be reused for many events in that region without retraining, further reducing the cost of population-level tests.","The same importance-sampling-corrected neural inference pipeline can be extended to longer signals (e.g., binary neutron stars, extreme-mass-ratio inspirals) and to other beyond-GR parametrizations, provided the proposal covers the target posterior."],"supporting_citations":[{"why":"Establishes the DINGO framework for neural posterior estimation and importance-sampling correction that DINGO-TGR extends.","marker":"[18, 19, 20]"},{"why":"Defines the TIGER parametrization of fractional PN phase deformations that DINGO-TGR adopts for the GR null test.","marker":"[8, 11]"},{"why":"Provides the IMRPhenomXPHM waveform model with spin precession and higher-order modes used for both training and inference.","marker":"[12]"},{"why":"Supplies the bilby inference library and dynesty nested sampler that serve as the conventional baseline for comparison.","marker":"[13, 14]"},{"why":"Introduces the PCA-based posterior compression method for multiparameter GR tests that DINGO-TGR applies to its neural posterior samples.","marker":"[41]"},{"why":"Provides an independent PCA analysis of GW150914 with which the paper compares its own results.","marker":"[56]"},{"why":"Provides the LIGO-Virgo-KAGRA analysis of GW250114 that the paper uses for qualitative comparison.","marker":"[57]"}],"fun_headline_variants":[],"cache_read_input_tokens":0,"weakest_assumption_plain":"The importance-sampling proposal must adequately cover the full 21-dimensional target posterior; if the neural flow fails to produce samples in regions where the true posterior has significant probability, the importance weights cannot correct the bias, and the inferred PCA modes and p-values would be unreliable.","fun_headline_variants_meta":{"raw":{"variants":[]},"model":"deepseek/deepseek-v4-flash","effort":"low","cost_usd":8.7e-05,"raw_usage":{"cost":8.67412e-05,"is_byok":false,"cost_details":{"upstream_inference_cost":8.67412e-05,"upstream_inference_prompt_cost":8.63772e-05,"upstream_inference_completions_cost":3.64e-07},"total_tokens":1156,"prompt_tokens":1154,"completion_tokens":2,"prompt_tokens_details":{"audio_tokens":0,"video_tokens":0,"cached_tokens":256,"cache_write_tokens":0},"completion_tokens_details":{"audio_tokens":0,"image_tokens":0,"reasoning_tokens":0}},"tokens_in":1154,"tokens_out":2,"cost_source":"openrouter_usage","duration_ms":1987,"image_pages":[],"temperature":1.0,"billed_attempts":1,"reasoning_tokens":0,"transport_provider":"openrouter","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-11-13T07:12:13.472447+00:00","model_set":{"reader":"deepseek/deepseek-v4-flash"},"pipeline_version":"pith-pipeline@v1.4.1-daily-deepseek-or-medium","input_kind":null,"input_chars":null,"input_total_chars":null,"passes_missing":null,"published_standing":null,"published_confidence":null,"objection_audit":null,"run_record":null,"basis":{"kind":"not_recorded","source":null,"chars_given":null,"chars_total":null},"receipt":null,"falsifier":"Run DINGO-TGR on a high-SNR event (e.g., GW250114) with a deliberately mismatched training prior that excludes the true binary parameters, and check whether the importance-sampling efficiency drops to near zero and the recovered PCA posteriors become biased relative to a conventional nested-sampling analysis; if the method still returns correct posteriors despite the mismatch, the proposal coverage assumption is not the weakest link.","supporting_citations":[{"cited_title":"Computationally efficient models for the dominant and sub-dominant harmonic modes of precessing binary black holes","cited_arxiv_id":"2004.06503","evidence_quote":"Provides the IMRPhenomXPHM waveform model with spin precession and higher-order modes used for both training and inference."},{"cited_title":"Mahapatraet al., Confronting general relativity with principal component analysis: Simulations and results from GWTC-3 events, Phys","cited_arxiv_id":null,"evidence_quote":"Introduces the PCA-based posterior compression method for multiparameter GR tests that DINGO-TGR applies to its neural posterior samples."},{"cited_title":"Accelerating parameter estimation for parameterized tests of general relativity with gravitational-wave observations","cited_arxiv_id":"2511.16879","evidence_quote":"Provides an independent PCA analysis of GW150914 with which the paper compares its own results."},{"cited_title":"Black Hole Spectroscopy and Tests of General Relativity with GW250114","cited_arxiv_id":"2509.08099","evidence_quote":"Provides the LIGO-Virgo-KAGRA analysis of GW250114 that the paper uses for qualitative comparison."}],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":1}