{"id":"23d724e5-ae4a-49dc-a82d-5ab33fcc49f6","arxiv_id":"2603.01735","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"KMT-2025-BLG-1314 and KMT-2025-BLG-1392 are modeled as binary-lens microlensing events with planet/brown-dwarf candidates, and HMC with the differentiable microlux code is shown to handle bimodal posteriors better than emcee.","lead":"This paper analyzes two new KMTNet microlensing events as binary-lens systems, finding planet/brown-dwarf companion candidates that remain degenerate between several interpretations. It also demonstrates that Hamiltonian Monte Carlo with a differentiable lensing model samples the bimodal posterior of one event more consistently than a standard ensemble MCMC sampler.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The HMC-vs-MCMC comparison in §5/Fig. 7 uses an untuned, short emcee run; the 'outperforms traditional MCMC' claim is therefore not established.","rationale":"The reader's weakest assumption is exactly the one I identify: the emcee run may not be a fair implementation of traditional MCMC, and the HMC-vs-MCMC comparison in §5/Fig. 7 is therefore not conclusive. This is the single most load-bearing concern because the paper's abstract and discussion frame HMC's superiority over traditional MCMC as a central result; if the comparison is unfair, that claim loses its evidential basis. The concrete test I propose is a longer, better-tuned, possibly tempered emcee run on the same posterior; this would directly settle whether the original emcee failure reflects a general property of traditional MCMC or merely a poor configuration. I agree with the reader's conditional verdict: the methodological claim is plausible but not fully established, and the physical results are accompanied by modeling choices (e.g., the log ρ cut) that also merit caution. However, since my concern matches the reader's stated weakest assumption rather than introducing a new one, the verdict should remain CONDITIONAL rather than being upgraded or downgraded.","tokens_in":16929,"tokens_out":5693,"duration_ms":53209,"concrete_test":"Re-run the Figure 7 comparison with emcee configured more fairly: e.g., 40 walkers, 50,000 warm-up steps, 50,000 sample steps, and optionally parallel tempering or an adaptive stretch proposal. Compute Rhat, effective sample size, and the marginal posteriors for log q, log s, and log ρ. If emcee then converges to the same posterior as HMC (Rhat < 1.1), the claimed HMC advantage is an artifact of the original emcee settings and the conclusion should be weakened. If emcee still fails to mix, the paper's claim survives this test.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central methodological claim—that HMC, enabled by microlux, outperforms traditional MCMC on the bimodal posterior of KMT-2025-BLG-1314—rests entirely on the comparison in §5 and Figure 7. There, emcee is run with 40 walkers, 2000 warm-up steps, and 4000 sample steps, with no indication of tuning, longer burn-in, or tempering. The resulting Rhat≈1.3 and poor mixing are then attributed to a general weakness of traditional MCMC. But this is not a fair or representative implementation: emcee is well known to require substantially longer burn-in and often tempering or adaptive strategies for multimodal posteriors. The HMC run, by contrast, uses NUTS with a Fisher-information reparameterization and a target acceptance of 0.9—a much more carefully preconditioned setup. The comparison does not isolate the sampling method from the implementation choices, so the abstract's 'outperforms traditional MCMC' is unsupported as a general statement. The event analyses may still be sound, but the headline methodological claim depends on this single, potentially unrepresentative benchmark. A second issue is the absence of effective-sample-size or wall-clock reporting, so 'outperforms' is inferred from Rhat alone.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents a light-curve analysis of two KMTNet microlensing events, KMT-2025-BLG-1314 and KMT-2025-BLG-1392, using the JAX-based differentiable code microlux and Hamiltonian Monte Carlo (NUTS). Both events show close/wide degeneracies; the first additionally shows planet/binary and point/finite degeneracies. The authors report mass-ratio estimates, reject the 1L2S hypothesis for both events (for the first using a μ_rel prior), and derive physical parameters via a Bayesian analysis. The paper's central methodological claim is that HMC 'outperforms traditional MCMC' on the bimodal posterior of KMT-2025-BLG-1314, based on a comparison with emcee in Fig. 7.","tokens_in":17346,"tokens_out":14651,"duration_ms":127076,"significance":"If the modeling results and the HMC comparison hold, the paper would be a useful demonstration that a differentiable microlensing code can handle real binary-lens events with multimodal posteriors, and it would add two candidate planetary/brown-dwarf systems. The paper is careful in presenting many degenerate solutions and in using Fisher-matrix reparameterization for HMC. The use of a public differentiable code (microlux) and the clear tables are strengths. However, the HMC-versus-emcee comparison, the 1L2S-rejection statistics, and a likely Jacobian error in Eq. (13) currently prevent the results from being fully accepted.","major_comments":[{"comment":"The claim that HMC 'outperforms traditional MCMC' rests on an emcee run with 40 walkers, 2000 warm-up and 4000 sample steps, with no tuning, tempering, or longer burn-in, and no effective-sample-size or wall-clock comparison. The observed Rhat≈1.3 for emcee shows non-convergence for this particular run, not a general property of ensemble MCMC. As the abstract's central methodological claim, this needs either a much more thorough benchmark (e.g., several emcee configurations, autocorrelation-time convergence, ESS/wall-clock) or substantial softening ('can be more robust in this instance').","section":"Section 5, Figure 7"},{"comment":"The rejection of the 1L2S model for KMT-2025-BLG-1314 uses a prior probability of 5e-4 converted to 'effective Δχ²~15' and added to the model Δχ²~10 to give total Δχ²~25. A tail probability is not a log-likelihood; adding it to Δχ² on the same scale is not a statistically justified model comparison. The physical-unplausibility argument is independent, but as written the combined Δχ² statement is unsupported. A proper computation would be a Bayes factor or full posterior predictive under a μ_rel prior.","section":"Section 3.2, 1L2S rejection"},{"comment":"The Jacobian transformation from (D_L, log M_L, μ_rel) to (M_L, θ_E, μ_rel) appears to be incorrect. Using π_rel ≈ au/D_L, the correct result is dΓ/(dM_L dθ_E dμ_rel) ∝ n D_L^4 θ_E^2 μ_rel^2 M_L^{-2} f_μ dξ/dlog M_L (up to constants), not M_L^{-1} as in Eq. (13). The extra M_L factor would bias the Bayesian mass estimates in Tables 6 and 7 toward higher masses. Please verify the derivation and rerun the analysis if needed.","section":"Section 4.2, Eq. (13)"}],"minor_comments":[{"comment":"The code is referred to as 'VBMicrolensing'; the standard name is 'VBBinaryLensing.' Please correct.","section":"Section 3.1"},{"comment":"The prior cut logρ∈[-4,-2] for unconstrained ρ is introduced without justification; please explain or quantify its effect on the derived physical parameters.","section":"Section 4.2"},{"comment":"The Δχ² between 'Planet Finite' and 'Planet Point' solutions is only ~2.5; the text calls the finite-source effect 'measurable.' Consider clarifying that the two classes are degenerate at this Δχ² level.","section":"Table 1, Section 3.2"},{"comment":"For emcee, 'two independent chains' should clarify that each chain is an ensemble of 40 walkers; also report the effective sample size and computation time for both methods.","section":"Figure 7 caption"}],"recommendation":"major_revision","confidential_remarks":"The paper is potentially useful, but the central methodological claim needs strengthening or softening, and the Eq. (13) error is a concrete technical issue. I would encourage the editor to send for revision rather than reject."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a careful, honest analysis of two new microlensing events, and it is the first real-event application of the microlux/HMC pipeline. That alone makes it worth a look. The event analyses are solid: thorough grid searches, a sensible treatment of the close/wide and planet/binary degeneracies, a third identified case of the 'Planet Point' sub-degeneracy, and a clean physical-property derivation with the standard Bayesian machinery. The tables with uncertainties are useful, and the authors are transparent about which solutions lack finite-source detections.\n\nThe main soft spot is the central methodological claim. The HMC-vs-MCMC comparison in §5 and Figure 7 rests on a single emcee run with 40 walkers, 2000 warm-up steps, and 4000 sample steps, with no evidence of tuning, tempering, or longer burn-in. The Rhat~1.3 for emcee and Rhat=1 for HMC is not a fair basis for saying HMC 'outperforms traditional MCMC methods' in general. The stress-test note is right: this compares a carefully preconditioned NUTS setup against an untuned, short emcee run. The comparison also lacks effective-sample-size or wall-clock reporting. I would not trust the abstract's wording as a general statement. What is demonstrated is that HMC can sample this particular bimodal posterior when initialized at different modes; that is a useful data point, but not a general superiority result.\n\nTwo smaller issues. For KMT-2025-BLG-1314, the 1L2S rejection uses a nonstandard conversion of a prior probability (5e-4) into an 'effective Δχ²' of ~15. That conversion is shaky, but the rejection also leans on a real Δχ²~10 and on the physically implausible near-contact binary geometry, so it is not load-bearing. The arbitrary logρ cut for the point/binary solutions is also a minor concern; it affects the derived physical parameters, and a sensitivity test would have been easy. They do not release data or code, which limits reproducibility, though the models themselves are standard.\n\nOn the whole, the event-level conclusions hold up: two candidate companions, the degeneracies mapped, physical parameters with honest uncertainties. The paper is for microlensing modelers, especially those who will face Roman/CSST-era data volumes. It deserves a serious referee, but the referee should push back on the HMC-vs-MCMC claim and ask for a fairer benchmark or a more carefully worded conclusion. I would take the event results as reliable; I would not take the 'outperforms' claim at face value.\n\nRecommendation: send it to peer review. It is a genuine first application, the events are new, and the methodological comparison is the kind of thing referees are supposed to interrogate.","headline":"Competent analysis of two new events, but the HMC-outperforms-MCMC claim in the abstract is not supported by the evidence they present.","tokens_in":17858,"tokens_out":2440,"would_cite":true,"duration_ms":25784,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper shows that Hamiltonian Monte Carlo, powered by a differentiable binary-lens model, robustly samples the bimodal posterior of a microlensing event where traditional MCMC gets stuck, and reports two new planet/brown-dwarf candidate","keywords":["gravitational microlensing","exoplanet detection","brown dwarfs","Hamiltonian Monte Carlo","Bayesian inference","degeneracy","binary lens","posterior sampling"],"falsifier":"Run a synthetic binary-lens event with a known bimodal posterior, sample it with both HMC and a traditional ensemble MCMC given substantially more steps and careful tuning, and check whether both converge to the same posterior; if the traditional sampler also converges, the claimed advantage of HMC in this setting is not general.","tokens_in":16866,"feed_emoji":"🪐","tokens_out":7067,"duration_ms":64989,"temperature":0.7,"pith_summary":"The paper analyzes two microlensing events, KMT-2025-BLG-1314 and KMT-2025-BLG-1392, whose light curves indicate a small companion to a lens star. Its central claim is methodological: when the posterior distribution is bimodal, Hamiltonian Monte Carlo sampling powered by a differentiable binary-lens model gives robust, reproducible parameter inference, while a traditional ensemble Markov chain Monte Carlo sampler becomes trapped in one mode. If true, gradient-based sampling is a practical tool for the many binary-lens events with close/wide, planet/binary, or point/finite degeneracies. The paper also finds that both events host planet or brown-dwarf candidates — mass ratios around log q ~ -3.5 and -1.3 respectively — and rejects the alternative single-lens binary-source interpretation for both.","feed_headline":"Gradient-based sampling beats MCMC on bimodal microlensing fits","feed_subtitle":"A differentiable lens model lets Hamiltonian Monte Carlo explore modes where traditional samplers stall.","key_machinery":"The key machinery is a differentiable binary-lens magnification model that computes accurate gradients of the light curve, enabling Hamiltonian Monte Carlo. The sampling is preconditioned by an information-matrix-based reparameterization: a triangular affine transformation of the latent parameters that approximates the local covariance and acts as a mass matrix, letting the HMC chains move efficiently across the strongly correlated posterior. An adaptive contour-integration error estimator keeps both the magnification and its derivatives accurate for planetary and extreme-binary configurations.","core_discovery":"The core discovery is that HMC, enabled by a differentiable binary-lens magnification model, can navigate the bimodal posterior of KMT-2025-BLG-1314 and converge to the same distribution from different starting modes, whereas an ensemble MCMC sampler with 40 walkers fails to mix and produces chain-dependent posteriors (standard convergence diagnostic ~1.3 versus 1.0). This is presented as the first application of differentiable modeling to real binary-lens microlensing events. The analysis identifies ten viable 2L1S solutions for KMT-2025-BLG-1314 — four planetary and six binary — including newly recognized 'Point' planetary solutions, and close/wide solutions for KMT-2025-BLG-1392 with a co","pith_inferences":["The same differentiable-model pipeline should transfer to other microlensing degeneracies, e.g., parallax versus xallarap or binary-source vs. binary-lens, where multi-modal posteriors are common.","If 'Planet Point' solutions prove common, previously published planet/binary events may need to be re-examined for missed point-source planetary solutions — a direct extension of the paper's finding.","The information-matrix reparameterization itself could be used to design observing strategies, since it reveals which parameter combinations are best constrained.","With next-generation surveys expected to deliver thousands of microlensing events, the computational cost of gradient-based sampling may make it the default, but its robustness on posteriors with more than two modes remains to be demonstrated."],"forward_implications":["HMC becomes a practical option for the large fraction of binary-lens events whose posteriors are multimodal, reducing the risk of mode trapping and yielding reproducible uncertainties.","The 'Planet Point/Finite' sub-degeneracy is now seen in a third event, so future planet/binary analyses should explicitly search for both point-source and finite-source planetary solutions.","For KMT-2025-BLG-1314, the 1L2S explanation is strongly disfavored statistically and physically, so the planet/binary interpretation remains viable until resolved by high-resolution imaging.","For KMT-2025-BLG-1392, the companion lies near the planet/brown-dwarf boundary; close and wide solutions remain nearly degenerate, so the projected separation is uncertain.","The 'Planet Finite' solutions for KMT-2025-BLG-1314 predict a low relative proper motion (~1 mas/yr), testable with future high-resolution imaging."],"fun_headline_variants":["Differentiable lens code lets HMC beat MCMC on bimodal fits","HMC with microlux finds modes MCMC stalls on","New microlensing solutions from gradient-based sampling","Microlensing degeneracy broken by Hamiltonian Monte Carlo","First differentiable binary-lens analysis unveils ten solutions"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The claim that HMC outperforms traditional MCMC assumes the traditional sampler was given a fair and adequately tuned run; the comparison in the paper uses one specific configuration, and a different tuning could narrow or erase the gap.","fun_headline_variants_meta":{"raw":{"variants":["Differentiable lens code lets HMC beat MCMC on bimodal fits","HMC with microlux finds modes MCMC stalls on","New microlensing solutions from gradient-based sampling","Microlensing degeneracy broken by Hamiltonian Monte Carlo","First differentiable binary-lens analysis unveils ten solutions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000152,"raw_usage":{"total_tokens":1065,"prompt_tokens":795,"completion_tokens":270,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":539,"completion_tokens_details":{"reasoning_tokens":189}},"tokens_in":539,"tokens_out":270,"duration_ms":3030,"temperature":1.0,"reasoning_tokens":189,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T19:31:27.780592+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a synthetic binary-lens event with a known bimodal posterior, sample it with both HMC and a traditional ensemble MCMC given substantially more steps and careful tuning, and check whether both converge to the same posterior; if the traditional sampler also converges, the claimed advantage of HMC in this setting is not general.","supporting_citations":[],"review_version":1}