{"id":"7f0f3662-04f0-46e2-b6be-ea19474e77c8","arxiv_id":"2412.11450","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":8,"one_line_summary":"GroupFace combines a multi-hop attention graph network with a reinforcement-learning margin scheduler for imbalanced face age estimation, reporting modest benchmark gains but with internal inconsistencies in the reward and evaluation tables.","lead":"This paper proposes GroupFace, a face age estimator that combines a multi-hop graph network with a reinforcement-learned margin schedule to improve accuracy on rare age groups. The method reports small accuracy gains on standard face age benchmarks, but the paper contains contradictory equations and evaluation numbers that weaken the claims.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline balanced-performance numbers are internally inconsistent: σ/AAR in Table VI and Section IV-E do not match Eq. (18) when recomputed from the listed group MAEs.","rationale":"The reader's weakest assumption targets Eqs. (11) and (12), where D_inter and D_intra are defined by nearly identical formulas, collapsing the two-term reward into one signal. That is a genuine mechanism-level flaw and would invalidate the claimed balance between inter-class separability and intra-class proximity. However, the central claim under stress is empirical: GroupFace achieves excellent and balanced performance on benchmarks. The most load-bearing evidence for that claim is the set of σ/AAR numbers. A direct arithmetic check shows those numbers are inconsistent with the group MAEs in the same tables and with the text. For MIVIA in Table VI, the group MAEs imply σ ≈ 0.79 and AAR ≈ 7.53, not the table's 1.17 and 7.15; the text's 0.77 and 7.54 are also irreconcilable with the table. For UTK-Face, the group MAEs imply σ ≈ 0.68, not 0.82. The AAR values are computed from σ and MAE by Eq. (18), so a discrepancy in σ propagates directly to the headline metric. Without code or data, no external check is possible, and the paper's own numbers do not cohere. Because the empirical claim fails this internal consistency test, the REJECT verdict stands; fixing the reward definition alone would not resolve this.","tokens_in":23719,"tokens_out":11506,"duration_ms":87900,"concrete_test":"Recompute σ and AAR for every row of Tables VI and VII from the listed group MAEs and overall MAE using Eq. (18), for both population and sample standard deviation. If the table entries and the Section IV-E claim (σ = 0.77, AAR = 7.54) do not match the recomputed values, the balanced-performance metrics are internally inconsistent and the central claim lacks quantitative support.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that GroupFace 'gains balanced performance in long-tailed group estimation,' supported by σ and AAR. The paper's own tables do not support that claim. In Table VI, GroupFace on MIVIA lists group MAEs 3.78, 2.31, 1.58, 2.56 and overall MAE 1.68. Plugging these into Eq. (18) with the population standard deviation gives σ ≈ 0.79 and AAR ≈ 7.53; with the sample standard deviation, σ ≈ 0.92 and AAR ≈ 7.40. The table reports σ = 1.17 and AAR = 7.15, while Section IV-E claims the lowest σ of 0.77 and highest AAR of 7.54. For UTK-Face, the same table lists group MAEs 4.49, 5.91, 4.08, 4.63 and overall MAE 4.32, which give σ ≈ 0.68 and AAR ≈ 5.00 (population), but the table reports σ = 0.82 and AAR = 4.86. Thus the paper's σ and AAR values are not reproducible from its own group MAEs and Eq. (18). The text and the tables also contradict each other. Since no code or data is provided, the reported balanced-performance numbers are the only evidence for the headline claim, and they are internally inconsistent.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GroupFace, a framework for imbalanced age estimation that combines an Enhanced Multi-hop Attention Graph Convolutional Network (EMAGCN) for feature extraction with a reinforcement-learning-based dynamic group-aware margin optimization. The authors claim that the method improves overall accuracy and balances performance across long-tailed age groups, supported by experiments on MORPH II, UTK-Face, ChaLearn LAP 2015, CACD, and MIVIA. The evaluation uses MAE, epsilon-error, and a composite metric AAR that combines overall MAE and the standard deviation sigma of groupwise MAEs.","tokens_in":24063,"tokens_out":9859,"duration_ms":75847,"significance":"If the claims were reliable, GroupFace would offer a useful integration of multi-hop graph attention with RL-tuned margins for long-tailed age estimation, and the paper provides a relatively compact architecture (8.6M parameters) with ablations for its components. The manuscript also makes the goal of balanced group performance explicit through the AAR metric and applies it across several benchmarks. However, the experimental evidence for the central claims is internally inconsistent, and the proposed reward function is not well-defined, so the current contribution is not verifiable from the manuscript.","major_comments":[{"comment":"The reward in Eq. (13) is the sum of two terms intended to quantify inter-class deviation and intra-class deviation, but the definitions do not deliver distinct quantities. Eq. (11) defines d^i_inter as 1/N_i sum_{i=1}^{N_i} max_{k=1:N_i} cos(x_k, x_i), and Eq. (12) defines d^i_intra as 1/N_i sum_{i=1}^{N_i} cos(x_k, x_i) (with the same notational problems: x_i is later called the feature center of group i, yet is also indexed by i in a sum over N_i samples). If x_k and x_i are sample embeddings, both expressions are averages of within-group pairwise cosine similarities; the max in Eq. (11) makes the inter-class term equal to the average of the maximal within-group similarity, which is not an inter-class distance, and if k is allowed to equal i, it is trivially 1. Consequently r(s_t,a_t) cannot balance inter-class separability against intra-class proximity; it is effectively a single within-group dispersion term. The authors should replace these definitions with a genuine between-group distance (e.g., between class centers) and a within-group dispersion, and then re-run the RL training.","section":"III-C, Eqs. (11)-(12)"},{"comment":"The reported sigma and AAR values are inconsistent with Eq. (18) and with each other. For MIVIA in Table VI, GroupFace reports group MAEs (3.78, 2.31, 1.58, 2.56) and overall MAE 1.68, yet the table lists sigma=1.17 and AAR=7.15. Using Eq. (18) with these four group MAEs gives sigma approximately 0.79 (population) or 0.91 (sample) and AAR approximately 7.53 or 7.40, not 7.15. The text in Section IV-E additionally claims 'the lowest sigma of 0.77 and the highest AAR of 7.54', which matches neither the table nor the recomputation. In Table VII, the three group MAEs (2.73, 1.58, 2.56) with overall MAE 1.68 imply sigma approximately 0.51 and AAR approximately 7.81, whereas the table lists sigma=1.71 and AAR=7.15; if sigma=1.71, Eq. (18) yields AAR=6.61. Since these numbers are the primary evidence for the abstract's claim of 'balanced performance in long-tailed group estimation', the discrepancies are load-bearing and must be resolved with corrected tables or corrected text.","section":"IV-E and Tables VI-VII"}],"minor_comments":[{"comment":"Eq. (10) has a typographical issue: the numerator's exponent uses theta_j where the true-class angle theta_{y_i} is intended, and the denominator's summation notation is nonstandard; please clarify the loss function.","section":"III-B, Eq. (10)"},{"comment":"There are two equations labeled (4); the second one, which defines the multi-head self-attention output, should be renumbered to avoid confusion.","section":"III-A, near Eq. (4)"},{"comment":"Eq. (7) refers to (I-Q)^{-1} converging to a series in M; the series should be in Q, and the convergence condition (spectral radius less than 1) should be stated explicitly.","section":"III-A, Eq. (7)"},{"comment":"The sentence claiming 'the lowest sigma of 0.77 and the highest AAR of 7.54' does not match any value in Table VI or Table VII; please align the text with the tables or add a table that contains those results.","section":"IV-E"},{"comment":"The dataset is called 'CLAP15' in Table I but 'CLAP 2015' in Section IV-D; please unify the naming.","section":"IV-C and Table I"},{"comment":"Several hyperparameters of the DQN and the margin strategy (kappa, gamma, lambda, n_D, n_M) are not specified; please provide them, along with the train/test splits and random seeds, to support reproducibility.","section":"IV-C and III-C"},{"comment":"The Acknowledgement section thanks the anonymous reviewers and the AE; this is inappropriate in a submitted manuscript and should be removed before submission.","section":"VI"}],"recommendation":"reject","confidential_remarks":"The paper leans heavily on self-citations of the authors' own prior work in the introduction and related work; some are peripheral to age estimation. More importantly, the internal inconsistencies in the main experimental tables are severe enough that I cannot recommend acceptance; the authors would need to provide code or detailed experiment logs to restore confidence. The acknowledgment of reviewers and the AE is also inappropriate for a submission and suggests a lack of editorial polish."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short take: this is a plausible engineering combo with real benchmarks and a thorough ablation, but the paper's own numbers and equations don't hold together. The text claims MIVIA has σ=0.77/AAR=7.54; Table VI gives 1.17/7.15; Table VII lists σ=1.71 alongside an AAR that actually corresponds to σ=1.17. The stress-test note says Table VI doesn't match Eq. (18) when recomputed, but that calculation used the mean of group MAEs as the center. Eq. (18) uses the overall test-set MAE; with that center, the four group MAEs in Table VI give σ≈1.18 and AAR≈7.14, so Table VI is internally consistent. The real inconsistency is between the text and tables, plus the σ=1.71 in Table VII. That's a serious reporting problem, not a cosmetic typo, because the headline claim rests on those balanced-performance numbers.\n\nOn the method: the components are mostly borrowed — MAGNA-style multi-hop attention diffusion, DropMessage, X2-Softmax-type margin loss, and Fair Loss/RL-RBN-style RL margin selection. The novelty is the specific combination and the four-age-group margin scheme. That's incremental but not vacuous; the ablation shows each piece contributes, and the parameter count is genuinely lower than the ResNet baselines. The MAE gains over the one-hop baseline are consistent across datasets.\n\nThe real soft spot is the RL reward. Eqs. (11) and (12) are not well-formed: the index i is used both for groups and for samples within a group, x_i is called the feature center in one place and a sample vector in another, and the “inter” vs “intra” distinction is not convincing as written. Even if the formulas are not literally the same, they don't establish two independent reward terms. That matters because the paper's story is that RL balances inter-class separability and intra-class proximity; that story is unsupported as written. Also, Eq. (7) hand-waves the power iteration and the claimed linear complexity, with no derivation. And there's no code or data release, so none of the reported numbers can be independently checked.\n\nWho gets value: people working on long-tailed age estimation or margin learning with RL will find the ablation and design choices useful as ideas, but not as a reliable benchmark. It deserves a serious referee — the system is coherent and the experiments are broad — but the current version should not be accepted. I'd tell the authors to fix the equations, reconcile σ/AAR across text and tables, and release code and data.","headline":"Plausible incremental system with a careful ablation, but the headline balanced-performance numbers don't reconcile across text and tables and the RL reward equations are too muddled to support the mechanism.","tokens_in":24632,"tokens_out":8143,"would_cite":false,"duration_ms":64349,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"GroupFace claims that training a multi-hop attention graph convolutional feature extractor with a reinforcement-learning-tuned per-group margin loss improves overall age-estimation accuracy while simultaneously narrowing the per-group…","keywords":["age estimation","long-tailed recognition","class imbalance","graph convolutional network","multi-hop attention","reinforcement learning","margin optimization","face analysis"],"falsifier":"Implement Eqs. (11) and (12) verbatim on any trained feature set: both formulas compute the same average cosine distance from each sample to its group center, so $D_{\\text{intra}} \\equiv D_{\\text{inter}}$ and the reward $r = -(D_{\\text{intra}} + D_{\\text{inter}})$ reduces to a single term; a GroupFace run whose reward keeps only one of the two terms should then produce the same margin policy as the full reward, showing that the balancing mechanism as written has no separate intra-class versus inter-class signal. Separately, recompute $\\sigma$ and AAR from the per-group MAE values in Table VI for MIVIA: those table entries give $\\sigma = 1.17$ and AAR $= 7.15$, while Section IV-E claims the lowest $\\sigma$ of 0.77 and highest AAR of 7.54, so reproducing the text's headline numbers requires checking the released model and evaluation protocol.","tokens_in":23492,"feed_emoji":"🎯","tokens_out":15984,"duration_ms":123283,"temperature":0.7,"pith_summary":"The paper proposes GroupFace, a framework for age estimation that treats the long-tail problem as two coupled tasks: learning features that separate age groups, and giving each group a classification margin matched to its sample count. The feature side is an enhanced multi-hop attention graph convolutional network (EMAGCN) that models a face as a graph of patches and fuses local and global aging cues with adaptive per-hop decay, random message dropping, and residual connections. The margin side is a deep-Q-learning agent that decides, for the children, teenager, adult, and senior groups, whether to shrink, keep, or expand each group's margin, with a reward meant to balance within-group compactness and between-group separability. The paper reports that joint training improves overall mean absolute error on MORPH II, UTK-Face, ChaLearn LAP 2015, and CACD, and, more distinctively, reduces the spread of per-group errors, reporting the lowest $\\sigma$ of 0.77 and highest AAR (accuracy-and-regularity metric) of 7.54 on MIVIA. If this holds, it gives practitioners a single end-to-end model that needs no separate balanced-sampling stage and reaches competitive accuracy with far fewer parameters than the CNN and ViT baselines it compares against.","feed_headline":"RL-tuned margins shrink long-tail age-estimation bias","feed_subtitle":"Multi-hop graph features plus per-group margins cut rare-age error while keeping overall accuracy.","key_machinery":"The machinery is a two-component joint learning loop. EMAGCN (Enhanced Multi-hop Attention Graph Convolutional Network) is the feature extractor: it builds a k-NN patch graph from a face image, computes a softmax attention matrix, diffuses it over K hops with adaptive learnable decay weights $\\delta_k$ rather than fixed decay, applies DropMessage to the message matrix, adds residual connections to prevent over-smoothing, and approximates the dense inverse with a power-iteration method so the higher-order model stays linear in cost. DGMO (Dynamic Group-aware Margin Optimization) is the controller: a deep Q-network observes state $s_t = \\{G, D_{\\text{inter}}, M\\}$, where $G$ is one of four age groups, $D_{\\text{inter}}$ is the deviation of inter-class distance from the head adult group, and $M$ is the current margin; it picks an action in $\\{-1, O, +1\\}$ that shrinks, keeps, or expands the margin by a constant $\\kappa$, and receives reward $r = -(D_{\\text{intra}} + D_{\\text{inter}})$, intended to tie the choice to both within-group cohesion and between-group separation. The chosen margins feed the dynamic group-aware margin loss $L_{\\text{DGM}}$, which replaces the cosine term in the softmax large-margin loss with $e^{a_i(t)(\\theta_j - h_i(t))^2 + k_i(t)}$, letting three learnable parameters per group reshape intra-class compactness and inter-class separation. The joint loss $\\lambda L_{\\text{CE}} + (1-\\lambda) L_{\\text{MAE}}$ couples the margin policy to final age accuracy.","core_discovery":"On its own terms, GroupFace's claim is that representation learning and margin calibration should be solved together, not in decoupled stages. The Enhanced Multi-hop Attention Graph Convolutional Network (EMAGCN) turns a face into a graph of image patches, computes attention at multiple hop distances with learnable per-hop decay weights so distant nodes contribute useful context rather than noise, drops messages randomly (DropMessage) to prevent over-fitting, and adds residual connections plus a power-iteration inverse approximation to keep deep training stable with linear complexity. The Dynamic Group-aware Margin Optimization (DGMO) partitions ages into children, teenager, adult, and senior, keeps the adult (head) margin as a fixed anchor, and runs deep Q-learning over a Markov decision process whose state is the triple {group, inter-class deviation, current margin}, with actions $\\{-1, O, +1\\}$ and reward $r = -(D_{\\text{intra}} + D_{\\text{inter}})$. The resulting margins enter a quadratic large-margin loss with per-group parameters $a_i(t)$, $h_i(t)$, $k_i(t)$, and the whole system is trained with $\\lambda L_{\\text{CE}} + (1-\\lambda) L_{\\text{MAE}}$. The paper reports that this joint optimization reaches competitive MAE across MORPH II, UTK-Face, ChaLearn LAP 2015, and CACD, the best $\\epsilon$-error of 0.239 on CLAP2015, and balanced long-tailed accuracy with the lowest $\\sigma$ of 0.77 and highest AAR (Age Accuracy and Regularity) of 7.54 on MIVIA, which it interprets as reducing representation bias and margin deviation simultaneously.","pith_inferences":["A natural next test is to freeze the converged per-group margin parameters and re-train without the DQN; if the AAR gains persist, the agent itself is not the active ingredient and the margin-loss shape or the EMAGCN features carry the improvement.","Because the stated reward collapses to a single term if Eqs. (11) and (12) are identical, the mechanism narrative is stronger than the equations; the inter/intra balancing should be re-derived with genuinely distinct distance measures, such as mean pairwise cosine distance for intra-class scatter and center-to-center cosine distance for inter-class separation.","The four-group discretization throws away fine-grained age structure; the same RL loop could instead output continuous margin offsets per estimated age, and the framework would apply to other skewed biometric attributes such as ethnicity, gender, or expression.","The AAR metric mixes accuracy (70%) and regularity (30%), so part of the reported 'balanced' improvement on MIVIA could come from a lower overall MAE rather than from leveling group errors; the per-group MAE tables, which the paper does provide, are the safer evidence."],"forward_implications":["The same end-to-end model can improve overall MAE and reduce per-group error spread simultaneously, eliminating the need for a separate balanced-sampling stage or a decoupled classifier retraining step.","A graph-based face representation with about 8.6M parameters can match or approach 138M-parameter CNN and ViT baselines on age benchmarks, making the accuracy-per-parameter trade-off favor the graph model.","The RL margin policy learned on one long-tailed age distribution carries over to other datasets with different group skews, shrinking $\\sigma$ and raising AAR on UTK-Face and MIVIA.","On ChaLearn LAP 2015 the method reports the lowest $\\epsilon$-error of 0.239, indicating it tolerates apparent-age label variance, not just class-count imbalance.","Adaptive-decay multi-hop diffusion with DropMessage and residual connections is claimed to prevent over-smoothing and distant-node noise, a design that transfers to graph networks used for other irregular visual inputs."],"supporting_citations":[{"why":"Supplies the multi-hop attention diffusion mechanism that EMAGCN extends with adaptive decay weights.","marker":"[51]"},{"why":"Provides the graph attention coefficient computation in Eq. (1) used to build the attention matrix.","marker":"[47]"},{"why":"Introduces the reinforcement-learning margin-adjustment paradigm with a head-class anchor that DGMO adapts to four age groups.","marker":"[58]"},{"why":"Precedent for margin-aware reinforcement learning as a loss; its fair-loss formulation motivates the group-aware margin loss.","marker":"[61]"},{"why":"The quadratic large-margin loss form that the dynamic group-aware margin loss $L_{\\text{DGM}}$ follows with per-group parameters.","marker":"[67]"},{"why":"The DropMessage strategy used to randomly drop messages and prevent over-fitting during graph propagation.","marker":"[63]"},{"why":"Provides the MIVIA dataset and the AAR evaluation metric used to measure balanced long-tailed performance.","marker":"[28]"},{"why":"The two-stage long-tailed age estimation framework whose AAR protocol is used for group-level evaluation.","marker":"[29]"},{"why":"The GLAE method is the main long-tailed age-estimation competitor on MIVIA and the comparison target for AAR.","marker":"[59]"}],"fun_headline_variants":["Multi-hop graph + RL margins fix long-tail age bias","Graph attention and RL margins shrink age estimation bias","GroupFace: joint graph and RL margin learning beats age imbalance","RL-tuned grouping cuts rare-age error without sacrificing accuracy","EMAGCN and DGMO: balanced age estimation for all groups"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The balancing story rests on the RL reward being a meaningful two-term signal $r = -(D_{\\text{intra}} + D_{\\text{inter}})$; in Section III-C, Eq. (11) and Eq. (12) define $D_{\\text{inter}}$ and $D_{\\text{intra}}$ with identical formulas, so the two terms collapse and the reward cannot distinguish within-group compactness from between-group separability.","fun_headline_variants_meta":{"raw":{"variants":["Multi-hop graph + RL margins fix long-tail age bias","Graph attention and RL margins shrink age estimation bias","GroupFace: joint graph and RL margin learning beats age imbalance","RL-tuned grouping cuts rare-age error without sacrificing accuracy","EMAGCN and DGMO: balanced age estimation for all groups"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001029,"raw_usage":{"total_tokens":4451,"prompt_tokens":1176,"completion_tokens":3275,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":792,"completion_tokens_details":{"reasoning_tokens":3193}},"tokens_in":792,"tokens_out":3275,"duration_ms":20770,"temperature":1.0,"reasoning_tokens":3193,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T14:56:28.572560+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Implement Eqs. (11) and (12) verbatim on any trained feature set: both formulas compute the same average cosine distance from each sample to its group center, so $D_{\\text{intra}} \\equiv D_{\\text{inter}}$ and the reward $r = -(D_{\\text{intra}} + D_{\\text{inter}})$ reduces to a single term; a GroupFace run whose reward keeps only one of the two terms should then produce the same margin policy as the full reward, showing that the balancing mechanism as written has no separate intra-class versus inter-class signal. Separately, recompute $\\sigma$ and AAR from the per-group MAE values in Table VI for MIVIA: those table entries give $\\sigma = 1.17$ and AAR $= 7.15$, while Section IV-E claims the lowest $\\sigma$ of 0.77 and highest AAR of 7.54, so reproducing the text's headline numbers requires checking the released model and evaluation protocol.","supporting_citations":[{"cited_title":"Fair loss: Margin-aware reinforcement learning for deep face recognition,","cited_arxiv_id":null,"evidence_quote":"Precedent for margin-aware reinforcement learning as a loss; its fair-loss formulation motivates the group-aware margin loss."},{"cited_title":"X2-softmax: Margin adaptive loss function for face recognition,","cited_arxiv_id":null,"evidence_quote":"The quadratic large-margin loss form that the dynamic group-aware margin loss $L_{\\text{DGM}}$ follows with per-group parameters."},{"cited_title":"Dropmessage: Unifying random dropping for graph neural networks,","cited_arxiv_id":null,"evidence_quote":"The DropMessage strategy used to randomly drop messages and prevent over-fitting during graph propagation."},{"cited_title":"Effective training of convolutional neural networks for age estimation based on knowledge distillation,","cited_arxiv_id":null,"evidence_quote":"Provides the MIVIA dataset and the AAR evaluation metric used to measure balanced long-tailed performance."},{"cited_title":"Lae: Long- tailed age estimation,","cited_arxiv_id":null,"evidence_quote":"The two-stage long-tailed age estimation framework whose AAR protocol is used for group-level evaluation."},{"cited_title":"General vs. long- tailed age estimation: An approach to kill two birds with one stone,","cited_arxiv_id":null,"evidence_quote":"The GLAE method is the main long-tailed age-estimation competitor on MIVIA and the comparison target for AAR."}],"review_version":1}