{"id":"cfc9926a-3a91-49db-9ea3-342825f13b16","arxiv_id":"2502.04852","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A differential age estimator that refines an initial age guess with learned age differences against retrieved reference faces reports MAE 2.47 on MORPH II and 5.27 on CACD.","lead":"This paper presents a face-age estimator that first makes a rough age guess, then refines it by comparing the face with similar-looking reference faces of known ages. The refinement step, plus an error-distribution sampling trick, improves accuracy on two standard age-estimation datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Test-time retrieval sentence in §III-A appears to use ground-truth age for reference selection; if literal, the reported MAE is invalid, and the paper must clarify the typo.","rationale":"The reader's weakest assumption matches my primary concern. The paper's central contribution is a refinement mechanism that must work without knowing the true age at test time. Section III-A is the only place describing the test-time retrieval protocol, and its sentence 'ensuring that baq = aq' contradicts Fig. 2, which explicitly says the BAR estimate âq is used for retrieval. Because the reported improvement over the prior SOTA is small (2.47 vs. 2.53 on MORPH II, 5.27 vs. 5.35 on CACD), even a modest protocol leak could account for the entire gain. The KDE generalization concern is secondary: it affects the magnitude of the benefit but not the validity of the experimental design. Without code or data, the ambiguity cannot be resolved from the paper alone. The correct disposition is therefore to keep the conditional verdict pending the authors' clarification and a reproduction of the test-time retrieval without access to ground-truth ages.","tokens_in":14082,"tokens_out":4695,"duration_ms":46784,"concrete_test":"Independently re-run the MORPH II subject-exclusive evaluation with two retrieval conditions on the same trained DAR: (A) reference age = BAR-predicted age (no Eq. 1 noise), and (B) reference age = ground-truth age. Compare the resulting MAE to the reported 2.47. If condition (A) does not reproduce 2.47 within expected variance, or if condition (B) is substantially better, the central claim depends on the leak. Also request the authors to confirm whether 'aq' in Section III-A is a typo for the predicted age 'âq'.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that at test time the reference set is retrieved using only deployment-available information: the BAR's predicted age and the query embedding. Section III-A states: 'During testing, the same retrieval process is repeated without using Eq. 1, ensuring that baq = aq.' Since Eq. 1 defines baq = aq + ε, omitting Eq. 1 cannot make baq equal the ground-truth age aq; it would make baq the raw BAR estimate. If 'aq' is literally the ground-truth age, then the test protocol uses the answer to choose references, and the reported MAE of 2.47 is not a valid estimate of deployment performance. The DAR refinement's advantage depends on retrieving references whose known ages are near the BAR prediction; if references are instead selected around the true age, the network receives a strong hint about the answer, and the claimed SOTA gain could vanish. The surrounding text and Fig. 2 suggest 'aq' is a typo for the predicted age âq, but as written this is the load-bearing weak point. A secondary concern is the KDE error distribution fitted on a 2% subject-exclusive subset, but the retrieval-protocol ambiguity is primary because it directly threatens the validity of the headline result.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a two-stage framework for facial age estimation. A Baseline Age Regressor (BAR) first produces an initial absolute-age estimate; a Differential Age Regressor (DAR) then refines it by comparing the query face with reference faces of similar age and appearance and estimating age differences. The reference set is retrieved using the BAR estimate and face-embedding nearest-neighbor similarity. Training uses an augmented age aq+epsilon sampled from a KDE-estimated error distribution, and an iterative refinement updates the error distribution over training epochs. Experiments on MORPH II and CACD under the subject-exclusive protocol report MAEs of 2.47 and 5.27, which the authors claim are state-of-the-art, and the paper includes ablations and a bias analysis.","tokens_in":14309,"tokens_out":5545,"duration_ms":56282,"significance":"The central idea of refining an absolute regressor with learned relative-age differences against retrieved reference faces is interesting, timely, and potentially model-agnostic. If the test-time protocol is genuinely deployment-valid, the reported gain over the Hierarchical-Attention BAR is a useful demonstration. The paper has concrete strengths: it uses the subject-exclusive protocol to avoid identity leakage, it isolates components through ablations, it reports results on two public datasets, and it includes a demographic bias analysis. However, the reported gains are small (0.06 MAE over the previous SOTA and 0.03–0.04 in ablations), and the absence of uncertainty estimates means the headline claim is not yet robustly supported. The stress-test concern about the test-time retrieval sentence is valid and must be resolved, because if ground-truth age is used for retrieval the reported MAE is not a valid deployment estimate.","major_comments":[{"comment":"The sentence “During testing, the same retrieval process is repeated without using Eq. 1, ensuring that baq = aq” is self-contradictory. Omitting Eq. 1 leaves baq as the BAR’s raw predicted age, not the ground-truth age aq. If the implementation literally sets baq = aq at test time, references are retrieved using the answer, which would invalidate the reported MAE of 2.47. If this is a typo and the intended statement is that the BAR-predicted age is used for retrieval, the main text must say so explicitly; Fig. 2’s caption suggests the latter, but the current sentence is load-bearing and must be corrected.","section":"Section III-A"},{"comment":"The claimed state-of-the-art result rests on very small MAE differences (2.47 vs 2.53 on MORPH II, and 2.47 vs 2.50 or 2.51 in ablations), yet no error bars, significance tests, or number of repeated runs are reported. Because the reference set is randomly sampled (Section III-A), the method has intrinsic run-to-run variability. Please report mean and standard deviation over multiple seeds and, ideally, paired comparisons against the BAR baseline. Without this, the headline improvement is not statistically established.","section":"Section IV-A / Tables II–VIII"},{"comment":"Several retrieval details needed for reproducibility and for assessing the method are missing. The size K of the initial age-retrieved pool is never specified, and it is unclear how “their age is baq” is interpreted when baq is continuous: exact integer equality, or an age range, and if so what bound? The KDE bandwidth for estimating Dε is not reported, and the 2% subject-exclusive subset used for the estimate is not analyzed for representativeness. Please specify K, the age tolerance, the KDE bandwidth, and the sensitivity of the final MAE to these choices.","section":"Section III-A / Section IV"}],"minor_comments":[{"comment":"The loss La_MSE appears both outside the sum and inside the sum over references, so it is double-counted and the relative weighting in the text is not matched by the equation. Please clarify whether this is a typographical duplication and state the intended loss weighting.","section":"Section III-C, Eq. (5)"},{"comment":"Several author names in the reference list and related-work text appear garbled (e.g., “Wang and Kamikaze” for [10], “Lieu” for [18]/[38], “Rote” for [30], “Tain” for [17]). These should be corrected for professional presentation.","section":"References"},{"comment":"The bias analysis would be more useful with a compact numeric table of MAE by gender and ethnicity; currently the text mentions a 24% gender gap and several qualitative claims, but the supporting numbers appear only in figures with no confidence intervals.","section":"Section V"},{"comment":"The tables contain formatting artifacts such as “2 .47” and “0 .0003”; these should be cleaned up before publication.","section":"Section IV-B, Tables IV–VIII"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and the core idea has merit. The main obstacle is the ambiguous test-time retrieval statement; my reading of Fig. 2 is that the authors intended the BAR prediction to be used, which would make the issue a clarity problem rather than a fatal flaw, but it must be fixed explicitly. I would also like to see variance reporting before acceptance, because the reported gains are small relative to the noise introduced by random reference sampling."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the refinement idea is real, the ablations hold together, and the SE protocol is the right one. But the paper has a sentence in Section III-A that, read literally, uses the true age to retrieve references at test time. That has to be fixed before the 2.47 MAE means anything.\n\nWhat's actually new: instead of predicting absolute age, the DAR predicts age differences between the query and retrieved reference faces, weighted and summed to refine the BAR estimate. The KDE-based sampling of the BAR error distribution and the iterative re-estimation of that distribution are a sensible combination. The authors are careful to use subject-exclusive splits and to compare against published SE results. The ablation tables are internally consistent, and the bias analysis, while descriptive, is honest about the dataset skew. They also use the same VGG-16 backbone as the prior SOTA they compare against, which is fair.\n\nThe soft spots. The test-time sentence in III-A: 'the same retrieval process is repeated without using Eq. 1, ensuring that baq = aq.' Eq. 1 adds noise to the true age for training. If test-time retrieval really uses aq, the references are chosen with knowledge of the answer and the reported MAE is not a deployment estimate. The surrounding text, Fig. 2, and the caption all say the BAR's predicted age is used at test, so I read the sentence as a typo for baq = âq. But as written it's a load-bearing ambiguity, and the authors have to state it in plain words. Second, the reported gains over the prior state of the art are 0.06–0.08 MAE, and there are no error bars, significance tests, or repeated-seed runs. For all the ablations, the differences are small; some configurations in Table IV differ by 0.04. That's the kind of margin that can vanish with a different seed. Third, the KDE error distribution is fit on 2% of the training data; the paper doesn't show how sensitive the results are to that split. None of these is evidence of fabrication, and the central mechanism doesn't rely on a fitted constant, so the circularity burden is low. But they need to ship code or at least repeated-seed results for the claim to be verified.\n\nWho this is for: people working on facial age estimation, and to a lesser extent anyone using retrieval-augmented refinement for continuous regression. It's worth a serious referee, but the current manuscript is not ready without the protocol clarification and variance reporting.","headline":"A plausible differential-refinement scheme with consistent small gains, but a test-time retrieval sentence that looks like ground-truth leakage must be clarified before the headline MAE is credible.","tokens_in":14850,"tokens_out":2934,"would_cite":false,"duration_ms":27743,"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":"Reference-face comparisons trim age error to 2.47 yr","keywords":["facial age estimation","differential regression","reference-based refinement","subject-exclusive protocol","mean absolute error","MORPH II","CACD","deep learning"],"falsifier":"Re-run the MORPH II subject-exclusive evaluation with reference retrieval driven only by the baseline's predicted age -- never by the ground-truth age -- and compare the resulting MAE to the reported 2.47; a material increase would show the published number depended on test answers.","tokens_in":13838,"feed_emoji":"👤","tokens_out":6503,"duration_ms":60287,"temperature":0.7,"pith_summary":"This paper argues that estimating a person's age from a face photo can be made more accurate by turning the task into a relative one: first a baseline regressor gives a rough age, then a second network estimates how much older or younger the face looks compared to a handful of reference faces of known, similar age and similar appearance, and the weighted average of those differences corrects the baseline. The correction is small in magnitude, so the network learns to predict residuals in a narrow range rather than the full span of human aging. The authors report that this differential, reference-based refinement achieves a mean absolute error of 2.47 years on MORPH II and 5.27 years on CACD under the identity-disjoint subject-exclusive protocol, improving on prior state-of-the-art results. Because the refinement is model-agnostic, any absolute age regressor can in principle be improved by wrapping it in this framework.","feed_headline":"Reference-face comparisons trim age error to 2.47 yr","feed_subtitle":"The differential network learns age gaps to matched reference photos, beating prior SE-protocol results on MORPH II and CACD.","key_machinery":"The carrying object is the Differential Age Regressor (DAR), a CNN that consumes the query image and reference images, concatenates their visual embeddings with numeric age embeddings of the known reference ages, and predicts each pairwise age difference as a softmax-weighted mixture of regression heads centered on offsets from $-20$ to $+20$. The final estimate is $\\hat{y} = \\hat{a}_q + \\sum_r w_r d_r$, where $\\hat{a}_q$ is the baseline age and $w_r$ are learned per-reference weights. A second-order refinement regresses residuals around each head's center, and training enforces the symmetry constraint $\\Delta(x_1,x_2) = -\\Delta(x_2,x_1)$ through multi-task losses (MSE, cross-entropy, and mean-variance terms). The retrieval machinery that feeds the DAR uses the baseline prediction for age-based candidate selection, a face-embedding nearest-neighbor step for visual similarity, and kernel density estimation of the baseline's error distribution to sample augmented ages during training.","core_discovery":"The central claim is that the residual error of a baseline age estimator is easier to model than absolute age itself. The paper's Differential Age Regressor (DAR) takes a query face and a retrieved reference set, predicts the age difference between the query and each reference, learns a weight for each reference, and forms the final estimate as the baseline prediction plus the weighted average of the predicted differences. Retrieval is age-constrained and appearance-based: candidates are faces whose known ages lie near the baseline estimate, then the visually closest are selected via face embeddings, with random sampling for diversity. The paper also models the baseline's out-of-sample error distribution by kernel density estimation and samples from it during training, and shows that iterating this refinement lowers the MORPH II MAE from 2.50 to 2.47. In the authors' experiments the full system outperforms the published subject-exclusive results it compares against on both MORPH II and CACD.","pith_inferences":["Editorial inference: the test-time sentence 'ensuring that $\\hat{a}_q = a_q$' is ambiguous. If it means the true age is used to choose reference faces at test time, the reported MAE would not reflect a real deployment; reproducing the protocol requires that only the baseline's predicted age drives retrieval.","Editorial inference: the same relative-refinement recipe could be applied to other fine-grained visual attributes where absolute scoring is noisy but pairwise comparisons are reliable, such as apparent health or facial attractiveness.","Editorial inference: because the framework samples references from a closed training database, its accuracy at the tails of the age range (60+ on MORPH II) appears to be limited by reference coverage as much as by the regressor, based on Table IX."],"forward_implications":["Any baseline age regressor can be plugged into the differential refinement loop, so the reported gain is not tied to one backbone architecture.","Under the subject-exclusive protocol, the paper reports MAE 2.47 on MORPH II and 5.27 on CACD, improving on the listed prior SE-protocol results.","The refinement's benefit depends on the quality of the estimated error distribution; KDE sampling outperformed a uniform distribution and converged faster in the ablation (Table VI).","Iterating the refinement gives a small but consistent accuracy gain, from 2.50 to 2.47 MAE with the KDE-based configuration (Table VIII).","The need for identity labels restricts the method to datasets like MORPH II and CACD where subject-exclusive splits are possible."],"supporting_citations":[{"why":"Supplies the Baseline Age Regressor and the previous SE-protocol SOTA numbers, and defines the evaluation protocol this paper compares against.","marker":"[22]"},{"why":"Provides MORPH II, the primary identity-labeled dataset enabling subject-exclusive training and test splits.","marker":"[26]"},{"why":"Provides CACD, the second identity-labeled evaluation dataset.","marker":"[27]"},{"why":"Supplies the VGG-16 backbone whose convolutional features are used for face embeddings and retrieval.","marker":"[45]"},{"why":"ArcFace loss used to fine-tune the backbone for face recognition before training the regressors.","marker":"[47]"},{"why":"Supplies the kernel density estimation method used to model the baseline error distribution for reference sampling.","marker":"[46]"},{"why":"Contributes the mean-variance loss components used in the DAR training objective.","marker":"[31]"},{"why":"One of the SE-protocol baselines on MORPH II whose published result is compared against to support the SOTA claim.","marker":"[20]"}],"fun_headline_variants":["Compare faces to refine age to 2.47 yr error","Age gaps to lookalikes beat absolute age nets","Retrieval-based age refinement shaves MAE to 2.47","Differential age regressor sets new SOTA on MORPH II","Iteratively sample error to refine face age guesses"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported accuracy assumes that at test time the reference faces are selected using the baseline model's predicted age, not the person's true age, and that the error distribution estimated on a small subject-exclusive subset generalizes to the test set.","fun_headline_variants_meta":{"raw":{"variants":["Compare faces to refine age to 2.47 yr error","Age gaps to lookalikes beat absolute age nets","Retrieval-based age refinement shaves MAE to 2.47","Differential age regressor sets new SOTA on MORPH II","Iteratively sample error to refine face age guesses"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000341,"raw_usage":{"total_tokens":1833,"prompt_tokens":857,"completion_tokens":976,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":473,"completion_tokens_details":{"reasoning_tokens":891}},"tokens_in":473,"tokens_out":976,"duration_ms":9691,"temperature":1.0,"reasoning_tokens":891,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T21:12:11.749003+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the MORPH II subject-exclusive evaluation with reference retrieval driven only by the baseline's predicted age -- never by the ground-truth age -- and compare the resulting MAE to the reported 2.47; a material increase would show the published number depended on test answers.","supporting_citations":[{"cited_title":"Hierarchical Attention-Based Age Estimation and Bias Analysis ,","cited_arxiv_id":null,"evidence_quote":"Supplies the Baseline Age Regressor and the previous SE-protocol SOTA numbers, and defines the evaluation protocol this paper compares against."},{"cited_title":"Morph: a longitudinal image database of normal adult age-progression,","cited_arxiv_id":null,"evidence_quote":"Provides MORPH II, the primary identity-labeled dataset enabling subject-exclusive training and test splits."},{"cited_title":"Cross-age reference coding for age-invariant face recognition and retrieval,","cited_arxiv_id":null,"evidence_quote":"Provides CACD, the second identity-labeled evaluation dataset."},{"cited_title":"Very deep convolutional networks for large-scale image recognition,","cited_arxiv_id":null,"evidence_quote":"Supplies the VGG-16 backbone whose convolutional features are used for face embeddings and retrieval."},{"cited_title":"Arcface: Additive angular margin loss for deep face recognition,","cited_arxiv_id":null,"evidence_quote":"ArcFace loss used to fine-tune the backbone for face recognition before training the regressors."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the kernel density estimation method used to model the baseline error distribution for reference sampling."},{"cited_title":"Mean-variance loss for deep age estimation from a face,","cited_arxiv_id":null,"evidence_quote":"Contributes the mean-variance loss components used in the DAR training objective."},{"cited_title":"Soft-ranking label encoding for robust facial age estimation,","cited_arxiv_id":null,"evidence_quote":"One of the SE-protocol baselines on MORPH II whose published result is compared against to support the SOTA claim."}],"review_version":1}