{"id":"cce151c5-4fc1-4b10-970c-9c32daffd276","arxiv_id":"2510.12976","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A neural network estimates full posterior distributions over phylogenetic trees without likelihood computation, beating likelihood-based methods in topology accuracy on simulated benchmarks.","lead":"Scientists built a neural network that guesses an entire family tree (phylogeny) and its uncertainty directly from aligned DNA or protein sequences, without ever computing the classical likelihood score. It is more accurate than standard maximum-likelihood methods in most tested cases, and works even for evolutionary models where the likelihood cannot be computed at all.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The posterior estimator may not be a normalized distribution over trees: factor-count mismatch (2N−3 vs N−2 merges), an unsampled final edge, and constraints that do not force sampling order to match the canonical order threaten the central calibration claim.","rationale":"I read the paper as making two central claims: (1) topological accuracy and speed, and (2) a well-calibrated posterior distribution over phylogenies. The first is supported by extensive simulation, though with caveats about model mismatch and test-set provenance. The second depends entirely on qψ being a proper, normalized probability distribution over trees and branch lengths. The OOD concern raised by the reader is legitimate but applies to every amortized neural posterior estimator and is explicitly acknowledged in §5; it does not test whether the method is correct within its training distribution. My concern is different and more fundamental: the manuscript's own specification of qψ is internally inconsistent. The number of factors in the defining equation (§3.2) does not match the loop in the algorithms; the branch-length parameterization omits the final internal edge; the stated Jacobian in the appendix contradicts the main text; and the constraint mechanism does not obviously force the sampled merge order to be the canonical order used during evaluation. If any of these inconsistencies reflects the actual implementation, then the object being optimized and the object being sampled are different, and the calibration plot in Fig. 3b cannot be trusted as evidence of posterior coverage. Because no code or data are released, the ambiguity cannot be resolved by inspection. I therefore recommend UNVERDICTED rather than CONDITIONAL: the authors should either release the implementation and demonstrate normalization for small N, or supply a corrected, self-consistent definition and proof that the sampling and evaluation procedures assign the same probability to every tree. This is not an attack on the authors' integrity; it is a request for the minimal verification that the central posterior claim requires.","tokens_in":20514,"tokens_out":19001,"duration_ms":161485,"concrete_test":"Implement Algorithms S.10 and S.11 exactly for N=4 (and N=5) with random or pretrained weights. For fixed input x, enumerate all tree topologies with discretized branch lengths and compute (i) the total probability Σθ qψ(θ|x) assigned by the evaluation procedure, and (ii) for trees generated by the sampling procedure, whether the canonical merge order recovered from the sampled tree equals the order in which merges were sampled. If the total deviates from 1 beyond numerical tolerance, or order mismatches occur with nonzero probability, the posterior estimator is not the sampling distribution and the calibration results in Fig. 3 cannot be interpreted as posterior probabilities.","verdict_should_be":"UNVERDICTED","load_bearing_attack":"The OOD caveat identified by the reader is real but generic to all amortized NPE; the more load-bearing issue is internal. In §3.2, qψ(θ) is defined as a product over 2N−3 merge factors, but Algorithms S.10/S.11 loop only k=1..N−2 and emit two branch lengths per merge, giving 2N−4 branch lengths. The final edge connecting the two remaining active nodes (e.g., u1–u2 for N=4) is never assigned a length, so the model does not define a density over the full θ ∈ R^{2N−3}_+. More subtly, the canonical-order constraints only require each sampled cherry sum to exceed previous sums. In the final tree, the distance between an early internal node and a leaf includes the unsampled final/internal edge and can be smaller than earlier merge sums. Thus Algorithm S.11 can generate a tree whose canonical merge order, as used by Algorithm S.10, differs from the sampling order. Then a tree may be counted via multiple merge orders, or not counted at all, and qψ is not a normalized probability distribution. Supporting inconsistencies: §3.3 states r = ℓ_i/(ℓ_i+ℓ_j) with a 1/s Jacobian factor, but Algorithm S.10 writes a Beta on ℓ_i/ℓ_j and omits the Jacobian; Appendix A.2 gives det = −1/r instead of −1/s. No code is released to resolve which definition is implemented, so the central 'well-calibrated posterior' claim is not currently verifiable.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents Phyloformer 2, a neural posterior estimator for phylogenies from aligned sequence data. It combines an evoPF encoder (an EvoFormer-like architecture) with a BayesNJ module that defines a factorized distribution over trees through a succession of pairwise merges. The authors claim this is the first likelihood-free posterior estimation method trained end-to-end beyond quartets, and report improved topological accuracy over likelihood-based tools, faster inference, and well-calibrated posterior branch supports on simulated data. The architecture, training procedure, and experiments are described in detail, with supplementary algorithms and extensive benchmarking.","tokens_in":20893,"tokens_out":10781,"duration_ms":81326,"significance":"If the mathematical inconsistencies in the definition of BayesNJ are resolved, this would be a substantial contribution: it demonstrates a scalable amortized likelihood-free approach to full phylogenetic posterior inference, with promising empirical performance and a clear path to use under intractable-likelihood models. The paper includes extensive experiments and comparisons, and the qualitative empirical claims are supported by the data in the relevant regimes. The issues identified here concern the formal correctness of the model distribution and are essential to address before the posterior calibration claims can be taken at face value.","major_comments":[{"comment":"The product defining qψ is over 2N−3 merges, but Algorithm S.10 and S.11 loop only over k=1..N−2 and emit two branch lengths per merge, i.e., 2N−4 branch lengths. The final edge connecting the two remaining active nodes is never assigned a length, so qψ does not define a probability density on R_+^{2N−3}. If the final edge is intended to be handled implicitly, it must be stated explicitly and included in the density; otherwise the parameter space of the model is not the space of phylogenies.","section":"§3.2, Eq. (1) and Algorithms S.10/S.11"},{"comment":"The constraint procedure does not guarantee that the sampling order matches the canonical merge order used for evaluation. The constraints only impose that each sampled cherry sum be at least the sum of any previous merge performed while that pair was available; they do not ensure that the sampled pair is the minimum-sum pair among all current candidates in the final tree. Concretely, for N=4, a tree with s(a,b)=20 and a later cherry (u1,c) with sum 12 would be sampled in the order (a,b) first, whereas the canonical order of the same tree would merge (a,c) first. Hence the sampled order and the evaluation order can disagree, and the same tree can receive different probabilities or be missed entirely. This breaks the normalization argument for qψ.","section":"§3.2 and Algorithm S.11"},{"comment":"The reparameterization of the branch-length distribution is inconsistent between the main text, the training loss, and the appendix. §3.3 defines r = ℓ_i/(ℓ_i+ℓ_j) and states that the Jacobian factor is 1/s. Algorithm S.10 evaluates the Beta density at ℓ_i/ℓ_j (not at r) and omits the Jacobian factor entirely. Algorithm S.11 samples r from the Beta and forms ℓ_i = r·s, ℓ_j = s−ℓ_i, which corresponds to the main-text reparameterization. Appendix A.2 gives the determinant as −1/r (should be −1/s) and concludes qℓ = q_s q_r / r. These three descriptions cannot all be correct, and without code it is impossible to determine which objective was actually optimized. This ambiguity affects the branch-length posterior and, through the training objective, the entire posterior approximation.","section":"§3.3, Algorithm S.10, Appendix A.2"}],"minor_comments":[{"comment":"The claim of 'more accurate tree topologies than existing methods' is unqualified; Figure 2a shows this holds for 10–175 leaves, but at 200 leaves the advantage over IQTree is not demonstrated. Please qualify the abstract accordingly.","section":"Abstract"},{"comment":"The test set is simulated under LG+GC while PF2 is trained on LG+G8. This is a model mismatch that should be explicitly acknowledged in the main text and accounted for when interpreting the comparison.","section":"§4.1"},{"comment":"The line 'ℓ(k)j ← s(k) − ℓ(k)j' appears to be a typo: it should read 'ℓ(k)j ← s(k) − ℓ(k)i'.","section":"Algorithm S.11"},{"comment":"The paper does not state code availability; given the discrepancies in the algorithms, releasing the implementation would greatly facilitate verification.","section":"Code availability"}],"recommendation":"major_revision","confidential_remarks":"The central idea is interesting and the empirical work is extensive, but the formal definition of BayesNJ has inconsistencies that touch the core of the paper's contribution. If the authors can fix the normalization/canonical-order issue and the branch-length reparameterization, and re-run the affected experiments, the paper could make a strong contribution. The OOD caveat is generic to amortized NPE and is already acknowledged in §5; the internal consistency issues are more pressing."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my honest take. The paper is genuinely new: it's the first end-to-end trained likelihood-free estimator of posterior distributions over full phylogenies, and the BayesNJ factorization and evoPF encoder are real contributions. The experiments are extensive, comparisons are against strong baselines, and the authors are transparent about the flawed Cherry dataset and their own limitations. For 10–175 leaves, the topological accuracy gains over IQTree/FastTree look real, and the amortized speedup is not a detail.\n\nThe soft spot that matters is internal. The formal definition of qψ says a product over 2N−3 merge factors, but Algorithms S.10/S.11 loop only N−2 times and emit two branch lengths per merge. That leaves the final edge connecting the last two nodes with no sampled length. An unrooted binary tree on N leaves has 2N−3 branch lengths; the model as written covers only 2N−4. So the object being trained is not, as it stands, a normalized density over the full θ. The canonical-order argument also has a hole: the constraints only compare cherry sums during the merge process, and the unconstrained final edge can break the partial order used for evaluation, meaning a tree might be weighted by multiple orders or not counted at all. The Jacobian appendix typo (−1/r vs −1/s) is minor by itself, but it's consistent with the rush here. These are fixable, but they directly affect the 'well-calibrated posterior' claim.\n\nOther issues are more familiar: the abstract's 'more accurate than ML' is unqualified but only holds in the tested range; the main topological benchmark is LG+GC while training is LG+G8; the calibration check is one-size (50 taxa) and limited; and there is no code or data released. The OOD caveat is real but generic to all amortized NPE, and the authors explicitly acknowledge it.\n\nWho is this for? Anyone working on simulation-based inference in phylogenetics or on scalable Bayesian tree inference. It deserves a serious referee — not a desk reject. But the referee needs to push for a corrected mathematical statement of what qψ covers, an explicit treatment of the final edge (or a redefinition of the output space), and code. If the math gets fixed, this is a strong paper; right now the empirical results are promising but the central formal claim is not fully verifiable.","headline":"Genuinely new likelihood-free posterior estimator for full phylogenies, but the missing final edge in the merge factorization leaves the normalization claim unproven — fixable, but central.","tokens_in":21393,"tokens_out":7403,"would_cite":true,"duration_ms":60279,"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":"Phyloformer 2 is the first likelihood-free method to estimate full posterior distributions over phylogenetic trees directly from aligned sequences, and it reconstructs topologies more accurately than likelihood-based state-of-the-art method","keywords":["phylogenetic inference","likelihood-free inference","neural posterior estimation","deep learning","Bayesian phylogenetics","sequence evolution","amortized inference","BayesNJ"],"falsifier":"Train Phyloformer 2 under a birth-death prior and LG+G8, then apply it to alignments simulated under a different prior (e.g., uniform topology with exponential branch lengths) or to real empirical alignments. If topological accuracy degrades sharply and calibration curves move off the diagonal while a likelihood-based method with model selection degrades gracefully, the claimed edge is an artifact of prior matching rather than a general property of the estimator.","tokens_in":20358,"feed_emoji":"🧬","tokens_out":5203,"duration_ms":42683,"temperature":0.7,"pith_summary":"The paper introduces Phyloformer 2, a neural network that estimates the entire posterior distribution over phylogenetic trees without ever computing a likelihood. It trains on simulated alignments and represents the posterior as a product of conditional probabilities over a sequence of subtree merges, with branch lengths modeled by parametric distributions. On tractable sequence-evolution models the method gives more accurate tree topologies than leading maximum-likelihood tools, provides well-calibrated branch-support estimates, and runs one to two orders of magnitude faster at inference time. Its main payoff is that it can be trained under evolutionary models with intractable likelihoods, where its accuracy advantage over misspecified likelihood-based methods grows.","feed_headline":"No likelihoods needed: neural net wins tree-posterior estimation","feed_subtitle":"It returns a full tree posterior in one pass, with calibrated branch supports and a widening edge on complex models.","key_machinery":"BayesNJ: a parameterized probability distribution over phylogenies factorized over successive pairwise merges, using a softmin over symmetric bilinear pair scores for the topological choice and shifted Gamma and Beta distributions for branch lengths, with a constraint mechanism enforcing a canonical merge order so the product is a proper distribution over trees. evoPF: an encoder maintaining one embedding per sequence and one per pair of sequences, sharing information through self-attention and outer-product means, which scales to hundreds of sequences.","core_discovery":"The authors claim that a neural posterior estimator can replace likelihood computation for Bayesian phylogenetic inference. Phyloformer 2 combines evoPF, an EvoFormer-style encoder that turns aligned sequences into per-sequence and per-pair embeddings, with BayesNJ, a factorization of the tree posterior over a canonical sequence of merges. Under a tractable model (LG+G8), its greedy MAP trees have lower normalized Robinson-Foulds distance to the true tree than maximum-likelihood methods; under models with intractable likelihoods (Cherry and SelReg), the gap widens. Comparison to a long MCMC run on a 50-taxon alignment shows strong agreement in split frequencies, and simulation-based calibrat","pith_inferences":["The claimed performance is conditional on the training prior; the authors themselves note that inputs far from training data would produce poor estimates with no warning, so practical deployment would need an out-of-distribution detector or uncertainty layer.","The canonical-merge trick could generalize to other recursive combinatorial objects in biology, such as gene-tree/species-tree reconciliations or phylogenetic networks, wherever a canonical decomposition can be enforced.","Posterior quality is currently evaluated mainly through split frequencies; a stricter test would compare full tree distributions or clade probabilities on larger alignments, which could reveal whether the merge-wise factorization loses higher-order correlations.","Because the ablation suggests most of the topological gain comes from the BayesNJ loss rather than the encoder alone, making the branch-length distributions more flexible than Gamma/Beta could further improve calibration."],"forward_implications":["Once trained, Phyloformer 2 samples from the posterior orders of magnitude faster than MCMC, making full posterior analysis practical on large datasets.","Because it is likelihood-free, it can be trained under richer evolutionary models—with site dependencies, heterogeneous selection, or other complex generative processes—where likelihood computation is infeasible.","Branch posterior probabilities from the network are well calibrated (slightly conservative), so they can be used as uncertainty estimates for reconstructed clades.","When the training prior matches the data-generating process, the method outperforms maximum-likelihood tools even under tractable models, and its edge grows when the likelihood model is misspecified."],"fun_headline_variants":["Neural net infers full tree posterior without likelihoods","Likelihood-free: neural net estimates tree posteriors more accurately","Bayes without likelihood: neural tree posterior estimator outperforms","Neural posterior estimator: accurate trees even without likelihoods","No likelihoods needed for neural tree posterior wins"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The whole pipeline assumes that the data were generated by the same prior over trees and the same sequence-evolution model used to simulate training data; the authors state that outside that distribution the network would currently produce poor estimates with no warning.","fun_headline_variants_meta":{"raw":{"variants":["Neural net infers full tree posterior without likelihoods","Likelihood-free: neural net estimates tree posteriors more accurately","Bayes without likelihood: neural tree posterior estimator outperforms","Neural posterior estimator: accurate trees even without likelihoods","No likelihoods needed for neural tree posterior wins"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0007,"raw_usage":{"total_tokens":2976,"prompt_tokens":704,"completion_tokens":2272,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":448,"completion_tokens_details":{"reasoning_tokens":2192}},"tokens_in":448,"tokens_out":2272,"duration_ms":15236,"temperature":1.0,"reasoning_tokens":2192,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T09:49:46.281859+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train Phyloformer 2 under a birth-death prior and LG+G8, then apply it to alignments simulated under a different prior (e.g., uniform topology with exponential branch lengths) or to real empirical alignments. If topological accuracy degrades sharply and calibration curves move off the diagonal while a likelihood-based method with model selection degrades gracefully, the claimed edge is an artifact of prior matching rather than a general property of the estimator.","supporting_citations":[],"review_version":1}