{"id":"1a5ab9c3-a0f9-454d-b455-e33b44c27990","arxiv_id":"2607.16546","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A multi-task affective system combining frozen AffectNet backbones, LoRA-adapted MAE for action units, temporal heads, fusion, and ensembling attains P=1.7302 on the s-Aff-Wild2 validation split.","lead":"This paper describes an emotion-recognition system for the ABAW multi-task challenge: frozen face encoders feed small task-specific heads, temporal models, and fusion layers, with a LoRA-adapted encoder for action units, reaching P=1.7302 on the official validation split. It is a useful benchmark entry because it systematically ablates where effort should go and shows frozen features can compete for valence/arousal and expression.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Validation-selected P likely inflated: Eq. (1) argmax and Eq. (8) calibration both use the same split used to report P=1.7302, with no held-out/test confirmation.","rationale":"The reader's weakest assumption exactly identifies the same issue: the validation split is used both for model selection and for reporting the final score. My stress-test finds no additional load-bearing defect beyond this. The concern is substantive because the candidate pool is built up over many ablations, and thresholds are fit on the same split, so P=1.7302 is an optimistic selected maximum rather than an unbiased estimate. The paper is honest about the risk but does not mitigate it. Since the reader already marked the paper CONDITIONAL on exactly this issue, my recommendation is UNCHANGED: the condition should remain that the authors provide a held-out evaluation or quantify selection bias before the central claim is treated as robust.","tokens_in":10266,"tokens_out":2804,"duration_ms":31412,"concrete_test":"Split the official 26,666-image validation set into two disjoint parts V_sel (e.g., 80%) and V_eval (20%). Use only V_sel to run Eq. (1) selection, fit b and θ_j in Eq. (8), and choose ensemble/seed configurations from Tables 3, 10, 11. Compute P once on V_eval with these fixed choices; repeat for 5 different splits. If the mean P_eval is materially below 1.7302 (e.g., >0.03 lower), the reported number is inflated by selection on the evaluation set.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that post-encoder adaptation reaches P=1.7302 and that the encoder is not the bottleneck. The most load-bearing assumption is that this validation number estimates real performance. Eq. (1) chooses each task predictor by argmax over validation metric M_t; Eq. (8) fits EXPR bias b and AU thresholds θ_j on validation; and Sec. 3.8 selects seed/top-K ensembles by validation scores. Every ablated decision (backbone, LoRA block count, fusion variant, window length, ensembling) is made on the same split. The paper even notes in Sec. 3 that 'every added module can overfit validation' but provides no nested split, no uncertainty estimate, and no test/leaderboard result. The reported P is therefore a selected maximum over a substantial candidate pool, not an unbiased estimate; the gap between validation-selected and true performance is typically largest when calibration parameters are also fit on the evaluation split. This directly threatens both the quantitative headline and the qualitative claim that frozen supervised features suffice to match or beat retrained encoders.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper describes AffectFuse, a system for the ABAW 2026 Multi-Task Learning (MTL) track. It builds on frozen AffectNet-supervised EfficientNet-B2 and DDAMFN encoders, a self-supervised MAE-Face encoder adapted via LoRA for action units, temporal GRU/LSTM heads, cross-task feature fusion from AU to valence/arousal and expression, per-unit expert routing with external DISFA data, and validation-selected ensembles. The final configuration achieves P=1.7302 on the official validation split (VA CCC 0.6596, EXPR macro-F1 0.5097, AU mean F1 0.5609). The central claim is that post-encoder adaptation — task heads, fusion, temporal modeling, and per-AU routing — can match or beat retraining large encoders on this benchmark, and that the encoder is not always the bottleneck.","tokens_in":10596,"tokens_out":5862,"duration_ms":65251,"significance":"If the reported validation result transfers to the test set, the contributions are practically significant: a data-efficient pipeline from public frozen checkpoints, a transparent ablation framework, and useful negative results (DISFA sequential transfer, flip-TTA, frozen MAE-Face). The paper is honest about its validation-based selection and includes per-stage diagnostics. However, the quantitative headline and the qualitative 'encoder is not the bottleneck' conclusion rest entirely on a validation-selected maximum with no held-out or test-set confirmation. The significance therefore depends on validation/test agreement, which is not demonstrated in the manuscript.","major_comments":[{"comment":"The reported P=1.7302 is obtained by optimizing the same validation metric used for evaluation. Eq. (1) selects each task predictor by argmax over validation M_t; Eq. (7) selects per-AU experts by validation F1; Eq. (8) fits the EXPR class bias b and the AU thresholds θ_j on the validation split after smoothing; Sec. 3.8 selects seeds and top-K ensembles by the same validation scores. With a candidate pool of a dozen-plus families and additional hyperparameters (window length, LoRA rank, unfrozen blocks, etc.), the quoted P is a selected maximum rather than an unbiased estimate of expected performance. The paper's own warning in Sec. 3 ('every added module can overfit validation') is not backed by a nested split, cross-validation, or test-set report. This is load-bearing for both the quantitative claim and the qualitative 'encoder is not the bottleneck' conclusion. Please provide test/le","section":"Eqs. (1), (7), (8); Secs. 3.1, 3.7, 3.8"},{"comment":"All reported metrics are point estimates on a single split. Multi-seed runs are used only to form ensembles; no standard deviations or confidence intervals are reported for any row. The final configuration is the argmax over a large validation-driven search, so the gap between validation-selected performance and true test performance is typically nonzero and can be substantial. Without test-set confirmation or an internal holdout, the claim that 'post-encoder adaptation ... provides a strong MTL pipeline' is not fully established. Please add error bars for the main rows and, if possible, a test-set result from the challenge.","section":"Tables 2–11; Sec. 4.3"},{"comment":"The ablation staging is inconsistent. The row labeled 'AU on concat. features + ASL' raises VA from 0.5294 to 0.5870 and EXPR from 0.3948 to 0.4165, even though the stage description indicates only an AU-branch change. The next AU-only row ('AU: LoRA fine-tuned MAE') leaves VA and EXPR unchanged. This suggests the rows do not isolate single candidate families as claimed, or that some joint retraining/feature-cache update is occurring off-stage. Please clarify what is retrained at each stage or correct the table; otherwise the incremental-gain story and the 'compact ablation' contribution are difficult to interpret.","section":"Table 3; Sec. 3.1"}],"minor_comments":[{"comment":"The text says 'the complete final four transformer blocks are also optimized,' but Table 5 reports the best result with 6 unfrozen blocks and parameter counts varying from 15.9M to ~44M. Please specify the exact trainable blocks in the final configuration and reconcile the 31.53M-of-89.0M statement with the table.","section":"Sec. 3.5; Table 5"},{"comment":"The code link is a placeholder ('Code is available here') with no URL. Also, 'landmark-aligned224×224 crops' is missing a space.","section":"Abstract; Sec. 3.7"},{"comment":"The 'Organizer score' row is undefined. State what this baseline represents (e.g., challenge baseline or previous best) and cite it.","section":"Table 2"},{"comment":"The 'Pretraining' column entries like 'AffectNet260' and 'AffectNet112' are ambiguous; these appear to be pretraining dataset plus input resolution. Please separate dataset name from resolution.","section":"Table 1"}],"recommendation":"major_revision","confidential_remarks":"The paper is competently organized and the ablation is transparent, but the central validity hinges on the validation/test gap. I would encourage requesting test-set or nested-validation evidence before acceptance. The inconsistent ablation staging in Table 3 should also be resolved."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing you should know: this is a textbook ABAW competition system report, but a careful and honest one. What's genuinely new is the per-AU expert routing across donor models (Eq. 7) and the staged ablation evidence that frozen AffectNet encoders plus post-encoder heads match or beat retraining on this benchmark. If the numbers hold up, the practical recipe -- freeze AffectNet, add temporal heads and cross-task fusion, LoRA-tune MAE for AUs only -- is useful for low-compute in-the-wild affect recognition.\n\nCredit where due: the ablation path in Table 3 is explicit, with negative results reported (POSTER++ no gain, frozen MAE weak, DISFA sequential transfer fails, flip-TTA hurts). The sigmoid-F1 collapse guard is a nice practical detail. The paper is transparent about the risk of validation overfitting in Sec. 3, which is more than most challenge papers do.\n\nThe soft spot is exactly what the stress-test note says: the reported P=1.7302 is a maximum over a candidate pool selected on the same validation split used to report it. Eq. (1) argmax, Eq. (8) fitted bias/thresholds, and Sec. 3.8 the top-K ensemble selection all use the official validation split. No test-set confirmation, no error bars. That makes the number an optimistic selected maximum, not an unbiased estimate. This does not undercut the qualitative message -- 'encoder is not always the bottleneck' -- because the relative comparisons are subject to the same bias, and the frozen-backbone baseline is already strong. But it does mean the exact P should be treated carefully.\n\nMinor issues: the code link is unverified in the preprint; Table 2 compares only against the organizer's score, not against prior systems; and the per-AU routing gain over the best single AU model is small (~+0.008 P). Still, the idea is the most conceptually interesting part of the paper.\n\nBottom line: this deserves a serious referee. The paper is coherent, the ablations are reproducible in structure, and the per-AU routing idea is worth testing elsewhere. A good referee should push for test-set confirmation or a nested validation protocol, and for the code to be published. I would not desk-reject it, and I'd cite the recipe as a credible baseline.","headline":"An honest, well-ablated ABAW system report; headline P is a validation-selected maximum, so treat the number with care, but it deserves a serious referee.","tokens_in":11057,"tokens_out":2567,"would_cite":true,"duration_ms":27144,"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":"Post-encoder adaptation with frozen face encoders reaches P=1.7302 on the ABAW multi-task affect recognition benchmark, without training a new face foundation model.","keywords":["multi-task learning","affective behavior analysis","ABAW","valence-arousal","facial action units","cross-task feature fusion","LoRA","per-unit expert routing"],"falsifier":"Take the exact validation-selected configuration and run it on the official ABAW test set, or on a held-out split never used in the described ablations: if P drops substantially below 1.7302, or a configuration chosen on a separate development split outperforms it, the central claim that post-encoder adaptation with frozen encoders is sufficient would be weakened.","tokens_in":10211,"feed_emoji":"🎭","tokens_out":5455,"duration_ms":56053,"temperature":0.7,"pith_summary":"The paper tries to establish that for multi-task affective behavior recognition, the encoder is not always the bottleneck: frozen, AffectNet-supervised face encoders can feed a strong multi-task system if the effort is spent after feature extraction. On the ABAW MTL track over s-Aff-Wild2, the proposed pipeline, combining task-specific temporal heads, cross-task feature fusion, a LoRA-adapted action-unit expert, and per-unit expert routing of external data, reaches a validation P of 1.7302 without training a new large-scale face foundation model. The paper argues that valence/arousal and expression gain most from temporal modeling and fusion, while action units repay parameter-efficient fine-tuning and per-unit selection. The result matters because it suggests a cheaper, reproducible route to competitive multi-task affect systems and identifies where frozen transfer saturates.","feed_headline":"Frozen encoders plus fusion hit 1.7302 on affect benchmark","feed_subtitle":"Task-specific temporal heads, cross-task fusion, and per-unit AU routing carry the gains—no new foundation model.","key_machinery":"The load-bearing mechanism is a staged post-encoder adaptation stack: frozen AffectNet-supervised encoders (EfficientNet-B2 and DDAMFN) provide cached multi-resolution features and remapped expression logits; a self-supervised MAE-Face ViT-B is LoRA-adapted (rank 16, alpha 32, six unfrozen blocks) with a masked asymmetric loss plus a collapse-guarded sigmoid-F1 surrogate to become the AU expert; a 48-frame GRU/BiGRU temporal head and a cross-task fusion module concatenate AU features into valence/arousal and expression representations; and a per-unit expert route (argmax over validation F1 per AU column) selects the best source for each of the twelve action units. Seven-frame smoothing and v","core_discovery":"On the official validation split of s-Aff-Wild2, the system reports P=1.7302, composed of average CCC 0.6596 for valence/arousal, macro-F1 0.5097 for eight-class expression, and mean F1 0.5609 across twelve action units, up from P=1.4529 for frozen backbones with plain heads. The central discovery is that after-encoder design choices, not new foundation-model training, drive most of the gain: temporal modeling with a 48-frame window, cross-task fusion that feeds AU features into valence/arousal and expression heads, LoRA fine-tuning of a self-supervised face autoencoder for AU, and per-unit expert routing that lets different AU columns pick different trained sources based on validation F1. T","pith_inferences":["Editorial inference: because the same validation split is used both to select every component and to fit the expression bias and AU thresholds, the quoted P=1.7302 is likely an optimistic selected maximum; a held-out test split would quantify how much validation-driven selection overfits.","Editorial inference: the per-unit expert-routing idea is not AU-specific; it could be applied to any sparse multi-label task where different labels benefit from different data sources, as long as per-column validation F1 is available.","Editorial inference: the flip-TTA negative result hints that horizontal-flip augmentation is harmful for expression and action units because facial asymmetry is informative; a direct test would be to train with horizontal-flip augmentation from the start rather than applying it only at test time.","Editorial inference: the frozen-feature cache makes the pipeline data-efficient; the same recipe could be transferred to other in-the-wild face benchmarks by swapping in any supervised face encoder and rerunning the candidate selection."],"forward_implications":["Competitive multi-task affective systems can be built without training a new face foundation model; frozen AffectNet-supervised features plus task heads, fusion, and temporal modeling account for most of the gains.","Action-unit recognition is the part of the pipeline that benefits most from parameter-efficient adaptation and per-unit expert routing; external lab data like DISFA helps as a feature donor but not as direct sequential transfer.","Temporal modeling and fusion gains are representation-dependent: increasing the window from 32 to 48 frames after fusion raises the overall score by 0.031.","Validation-guided selection with explicit candidate pools and documented negative results, such as flip test-time augmentation hurting all tasks, yields a reproducible ablation path for multi-task affect systems.","Unfreezing the encoder for expression and valence/arousal saturates below frozen supervised features, so compute is better spent after the encoder."],"fun_headline_variants":["Post-encoder tweaks beat new foundation model on affect","Cross-task fusion and LoRA lift affect benchmark to 1.73","No new model, but 1.73 P on s-Aff-Wild2 via fusion","Fusion and temporal heads boost affect AI to 1.7302"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that choosing every component and fitting the final calibration parameters on the official validation split yields a configuration that transfers to unseen test data; if validation-selected tuning does not generalize, the reported P=1.7302 is an optimistic selected maximum rather than an estimate of true performance.","fun_headline_variants_meta":{"raw":{"variants":["Post-encoder tweaks beat new foundation model on affect","Cross-task fusion and LoRA lift affect benchmark to 1.73","No new model, but 1.73 P on s-Aff-Wild2 via fusion","Fusion and temporal heads boost affect AI to 1.7302"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000171,"raw_usage":{"total_tokens":1114,"prompt_tokens":758,"completion_tokens":356,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":502,"completion_tokens_details":{"reasoning_tokens":287}},"tokens_in":502,"tokens_out":356,"duration_ms":4620,"temperature":1.0,"reasoning_tokens":287,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T20:38:39.432064+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the exact validation-selected configuration and run it on the official ABAW test set, or on a held-out split never used in the described ablations: if P drops substantially below 1.7302, or a configuration chosen on a separate development split outperforms it, the central claim that post-encoder adaptation with frozen encoders is sufficient would be weakened.","supporting_citations":[],"review_version":1}