{"id":"e20d7f57-d1ae-4de4-b953-f6759d9fcf9f","arxiv_id":"2602.03824","paper_version":2,"verdict":"REJECT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"Features extracted from a 10,000-species bird classifier form a visual morphospace that shows a rapid early burst of morphological disparity after the K-Pg extinction.","lead":"A bird-classification AI's internal image features are used as a new kind of 'shape space' to study evolution, and the space shows a rapid burst of visual diversity right after the dinosaur-killing asteroid. The method could let biologists measure hard-to-quantify traits like plumage and texture across thousands of species at once.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (2) weights the reconstructed ancestor toward the longer branch, opposite to the Brownian-motion precision weighting; without validation the DTT early-burst may be an ASR artifact.","rationale":"The paper makes an interesting, falsifiable claim, and it provides some independent-looking support: the hierarchical clustering recovers many taxonomic groups, the Grad-CAM maps are plausible, and the adversarial 'chicken-plumaged penguin' example is a nice sanity check. Those supports bear on the embedding's biological relevance, but they do not validate the disparity-through-time reconstruction. The strongest claim (an early burst after K-Pg with ~50% of modern disparity reached quickly) is obtained through a new spherical ASR that is presented in §2.5 with no derivation and no simulation check. The specific formula in Eq. (2) weights the ancestor toward the longer-branch child; in the flat limit of Brownian motion the correct weighting is inverse-variance (shorter branch dominates), so the estimator is at least not the BM estimator it claims to be. Since the DTT curve and the MDI are computed from reconstructed ancestors, a bias in ASR is load-bearing. The reader's verdict is REJECT; my concern does not weaken that verdict, and I would keep it. I do not think this is a case of fraud or sloppiness—it is a missing validation step that should be routine for a new phylogenetic comparative method. A simple simulation on the actual tree with known ancestral states would settle whether Eq. (2) is biased and whether the early burst persists under a corrected ASR.","tokens_in":12884,"tokens_out":9472,"duration_ms":106528,"concrete_test":"Simulate Brownian motion on the unit 512-sphere on the Stiller et al. tree with known ancestral states and branch lengths. Run the paper's ASR (Eqs. 2–3) on the simulated tip vectors and compare reconstructed ancestral states and the reconstructed DTT curve to the true values. Include asymmetric sister branches (e.g., la=10, lb=1): if the reconstructed parent is closer to the long-branch child, Eq. (2) is biased. Then recompute the empirical DTT using a precision-weighted spherical estimator (or corrected Eq. 2 with t=la/(la+lb)) and check whether the early burst and MDI persist. Report mean reconstruction error and DTT curves for both estimators.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline DTT claim depends entirely on the new spherical ASR in §2.5. That algorithm is stated without derivation or simulation test, and Eq. (2) is suspect: for sister nodes with branch lengths la, lb, it sets t=lb/(la+lb) and places the parent by slerp, so when lb << la the ancestor is pulled toward va, the lineage with the longer branch. Under Brownian motion, the conditional expectation of the parent given two descendant states is precision-weighted by inverse branch lengths, (va/la + vb/lb)/(1/la + 1/lb); the shorter branch should dominate. So Eq. (2) is backwards unless an alternative model is specified. Eq. (3)'s curvature correction and the 'equivalent branch lengths' for internal nodes are also undefined, so the estimator is not reproducible as written. This is not a mere detail: the empirical DTT curve (Fig. 7) is reconstructed with this ASR, and the '50% of modern disparity shortly after K-Pg' could be an artifact of the estimator's bias. The null simulations use the same ASR, so the comparison may not be invalidated simply by shared bias, but the magnitude and timing of the empirical burst are not trustworthy until the ASR is checked against known ancestral states.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a deep-learning-based morphometric framework for birds. A ResNet34 is trained on the DIB-10K dataset to classify over 10,000 species; the weights of the final fully connected layer are treated as 512-dimensional morphological trait vectors. The authors show that hierarchical clustering of these vectors recovers much of the traditional taxonomy, that taxonomic similarity correlates with morphological disparity, and that an analysis of disparity through time (DTT) reveals an early-burst pattern after the K-Pg extinction. The early-burst result is obtained with a newly proposed 'spherical ancestral state reconstruction' (ASR) algorithm and compared with a Brownian-motion null simulation.","tokens_in":13208,"tokens_out":5973,"duration_ms":63112,"significance":"If the results were robust, this would be a useful high-throughput tool for macroevolutionary studies and a striking demonstration that visual morphospace expanded rapidly after the K-Pg boundary. The paper makes its code and model weights publicly available and includes an adversarial-example analysis arguing that the network captures shape rather than texture. However, the central DTT conclusion is not adequately supported: the spherical ASR is not derived or validated, the null simulation lacks confidence intervals, and the removal of 249 species is not transparently reported. These issues are fixable in principle but require substantial additional work.","major_comments":[{"comment":"The spherical ASR is stated without derivation. Eq. (2) sets v_p = slerp(v_a, v_b; t) with t = l_b/(l_a+l_b), so the parent is pulled toward the lineage with the longer branch. Under Brownian motion, the conditional expectation of the parent is precision-weighted by inverse branch lengths: (v_a/l_a+v_b/l_b)/(1/l_a+1/l_b), so the shorter branch should dominate. Thus Eq. (2) is backwards for BM. Since the empirical DTT curve (Fig. 7) is reconstructed with this ASR, the early-burst magnitude/timing may be an artifact. A simulation study with known ancestral states is needed to validate or correct this.","section":"§2.5, Eq. (2)"},{"comment":"The correction factor Contrast2 = Euclidean variance × θ²/||va-vb||² is ad hoc; the 'equivalent branch length' for internal nodes is undefined. Without precise definitions, the ASR is not reproducible. Please provide a full derivation and pseudocode for all steps, including how internal branch lengths are assigned.","section":"§2.5, Eq. (3) and 'equivalent branch lengths'"},{"comment":"The null simulation is shown only as the mean of 100 runs, without confidence intervals or the number of species used. The claim that the empirical curve 'deviates extremely' cannot be assessed without a distribution. Report the full null distribution, the number of tips after pruning the Stiller et al. (2024) tree, and how missing species were handled.","section":"§2.5, Fig. 7"},{"comment":"249 species were removed after 'manual review' because they 'lacked biological significance.' The criteria are subjective and the list is not given. This removal could bias both the disparity and DTT analyses. Provide the list, explicit a priori criteria, and a sensitivity analysis including/excluding these species.","section":"§2.4 / Discussion"},{"comment":"The 'morphological trait' vector is the row of the final fc-layer weight matrix (a per-class template), not an embedding of an input image. The paper should justify why this is a valid phenotypic measurement; the Grad-CAM evidence does not establish that the weight vectors capture morphology. Compare with image-derived features or use an unsupervised embedding to address the circularity of recovering taxonomy from a classifier trained on those very labels.","section":"§2.2"}],"minor_comments":[{"comment":"Typo: 'aenerative' should be 'generative'.","section":"§2.2"},{"comment":"Typo: 'futherly' should be 'further'; 'disaprity' should be 'disparity' in the residual discussion.","section":"§4.3"},{"comment":"The caption says 'red dash line' — should be 'red dashed line'.","section":"Fig. 4 caption"},{"comment":"Many labels render as '??', making the figure unreadable. Provide a vector figure or higher-resolution image.","section":"Fig. 5"},{"comment":"The title in the PDF body differs from the arXiv metadata title; unify them.","section":"Title"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses an important and timely topic, and the scale of the data is impressive. However, the DTT section is the core contribution and currently rests on an unvalidated and likely incorrect ASR. The missing confidence intervals and the opaque removal of 249 species further undermine the main conclusion. If the authors can correct the ASR, add proper uncertainty quantification, and make the data-cleaning transparent, the paper could be suitable. As submitted, the central claim is not supported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The most interesting thing in this paper is the scale and the framing: a 10,000-species visual morphospace from a CNN classifier, with taxonomy and convergence recovered without any phylogenetic input. The clustering results are genuinely worth a look, and the author deserves credit for shipping code and model weights. The disparity–diversity scaling and the residual analyses are also reasonable as descriptive results, though they are not the main event.\n\nThe main event is the DTT early-burst claim. That claim is not supported as written, and I think the reader's report is right about why. The new spherical ASR algorithm in §2.5 is stated without derivation or any simulation check. More concretely, Eq. (2) weights the reconstructed ancestor toward the longer branch: t = lb/(la+lb), so when lb << la, the ancestor is pulled toward va, the lineage with the longer branch. Under Brownian motion, the conditional expectation should be precision-weighted by inverse branch lengths, which would pull the ancestor toward the shorter branch. That is backwards unless the author is deliberately modeling something else, and no alternative model is given. The curvature correction in Eq. (3) and the 'equivalent branch lengths' are also undefined, so the estimator is not reproducible as written. This is not a minor detail: the empirical DTT curve in Fig. 7 is computed with this ASR, and the '50% of modern disparity shortly after K-Pg' could easily be an artifact of the estimator's bias. The null simulations use the same biased estimator, so the comparison may not be entirely misleading, but the magnitude and timing of the empirical burst cannot be trusted until the ASR is validated against known ancestral states.\n\nThe other soft spots are in proportion. The null simulation is reported only as a mean, no confidence intervals, and the number of species in the timetree subset is not stated. The removal of 249 species after manual review is potentially circular, since the same clustering is used to infer biological signal. The embedding itself comes from a classifier trained on the same species labels, so recovering taxonomy from the vectors is partly by construction; that is a real limitation for claiming independent evidence of convergence, though the specific convergence cases (e.g., penguins/auks/petrels) are plausible.\n\nIs this paper worth serious referee time? Yes. The framework is new in combination, the scale is impressive, and the problems are fixable in principle. But the DTT conclusion should be withdrawn or heavily caveated until the ASR is corrected and tested. I would send it to peer review, with the explicit expectation of major revision and a request for simulation-based validation of the ancestral-state reconstruction.","headline":"The DTT early-burst claim rests on an unvalidated spherical ASR whose weighting is backwards; the framework is promising but the headline result is not yet supported.","tokens_in":13653,"tokens_out":1554,"would_cite":false,"duration_ms":19334,"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":"The visual traits learned by a bird-recognition network show that avian form exploded soon after the K-Pg extinction, then decelerated — evidence for the niche-filling model of adaptive radiation.","keywords":["morphological evolution","deep learning","representation learning","disparity through time","adaptive radiation","birds","visual morphospace","ancestral state reconstruction"],"falsifier":"A concrete check: run a landmark-based geometric-morphometric disparity-through-time analysis on the same bird clades using the same timetree and comparable sampling; if that curve shows gradual disparity accumulation rather than an early burst, the classifier-embedding result would be contradicted. Alternatively, a simulation study demonstrating that the spherical ASR systematically inflates early disparity under a Brownian-motion null would undermine the method.","tokens_in":12749,"feed_emoji":"🐦","tokens_out":3821,"duration_ms":40270,"temperature":0.7,"pith_summary":"This paper tries to establish that the high-dimensional feature vectors of a deep neural network trained only to classify bird species can serve as a valid, landmark-free morphospace for studying morphological evolution. By treating the network's final-layer weights for each species as morphological trait vectors, the author shows that this space recovers taxonomic hierarchy and captures convergent evolution without any a priori taxonomic input. The central macroevolutionary finding is a disparity-through-time curve that rises steeply after the K-Pg mass extinction, reaching roughly half of modern avian visual disparity shortly afterward, then decelerates — a pattern that deviates strongly from neutral Brownian-motion expectations. If correct, this supports the niche-filling hypothesis: surviving bird lineages rapidly diversified into vacated ecological niches. The work also argues that the network encodes whole-body shape rather than local texture, challenging a common assumption about convolutional networks.","feed_headline":"Birds filled half their visual space right after the K-Pg extinction","feed_subtitle":"A bird-recognition network's learned traits recover taxonomy and show an early-burst rise in disparity, backing the niche-filling model.","key_machinery":"The load-bearing mechanism is the final-layer weight vector of a fine-grained image classifier, projected onto a unit hypersphere, as a high-dimensional morphological phenotype. These vectors define a cosine-similarity morphospace in which taxonomic and convergent structure emerges. The disparity-through-time analysis runs on a spherical Brownian-motion model: ancestral states are inferred by interpolating along the great-circle arc between sister nodes, with a correction factor that scales Euclidean contrast variance by the squared arc-to-chord ratio, and null simulations are generated by projecting Gaussian noise onto the tangent space of the sphere and moving along geodesics.","core_discovery":"On the paper's own terms, the central discovery is that a convolutional neural network trained to recognize over 10,000 bird species implicitly learns a biologically meaningful morphospace. The 512-dimensional weight vectors extracted from the final fully connected layer, when L2-normalized and treated as points on a unit hypersphere, cluster by taxonomic group and place convergent taxa close together. Disparity-through-time analysis using a novel spherical ancestral-state-reconstruction algorithm reveals an empirical early-burst trajectory: after the K-Pg extinction, relative morphological disparity jumps to about 50% of the present-day value far faster than a Brownian-motion null model pre","pith_inferences":["The early-burst result should be tested against classical landmark-based morphometrics on a subset of the same bird clades; if landmark-based DTT shows a different tempo, the classifier-embedding geometry may be capturing something other than morphology.","The same classifier-embedding pipeline could be applied to other imageable taxa (e.g., insects, plants, fungi) to build a common visual morphospace, enabling cross-taxon disparity comparisons that landmark data cannot easily provide.","The shape-over-texture claim could be strengthened by generating a broader suite of texture-shape conflict images across many orders and measuring whether classification remains shape-driven; the current single adversarial example is suggestive but not exhaustive.","The spherical ancestral-state-reconstruction algorithm is central to the DTT result; a formal validation against known evolutionary simulations would clarify whether its arc-versus-chord correction introduces any bias in early disparity estimates."],"forward_implications":["If the embedding morphospace is valid, species richness is the primary driver of morphospace expansion, with strong rank correlations at both order and family levels.","The recovered clustering reinforces that deep-network embeddings can rediscover established taxonomy and expose convergent phenotypes, such as ground-dwelling birds grouping together or penguins, auks, and procellariiforms sharing a visual region.","The disparity-through-time curve implies birds occupied about half of their current visual form space shortly after the K-Pg extinction, supporting an early-burst tempo for avian morphological diversification.","The model's apparent focus on whole-body plan rather than texture suggests that fine-grained visual classification may engage shape representations more strongly than generic object recognition.","The approach provides a high-throughput, anatomy-independent morphometric tool that could be extended to other imageable organismal groups."],"fun_headline_variants":["AI morphospace reveals birds' post-K-Pg explosion","Deep learning maps bird evolution: early burst after dinosaurs","CNN-trained morphospace shows birds exploded after K-Pg","Birds surged in disparity right after K-Pg, AI shows","Neural network decodes bird shapes, finds early burst"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The whole argument rests on the premise that the classifier's final-layer vector for each species is a faithful, unbiased measurement of its morphology, and that the spherical ancestral-state-reconstruction algorithm correctly models trait evolution on a hypersphere — if either fails, the early-burst pattern could be an artifact.","fun_headline_variants_meta":{"raw":{"variants":["AI morphospace reveals birds' post-K-Pg explosion","Deep learning maps bird evolution: early burst after dinosaurs","CNN-trained morphospace shows birds exploded after K-Pg","Birds surged in disparity right after K-Pg, AI shows","Neural network decodes bird shapes, finds early burst"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000632,"raw_usage":{"total_tokens":2738,"prompt_tokens":710,"completion_tokens":2028,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":454,"completion_tokens_details":{"reasoning_tokens":1960}},"tokens_in":454,"tokens_out":2028,"duration_ms":12380,"temperature":1.0,"reasoning_tokens":1960,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T04:48:48.780593+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete check: run a landmark-based geometric-morphometric disparity-through-time analysis on the same bird clades using the same timetree and comparable sampling; if that curve shows gradual disparity accumulation rather than an early burst, the classifier-embedding result would be contradicted. Alternatively, a simulation study demonstrating that the spherical ASR systematically inflates early disparity under a Brownian-motion null would undermine the method.","supporting_citations":[],"review_version":1}