{"id":"67653c48-84b2-41e9-bd41-22a1e347547e","arxiv_id":"2412.18219","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"ACMap achieves exemplar-free class-incremental learning with constant inference time by averaging task-specific adapters and shifting previous prototypes with the current task's centroid drift.","lead":"ACMap merges the per-task adapters of a pre-trained vision model into one shared adapter by averaging weights, with a centroid-shift step that re-aligns old class prototypes in the new shared feature space. The method keeps class-incremental learning inference time constant as tasks grow, while matching the accuracy of the slower state-of-the-art on most benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Centroid prototype mapping assumes a per-task shift that is never validated online; VTAB Table 1 (87.56 vs EASE 93.55) is a direct counterexample, so the abstract's unqualified 'matches SOTA' claim needs qualification.","rationale":"I read ACMap as a heuristic but well-scoped CIL method. The main contribution is the combination of adapter merging with centroid prototype mapping; the code is released and ablations confirm both components help, so the work is reproducible enough for conditional acceptance. My concern does not challenge the internal consistency of the algorithm; it challenges the unqualified title claim. The mapping's constant-shift assumption is the point where the method is least secure: it is a heuristic validated on four benchmarks and contradicted on the standard VTAB protocol. The paper itself flags this in Section 6 and Appendix E.2, which is a strength, but the abstract and introduction do not carry the qualification. A conditional acceptance with a request to qualify the claim and add error bars is the appropriate outcome. I agree with the reader's identification of the centroid mapping as the weakest assumption, and I add that the VTAB failure is not merely an outlier: it is the expected symptom of the assumption failing when task distributions are heterogeneous.","tokens_in":18747,"tokens_out":7604,"duration_ms":72939,"concrete_test":"Run an offline diagnostic on VTAB B0 Inc10 with all validation sets available: for each previous task i, compute the true shift Delta_i,t = mean_c(P_i,c(Abar_t) - P_i,c(Abar_i)) and compare it with the Algorithm 1 estimate Delta_est_i,t = mean_c(P_t,c(Abar_t) - P_t,c(Abar_i)); if the mean cosine or relative L2 error between Delta_i,t and Delta_est_i,t is large (e.g., > 0.1) for early tasks, the constant-shift approximation is the cause of the VTAB gap, whereas if the shifts match but accuracy still differs, the gap must be attributed to another component such as adapter merging or data imbalance.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—ACMap matches SOTA accuracy with constant inference time—requires the centroid prototype mapping (Eqs. 6–7) to be an accurate approximation of the feature-space drift for every previous task. The paper estimates the shift from current-task prototypes and applies it to all old prototypes. This shift-invariance is only verified indirectly: Figures 6 and H compare mapped prototypes against true prototypes computed with previous-task validation data, which is unavailable in the exemplar-free setting, and the comparison is reported as class-level cosine similarity rather than as a check that the shift vector is actually constant across tasks. The standard VTAB B0 Inc10 result is the concrete failure mode: ACMap reaches 87.56 final accuracy versus EASE's 93.55 (average 91.21 vs 93.61). The balanced-VTAB experiment in Appendix E.2 shows parity when task sizes are equalized, but the imbalanced standard benchmark is the protocol used in Table 1. The method therefore matches SOTA on four benchmarks and misses on the fifth, so the abstract's unqualified claim is not supported as written. The concern is not that the method is internally unsound; it is that the load-bearing condition for the headline claim is an empirical heuristic whose failure mode is already visible in the paper's own results.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ACMap, an exemplar-free class-incremental learning method that trains a task-specific adapter for each task, incrementally averages their weights into a single merged adapter, and uses a centroid prototype mapping to align previously computed prototypes to the current merged-adapter subspace. The claimed benefits are constant (O(1)) inference time independent of the number of tasks, competitive accuracy with state-of-the-art methods such as EASE, and scalability via an early-stopping threshold. The method is evaluated on five benchmarks (CIFAR-100, CUB, ImageNet-R, ImageNet-A, VTAB) under the exemplar-free protocol, with comparisons to prompt- and adapter-based baselines. The central empirical claim is that ACMap matches EASE's accuracy on most benchmarks while being roughly 39x faster at inference on 40-task ImageNet-R.","tokens_in":18959,"tokens_out":3959,"duration_ms":35476,"significance":"If the results hold, ACMap is a valuable contribution to the continual learning literature: it directly addresses the accuracy-inference-time trade-off in pre-trained-model-based CIL, achieves strong performance with a simple and reproducible mechanism, and provides code. The paper's strengths include clean ablation studies (Table 3 and Table 4), a clear complexity argument for O(1) inference, and a thorough set of experiments across diverse datasets. The centroid prototype mapping is an original idea, and the analysis of the loss landscape (Figure 3 and Appendix G) provides intuitive support for why weight averaging works in this setting. The method is also genuinely exemplar-free, which is important for privacy-sensitive applications. However, the significance is moderated by the paper's own results showing a substantial accuracy drop on VTAB, and by the absence of error bars, which makes it difficult to assess whether the small observed differences from EASE are meaningful.","major_comments":[{"comment":"The abstract states that ACMap 'matches state-of-the-art accuracy,' but the results in Table 1 do not support an unqualified claim. On VTAB B0 Inc10, ACMap achieves an average accuracy of 91.21 and final accuracy of 87.56, versus EASE's 93.61 and 93.55, a gap of 2.4 and 6.0 percentage points respectively. On CUB B0 Inc10, IN-R B0 Inc5, and IN-A B0 Inc20, ACMap is also slightly below EASE in final accuracy. The only benchmark where ACMap clearly exceeds EASE on both metrics is CIFAR-100 B0 Inc5. The abstract and introduction should qualify the claim, for example by stating that ACMap matches state-of-the-art accuracy on most benchmarks while providing a large inference-speed advantage, and the VTAB exception should be explicitly acknowledged in the abstract.","section":"Abstract and Section 5.2, Table 1"},{"comment":"The centroid prototype mapping assumes that a single shift vector ∆p, estimated from current-task prototypes computed in the current and an older adapter, applies unchanged to all previous tasks' prototypes. This assumption is load-bearing for the method's accuracy, but the paper's validation in Figure 6 and Appendix H relies on computing true prototypes with previous-task validation data, which is not available in the exemplar-free CIL setting. The VTAB result in Table 1 is a concrete failure case: the mapping does not adequately correct the prototypes, leading to a large accuracy drop relative to EASE. While the authors attribute this to data imbalance and show parity on a balanced VTAB variant (Appendix E.2), the standard imbalanced protocol is the one used in the main comparison, and the robustness of the mapping to realistic imbalances is therefore not established. I recommend that the authors either (a) restrict the central claim to settings where the task distribution is reasonably balanced, or (b) provide additional evidence—beyond cosine similarity on held-out validation data—that the shift vector is stable across tasks, such as measuring the variance of ∆p across different current tasks.","section":"Section 4.2, Eqs. (6)-(7), and Table 1"},{"comment":"All ACMap results are reported as averages over five runs without standard deviations, confidence intervals, or significance tests. Since the method's accuracy is within 0.1–1.5 percentage points of EASE on several benchmarks (e.g., IN-R final accuracy 70.49 vs. 70.58, IN-A final accuracy 56.19 vs. 55.04), the reader cannot judge whether these differences are statistically meaningful. The claim of 'comparable' or 'slightly better' performance should be supported by error bars or at least by a statement of variance across seeds. This is a standard expectation for empirical papers and is necessary to assess the robustness of the headline result.","section":"Section 5.1 and Table 1"}],"minor_comments":[{"comment":"Typo: 'We follow the the protocol in [59]' should be 'We follow the protocol in [59]'.","section":"Section 5.1"},{"comment":"Typo: 'In examplar-based CIL' should be 'In exemplar-based CIL'.","section":"Section 3.1"},{"comment":"In the paragraph beginning 'CIL with model merging:', there is a missing space before the colon: 'CIL with model merging:Several' should read 'CIL with model merging: Several'.","section":"Section 2"},{"comment":"The caption 'The curve showing the differences in cosine similarity that arise when earlier task prototypes are substituted for prototypes in subsequent subspaces' is grammatically awkward; consider rewording to 'Cosine similarity between the true prototypes in the current subspace and the substitutes used when earlier task prototypes are reused.'","section":"Figure 4 caption"},{"comment":"The notation ∆P in Eq. (7) and in Algorithm 1 is defined as a matrix formed by repeating the vector ∆p, but Eq. (6) writes Pi(¯At) ≈ Pi(¯Ai) + ∆P, which is ambiguous because Pi(¯Ai) is a matrix while ∆p is a vector. Clarify the broadcasting convention, for instance by writing Pi(¯At) ≈ Pi(¯Ai) + 1⊤⊗∆p.","section":"Section 4.2, Algorithm 1"},{"comment":"The balanced-VTAB experiment is described in the text but the figure caption (Figure F) does not state the number of runs or whether the same seeds are used; please specify the experimental setup for this ablation.","section":"Appendix E.2"}],"recommendation":"major_revision","confidential_remarks":"The paper presents a simple and effective method for scalable exemplar-free CIL, but the authors overstate the accuracy claim in the abstract. The VTAB gap is not just a minor outlier; it is a 6-point final-accuracy drop that the current explanation (data imbalance) only partially justifies. The missing error bars are also a significant weakness, as the paper essentially claims parity with EASE based on differences that are likely within noise. I would encourage the editor to request a revision that qualifies the claims, adds statistical detail, and critically examines the conditions under which the centroid mapping assumption holds. The method itself is a solid contribution and deserves a chance to be published after these issues are addressed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"ACMap is the kind of paper that doesn't try to be a paradigm shift but solves a real engineering problem. It merges task-specific adapters into one by simple weight averaging, starting every adapter from the first task's weights, and then realigns old prototypes with a single centroid shift. Net result: EASE-level final accuracy on four of five benchmarks at O(1) inference time, up to 39x faster on ImageNet-R. That is a real, useful outcome, and the recipe is new even though each ingredient exists elsewhere.\n\nThe paper is also honest where it matters. The ablation shows both initial-weight replacement and centroid mapping contribute; the early stopping threshold is essentially free, with L=10 matching L=infinity. Code is out. The loss landscape analysis is a nice touch.\n\nSoft spots are real but not fatal. The centroid prototype mapping assumes the drift between current-task prototypes in the old and new adapter is the same for every previous task. That's an empirical heuristic. It's verified only with cosine similarity computed using previous-task validation data, which an exemplar-free deployment doesn't have. And VTAB is a direct counterexample: ACMap's final accuracy is 87.56 vs EASE's 93.55, a 6-point gap. The authors point at the data imbalance and a 'balanced VTAB' experiment shows parity, but Table 1 is the standard imbalanced protocol. So the abstract's claim that ACMap 'matches state-of-the-art accuracy' without qualification is not supported as written; 'on four of five benchmarks' would be accurate.\n\nTwo smaller issues. The five-run averages have no error bars, so we don't know if the remaining small gaps to EASE are noise. And the baseline numbers are taken from the EASE paper, not recomputed. Both are common in this literature and not disqualifying.\n\nWho should read it: anyone working on exemplar-free CIL with pre-trained models who cares about inference cost scaling with T. It is a solid follow-up to the EASE line, not a break with it. A serious referee should spend time on it; the main thing to fix is the overbroad abstract and ideally some variance estimates or a clearer statement on when the centroid shift can be expected to break.","headline":"Useful, honest adapter-merging CIL paper with real O(1) inference gains; the abstract overclaims SOTA parity because VTAB clearly breaks the centroid-mapping assumption.","tokens_in":19532,"tokens_out":2215,"would_cite":true,"duration_ms":20430,"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":"ACMap claims to break the accuracy–inference trade-off in exemplar-free class-incremental learning by merging per-task adapters into a single shared adapter and re-aligning old prototypes with one centroid shift measured on the current…","keywords":["class-incremental learning","exemplar-free continual learning","adapter merging","centroid prototype mapping","catastrophic forgetting","parameter-efficient fine-tuning","vision transformer","cosine classifier"],"falsifier":"A decisive check is to measure the per-task drift on a benchmark with heterogeneous domains and balanced task sizes: compute $\\Delta P_i = P_i(\\bar{A}_t) - P_i(\\bar{A}_i)$ for each old task $i$ and compare the spread of these shifts to the single centroid shift $\\Delta P$. If the spread is comparable to inter-class prototype distances, the re-aligned prototypes will point at wrong classes and final accuracy should fall, as the paper's own VTAB B0 Inc10 result (Section 5.2, Table 1) hints by dropping after the fourth task.","tokens_in":18510,"feed_emoji":"🧩","tokens_out":8743,"duration_ms":70977,"temperature":0.7,"pith_summary":"Class-incremental learning asks a model to keep learning new classes without forgetting old ones, and privacy rules often forbid keeping old images. Existing exemplar-free methods built on pre-trained models either retrain a fresh adapter per task, which is accurate but slow at inference because every task's adapter must run, or reuse a single adapter, which is fast but less accurate. This paper claims a middle path: train one adapter per task but merge them by simple weight averaging into a single adapter, so inference costs stay constant no matter how many tasks have arrived. To keep old class prototypes comparable inside the merged space, the method shifts them by a single translation vector estimated from the current task's class centroids. On five benchmarks the method matches the accuracy of the strongest adapter-ensemble baseline while running at the speed of the fastest prototype baselines, for example a 39-fold inference speedup on the 40-task ImageNet-R setting.","feed_headline":"Merging adapters cuts continual learning inference 39x, same accuracy","feed_subtitle":"Class-incremental model collapses per-task adapters into a shared subspace and re-aligns old prototypes with one centroid shift.","key_machinery":"The central objects are the task-specific adapters $A_t$, the merged adapter $\\bar{A}_t = (1 - 1/t)\\bar{A}_{t-1} + (1/t)A_t$, and the centroid prototype mapping that approximates unavailable old prototypes by $P_i(\\bar{A}_t) \\approx P_i(\\bar{A}_i) + \\Delta P$, where $\\Delta P$ is the mean difference between the current task's prototypes computed in $\\bar{A}_t$ and in $\\bar{A}_i$. The first task's trained weights replace the random initialization for all later adapters, creating a shared low-loss basin that makes weight averaging safe. Early stopping sets a threshold $L$ after which no new adapter is trained, so the merged adapter stops changing. These pieces together reduce inference from $O(T)$ to $O(1)$ while keeping the classifier a cosine classifier over re-aligned prototypes.","core_discovery":"The paper's central claim is that catastrophic forgetting and scalability in exemplar-free class-incremental learning can be handled together by consolidating task-specific adapters into one merged adapter. The consolidation is average merging of adapter weights, made effective by initializing every task's adapter from the first task's trained weights so that the tasks occupy a shared low-loss basin. Because old data is unavailable, previous prototypes cannot be recomputed in the merged space; the paper claims that the drift of the current task's prototypes between an old adapter and the merged adapter is a good proxy for the drift of every earlier task's prototypes, so a single affine shift $\\Delta P = E[P_t(\\bar{A}_t) - P_t(\\bar{A}_i)]$ re-aligns all old prototypes. With this mapping, ACMap reports final-accuracy parity with the EASE ensemble (70.49 versus 70.58 on ImageNet-R task 40) while reducing inference time from 916.5 seconds to 23.5 seconds. The paper also claims that stopping the merging after about ten tasks saves training computation with essentially no accuracy loss.","pith_inferences":["Editorial extension: if the drift of prototypes is roughly task-invariant, the same centroid-shift trick could be applied to other parameter-efficient modules, such as prompt pools, to obtain constant inference time there too.","Editorial extension: the dependence of the shift on task order suggests a testable refinement — instead of a single global shift, use a shift estimated from the most recent tasks and interpolate for older tasks, which might recover the VTAB gap while keeping $O(1)$ inference.","Editorial extension: the early-stopping result implies that after roughly ten tasks the merged adapter has stabilized, so the method could be extended to never-ending streams by freezing the merged adapter and training new adapters only when a drift signal, such as prototype cosine similarity, exceeds a threshold.","Editorial extension: the paper does not report total stored-parameter counts against ensemble methods; a direct comparison would clarify the memory advantage of storing one merged adapter rather than $T$ adapters."],"forward_implications":["Inference time stays constant as tasks grow: on ImageNet-R B0 Inc5 task 40, ACMap runs in 23.5 seconds versus 916.5 seconds for the strongest ensemble baseline, a 39-fold speedup, with final accuracy 70.49 versus 70.58.","On five benchmarks, ACMap matches or slightly exceeds the state-of-the-art adapter-ensemble accuracy on all settings except VTAB B0 Inc10, where the ensemble baseline keeps separate adapters per domain.","Early stopping at $L=10$ achieves accuracy nearly identical to $L=\\infty$ on CIFAR-100 and ImageNet-R, so training cost can be cut without hurting accuracy.","Both initial-weight replacement and centroid prototype mapping contribute: the ablation shows each raises accuracy, and their combination gives the best results.","Because only one merged adapter is used at inference, the method stays exemplar-free and privacy-compatible."],"supporting_citations":[{"why":"Supplies the EASE baseline and the evaluation protocol that ACMap matches in accuracy while replacing its O(T) inference with O(1).","marker":"[59]"},{"why":"Supplies the SimpleCIL and APER prototype baselines whose inference time and accuracy ACMap is compared against.","marker":"[61]"},{"why":"Model soups average merging is the weight-averaging technique that adapter merging builds on.","marker":"[51]"},{"why":"Supports the claim that weight-averaging models from a shared initialization produces a low-loss basin that makes merging effective.","marker":"[24]"},{"why":"Semantic drift compensation is the earlier prototype-shift method that centroid prototype mapping is compared with and shown to improve on.","marker":"[55]"},{"why":"Provides the adapter architecture used for the per-task adapter module, and its fine-tuning baseline appears in the comparisons.","marker":"[5]"},{"why":"Provides the training hyperparameters and protocol followed for all datasets, so the reported numbers are comparable to the EASE baseline.","marker":"[62]"}],"fun_headline_variants":["Adapter merging with centroid mapping achieves constant inference time in CIL","Merging adapters across tasks cuts inference time 39x with no accuracy drop","Exemplar-free CIL: one adapter shift re-aligns old prototypes accurately","ACMap consolidates adapters, matching EASE accuracy at 39x lower cost","Class-incremental learning without exemplars: adapter merge + prototype shift"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole method depends on the assumption that the shift between the current task's class prototypes computed in an old adapter and in the merged adapter is about the same for every earlier task, so one average shift can re-align all stored prototypes; if older tasks drift in different directions, those prototypes point to the wrong places.","fun_headline_variants_meta":{"raw":{"variants":["Adapter merging with centroid mapping achieves constant inference time in CIL","Merging adapters across tasks cuts inference time 39x with no accuracy drop","Exemplar-free CIL: one adapter shift re-aligns old prototypes accurately","ACMap consolidates adapters, matching EASE accuracy at 39x lower cost","Class-incremental learning without exemplars: adapter merge + prototype shift"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000471,"raw_usage":{"total_tokens":2330,"prompt_tokens":921,"completion_tokens":1409,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":537,"completion_tokens_details":{"reasoning_tokens":1317}},"tokens_in":537,"tokens_out":1409,"duration_ms":10474,"temperature":1.0,"reasoning_tokens":1317,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T04:54:39.459273+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive check is to measure the per-task drift on a benchmark with heterogeneous domains and balanced task sizes: compute $\\Delta P_i = P_i(\\bar{A}_t) - P_i(\\bar{A}_i)$ for each old task $i$ and compare the spread of these shifts to the single centroid shift $\\Delta P$. If the spread is comparable to inter-class prototype distances, the re-aligned prototypes will point at wrong classes and final accuracy should fall, as the paper's own VTAB B0 Inc10 result (Section 5.2, Table 1) hints by dropping after the fourth task.","supporting_citations":[{"cited_title":"Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time","cited_arxiv_id":null,"evidence_quote":"Model soups average merging is the weight-averaging technique that adapter merging builds on."},{"cited_title":"Expandable Subspace Ensemble for Pre-Trained Model-Based Class-Incremental Learning","cited_arxiv_id":null,"evidence_quote":"Supplies the EASE baseline and the evaluation protocol that ACMap matches in accuracy while replacing its O(T) inference with O(1)."},{"cited_title":"Revisiting Class-Incremental Learning with Pre-Trained Models: Generalizability and Adaptivity are All You Need","cited_arxiv_id":null,"evidence_quote":"Supplies the SimpleCIL and APER prototype baselines whose inference time and accuracy ACMap is compared against."},{"cited_title":"Semantic Drift Compensation for Class-Incremental Learning","cited_arxiv_id":null,"evidence_quote":"Semantic drift compensation is the earlier prototype-shift method that centroid prototype mapping is compared with and shown to improve on."},{"cited_title":"AdaptFormer: Adapting Vision Transformers for Scalable Visual Recogni- tion","cited_arxiv_id":null,"evidence_quote":"Provides the adapter architecture used for the per-task adapter module, and its fine-tuning baseline appears in the comparisons."},{"cited_title":"https://github.com/sun-hailong/CVPR24-Ease, 2024","cited_arxiv_id":null,"evidence_quote":"Provides the training hyperparameters and protocol followed for all datasets, so the reported numbers are comparable to the EASE baseline."}],"review_version":1}