{"id":"e47e60bf-7922-4c83-b84a-3eff9c0c341a","arxiv_id":"2607.08755","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":3,"one_line_summary":"A pipeline combining mutual information screening with alternating symbolic regression recovers closed-form degeneracy equations from posterior samples.","lead":"DegenDetector is a software pipeline that automatically discovers the mathematical equations describing parameter degeneracies in Bayesian posterior distributions. Scientists can use it to find hidden relationships between fitted parameters that standard visualization tools miss entirely.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"The R²_perp diagnostic for flagging non-separable degeneracies is stated as fact but never tested; if it fails to catch false positives, the 'without domain-specific input' claim weakens.","rationale":"The reader correctly identified the separability assumption as the weakest point, and the CONDITIONAL verdict with MODERATE confidence is appropriate. My concern is a sharpening of the reader's: the specific load-bearing issue is not just that the method fails for non-separable degeneracies (which the authors acknowledge), but that the paper makes an untested claim about R²_perp serving as a reliable self-diagnostic for that failure. This is the method's only guardrail against silently returning wrong equations, and it is never validated. However, this does not change the verdict because: (a) the reader already flagged the lack of systematic failure-mode analysis, which encompasses this issue; (b) the benchmarks that are tested are genuinely successful with R²_perp > 0.98; (c) the Planck recovery, while imperfect (ratio ~2.947 vs expected ~3), is reasonable. The concern strengthens the case for addressing failure modes before full trust, which the CONDITIONAL verdict already captures. The paper's code being publicly available (GitHub) is a positive factor — the concrete test I propose could be run straightforwardly.","tokens_in":6983,"tokens_out":3460,"duration_ms":128951,"concrete_test":"Construct a synthetic posterior with a known non-separable degeneracy — e.g., θ1·θ2 + θ3 = c with Gaussian noise perpendicular to the manifold, plus 4 nuisance parameters drawn from an isotropic Gaussian. Run DegenDetector (both with and without LOGDEGEN). Check: (1) Does R²_perp drop below a useful threshold (e.g., 0.9)? (2) Does the method return a spurious separable equation with high R²_perp? If R²_perp remains high (>0.95) for a genuinely non-separable degeneracy, the self-diagnostic is unreliable and the central claim of automated, domain-agnostic detection is overstated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper claims in §2.2 that 'when true degeneracy is not separable, the orthogonal R² will return a low score, indicating that the functional form has not been captured.' This is the method's only safeguard against false positives — its ability to self-diagnose failure when the separable model (Eq. 2) is wrong. Yet all four benchmarks (S-curve, Banana, Trig, Cubic) and the Planck application involve degeneracies that are either directly separable or log-separable. There is no test case with a genuinely non-separable degeneracy (e.g., θ1·θ2 + θ3 = c, or sin(θ1·θ2) + θ3 = c) to verify that R²_perp actually drops. This matters because a separable function can approximate a non-separable one well over a bounded region of parameter space, potentially yielding a deceptively high R²_perp. If the diagnostic does not reliably flag non-separable degeneracies, the method could silently return incorrect equations, directly undermining the 'without domain-specific input' part of the central claim. The authors acknowledge separability as a limitation in the Conclusion but never test whether their stated mitigation (low R²_perp as a warning signal) actually works.","agreement_with_reader":"partial"},"referee_report":{"model":"glm-5.2","summary":"The manuscript introduces DegenDetector, a framework for identifying parameter degeneracies in Bayesian posteriors and expressing them as closed-form symbolic equations. The pipeline proceeds in two stages: (1) mutual information (MI) screening ranks k-tuples of parameters by their aggregate pairwise MI to pinpoint candidate degenerate subsets, and (2) alternating symbolic regression fits a separable level-set model g_1(theta_j1) + ... + g_k(theta_jk) = c using PySR, cycling through one-dimensional subproblems. The method is validated on four synthetic benchmarks with known ground-truth degeneracies (S-curve, Banana, Trig, Cubic) and applied to Planck 2018 posteriors, where it recovers the CMB horizon-angle degeneracy Omega_m h^3 ~ const in log-space with R^2_perp ~ 0.98. The approach is motivated by the limitation that corner plots reveal pairwise correlations but not the functional form of multi-parameter degeneracies.","tokens_in":7196,"tokens_out":1253,"duration_ms":185879,"significance":"The problem addressed is real and practical: high-order degeneracies that do not manifest in pairwise projections are a known challenge in cosmological inference, and automated recovery of interpretable closed-form expressions is a useful contribution. The strengths of the manuscript include the clean algorithmic design (MI screening to avoid combinatorial explosion, followed by alternating optimization that decomposes a k-dimensional problem into k one-dimensional symbolic regressions), the use of synthetic benchmarks with pre-defined ground truth, and the application to a well-established cosmological degeneracy. The public release of code on GitHub is a positive step for reproducibility. The central claim—that the pipeline can identify degenerate parameter subsets and recover their functional form without domain-specific input—is supported by the benchmarks, though its scope is constrained by the separability assumption discussed below.","major_comments":[{"comment":"The manuscript states that 'when true degeneracy is not separable, the orthogonal R^2 will return a low score, indicating that the functional form has not been captured.' This self-diagnostic property is the method's only safeguard against false positives when the separable model (Eq. 2) is misspecified, and it underpins the claim of operation 'without domain-specific input.' However, all four benchmarks (S-curve, Banana, Trig, Cubic) and the Planck application involve degeneracies that are either directly separable or log-separable. There is no test case with a genuinely non-separable degeneracy (e.g., theta_1 * theta_2 + theta_3 = c, or sin(theta_1 * theta_2) + theta_3 = c) to verify that R^2_perp actually drops as claimed. This matters because a separable function can approximate a non-separable one well over a bounded region of parameter space, potentially yielding a deceptively high","section":"§2.2 (Orthogonal Fit Quality) and §3"},{"comment":"The Planck recovery yields a coefficient ratio of 123.97/42.07 ~ 2.947 rather than the expected value of 3 for the Omega_m h^3 degeneracy. The manuscript describes this as 'reasonably consistent' but provides no error bars, uncertainty quantification, or goodness-of-fit assessment on the recovered coefficients themselves. Without a measure of statistical uncertainty on the ratio, it is unclear whether the deviation from 3 is within expected noise or indicates a systematic bias in the fitting procedure. The residual standard deviation sigma_perp = 0.128 in log-space is reported, but this characterizes the scatter of samples around the fitted plane, not the uncertainty on the coefficient ratio. Adding a bootstrap or jackknife estimate of the uncertainty on the recovered exponents would substantially strengthen the quantitative claim.","section":"§4, Eq. (4)"}],"minor_comments":[{"comment":"The MI-based ranking assigns each k-tuple a score equal to the sum of all pairwise MI values among its elements. For k > 2, this pairwise-sum score may not capture higher-order interactions that only manifest jointly (e.g., three parameters that are pairwise independent but jointly dependent). The authors should briefly acknowledge this limitation of the screening step.","section":"§2.1"},{"comment":"The figure captions and axis labels are difficult to parse in places. For instance, in panel (d), the equation label appears garbled ('2 2 / 1 + 2 / 2 3 = 0.5'), and in panel (f), the label reads '1 (10 2)3 0'. These should be corrected to clearly state the ground-truth degeneracy equations.","section":"Figure 1"},{"comment":"The epsilon in the denominator of the orthogonal loss is introduced without comment. A brief note on its role (numerical stabilization near critical points of F) and its typical magnitude would help reproducibility.","section":"§2.2, Eq. (3)"},{"comment":"The Trig benchmark tests whether the method can approximate periodic functions 'through Taylor expansions or Gaussian-envelope estimations.' It would be useful to state explicitly what functional form PySR actually returned for this case, since the recovered expression is not shown in the text.","section":"§3"},{"comment":"The sentence 'All code is publicly available on Github' should include the URL in the main text body (it appears only in a footnote) for visibility.","section":"§5 (Conclusion)"}],"recommendation":"minor_revision","confidential_remarks":"The stress-test concern about the untested R^2_perp diagnostic for non-separable degeneracies is the most substantive issue. It does not undermine the central claim for the cases tested, but it does mean the scope of the 'without domain-specific input' claim is untested for the failure mode the authors themselves identify as the primary limitation. I recommend requiring at least one non-separable negative-control benchmark before publication, but this is a local fix rather than a fundamental flaw. The Planck coefficient ratio issue is also worth flagging to the authors but is not load-bearing for the methodological contribution."},"author_rebuttal":null,"desk_editor":{"model":"glm-5.2","letter":"This paper presents DegenDetector, a pipeline that takes posterior samples and returns closed-form symbolic equations for parameter degeneracies. The core idea—use k-NN mutual information to screen candidate parameter tuples, then fit an implicit separable surface via alternating symbolic regression with PySR—is genuinely new as an integrated tool. The synthetic benchmarks have known ground truth, and the Planck 2018 application recovers the established Ω_m h³ ≈ const degeneracy without being told the physics. Code is public on GitHub. The orthogonal R² metric is a sensible choice for implicit surfaces, and the alternating optimization trick that reduces a k-dimensional fit to k one-dimensional symbolic regressions is clean and practical. Credit earned here: the pipeline works, it ships, and it addresses a real gap—corner plots show correlations but not functional forms, and high-order degeneracies don't show up in pairwise projections at all. The stress-test concern about R²_perp as a false-positive safeguard is the right thing to worry about. The paper states that non-separable degeneracies will produce low R²_perp, but every benchmark is separable or log-separable. There's no test case with a genuinely non-separable degeneracy (say, θ₁·θ₂ + θ₃ = c) to verify the diagnostic actually catches failure. A separable function can approximate a non-separable one well over a bounded region, potentially yielding a deceptively high R²_perp. This is the main soft spot, and it's real—the self-diagnostic is untested. The Planck coefficient ratio of ~2.947 vs. the expected 3 is minor; the paper notes the empirical literature finds ~3.4, and the fit is in log-space with finite samples. More bothersome is the absence of error bars on recovered coefficients and underspecified details like PySR operator sets and convergence criteria. These are fixable but matter for reproducibility. The separability assumption is acknowledged as the primary limitation, which is honest. The paper is for cosmologists and SBI practitioners who hit degeneracies and want interpretable equations rather than just corner plots. It's a methods paper with a working tool and clear validation. The missing failure-mode analysis is the one thing that genuinely weakens the central claim of operating 'without domain-specific input.' Recommend serious peer review. The referee should push for a non-separable test case and error quantification, but the approach is sound and the contribution is real.","headline":"Useful pipeline for symbolic degeneracy recovery, but missing failure-mode tests","tokens_in":7700,"tokens_out":565,"would_cite":false,"duration_ms":78818,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"glm-5.2","headline":"Pipeline extracts degeneracy equations from Bayesian posteriors","keywords":["parameter degeneracy","symbolic regression","mutual information","Bayesian inference","posterior distribution","cosmological parameters","Planck","CMB"],"falsifier":"A posterior with a known non-separable degeneracy (e.g., one involving products or compositions of parameters that no coordinate transform can render separable) would yield a low orthogonal R-squared and no recovered equation, exposing the boundary of the method's applicability.","tokens_in":7042,"feed_emoji":"📐","tokens_out":2637,"duration_ms":150072,"temperature":0.7,"pith_summary":"The paper argues that parameter degeneracies in Bayesian posteriors — the situation where observed data constrains a combination of parameters more tightly than the individual values — can be automatically discovered and expressed as human-readable equations. The method works in two stages: first, mutual information screening identifies which parameters are coupled, sidestepping the combinatorial explosion of checking all possible parameter subsets; second, alternating symbolic regression fits the degeneracy as a level-set equation of the form g1(θ1) + g2(θ2) + ... + gk(θk) = c, decomposing a k-dimensional symbolic search into k independent one-dimensional problems. The decomposition is valid under the assumption that the degeneracy manifold is separable — expressible as a sum of univariate functions. On four synthetic benchmarks with known analytical degeneracies (including a Rosenbrock banana, a trigonometric surface, and an S-curve) and on Planck 2018 CMB posterior chains, the method recovers the correct functional forms with orthogonal R-squared above 0.98. Notably, applied to Planck data without any physics input, it independently recovers the cosmological horizon-angle degeneracy Ω_m h³ ≈ const, yielding an exponent ratio of approximately 2.95, close to the theoretically expected value of 3.","feed_headline":"Pipeline extracts degeneracy equations from Bayesian posteriors","feed_subtitle":"Mutual information plus alternating symbolic regression finds multi-parameter relationships that corner plots miss, recovering known cosmolo","key_machinery":"Alternating symbolic regression on separable level sets, with mutual information screening for candidate selection and an orthogonal R-squared metric for goodness of fit.","core_discovery":"The central mechanism is the alternating optimization loop: by modeling a degeneracy as a separable level set and cycling through each component function one at a time while holding the others fixed, the method converts an intractable multivariate symbolic regression into a sequence of tractable univariate ones. Combined with mutual information screening to pre-select which parameter tuples are worth fitting, this pipeline can take raw posterior samples and return a closed-form equation describing the degeneracy — including recovering known cosmological relationships from real data without domain-specific priors.","pith_inferences":["If separability is the main bottleneck, a natural extension would be iterative: fit a separable model, examine residuals for remaining structure, and re-apply the pipeline to the residuals — potentially recovering degeneracies that are sums of separable terms plus corrections.","The method's success on Planck data suggests it could serve as an automated consistency check on published cosmological results: if it independently recovers known degeneracies from posterior chains, it could also surface previously unrecognized ones in newer datasets.","The orthogonal R-squared metric defined here could become a standard diagnostic for degeneracy characterization quality, analogous to how Gelman-Rubin statistics are used for MCMC convergence — giving analysts a single number to decide whether a degeneracy has been adequately captured."],"forward_implications":["Degeneracy equations discovered by this method could directly inform reparameterization, letting analysts define new parameters aligned with and orthogonal to degenerate directions for tighter, more honest constraints.","The mutual information screening step serves as a standalone diagnostic: it flags which parameters are coupled even when the subsequent symbolic regression fails on non-separable degeneracies.","Applied to simulation-based inference pipelines, the tool could act as an automated post-processing step that surfaces hidden degeneracies in neural posterior estimators, where such degeneracies are common but harder to diagnose than in traditional MCMC.","The alternating optimization strategy — decomposing a multivariate symbolic regression into sequential univariate fits — could generalize beyond degeneracy detection to any implicit surface fitting problem where approximate separability holds."],"fun_headline_variants":["DegenDetector recovers closed-form degeneracy equations from posteriors","Symbolic regression pipeline extracts multi-parameter degeneracies as equations","Alternating optimization turns posterior degeneracies into closed-form formulas","Pipeline finds symbolic degeneracy equations without domain-specific priors","Mutual information plus symbolic regression recovers cosmological degeneracies"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The method assumes that any degeneracy can be written as a sum of single-parameter functions set equal to a constant — that is, the degeneracy is separable. Many physical degeneracies are multiplicative or otherwise non-separable in the native parameter space. A log-transform workaround handles some of these cases, but if a degeneracy is genuinely non-separable and cannot be linearized by such a transform, the method returns a poor fit and fails to recover the functional form","fun_headline_variants_meta":{"raw":{"variants":["DegenDetector recovers closed-form degeneracy equations from posteriors","Symbolic regression pipeline extracts multi-parameter degeneracies as equations","Alternating optimization turns posterior degeneracies into closed-form formulas","Pipeline finds symbolic degeneracy equations without domain-specific priors","Mutual information plus symbolic regression recovers cosmological degeneracies"]},"model":"glm-5.2","effort":"low","cost_usd":0.0,"raw_usage":{"total_tokens":492,"prompt_tokens":404,"completion_tokens":88,"prompt_tokens_details":null},"tokens_in":404,"tokens_out":88,"duration_ms":36907,"temperature":1.0,"reasoning_tokens":null,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-10T01:43:40.014423+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"A posterior with a known non-separable degeneracy (e.g., one involving products or compositions of parameters that no coordinate transform can render separable) would yield a low orthogonal R-squared and no recovered equation, exposing the boundary of the method's applicability.","supporting_citations":[],"review_version":1}