{"id":"d8e43b75-6452-4525-b515-818bc9c87569","arxiv_id":"2411.16132","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"TreeFormer adds a training-time minimum spanning tree projection to a transformer graph generator, improving single-view plant skeleton estimation.","lead":"TreeFormer trains a graph-generation network to output tree-shaped plant skeletons by projecting its raw edge predictions through a minimum spanning tree during training. It beats unconstrained and two-stage baselines on synthetic trees, real Arabidopsis roots, and grapevine images.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 1 reports single-run point estimates without error bars or significance tests, so the claimed significant improvements over the test-time MST baseline are not yet quantified.","rationale":"The paper proposes a clean, practical mechanism: project raw edge probabilities through an MST during training, suppress features for edges where the projection disagrees with the raw output, and backpropagate through the remaining feature. The forward pass does produce the constrained tree, and Equation (10) plus Table S1 give a coherent local analysis. The train-time-only ablation (Table S5) and the Lambda ablation (Table S6) provide some evidence that the training signal, not merely test-time MST, is responsible for the gains. These are real, independent supporting observations. My concern is not that the method is internally inconsistent; it is that the central empirical claim is not yet reproducible from the reported numbers. All comparisons in Table 1 are single runs. The claimed 'significant improvement' is a statistical statement, and the paper does not supply the statistics. Because the improvements are small in absolute SMD units and nearly zero in grapevine TOPO F1, this is the most load-bearing weak point. The reader identified the SFS gradient approximation as the weakest assumption; I partially agree, but a strong empirical result could mitigate that theoretical gap. The missing seed-level statistics prevent that mitigation from being credible. A multi-seed rerun directly settles whether the concern lands: if the gaps are stable, the conditional verdict can move toward ACCEPT; if not, the claimed advantage is unsupported. For these reasons I keep the verdict at CONDITIONAL rather than moving it, since there is no evidence of error, only missing evidence of stability.","tokens_in":18,"tokens_out":5309,"duration_ms":118544,"concrete_test":"Rerun the unconstrained, test-time-constraint, and TreeFormer methods from Table 1 for at least five random seeds on the root and grapevine datasets, using the same data splits, augmentation, and early-stopping criterion. Report mean and standard deviation for SMD and TOPO F1, and compute a paired bootstrap or sign test between TreeFormer and test-time constraint. If the SMD gap is less than one pooled standard deviation on either real dataset, the claimed significant improvement is not established and the verdict should remain CONDITIONAL; if the gap exceeds the variance consistently, the central claim is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim rests on Table 1, where every cell is a single point estimate with no seeds, no standard deviations, and no significance testing. The headline SMD gaps are in absolute terms small: synthetic 6.26e-6 vs 4.78e-6, root 1.52e-4 vs 8.82e-5, grapevine 1.47e-4 vs 1.03e-4. On grapevine the TOPO F1 gap between test-time constraint and TreeFormer is only 0.867 vs 0.870 (0.003), and on root the gap is 0.799 vs 0.833 (0.034). With early stopping and stochastic training, such differences can plausibly arise from run-to-run variance, especially for SMD values in the 1e-5 to 1e-4 range. The paper even uses the word 'significantly' in the Table 1 caption, but no statistical evidence is provided. This is load-bearing because the entire advantage of SFS over the test-time MST baseline is these numerical gaps; if they are within training noise, the empirical support for the method collapses. The reader's concern about the local straight-through-style gradient is legitimate, but the ablation in Table S5 would be the natural mitigation; however, without repeated runs, that ablation also cannot be evaluated statistically.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"TreeFormer proposes a tree-constrained graph generation method for single-image plant skeleton estimation. The paper's key idea is to insert a selective feature suppression (SFS) layer into an off-the-shelf graph generator (RelationFormer). During training, raw edge probabilities are projected onto a minimum spanning tree via Kruskal's algorithm; for edges that the MST adds or removes, the corresponding logit is replaced by a large negative constant so that the projected probabilities can be trained with cross-entropy while a gradient path to the network is retained. At inference the output is again projected to a tree. The method is evaluated on a synthetic L-system dataset, real Arabidopsis root images, and the 3D2cut grapevine dataset against a re-implemented two-stage method, unconstrained RelationFormer, and a test-time-MST baseline. The paper reports lower SMD and higher TOPO scores for TreeFormer on all three datasets, with 100% tree rate.","tokens_in":21572,"tokens_out":9458,"duration_ms":93970,"significance":"The SFS layer is simple, model-agnostic in principle, and accompanied by public code and several useful ablations (train-time vs test-time constraint, Lambda sensitivity, and a comparison with the autoregressive GGT baseline). The paper also explicitly discloses that tree rate is trivially 100% for all constrained methods, which is the correct framing. If the reported improvements are statistically robust, the method is a practical contribution to plant phenotyping and a useful template for imposing discrete graph constraints in end-to-end training. The main qualification is that the empirical advantage over the test-time-constraint baseline is small on some metrics and is currently supported only by single-run numbers.","major_comments":[{"comment":"The central empirical claim rests on Table 1, but every cell is a single point estimate from one training run with early stopping. The caption uses the word 'significantly,' yet no error bars, number of seeds, or significance tests are reported. On several load-bearing comparisons the gaps are small: on grapevine the TOPO F1 gap versus test-time constraint is 0.867 vs 0.870, and on root the SMD gap is 1.52e-4 vs 8.82e-5. With stochastic training and early stopping, these differences could plausibly be within run-to-run variance. Please report multiple seeds with standard deviations and, where possible, paired statistical tests for at least the comparison against the test-time-constraint baseline; without this, the claimed advantage of SFS over simple test-time MST is not established.","section":"Table 1, §5.5"},{"comment":"The SFS layer uses a straight-through-style gradient: for edges in E+ ∪ E− it backpropagates through the kept feature while treating the MST solution as fixed, and for all other edges it backpropagates the ordinary cross-entropy gradient. This ignores how changes in edge weights would alter which tree the MST selects. The approximation is not discussed as a limitation, and Table S5 shows its reach: when the SFS layer is used only during training and MST is not applied at test time, the raw output has tree rate 0% on the grapevine dataset (56.5% on synthetic, 37.2% on root). Please state this approximation explicitly and provide evidence that it is a sufficient training signal, e.g., by comparing with a differentiable surrogate of MST or by analyzing gradient quality on a small controlled problem.","section":"§3.3, Eq. (10), Table S5"},{"comment":"For edge pairs not in E+ ∪ E−, the constrained term in Eq. (11) is identical to the unconstrained term, so the total loss gives those edges twice the weight of modified edges. This asymmetric weighting is not discussed. Please state whether this is intentional and, if so, explain its effect on the learned edge weights; if it is unintentional, the loss formulation should be adjusted.","section":"§4.2, Eq. (11)"},{"comment":"The two-stage baseline is a re-implementation of ViNet with per-dataset heuristic parameters (d, τm, τn) selected by best SMD. Because these parameters are tuned for each dataset, the comparison between TreeFormer and the two-stage baseline may be affected by tuning bias. Please clarify how the thresholds were selected for the other methods and whether the selection was blinded to the test set.","section":"§5.3, §C.1, Table S3"}],"minor_comments":[{"comment":"There is a typographical issue in the otherwise branch of Eq. (7), where a stray character appears where a bracket is intended; please correct the notation.","section":"Eq. (7)"},{"comment":"The TOPO score is computed only over nodes with degree != 2, which is a modification of the Sat2Graph implementation. Please justify this choice explicitly and note how it affects comparability with previously reported TOPO numbers.","section":"§5.2"},{"comment":"The claim of 'approximately 30% improvement on edge accuracy' should be phrased as a relative reduction in SMD, not an improvement in edge accuracy, to avoid conflating the two metrics.","section":"§5.6"},{"comment":"The paper uses validation-based early stopping but does not state how many random seeds were used. Given the absence of error bars, a statement about seed variability is important.","section":"§5.4"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal. My main concern is statistical: the headline comparison to test-time MST is based on single runs, and the margins are small on some metrics. I recommend major revision because the missing experiments (multiple seeds, error bars, significance testing) are feasible and would likely settle the question. The gradient-approximation issue in §3.3 should also be addressed in the revision rather than treated as a purely theoretical aside."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid, well-scoped method paper. The core idea is simple and effective: during training, project the raw edge predictions to the nearest tree via MST, then modify the logits for edges where the MST disagrees so the loss sees a tree. The selective feature suppression (SFS) layer keeps one of the two logits untouched, so gradients still flow to the graph generator. That's a clean straight-through trick, and the paper analyzes which cases get stronger gradients (Table S1).\n\nThe experiments are reasonable. Three datasets, a proper test-time MST baseline, a train-time-only ablation, and a Lambda sensitivity check. Results consistently favor the full method over the baselines. The train-time-only ablation is particularly useful: on synthetic it helps, on real images it doesn't make the raw output a tree, which tells you the method is really about learning better edge weights for the MST, not about fitting the constraint end-to-end. Good that they disclose tree rate is 100% for any method that applies MST.\n\nThe main soft spot is exactly what the stress-test flags: Table 1 is single-run. No seeds, no error bars, no significance tests. Some of the gaps are small, especially grapevine TOPO F1 (0.867 vs 0.870). With stochastic training and early stopping, I can't rule out run-to-run noise. That said, the SMD improvements are larger and consistent in direction across three datasets, and the ablations support the story. So it's a missing robustness check, not a fatal flaw.\n\nThe second soft spot is the gradient approximation. The SFS layer ignores the global effect of MST selection on the loss. The paper gives a case-by-case analysis showing sensible gradient behavior, but it's still a heuristic. Not a dealbreaker, but they should be more explicit that this is a straight-through estimator with no guarantee of matching the true gradient.\n\nThe two-stage baseline is a re-implementation with per-dataset hyperparameters. They show it beats a naive re-implementation, so it's not obviously unfair, but it's a potential source of bias.\n\nWho is this for? Anyone working on structured graph prediction from images, especially plant phenotyping. The SFS layer is reusable. I'd send it to review and ask for multi-seed variance and a short discussion of the gradient approximation. That would bring it to accept.\n\nRecommendation: send to peer review; it deserves a serious referee.","headline":"A clean, well-evaluated method for tree-constrained graph generation; the main fix needed is error bars for the headline numbers.","tokens_in":22131,"tokens_out":3792,"would_cite":true,"duration_ms":35350,"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":"TreeFormer claims that a graph generator trained with a feature-suppression layer plus MST projection can output exact plant-skeleton trees from a single image, with better edge accuracy than unconstrained or post-hoc-constrained baselines.","keywords":["plant skeleton estimation","graph generation","minimum spanning tree","tree constraint","selective feature suppression","differentiable reparameterization","transformer","plant phenotyping"],"falsifier":"Train the SFS layer on a dataset where, at many training examples, the MST projection disagrees with the ground-truth tree on a large fraction of edges; if the learned edge weights do not reduce this disagreement beyond what the unconstrained baseline achieves, the training signal is too weak. A simpler check is the paper's own train-time-only setting: if a model trained with the SFS layer but evaluated without test-time MST still outputs non-tree graphs (0 percent tree rate on grapevine), then the constraint is not fully learned by the network.","tokens_in":21143,"feed_emoji":"🌿","tokens_out":6548,"duration_ms":56640,"temperature":0.7,"pith_summary":"The paper sets out to solve a mismatch: neural graph generators are good at finding thin structures in images, but nothing forces their output to be a tree, and plant skeletons are trees. TreeFormer addresses this by projecting the network's raw edge predictions onto a minimum spanning tree during every training step, then using a selective feature suppression (SFS) layer to push gradients through that discrete projection. The authors claim this makes the whole pipeline end-to-end trainable while guaranteeing tree-structured output, and that it improves both edge position and topology over unconstrained generation, test-time MST post-processing, and a two-stage method on synthetic trees, real roots, and grapevine branches. If correct, this would give plant scientists a practical single-image tool for measuring branching architecture without manual effort.","feed_headline":"TreeFormer makes graph networks output exact plant-skeleton trees","feed_subtitle":"Projecting edge scores onto minimum spanning trees during training cuts edge errors on grapevine and root images.","key_machinery":"The selective feature suppression (SFS) layer is the load-bearing mechanism: it turns the output of Kruskal's minimum spanning tree, a discrete combinatorial operation, into a differentiable reparameterization. For every edge whose availability the MST flips, the layer overwrites the unwanted logit with $-\\Lambda$, keeps the other logit, and leaves untouched edges unchanged; the training loss is the sum of the usual unconstrained edge cross-entropy and a constrained cross-entropy evaluated on the rewritten probabilities. This lets the gradient descend toward weights whose MST edges agree with the ground-truth tree.","core_discovery":"The central claim is that a non-differentiable combinatorial projection can be absorbed into neural-network training by rewriting only the logits of edges the projection would change. Concretely, the SFS layer computes the unconstrained edge probabilities, runs Kruskal's MST algorithm using non-existence probabilities as costs, and then replaces the suppressed feature of each disagreed edge with $-\\Lambda$ (e.g., $\\Lambda=10$) while leaving the other logit intact. Softmax then produces probabilities that match the MST output, and backpropagation flows through the surviving logit. The authors report that this training-time constraint yields 100 percent tree rate on all three test sets and beats the test-time-MST baseline on street-mover distance and TOPO scores, with the largest gains on grapevine images.","pith_inferences":["The same logit-suppression trick could generalize to any constraint with a fast projection oracle, such as planar graphs, bounded-degree graphs, or graphs with specified connected components, not just trees.","The gradient is a straight-through approximation: it flows only through the surviving logit and ignores how changing edge weights would re-arrange the MST itself, so a decision-focused loss that differentiates through tree selection might be a stronger successor; the paper's Table S5, where train-time-only SFS gives 0 percent tree rate on grapevine, is consistent with that limitation.","Part of the measured gain could come from the extra constrained loss term acting as a regularizer, so an ablation that adds an auxiliary MST-supervised head without rewriting logits would isolate the reparameterization's contribution.","The 100 percent tree guarantee makes longitudinal tracking of root growth from time-series images straightforward, but undetected nodes remain the practical bottleneck for dense canopies."],"forward_implications":["Training with the SFS layer instead of applying MST only at test time reduces SMD from $6.26\\times10^{-6}$ to $4.78\\times10^{-6}$ on synthetic trees, from $1.52\\times10^{-4}$ to $8.82\\times10^{-5}$ on roots, and from $1.47\\times10^{-4}$ to $1.03\\times10^{-4}$ on grapevine.","Every output is guaranteed to be a tree at inference, while unconstrained RelationFormer produces trees only 36.2 percent, 35.9 percent, and 0.0 percent of the time on the three datasets even though all training graphs are trees.","The layer is model-agnostic: attaching it to the autoregressive GGT baseline also improves SMD and TOPO scores over the same model with test-time MST.","Because node detection and edge prediction are optimized together end-to-end, the method avoids the accumulated node-detection errors that hurt the two-stage ViNet-style baseline.","A model trained only on grapevine images generalizes to grapevine photos with natural backgrounds and to other tree species, suggesting the learned representation is not dataset-specific."],"supporting_citations":[{"why":"Supplies the transformer-based graph generator (node detector plus edge predictor) that TreeFormer constrains.","marker":"[55]"},{"why":"Provides the minimum spanning tree algorithm used to project unconstrained graphs onto trees.","marker":"[34]"},{"why":"Provides the grapevine dataset with annotated branch patterns and the two-stage baseline to beat.","marker":"[18]"},{"why":"Serves as the autoregressive graph generator baseline and the source of the SMD evaluation metric.","marker":"[7]"},{"why":"Forms the deformable-DETR node detection backbone inherited by RelationFormer.","marker":"[67]"},{"why":"Provides the L-system grammar used to synthesize the large synthetic tree dataset.","marker":"[42]"}],"fun_headline_variants":["MST-constrained training makes plant skeletons exactly tree-shaped","Plant skeleton AI now outputs true trees via MST projection","TreeFormer enforces tree structure with non-differentiable trick","Training trick yields 100% tree rate for plant skeleton graphs","Absorbing MST into backprop for perfect plant-branch trees"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that backpropagating a gradient through the one surviving feature of each flipped edge is enough for the network to learn edge weights whose minimum spanning tree matches the true tree, even though that gradient ignores how edge-weight changes would change which tree is selected.","fun_headline_variants_meta":{"raw":{"variants":["MST-constrained training makes plant skeletons exactly tree-shaped","Plant skeleton AI now outputs true trees via MST projection","TreeFormer enforces tree structure with non-differentiable trick","Training trick yields 100% tree rate for plant skeleton graphs","Absorbing MST into backprop for perfect plant-branch trees"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000201,"raw_usage":{"total_tokens":1343,"prompt_tokens":872,"completion_tokens":471,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":488,"completion_tokens_details":{"reasoning_tokens":387}},"tokens_in":488,"tokens_out":471,"duration_ms":4448,"temperature":1.0,"reasoning_tokens":387,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T13:31:00.585291+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the SFS layer on a dataset where, at many training examples, the MST projection disagrees with the ground-truth tree on a large fraction of edges; if the learned edge weights do not reduce this disagreement beyond what the unconstrained baseline achieves, the training signal is too weak. A simpler check is the paper's own train-time-only setting: if a model trained with the SFS layer but evaluated without test-time MST still outputs non-tree graphs (0 percent tree rate on grapevine), then the constraint is not fully learned by the network.","supporting_citations":[{"cited_title":"Relation- former: A unified framework for image-to-graph generation","cited_arxiv_id":null,"evidence_quote":"Supplies the transformer-based graph generator (node detector plus edge predictor) that TreeFormer constrains."},{"cited_title":"On the shortest spanning subtree of a graph and the traveling salesman problem","cited_arxiv_id":null,"evidence_quote":"Provides the minimum spanning tree algorithm used to project unconstrained graphs onto trees."},{"cited_title":"Towards smart pruning: ViNet, a deep-learning approach for grapevine structure estima- tion","cited_arxiv_id":null,"evidence_quote":"Provides the grapevine dataset with annotated branch patterns and the two-stage baseline to beat."},{"cited_title":"Image-conditioned graph generation for road network extraction","cited_arxiv_id":null,"evidence_quote":"Serves as the autoregressive graph generator baseline and the source of the SMD evaluation metric."},{"cited_title":"Deformable DETR: Deformable transform- ers for end-to-end object detection","cited_arxiv_id":null,"evidence_quote":"Forms the deformable-DETR node detection backbone inherited by RelationFormer."},{"cited_title":"Mathematical models for cellular in- teractions in development I","cited_arxiv_id":null,"evidence_quote":"Provides the L-system grammar used to synthesize the large synthetic tree dataset."}],"review_version":1}