{"id":"3fb6bccf-8c51-409f-a371-08505a0635e2","arxiv_id":"2506.21260","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"DuET is an exemplar-free task-arithmetic model-merging framework that performs simultaneous class- and domain-incremental object detection, validated on YOLO11 and RT-DETR, with a new directional consistency loss and a retention-adaptability metric.","lead":"DuET is a method for object detection that learns new object categories and adapts to new visual domains at the same time, without storing old training images. It merges task-specific model updates with task arithmetic and adds a directional consistency loss, reporting better retention and adaptability than existing baselines on Pascal and weather-shift benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The +13.12% RAI claim rests on Avg GI, whose denominator is an unreported reference mAP; if the reference models are weak, DuET's adaptability score is inflated and the headline comparison is not reproducible.","rationale":"The reader's conditional verdict already identifies the unreported reference mAPs in Eq. 18 / Appendix A.2 as the primary weakness, and my stress-test agrees: this is the single most load-bearing unverified assumption because RAI is the headline metric and Avg GI is half of RAI. Without mAPref, the adaptability component of the central claim cannot be checked or reproduced, and weak reference models would directly inflate the reported RAI improvements. The concern is not an internal contradiction in the method; it is an evaluation-anchoring gap. The paper's own supplementary results provide partial same-backbone evidence (Appendix E.1), which weakens but does not remove the cross-backbone concern; the reference-mAP issue is more fundamental because it affects the meaning of the metric itself. I therefore recommend no change to the reader's conditional verdict: the paper should be accepted only if the authors release the reference mAP values and details, or provide code that lets reviewers compute them. This is consistent with the reader's assessment and with the stated limitations in Appendix H, which do not address the metric gap.","tokens_in":32042,"tokens_out":7760,"duration_ms":105387,"concrete_test":"Compute and report mAPref for every (Di,Cunseen) pair in Table S1 under the same protocol as DuET: same pre-trained YOLO11n, AdamW optimizer, 100 epochs, batch size 64, and mAP@0.5, with reference models trained only on that domain's unseen class set. If any reference mAP is low (e.g., below roughly 40 on Pascal domains or 25 on weather domains), retrain with a longer schedule until a stable ceiling is reached, then recompute Avg GI and RAI for all methods. The concern is settled if DuET's RAI advantage persists using published, stable reference values; it is not settled while the reference values are withheld.","verdict_should_be":"UNCHANGED","load_bearing_attack":"DuET's central quantitative claim is the RAI improvement, and RAI is defined as the mean of Avg RI and Avg GI (Eq. 19). Avg GI (Eq. 18, Appendix A.2) divides the model's mAP on unseen class-domain pairs by mAPref obtained by training “solely on these unseen classes” on that domain. The paper never reports mAPref for any (Di,Cunseen) pair, nor the reference models' training epochs, schedules, backbones, or hyperparameters. Because several reference domains are small (Clipart: 500 train images; Watercolor/Comic: 1000; Foggy Cityscapes: 1829), a weakly trained reference would make even modest raw unseen mAP look like high generalization. The denominator is shared across methods, so rankings can survive weak references, but the absolute RAI values and the claimed “+13.12% RAI improvement” are not interpretable or reproducible without it. The paper's statement that joint-training upper bounds are omitted does not remove this need; mAPref is exactly an upper-bound-style anchor. Since the headline Tables 1 and 2 combine Avg GI into RAI, this missing quantity is the most load-bearing unverified assumption. The cross-backbone comparisons in the main text and the unspecified new-class head initialization are additional concerns, but Appendix E.1 partially addresses the backbone issue, and the metric issue is more direct.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces DuIOD, an exemplar-free dual incremental object detection setting that combines class and domain shifts, and proposes DuET, a detector-agnostic task-arithmetic framework. DuET computes shared and task-specific parameters, merges shared task vectors through layer-wise retention and adaptation factors, concatenates task-specific head parameters in an Incremental Head, and adds a Directional Consistency Loss. A new metric, RAI, is defined as the average of Retention Index and Generalization Index. Experiments on Pascal Series and Diverse Weather Series report large RAI gains over sequential fine-tuning, LwF, ERD, LDB, and CL-DETR, with an ablation study showing each DuET component contributes.","tokens_in":32426,"tokens_out":5794,"duration_ms":58528,"significance":"If the empirical claims hold, DuET would be a meaningful contribution: it is the first task-merging approach for the combined dual incremental setting, it is detector-agnostic across YOLO11, RT-DETR, Deformable DETR, and ViTDet, and the ablations (Table 4, Appendix C) carefully demonstrate the role of each module and loss. The problem setting is timely and the core merging equations are simple and clearly presented. However, the central quantitative claims currently depend on an unreported reference denominator in the adaptability metric and on cross-backbone baseline comparisons, so the empirical significance is not yet established.","major_comments":[{"comment":"The central RAI claim is not reproducible because the reference mAP values mAPref(Di[Cunseen]) used as denominators in Avg GI are never reported, and no training details (epochs, schedule, backbone, hyperparameters) are given for the reference models. Since RAI is the headline metric in the abstract and Tables 1 and 2, and several reference domains are small (Clipart has 500 training images, Watercolor/Comic 1000, Foggy Cityscapes 1829), weakly trained reference models would inflate Avg GI and make the reported +13.12% and +11.39% RAI improvements uninterpretable. The authors should add a table of all mAPref values and a precise description of how those models were trained, or replace the ratio with raw unseen-class mAP.","section":"Sec. 4.2 and Appendix A.2, Eq. (18)/(24)"},{"comment":"The main comparisons are confounded by backbone choice: DuET is evaluated on YOLO11n (2.58M trainable parameters) while LDB uses ViTDet (110.52M) and CL-DETR uses Deformable DETR (39.85M). Consequently, the headline superiority cannot be attributed to the method rather than the architecture. Appendix E.1 partially addresses this with same-backbone results in Tables S7-S14, but the main text should either present controlled comparisons on shared backbones or explicitly restrict the claimed improvement to the YOLO11n comparisons.","section":"Tables 1 and 2 and Sec. 5.1"},{"comment":"The Incremental Head concatenates task-specific parameters from past and current tasks, but the paper never specifies how the detection head output dimension is expanded for new classes, how new-class weights are initialized, or how class-ID alignment is maintained across tasks. This is essential for reproducibility, especially because YOLO11 head output channels depend on the cumulative class count, and the ablation in Table 4 attributes substantial gains to the Incremental Head.","section":"Sec. 3.4, Eq. (13) and Fig. 3"},{"comment":"The key hyperparameters alpha_base=0.5, gamma=0.1, lambda_Distill=0.01 and lambda_DC=0.01 are selected by maximizing RAI on the Pascal Series and Diverse Weather Series, which are the same datasets used for the final reported results. No held-out validation split or nested selection procedure is described. This risks overfitting the evaluation protocol and makes the reported gains less trustworthy as a fair model-selection outcome.","section":"Appendix C.2, Fig. S2"},{"comment":"The statement that DuET outperforms CL-DETR and LDB on their own backbones with +5.97% and +11.98% RAI gains does not match the data. Averaging the seven paired two-phase and multi-phase experiments gives approximately +5.2 RAI for Deformable DETR and +8.4 RAI for ViTDet. Moreover, in Table S12, DuET on ViTDet has RAI 53.01 vs. LDB's 52.83 but Avg RI drops from 86.08 to 65.57, so the claim of consistent superiority across all experiments and backbones is not supported.","section":"Appendix E.1, Tables S7-S14"}],"minor_comments":[{"comment":"The Directional Consistency Loss uses tau^(i)_{st-2}, but for t=2 this term is not defined in the main text; the supplementary should clarify that tau_{s0} = 0, since theta_{s0} - theta_{s0} = 0.","section":"Sec. 3.5, Eq. (16)"},{"comment":"The column header 'mAP@0.5%' should read 'mAP@0.5' or 'mAP@0.5 (%)' for consistency with the metric definition.","section":"Table 1"},{"comment":"The caption reads 'Diverse Series' but should be 'Diverse Weather Series' to match the terminology used elsewhere.","section":"Fig. S5 caption"},{"comment":"The clamping operation on delta_l is redundant because delta_l = gamma * tanh(p_l) already lies within [-gamma, gamma]; consider removing it or explaining why it is retained.","section":"Sec. 3.4, Eq. (10)-(11)"}],"recommendation":"major_revision","confidential_remarks":"The most urgent issue is the missing mAPref denominators in the RAI metric; without them the headline numbers are not interpretable and the paper should not be accepted in its current form. I also noticed that the appendix's average RAI gain claims do not match the numbers in Tables S7-S14, which suggests the authors should re-check their aggregations. The same-backbone comparisons in the appendix should be promoted to the main text to support the cross-method claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Tim,\n\nQuick take on arXiv:2506.21260 (DuET). The paper introduces a genuinely practical problem formulation — dual incremental object detection (DuIOD), where class and domain shift simultaneously and no exemplars are stored. That setup has been under-served; VIL is classification-only, and neither CIOD nor DIOD covers both. The method is refreshingly simple: decompose the detector into shared and task-specific parameters, merge shared task vectors with per-layer retention/adaptation factors, and concatenate task-specific heads. The Directional Consistency Loss is modest but reasonable, and the ablations show it adds a few points of RAI across merging methods. Detector-agnostic is a real plus: they demonstrate YOLO11n/x, RT-DETR-l/x, Deformable DETR, and ViTDet.\n\nThe experiments are extensive — seven DuIOD scenarios, multiple backbones, clean ablation table. Each component earns its place. I also appreciate the explicit limitations section; they don't hide that task-vector methods need a pretrained base and may struggle with extreme shifts.\n\nNow the soft spots, in order of importance. First, the Avg GI denominator. RAI is the mean of Avg RI and Avg GI, and Avg GI divides the model's mAP on unseen class-domain pairs by a reference mAP obtained by training \"solely on these unseen classes.\" Those reference numbers are never reported, and several reference domains are small (Clipart 500 train images, Watercolor/Comic 1000, Foggy Cityscapes 1829). If the reference models are weak, the GI numbers and the headline \"+13.12% RAI\" become hard to interpret. Rankings are less affected because the denominator is shared across methods, but the absolute claims are not reproducible as reported. This is the one load-bearing missing quantity. Second, the main tables compare DuET on YOLO11n against LDB on ViTDet and CL-DETR on Deformable DETR; that is apples-to-oranges. The appendix does include DuET on those same backbones and shows it still beats the baselines on their own architectures, so the concern is partly mitigated, but the main-text presentation invites the problem. Third, the hyperparameters (alpha_base, gamma, loss weights) are selected via sensitivity analysis on the same evaluation benchmarks; not unusual, but worth flagging. No code is released, which makes the reference-mAP omission harder to work around.\n\nNet: the central idea is sound, and the evidence, on inspection, supports the general conclusion that DuET does well on its own setting. The reference-mAP issue should be fixed before publication — report all mAPref values, ideally with training regimes. This paper deserves a serious referee. I'd send it to review, and would cite the DuIOD setting if I work on continual detection.\n\nRecommendation: engage. Send to a careful reviewer, ask for the missing reference numbers and same-backbone comparisons in the main text.","headline":"A new dual-incremental detection setting and a simple, detector-agnostic task-arithmetic solution that mostly holds up; the headline RAI gains are plausible but rest on an unreported reference denominator that needs to be disclosed.","tokens_in":32872,"tokens_out":2632,"would_cite":true,"duration_ms":29582,"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":"Merging task vectors on shared weights lets one detector learn new object classes in new domains, without storing old images.","keywords":["dual incremental object detection","task arithmetic","model merging","catastrophic forgetting","domain shift","exemplar-free incremental learning","retention-adaptability index","directional consistency loss"],"falsifier":"Train reference models for every unseen-class domain pair using the same pretrained checkpoint, epochs, and data splits used for DuET, then recompute Avg GI and RAI for all methods on a shared backbone such as YOLO11n; if the reported +13.12% and +11.39% RAI margins shrink substantially or reverse, the adaptability gain is an artifact of the normalization or backbone choice rather than the merging algorithm.","tokens_in":31861,"feed_emoji":"🎯","tokens_out":5298,"duration_ms":56987,"temperature":0.7,"pith_summary":"This paper proposes a new learning setting, Dual Incremental Object Detection (DuIOD), in which a detector must add new object categories and adapt to a new visual domain at every task, without access to previous training data. The authors argue that existing class-incremental and domain-incremental detectors each miss one half of this problem, and they introduce DuET, a task-arithmetic-based model merging framework to handle both at once. DuET splits a detector into shared parameters (backbone and neck) and task-specific parameters (detection head), then fuses old and new task vectors—the weight differences from a pretrained model—layer by layer. A per-layer retention factor decides how much old knowledge to keep versus how much new signal to adopt, and a Directional Consistency Loss penalizes weight updates that oppose earlier ones. If the paper is right, DuET delivers the first exemplar-free way to make a standard real-time detector (YOLO11, RT-DETR) continuously learn both classes and domains, with a combined score, the Retention-Adaptability Index, to measure retention and adaptation together.","feed_headline":"Task-vector merging lets detectors add classes across new domains","feed_subtitle":"Exemplar-free incremental detection keeps ~89% of old knowledge while generalizing to unseen class–domain pairs.","key_machinery":"The machinery is task-vector arithmetic on shared weights. A task vector $\\tau = \\theta_{\\mathrm{finetuned}} - \\theta_{\\mathrm{pretrained}}$ records how a pretrained model's weights must shift for one task; DuET stores just two such vectors at a time, $\\tau_{\\mathrm{old}}$ and $\\tau_{\\mathrm{curr}}$, and fuses them with layer-wise weights $\\alpha_l$, $\\beta_l$ chosen by a $p$-factor $p_l = (\\|\\tau_{\\mathrm{old}}^l\\| - \\|\\tau_{\\mathrm{curr}}^l\\|) / (\\|\\tau_{\\mathrm{old}}^l + \\tau_{\\mathrm{curr}}^l\\| + \\epsilon)$, mapped through $\\gamma \\tanh$ to keep $\\alpha$ in $[\\alpha_{\\mathrm{base}} - \\gamma, \\alpha_{\\mathrm{base}} + \\gamma]$. The Directional Consistency Loss (ReLU of the negative dot product of successive shared-weight updates) stabilizes the trajectory of these vectors. The Incremental Head concatenates task-specific head weights so that new classes get their own parameters rather than overwriting old ones.","core_discovery":"The central claim is that simultaneous class and domain incremental learning can be handled by merging task vectors on the shared part of a detector, without exemplars. In DuET the shared parameters after the current task are rebuilt as $\\theta_{s0} + \\alpha_l \\tau_{\\mathrm{old}} + \\beta_l \\tau_{\\mathrm{curr}}$ per layer, where $\\tau_{\\mathrm{old}}$ is the cumulative task vector from previous phases and $\\tau_{\\mathrm{curr}}$ is the current phase's vector; $\\alpha_l$ and $\\beta_l$ are derived from the ratio of the vectors' $\\ell^1$ norms through a tanh-scaled $p$-factor, with $\\beta_l = 1 - \\alpha_l$. The task-specific head parameters are simply concatenated across tasks. A Directional Consistency Loss adds a ReLU penalty on negative dot products between consecutive shared-weight updates, reducing sign conflicts during merging. On the Pascal Series (4 tasks) the method reports 89.30% average retention and a +13.12% RAI improvement over baselines, and on the Diverse Weather Series (3 tasks) 88.57% Avg RI and +11.39% RAI, across detectors including YOLO11 and RT-DETR.","pith_inferences":["A testable consequence the authors do not pursue: the $p$-factor uses $\\ell^1$ norms; variants using cosine similarity or per-channel statistics might shift the stability-plasticity trade-off, especially for extreme weather or style changes.","Because DuET requires a pretrained detector and a common initialization, it cannot start from scratch; extending the decomposition to a multi-stage pretrain-adapt loop would broaden its applicability.","The RAI metric's validity rests on the reference mAP models used to normalize the Average Generalization Index; publishing those reference values and their training protocols would make cross-paper comparisons meaningful.","The Incremental Head grows wider with each task; the paper does not study head-capacity saturation over long task sequences, which is a natural stress test for a ten- or twenty-task deployment."],"forward_implications":["A detector can be turned into an incremental one without changing its architecture; the same pipeline works for YOLO11-based and RT-DETR-based detectors and preserves real-time inference speed.","Retention stays high across multi-phase sequences: above 88% Avg RI on both the four-task Pascal sequence and the three-task weather sequence, compared with near-total forgetting for sequential fine-tuning.","The merged model generalizes to unseen class–domain pairs—for example, detecting VOC classes rendered in Clipart style and vice versa—which neither class-only nor domain-only incremental detectors provide.","The Retention-Adaptability Index offers a single number combining retention of old classes and adaptability to new ones, which could serve as a standard evaluation metric for this setting.","The method is exemplar-free and keeps memory footprint roughly constant by storing only two shared task vectors plus the pretrained weights, rather than a growing history of task vectors."],"supporting_citations":[{"why":"Defines task vectors as weight differences from a pretrained model, the base operation DuET builds on.","marker":"[13]"},{"why":"Establishes the sequential fine-tuning plus model-merging paradigm in continual learning, which DuET adopts and extends.","marker":"[31]"},{"why":"Identifies sign conflicts in model merging and motivates the Directional Consistency Loss.","marker":"[46]"},{"why":"Provides the distillation baseline (LwF) whose loss DuET modifies with a thresholded variant.","marker":"[25]"},{"why":"Supplies the ERD baseline, a replay-free distillation method for incremental object detection that DuET compares against.","marker":"[8]"},{"why":"Supplies the domain-incremental baseline (LDB) that DuET beats; LDB is evaluated on its own ViTDet backbone.","marker":"[42]"},{"why":"Supplies the class-incremental transformer baseline (CL-DETR) that DuET compares against on Deformable DETR.","marker":"[28]"},{"why":"YOLO11 is the primary real-time detector on which DuET is instantiated and whose backbone/neck/head decomposition is used.","marker":"[17]"},{"why":"Source of the Watercolor, Comic, and Clipart domains in the Pascal Series, which provide the domain shifts in the benchmarks.","marker":"[14]"},{"why":"Source of the BDD100k daytime and night sunny weather domains used in the Diverse Weather Series.","marker":"[49]"}],"fun_headline_variants":["Exemplar-free task arithmetic handles class and domain shifts","Merging task vectors adds classes and domains without forgetting","Dual incremental detection: class and domain shifts, exemplar-free","Task arithmetic merges old and new knowledge for dual shifts","Exemplar-free dual incremental detection via task vectors"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported adaptability advantage assumes that the reference mAP models used to normalize the Average Generalization Index are strong and fairly trained; the paper does not report their mAP values or training details, so weak reference models would inflate DuET's RAI advantage, and some headline comparisons also mix backbones across methods.","fun_headline_variants_meta":{"raw":{"variants":["Exemplar-free task arithmetic handles class and domain shifts","Merging task vectors adds classes and domains without forgetting","Dual incremental detection: class and domain shifts, exemplar-free","Task arithmetic merges old and new knowledge for dual shifts","Exemplar-free dual incremental detection via task vectors"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000672,"raw_usage":{"total_tokens":3126,"prompt_tokens":1078,"completion_tokens":2048,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":694,"completion_tokens_details":{"reasoning_tokens":1966}},"tokens_in":694,"tokens_out":2048,"duration_ms":16343,"temperature":1.0,"reasoning_tokens":1966,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:29:18.007193+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train reference models for every unseen-class domain pair using the same pretrained checkpoint, epochs, and data splits used for DuET, then recompute Avg GI and RAI for all methods on a shared backbone such as YOLO11n; if the reported +13.12% and +11.39% RAI margins shrink substantially or reverse, the adaptability gain is an artifact of the normalization or backbone choice rather than the merging algorithm.","supporting_citations":[{"cited_title":"Magmax: Leveraging model merging for seamless continual learning","cited_arxiv_id":null,"evidence_quote":"Establishes the sequential fine-tuning plus model-merging paradigm in continual learning, which DuET adopts and extends."},{"cited_title":"Ties-merging: Resolving interference when merging models","cited_arxiv_id":null,"evidence_quote":"Identifies sign conflicts in model merging and motivates the Directional Consistency Loss."},{"cited_title":"Learning without forgetting","cited_arxiv_id":null,"evidence_quote":"Provides the distillation baseline (LwF) whose loss DuET modifies with a thresholded variant."},{"cited_title":"Overcoming catastrophic forgetting in incremental object detection via elastic response distillation","cited_arxiv_id":null,"evidence_quote":"Supplies the ERD baseline, a replay-free distillation method for incremental object detection that DuET compares against."},{"cited_title":"Non-exemplar domain incremental object detection via learning domain bias","cited_arxiv_id":null,"evidence_quote":"Supplies the domain-incremental baseline (LDB) that DuET beats; LDB is evaluated on its own ViTDet backbone."},{"cited_title":"Continual detection transformer for incremen- tal object detection","cited_arxiv_id":null,"evidence_quote":"Supplies the class-incremental transformer baseline (CL-DETR) that DuET compares against on Deformable DETR."},{"cited_title":"Ultralytics yolo11, 2024","cited_arxiv_id":null,"evidence_quote":"YOLO11 is the primary real-time detector on which DuET is instantiated and whose backbone/neck/head decomposition is used."},{"cited_title":"Cross-domain weakly-supervised object de- tection through progressive domain adaptation","cited_arxiv_id":null,"evidence_quote":"Source of the Watercolor, Comic, and Clipart domains in the Pascal Series, which provide the domain shifts in the benchmarks."},{"cited_title":"Bdd100k: A diverse driving dataset for heterogeneous multitask learning","cited_arxiv_id":null,"evidence_quote":"Source of the BDD100k daytime and night sunny weather domains used in the Diverse Weather Series."}],"review_version":1}