{"id":"195be0d4-00e0-4e74-841e-dc6628d9b005","arxiv_id":"2501.13420","paper_version":3,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"LVFace is a ViT-based face recognition model whose Progressive Cluster Optimization three-stage training beats prior CNN and ViT baselines on MFR-Ongoing, IJB-B, and IJB-C.","lead":"A ByteDance team trains Vision Transformer models for face recognition with a three-stage curriculum that first aligns features, then stabilizes class centroids, then sharpens cluster boundaries, and reports state-of-the-art scores on IJB-C, IJB-B, and the MFR-Ongoing benchmark. The method matters because it suggests CNN-style loss recipes are not enough for transformer face recognition and offers a staged alternative that practitioners can copy.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Cosine Stage Scheduler monitors similarity to classifier weights, but Stage 2 optimizes feature-expectation prototypes; without evidence that these track each other, the fixed thresholds δ1=0.2, δ2=0.35 may not transfer, leaving the stage-wise gains and final SOTA unverified.","rationale":"The central claim is that LVFace achieves SOTA through a three-stage Progressive Cluster Optimization governed by the Cosine Stage Scheduler. The reader identified the scheduler's reliability and threshold transferability as the weakest assumption; my reading agrees. In fact, there is a sharper version of the concern: the scheduler in Eq. 13 measures alignment with classifier weights w_yi, but the centroid-stabilization loss (Eqs. 9–11) centers on feature-expectation prototypes e_i. The paper never establishes that w_yi and e_i move together under the proposed update rules, nor does it justify that the chosen squared-cosine thresholds (0.2 and 0.35) are intrinsic rather than artifacts of the authors' hyperparameter choices. Without this justification, the ablation gains in Table 5 could be sensitive to the exact threshold values or to the transition dynamics, making the method difficult to reproduce on other backbones or datasets. This is a correctness risk, but it is not a fundamental invalidation: the method is plausible, the ablations show consistent monotone improvements, and comparisons with TransFace at matched backbone sizes in Table 2 give partial support. Thus the verdict should remain conditional, pending the concrete test above. I agree with the reader's identification of the same weakest assumption.","tokens_in":13480,"tokens_out":6816,"duration_ms":56932,"concrete_test":"Run the full LVFace-L training on Glint360K with δ1 and δ2 swept over a small grid, e.g., δ1 ∈ {0.1, 0.2, 0.3}, δ2 ∈ {0.25, 0.35, 0.45}, fixing all other hyperparameters, and re-measure IJB-C TAR@1e-5 and MFR-All. If final accuracy varies by more than 0.2% across the grid, the thresholds are load-bearing; also record s(t) curves over training to check monotonicity and the actual transition epochs. Alternatively, replace the s(t) criterion with fixed epochs (e.g., 40/80/120) and compare: if the threshold-free schedule matches the PCO results, the specific CSS mechanism is not essential to the reported gains.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.3 defines the scheduler signal s(t) (Eq. 13) as the batch mean of squared cosine similarity between features x_i and classifier weights w_yi. However, the Centroid Stabilization stage (Eqs. 9–11) explicitly replaces the centroid with a feature expectation e_i, updated by a moving average with similarity-dependent coefficient α_i (Eq. 10). The paper provides no analysis, bound, or experiment showing that similarity to w_yi is a reliable proxy for cluster compactness around e_i. The thresholds δ1=0.2 and δ2=0.35 are fixed globally and described as empirical; the paper does not report s(t) curves, the epochs at which transitions fire, or the sensitivity of final metrics to these thresholds. Because the central claim includes stage-wise gains in Table 5 (e.g., MR-All 97.27→98.49), if the scheduler fires prematurely or late on another dataset or backbone, the PCO curriculum is not reproduced. Algorithm 1 also breaks the stage loop when s(t) ≥ δ, but since s(t) is a stochastic batch average, it may fluctuate around the threshold, causing unstable transitions. The paper does not specify the transition check frequency, how the batch-size schedule (384→128 at epoch 60, Sec. 5.2) aligns with stage boundaries, or the activation σ in Eq. 10. These under-specifications make the scheduler the least-secure link between the described method and the reported SOTA.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes LVFace, a ViT-based face recognition model that replaces the standard single-stage margin-based training with a three-stage Progressive Cluster Optimization (PCO): feature alignment with negative class sub-sampling, centroid stabilization via feature-expectation penalties, and boundary refinement with the full negative class set. A Cosine Stage Scheduler (CSS) monitors the batch-averaged squared cosine similarity between features and classifier weights and switches stages when fixed thresholds are reached. The authors report state-of-the-art accuracy on MFR-Ongoing, IJB-B, and IJB-C using ViT-S/B/L backbones trained on Glint360K or WebFace42M, with ablations showing monotonic gains from each PCO stage.","tokens_in":13884,"tokens_out":7304,"duration_ms":58144,"significance":"If the reported results are reproducible, the paper offers a practical contribution to ViT-based face recognition: it shows that a staged curriculum with explicit stabilization can reduce training instability and improve accuracy over single-stage CosFace/ArcFace baselines. The public code release and the monotonic stage-wise ablations (Table 5) are strengths. However, the central quantitative claims depend on the Cosine Stage Scheduler, whose monitoring signal is not shown to track the feature-expectation prototypes used in Stage 2, and several hyperparameters (m1, m2, sigma, transition thresholds) are not specified. These are serious but fixable reproducibility issues; they do not undermine the core idea itself.","major_comments":[{"comment":"The scheduler signal s(t) is the mean squared cosine similarity between features x_i and classifier weights w_yi, but Stage 2 (Eqs. 9-11) optimizes feature-expectation prototypes e_i, updated by a moving average with coefficient sigma(sim(e_i, x_i)). The paper provides no analysis or experiment showing that similarity to w_yi is a reliable proxy for compactness around e_i. The thresholds delta1=0.2 and delta2=0.35 are stated as fixed, but no s(t) curves, transition epochs, or sensitivity analysis are reported. Because Table 5 attributes a 1.22-point MR-All improvement to the stage transitions, the absence of this evidence makes the scheduler the least-secure link between the method and the reported SOTA. Moreover, since s(t) is a stochastic batch average, Algorithm 1's break condition may fire unstably; the transition check frequency is also unspecified.","section":"Sec. 4.3, Eq. (13)"},{"comment":"The method as written is not fully reproducible. The margin hyperparameters m1 and m2 in Eqs. (11) and (12) are never given values, and the activation sigma in Eq. (10) is not defined. Eq. (11) appears to have a misplaced fraction, making the loss formula ambiguous. Algorithm 1 calls NCS with Y_train rather than the class count C, and the Stage 3 loop samples features for i in [1,C] per batch, which contradicts the batch size of 128 stated in Sec. 5.2. The text also alternates between 'full-batch' and 'full-sample' training in Stage 3. These omissions and inconsistencies must be corrected before the stage-wise gains in Table 5 can be independently verified.","section":"Sec. 4.2, Eqs. (10)-(12) and Algorithm 1"},{"comment":"All reported numbers come from single training runs with no error bars or significance tests, and the improvements are often small (e.g., IJB-C 1e-5 97.00 vs 97.02 in Table 3; IJB-C 1e-5 97.25 vs 97.23 in Table 1). The paper does not state whether the ViT-L baseline and the PCO variants use the same total iteration budget and the same batch-size schedule, so the cumulative gains in Table 5 might partly reflect extra training rather than the PCO stages themselves. A compute-controlled comparison with multiple seeds (or at least multiple evaluation trials) is needed to substantiate the claim that each PCO stage adds measurable, significant accuracy.","section":"Sec. 5.4, Tables 3-5"}],"minor_comments":[{"comment":"The abstract contains a spacing typo: 'instability.To address this challenge' should be 'instability. To address this challenge'.","section":"Abstract"},{"comment":"The notation for the squared cosine similarity is awkward; please write it as (cos theta_yi)^2 or define it explicitly. Also, 'mean-square' should be 'mean squared'.","section":"Eq. (13)"},{"comment":"The subroutine call NCS(Y_train; C, r) is inconsistent with the definition NCS(C, r) = C*r in Eq. (7); it should be NCS(C, r).","section":"Algorithm 1"},{"comment":"The description 'two fully-connected layers (512-d each)' is unclear; the embedding dimension d is not defined, and the meaning of '512-d' should be clarified.","section":"Sec. 5.2"},{"comment":"The abstract and contributions claim 'compatibility with mainstream VLMs and LLMs,' but no experiment or analysis in the paper supports this claim; either add supporting experiments or remove the claim.","section":"Abstract and Sec. 1"},{"comment":"The statement that PCO has 'comparable total training costs' is not backed by actual training time or FLOPs measurements; please provide a quantitative comparison.","section":"Sec. 5.5"},{"comment":"The statement that LVFace 'ranks first' on the MFR-Ongoing leaderboard is time-sensitive; please add the date and note that leaderboard positions may change.","section":"Sec. 5.3.1"}],"recommendation":"major_revision","confidential_remarks":"The under-specification of m1, m2, sigma, and the scheduler thresholds is the main barrier to reproducibility. I would suggest the editor ask for an appendix with the exact training configuration and s(t) curves. The VLM/LLM compatibility claim in the abstract is not backed by experiments and should be de-emphasized."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"LVFace is a ViT face recognition training recipe with a three-stage curriculum: negative class sub-sampling with CosFace for early alignment, feature-expectation centroid stabilization borrowed from EPL, and final full-batch boundary refinement. The contribution is the staged arrangement plus the Cosine Stage Scheduler (CSS) controlling transitions. The paper reports consistent stage-wise gains in Table 5 (MR-All from 97.27 to 98.49) and SOTA numbers on IJB-C/B and MFR-Ongoing for ViT backbones. That combination is not in the cited literature.\n\nWhat's good: the ablations are clean. Each stage adds a measurable gain, and the gains replicate across backbone sizes and datasets. The paper is honest about TopoFR being better on the Mask subset. The result is a real engineering recipe that likely will be useful to practitioners wanting to train ViTs for FR.\n\nSoft spots: the CSS is under-specified. s(t) in Eq. 13 is the mean-square cosine similarity between sample features and the classifier weight w_yi. But Stage 2's loss (Eq. 11) uses the feature expectation e_i as the centroid, and the adaptive update rule in Eqs. 9–10 defines a different prototype. The paper provides no evidence that similarity to w_yi tracks compactness around e_i. The thresholds 0.2 and 0.35 are fixed and empirical, with no s(t) curves, transition epochs, or sensitivity analysis. Since s(t) is a stochastic batch average, the break condition in Algorithm 1 could fire prematurely or late due to fluctuation, and the paper does not specify the check frequency. It also doesn't say how the batch-size schedule (384 then 128 at epoch 60) aligns with data-dependent stage boundaries. For a recipe paper, that's a reproducibility issue, not a fatal flaw.\n\nMinor: m1 and m2 in Eqs. 11–12 are never given numeric values, and the activation sigma in Eq. 10 is undefined. The referee should ask for these.\n\nVerdict: a solid within-subfield contribution with a plausible mechanism and real empirical support. It changes practice more than theory. A serious referee should engage; the required revisions are mainly specification and analysis of the scheduler. With code/checkpoints and scheduler curves, the SOTA claim would be credible.","headline":"A solid, well-executed ViT face recognition recipe with a three-stage curriculum and a Cosine Stage Scheduler, but the scheduler is under-specified and its monitoring signal may not match what Stage 2 optimizes; still worthy of peer review.","tokens_in":14361,"tokens_out":3044,"would_cite":true,"duration_ms":27444,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A Vision Transformer trained with three progressive stages surpasses CNN-based face recognition baselines, scoring 98.49% on MFR-Ongoing.","keywords":["face recognition","vision transformer","progressive cluster optimization","negative class sub-sampling","cosine stage scheduler","centroid stabilization","boundary refinement","MFR-Ongoing benchmark"],"falsifier":"Train LVFace on a different large-scale face dataset (for example WebFace260M or MS1MV3) with the same δ1=0.2 and δ2=0.35, logging when each stage activates and the final MFR-Ongoing and IJB-C accuracy; if the stages trigger too early or too late, or if s(t) is non-monotone for some backbone, the scheduler's transferability claim fails. A second test: re-run the Table 5 ablation with the thresholds replaced by per-dataset tuned values and check whether the reported gains shrink.","tokens_in":13306,"feed_emoji":"👤","tokens_out":6398,"duration_ms":60854,"temperature":0.7,"pith_summary":"This paper claims that the reason Vision Transformers underperform in face recognition is not the architecture itself but the single-stage, CNN-era training recipe applied to it. It introduces LVFace, which splits training into three explicit phases: feature alignment using a CosFace loss on a subsample of negative classes, centroid stabilization using a feature-expectation penalty, and boundary refinement using all classes. A Cosine Stage Scheduler decides when to move between phases by monitoring the average squared cosine similarity between features and their class centers. With this schedule, LVFace-L reaches 98.49% MR-All on MFR-Ongoing and outperforms leading CNN and ViT baselines on IJB-C and IJB-B.","feed_headline":"LVFace's staged training wins MFR-Ongoing at 98.49%","feed_subtitle":"Progressive cluster optimization lifts ViT past UniFace and TopoFR on masked, cross-race, and IJB benchmarks.","key_machinery":"PCO is carried by three mechanisms: negative class sub-sampling (NCS), which keeps only a random proportion r=0.1 of negative classes to make early feature alignment robust; a feature-expectation prototype e_i, updated as e_new = α e_old + (1-α)x_i with α=σ(cos(θ_e)), which anchors each class centroid and prevents hard positives from dragging it away; and a final full-batch stage that releases all negatives to tighten cluster boundaries. The Cosine Stage Scheduler (CSS) controls the transitions by computing s(t), the mean squared cosine similarity between features and their class centers, and switches stages when s(t) crosses fixed thresholds δ1=0.2 and δ2=0.35.","core_discovery":"LVFace's central claim is that a three-stage Progressive Cluster Optimization (PCO) schedule unlocks the accuracy that Vision Transformer backbones already contain for face recognition, making them competitive with or better than CNN-based state-of-the-art. On WebFace42M, LVFace-L reports 98.49% MR-All on MFR-Ongoing, 97.25% TAR@FAR=1e-5 on IJB-C, and, trained on Glint360K, 96.51% TAR@FAR=1e-4 on IJB-B; the same model edges out Partial FC, UniFace, TopoFR, and TransFace on these benchmarks. The ablation in Table 5 shows each phase matters: the complete PCO raises MFR-All from 97.27% for a single-stage ViT-L to 98.49%.","pith_inferences":["The same cosine-similarity staging signal could serve as a generic curriculum trigger for other large-class metric-learning problems, though the fixed thresholds would probably need re-calibration per task.","Because PCO keeps the backbone untouched, it could be combined with vision-language pretraining; whether the schedule still helps when starting from CLIP-style weights is a testable extension.","The fixed thresholds δ1=0.2 and δ2=0.35 look dataset-dependent; a relative-progress criterion, such as switching when s(t) plateaus, might make the method more portable.","The claim that CNN-inspired training paradigms 'fail to unlock ViT's potential' could be tested directly by running the same PCO recipe on a CNN backbone; the paper does not report that comparison."],"forward_implications":["A ViT encoder trained with PCO can replace CNN backbones in face-recognition systems without changing the inference architecture or latency.","Adding data scale compounds the benefit: LVFace-L improves on IJB-C TAR@1e-5 from 97.02% (Glint360K) to 97.25% (WebFace42M).","The three-stage schedule is loss-agnostic: PCO improves both ArcFace and CosFace baselines, with CosFace giving the largest gains.","Each PCO stage contributes on MFR-Ongoing; removing boundary refinement alone drops MR-All from 98.49% to 98.22%.","As of March 2025, LVFace ranks first on the academic track of the MFR-Ongoing challenge leaderboard."],"supporting_citations":[{"why":"Supplies the negative class sub-sampling (Partial FC) strategy that PCO's first stage builds on.","marker":"[2]"},{"why":"CosFace loss is the base margin loss used in all three PCO stages.","marker":"[36]"},{"why":"Introduces the feature-expectation prototype update that anchors centroids in stage two.","marker":"[15]"},{"why":"WebFace42M is the large-scale training dataset for LVFace-L.","marker":"[50]"},{"why":"Defines the MFR-Ongoing benchmark and the evaluation protocol for masked, cross-race, and age-gap verification.","marker":"[10]"},{"why":"IJB-C is one of the three benchmarks used to measure verification accuracy.","marker":"[27]"},{"why":"IJB-B is the third benchmark, focused on image-to-video matching.","marker":"[39]"},{"why":"TransFace is the prior ViT-based FR method that LVFace compares against and improves on.","marker":"[6]"},{"why":"UniFace is a state-of-the-art baseline whose scores LVFace surpasses on MFR-Ongoing and IJB-C.","marker":"[48]"},{"why":"TopoFR is the leading baseline on the MFR mask subset that LVFace's overall MR-All surpasses.","marker":"[7]"}],"fun_headline_variants":["Staged training gives ViT face recognition win at 98.49%","ViT for face recognition: progressive clusters win at 98.49%","LVFace: three-phase training tops face recognition benchmarks","How progressive clusters put ViT ahead in face recognition","98.49%: LVFace's staged training wins MFR-Ongoing"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The scheduler assumes that the mean squared cosine similarity between features and class centers rises smoothly and monotonically during training, and that fixed thresholds 0.2 and 0.35 transfer across datasets and backbone sizes.","fun_headline_variants_meta":{"raw":{"variants":["Staged training gives ViT face recognition win at 98.49%","ViT for face recognition: progressive clusters win at 98.49%","LVFace: three-phase training tops face recognition benchmarks","How progressive clusters put ViT ahead in face recognition","98.49%: LVFace's staged training wins MFR-Ongoing"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000696,"raw_usage":{"total_tokens":3148,"prompt_tokens":948,"completion_tokens":2200,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":564,"completion_tokens_details":{"reasoning_tokens":2108}},"tokens_in":564,"tokens_out":2200,"duration_ms":13789,"temperature":1.0,"reasoning_tokens":2108,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T15:57:50.585018+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train LVFace on a different large-scale face dataset (for example WebFace260M or MS1MV3) with the same δ1=0.2 and δ2=0.35, logging when each stage activates and the final MFR-Ongoing and IJB-C accuracy; if the stages trigger too early or too late, or if s(t) is non-monotone for some backbone, the scheduler's transferability claim fails. A second test: re-run the Table 5 ablation with the thresholds replaced by per-dataset tuned values and check whether the reported gains shrink.","supporting_citations":[{"cited_title":"Killing two birds with one stone: Efficient and robust training of face recognition cnns by partial fc","cited_arxiv_id":null,"evidence_quote":"Supplies the negative class sub-sampling (Partial FC) strategy that PCO's first stage builds on."},{"cited_title":"Cosface: Large margin cosine loss for deep face recognition","cited_arxiv_id":null,"evidence_quote":"CosFace loss is the base margin loss used in all three PCO stages."},{"cited_title":"Webface260m: A benchmark for million- scale deep face recognition.IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(2):2627–2644, 2022","cited_arxiv_id":null,"evidence_quote":"WebFace42M is the large-scale training dataset for LVFace-L."},{"cited_title":"Masked face recognition challenge: The insight- face track report","cited_arxiv_id":null,"evidence_quote":"Defines the MFR-Ongoing benchmark and the evaluation protocol for masked, cross-race, and age-gap verification."},{"cited_title":"Iarpa janus benchmark-c: Face dataset and protocol","cited_arxiv_id":null,"evidence_quote":"IJB-C is one of the three benchmarks used to measure verification accuracy."},{"cited_title":"Jain, James A","cited_arxiv_id":null,"evidence_quote":"IJB-B is the third benchmark, focused on image-to-video matching."},{"cited_title":"Transface: Calibrating trans- former training for face recognition from a data-centric per- spective, 2023","cited_arxiv_id":null,"evidence_quote":"TransFace is the prior ViT-based FR method that LVFace compares against and improves on."},{"cited_title":"Uniface: Unified cross-entropy loss for deep face recognition","cited_arxiv_id":null,"evidence_quote":"UniFace is a state-of-the-art baseline whose scores LVFace surpasses on MFR-Ongoing and IJB-C."},{"cited_title":"Topofr: A closer look at topology alignment on face recognition.arXiv preprint arXiv:2410.10587, 2024","cited_arxiv_id":null,"evidence_quote":"TopoFR is the leading baseline on the MFR mask subset that LVFace's overall MR-All surpasses."}],"review_version":1}