{"id":"594dc7a2-e599-4006-a305-c2bb91bbfb72","arxiv_id":"2510.17459","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Warm-starting parallel-tempered MCMC with flow-matching posterior proposals infers β Pictoris b's orbit about 78-365× faster than conventional samplers with comparable posteriors, though the comparison is not fully apples-to-apples.","lead":"This paper proposes FM-MCMC, a hybrid that uses a flow-matching neural network trained on 16 million simulated orbits to seed a parallel-tempered MCMC sampler, and applies it to infer the orbit of β Pictoris b. The authors report a 78-365× speedup over standard samplers while matching their posterior estimates, but the comparison omits training time and several methodological checks.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported mean log-likelihoods (−133 vs −285 vs −234) are mutually incompatible for samplers targeting the same posterior; this is internal evidence that FM-MCMC's 1%-burn-in warm start is not converged, so the 'comparable accuracy' claim is unsupported.","rationale":"The reader's weakest assumption focused on synthetic training-set fidelity and the possibility of biased warm-started chains. That is a valid concern, but Table 2 provides a more direct and internal problem: the reported mean log-likelihoods are mutually inconsistent if all three samplers are targeting the same posterior. A converged MCMC estimator of E[ln L] should give the same value for any sampler; the ~150-nat gap between FM-MCMC and PTMCMC, and the ~51-nat gap between PTMCMC and nested sampling, cannot be sampling noise. This suggests that the FM-MCMC chain has not burned in and is still stuck near the flow-initialized high-likelihood region. The paper's use of 'highest average log-likelihood' as a success metric is therefore misleading, and the claimed 'statistical consistency' with PTMCMC/nested sampling is directly contradicted by the paper's own table. I agree with the reader that more diagnostics are needed, but I regard this as more than a missing check: unless the authors can show the discrepancy is a definitional artifact or rerun with adequate burn-in and recompute the speed comparison on an equal footing, the central claim fails. Hence I would move the verdict from CONDITIONAL to REJECT, conditional on the proposed test confirming the discrepancy.","tokens_in":16714,"tokens_out":9938,"duration_ms":88230,"concrete_test":"Compute the sample mean of the log-likelihood over the 10,000 retained FM-MCMC samples and over a long, cold-started PTMCMC chain run to formal convergence (Gelman-Rubin < 1.01 for all eight parameters, >25% burn-in). If the two means differ by more than, say, 5 nats (the paper reports ~150 nats), the FM-MCMC samples are not draws from the same posterior, and the consistency/speedup claim fails. This direct check uses only the paper's own likelihood and priors.","verdict_should_be":"REJECT","load_bearing_attack":"Central claim: FM-MCMC yields β Pictoris b posteriors statistically consistent with PTMCMC/nested sampling while being 77.8×/365.4× faster. A necessary condition is that FM-MCMC's output is a set of draws from the target posterior. Table 2 reports the sample mean log-likelihood as −133.0 (FM-MCMC), −285.4 (PTMCMC), −234.3 (nested). For any two correctly converged samplers targeting the same distribution, the Monte Carlo means of a fixed function of the parameters must agree to within sampling error (the max log-likelihoods in the same table are essentially identical, ≈−129.6). A ~150-nat gap between FM-MCMC and PTMCMC, and a ~51-nat gap between PTMCMC and nested sampling, cannot be reconciled with all three outputting posterior samples. It indicates at least the FM-MCMC chains are still concentrated near the flow-initialized high-likelihood region after the 1% burn-in; the 'higher average log-likelihood' is an artifact of non-stationarity, not a demonstration of sampling efficiency. The paper provides no trace plots, effective sample sizes, Gelman-Rubin statistics, or coverage checks on real data, so the central accuracy claim is not established. Note also that the speed comparison excludes the 22-minute training time and the warm-vs-cold comparison, but the convergence issue is more load-bearing: if the FM-MCMC samples are not from the posterior, both the speedup and consistency claims fail.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FM-MCMC, a two-stage inference method for orbital parameter estimation of directly imaged exoplanets. First, a continuous normalizing flow is trained via flow matching on 16 million simulated astrometric datasets drawn from priors tailored to β Pictoris b. Second, the trained flow generates initial proposals that are injected into a parallel-tempered MCMC sampler, with the claim that burn-in can be reduced to 1% of the traditional requirement. Applying this to β Pictoris b, the authors report posterior distributions consistent with PTMCMC and nested sampling (Fig. 3), a 77.8× speedup over PTMCMC and 365.4× over nested sampling, and the highest mean log-likelihood among the three methods (Table 2). The paper also compares against NPE and concludes that FM-MCMC yields more concentrated posteriors.","tokens_in":17176,"tokens_out":5561,"duration_ms":49230,"significance":"The application is timely: direct-imaging orbit fitting is computationally expensive, and a learned-proposal plus MCMC hybrid is a plausible route to acceleration. The paper does train a large simulation set, uses a physically motivated Keplerian simulator via orbitize!, and provides a P-P calibration plot for the flow (Fig. 4). If the central claims held, the method would be a useful contribution to exoplanet orbit fitting and to the broader learned-MCMC literature. However, the current manuscript does not establish equivalence of the FM-MCMC posterior with reference samplers: Table 2 contains an internal inconsistency in reported mean log-likelihoods that cannot be reconciled with all three samplers targeting the same posterior, and no convergence diagnostics are provided. The significance of the contribution therefore hinges on whether the FM-MCMC chains are actually converged; as presented, this is not demonstrated.","major_comments":[{"comment":"For any two correctly converged samplers targeting the same posterior, the Monte Carlo mean of a fixed function (here log-likelihood) must agree to within sampling error. Table 2 reports mean log-likelihoods of -133.0 (FM-MCMC), -285.4 (PTMCMC), and -234.3 (nested sampling), with maxima all near -129.6. These differences, of order 50-150 nats, are wildly incompatible with the hypothesis that all three outputs are draws from the same posterior. The likely explanation is that the FM-MCMC chain has not converged after the claimed 1% burn-in and remains in the high-likelihood region seeded by the flow. Please provide trace plots, effective sample sizes, and Gelman-Rubin statistics for every parameter and sampler, and recompute all likelihood statistics on properly thinned, post-burn-in samples. The '53.4% better' and '43.2% better' statements in the text are not meaningful for negative log-l","section":"Section 2.2, Table 2"},{"comment":"The speedup comparison is not apples-to-apples. FM-MCMC's reported 348.9 s excludes the 22-minute flow training time (Sec. 5.2) and uses warm starts produced by the trained flow, whereas PTMCMC and nested sampling are run from cold starts. Including the one-time training for a single analysis gives about 1669 s, which changes the speedups from 77.8×/365.4× to roughly 16×/76×. If the training is intended to be amortized over many objects, state that assumption explicitly. Please also report the number of likelihood evaluations, the exact burn-in lengths, and whether the PTMCMC/nested-sampling settings were identical (temperature ladder, number of walkers, stopping criteria) for all three methods.","section":"Sections 2.2 and 5.2"},{"comment":"The training priors are tightly centered on the known β Pictoris b solution: i is restricted to 81°-99°, Ω to 25°-85°, π to N(51.44, 0.12) mas, and M_T to N(1.75, 0.05) M_sun. The posterior medians for π and M_T in Table 3 essentially coincide with the prior means, indicating that part of the reported agreement reflects prior information rather than data information. The claim in Sec. 3 that the flow is 'trained once on simulated astrometric datasets covering extensive parameter spaces' and can work 'with limited prior knowledge' is not supported by these narrow, system-specific priors. Please state the noise scales ϵ_RA and ϵ_Dec used in the training set, report a sensitivity test with wider priors, or at minimum remove the over-general claim.","section":"Section 5.1, Table 1, Section 3"},{"comment":"The P-P plot in Fig. 4 validates the calibration of the flow network on simulated training data; it does not validate the full FM-MCMC pipeline on real observations. The assertion that burn-in is 'reduced to 1% of the original requirement' is asserted without any quantitative definition or diagnostic support. Please provide simulation-based calibration or coverage checks for the full FM-MCMC algorithm on held-out synthetic datasets, and report whether the real-data run passes any such consistency test. Without this, the 'comparable accuracy' claim rests on the qualitative overlap in Fig. 3 only, which is contradicted by the Table 2 likelihood discrepancy.","section":"Fig. 4 and Section 4.2"}],"minor_comments":[{"comment":"Text reads 'up to a 77.8× speedup over nested sampling, and a 365.4×speedup over nested sampling.' The first speedup is over PTMCMC according to the Abstract and Table 2; this is a typo.","section":"Introduction, last paragraph"},{"comment":"Typo: '12sisting' should be 'consisting'; 'Kapler's law' should be 'Kepler's law'.","section":"Section 5.1"},{"comment":"Percentage improvements of negative log-likelihoods are not interpretable; report raw differences in nats or use a proper accuracy metric such as expected log posterior score.","section":"Section 2.2"},{"comment":"The sine prior for inclination should be defined explicitly, e.g., p(i) ∝ sin i over [81°,99°], so the reader knows the normalization.","section":"Table 1"},{"comment":"Section 4.2 says the network is trained on O(10^6) simulated observations, while Section 5.1 says 16 million. These should be made consistent.","section":"Section 4.2 vs 5.1"}],"recommendation":"major_revision","confidential_remarks":"The mean-log-likelihood inconsistency in Table 2 is the most serious issue. If the FM-MCMC chain is not converged, both the accuracy and speedup claims fail. I would not recommend acceptance until the authors provide rigorous convergence diagnostics (trace plots, ESS, Gelman-Rubin) and an apples-to-apples timing comparison that includes training time. The flow+MCMC hybrid itself is not novel relative to existing work (e.g., flowMC, Refs. [57]-[59]); the contribution is mainly the application to direct-imaging orbit fitting and the claimed speedup. That application could be publishable, but only if the convergence issue is resolved."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nYou should know two things before spending time on this paper. The core idea—using flow-matching posterior estimation to initialize a PTMCMC sampler—is a reasonable and potentially useful trick for orbit fitting with direct imaging astrometry. The second is that the paper's central quantitative claim, '77.8×/365.4× faster with comparable accuracy,' is not supported by the paper's own numbers.\n\nThe new content: a continuous normalizing flow trained via flow matching on 16 million simulated β Pic b-like astrometric datasets provides starting points for PTMCMC. The posterior medians and contours in Fig. 3 look similar to those from PTMCMC and nested sampling, and the P-P plot in Fig. 4 is a decent calibration check. If the method holds up, it could save real time in survey-scale analyses.\n\nBut there is a hard problem. Table 2 reports mean log-likelihoods of −133.0 for FM-MCMC, −285.4 for PTMCMC, and −234.3 for nested sampling. FM-MCMC and PTMCMC are both supposedly sampling the same posterior, so the Monte Carlo mean of the log-likelihood must agree within sampling error. A 152-nat gap cannot happen if both chains have converged. The max log-likelihoods are nearly identical, so all methods find the same high-likelihood region. That pattern says FM-MCMC's samples are still clustered near the flow's initial high-likelihood region after the 1% burn-in; the 'higher average log-likelihood' is not a sign of efficiency, it's a sign that the chain has not mixed out. No trace plots, effective sample sizes, or Gelman-Rubin statistics are provided to counter that.\n\nThere are also comparison problems. The 348.9 s runtime excludes the 22-minute training time, compares a warm start against cold starts, and there is no benchmark against flowMC—which is the same normalizing-flow-plus-MCMC idea. The priors are centered on the known β Pic b values, so for π and M_tot the posteriors essentially reproduce the prior means. And there is no released code or data, just 'available upon request.'\n\nThe paper is not nonsense; the idea is worth exploring and the visual posterior agreement is encouraging. But the accuracy claim in the abstract is not established, and likely false as measured. Someone working on SBI for orbit fitting will want to read it with a skeptical eye. I would send it to peer review—a referee could force a proper comparison to flowMC, convergence diagnostics, and a corrected runtime accounting—but the headline numbers should not be cited as is.\n\nRecommendation: engage, but do not take the speedup/accuracy claim at face value.","headline":"Sensible hybrid sampler idea, but the headline speedup/accuracy claims are undercut by the paper's own inconsistent likelihood numbers.","tokens_in":17672,"tokens_out":4760,"would_cite":false,"duration_ms":38670,"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":"A neural-network warm start makes exoplanet orbit fitting up to 365 times faster while reproducing full posteriors.","keywords":["exoplanet orbit fitting","direct imaging","flow matching","Markov chain Monte Carlo","normalizing flows","Bayesian inference","Beta Pictoris b","simulation-based inference"],"falsifier":"Take the same β Pictoris b astrometric data, remove the flow warm start (so PTMCMC starts from random draws), and compare the resulting posteriors to the FM-MCMC ones; if they differ by more than the Monte Carlo noise at identical chain lengths, the flow warm start is not just accelerating but also changing the target. Alternatively, apply FM-MCMC to a synthetic dataset with non-Gaussian noise (e.g., Student-t) and check whether posterior coverage falls below the nominal 68%.","tokens_in":16628,"feed_emoji":"🪐","tokens_out":3821,"duration_ms":31206,"temperature":0.7,"pith_summary":"The paper claims that a hybrid sampler—a flow-matching neural network that proposes starting parameters, followed by a tempered MCMC chain—infers the orbital elements of directly imaged exoplanets just as accurately as standard Bayesian samplers while running tens to hundreds of times faster. Applied to β Pictoris b, the method reaches posteriors consistent with parallel-tempered MCMC and nested sampling in about 5.8 minutes instead of 7.5 or 35.4 hours, and reports a higher mean log-likelihood. The authors argue this makes the approach suitable for the large data volumes expected from upcoming surveys and for daily orbit updates during observing campaigns.","feed_headline":"Neural warm start makes orbit fitting 365x faster","feed_subtitle":"Flow-matching network seeds MCMC to recover Beta Pictoris b's full posterior in under six minutes, matching hour-long samplers.","key_machinery":"Flow-matching posterior estimation (FMPE) trains a continuous normalizing flow—an invertible neural network parametrizing a time-dependent velocity field that transports a simple base distribution to the posterior—by regressing on conditional velocity targets rather than backpropagating through an ODE solver. The trained flow is then used as the proposal and initialization for a parallel-tempered MCMC sampler, reducing burn-in to about 1% of its normal length. The key work the flow does is to place the MCMC chains near the high-probability region of parameter space immediately, so the sampler spends its iterations refining rather than exploring.","core_discovery":"The central claim is that a proposal distribution learned by flow matching can be injected into a parallel-tempered MCMC sampler to eliminate most of its burn-in, so that the sampler converges to the true posterior in a fraction of the wall-clock time. For the single-planet direct-imaging case of β Pictoris b, the flow network is trained on 16 million simulated astrometric observations drawn from Keplerian orbits with Gaussian noise; its outputs seed a PTMCMC chain with 20 temperature tiers and 1,000 chains. The resulting posterior medians and credible intervals for all eight orbital parameters (semi-major axis, eccentricity, inclination, argument of periastron, longitude of ascending node,","pith_inferences":["If the flow's proposal quality degrades as the true parameters move outside the synthetic training distribution (e.g., a wider prior or different noise model), the apparent convergence after a 1%-burn-in could mask bias; a useful test would be to run FM-MCMC with the prior ranges widened substantially and compare against a fully converged PTMCMC run.","The framework generalizes to any likelihood that is cheap to evaluate; the same flow-seeding trick should carry over to radial-velocity + astrometry joint fits or multi-planet systems, where the proposal distribution would need to encode the stronger degeneracies.","The reported likelihood gain may partly reflect the narrow, system-centered priors used in training; on a truly unknown system, the amortized flow would provide less of an advantage, so the speed-up is likely prior-dependent.","One could test the claim that burn-in shrinks to 1% by measuring the autocorrelation time and the Gelman-Rubin statistic as a function of chain length on the real data; if the chains are not at stationarity, the posteriors, though overlapping the baselines here, might diverge for other systems."],"forward_implications":["Posterior distributions for β Pictoris b's eight orbital parameters are reproduced within the same 1σ credible regions as PTMCMC and nested sampling, so claims made from the fast posteriors carry the same statistical weight.","Orbit fits that once took 7.5 hours (PTMCMC) or 35.4 hours (nested sampling) can be done in under 6 minutes on a consumer-class CPU/GPU setup, enabling same-day updates during observing runs.","With a mean log-likelihood higher than either baseline while the maximum likelihood matches, the sampler explores at least as well as the baselines, suggesting the speed-up does not come at the cost of missing high-probability regions.","The method is targeted to single-planet direct-imaging systems, the dominant confirmed population, so it applies immediately to most known directly imaged planets, not just β Pictoris b."],"fun_headline_variants":["Flow-matching MCMC cuts exoplanet orbit fit time 365x","Neural net seeds MCMC: β Pic b orbits in minutes, not hours","AI-accelerated MCMC: 365x speedup for exoplanet orbits","Deep generative MCMC: orbit parameters 365x faster","Flow matching + MCMC: exoplanet orbit estimation 365x speedup"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The trained flow network gives an unbiased warm start for the real β Pictoris b data, which depends on the synthetic training set—Keplerian orbits plus Gaussian noise calibrated to that star—and on the prior ranges being wide enough to contain the true parameters; if the real noise or astrometric calibration differs from the simulation, the fast chains could converge to a biased posterior.","fun_headline_variants_meta":{"raw":{"variants":["Flow-matching MCMC cuts exoplanet orbit fit time 365x","Neural net seeds MCMC: β Pic b orbits in minutes, not hours","AI-accelerated MCMC: 365x speedup for exoplanet orbits","Deep generative MCMC: orbit parameters 365x faster","Flow matching + MCMC: exoplanet orbit estimation 365x speedup"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000177,"raw_usage":{"total_tokens":1133,"prompt_tokens":751,"completion_tokens":382,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":495,"completion_tokens_details":{"reasoning_tokens":281}},"tokens_in":495,"tokens_out":382,"duration_ms":3771,"temperature":1.0,"reasoning_tokens":281,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T09:01:45.385900+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the same β Pictoris b astrometric data, remove the flow warm start (so PTMCMC starts from random draws), and compare the resulting posteriors to the FM-MCMC ones; if they differ by more than the Monte Carlo noise at identical chain lengths, the flow warm start is not just accelerating but also changing the target. Alternatively, apply FM-MCMC to a synthetic dataset with non-Gaussian noise (e.g., Student-t) and check whether posterior coverage falls below the nominal 68%.","supporting_citations":[],"review_version":1}