{"id":"ddb30c0d-833d-4841-a3d7-326cfb02e109","arxiv_id":"2412.10900","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"PEARL uses an input-agnostic global prompt with a negative-feedback adaptive momentum update to reduce catastrophic forgetting in pre-trained-model class-incremental learning, achieving state-of-the-art average accuracy of 86.41% across six benchmarks.","lead":"A new method for class-incremental learning replaces the usual 'query-select' prompt mechanism with one global prompt that is adapted by an output-feedback momentum rule, reporting gains over previous state-of-the-art on six benchmarks. It matters because continual learning systems need to add new classes without forgetting old ones, and this method is a candidate for making prompt-based adaptation more stable.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 10's forgetting proxy is unvalidated: MAE on new-task old-class logits may not track real retention, so the NKA mechanism's role in achieving SOTA is not established.","rationale":"The strongest claim has two parts: the empirical SOTA results and the mechanistic explanation via negative feedback. The empirical results are substantial, but the mechanism is the paper's central novelty. The feedback signal in Eq. 10 is the only link between the observed outputs and the adaptive momentum weight in Eq. 11. If that signal does not measure forgetting, the negative-feedback regulation is not doing what the paper claims, and the method's success would have to be explained by a different mechanism (e.g., an implicitly learned alpha schedule). The reader's weakest assumption identifies exactly this proxy concern, and I agree it is the most load-bearing: it is conceptually unvalidated, and the paper provides no evidence that mae correlates with actual retention. A direct correlation test on standard benchmarks would settle this. The evaluation-protocol issues (no error bars, hyperparameter tuning without a validation split) are real but secondary because the SOTA margins are large and the methods follow common practice; the NKA proxy is more fundamental to the contribution. Since the reader already assigned CONDITIONAL and this concern reinforces that condition, the verdict should remain unchanged.","tokens_in":21897,"tokens_out":11071,"duration_ms":99782,"concrete_test":"Run PEARL on CIFAR-100 and CUB-200 with 10 sessions. At each session t, record the mae from Eq. 10 and independently measure true forgetting: compute accuracy on a held-out test set of all classes seen before session t, before and after the session's prompt update, and take the difference. Repeat for 3 random seeds, then compute the Spearman correlation between the per-session mae values and the true forgetting values. If the correlation is near zero or negative, Eq. 10 does not track retention, invalidating the negative-feedback claim. As a complementary check, replace the mae input to sigma with a constant matching the converged alpha in Fig. 4; if final accuracy is unchanged, the adaptive signal is not the cause of the improvement.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that PEARL achieves state-of-the-art performance through negative-feedback regulation (Eqs. 3 and 11) depends on the feedback signal in Eq. 10: mae = MAE(l_t[0:K(t-1)]*lambda, l_{t-1}*lambda). Both logits are computed on the current session's new-task images, and the MAE is restricted to old-class logits. For these inputs, the old-class logits are out-of-distribution responses; their stability under a prompt update does not measure whether the model has forgotten previously learned classes. True forgetting is the drop in accuracy on old-task test data after learning a new task, which the paper never measures or correlates with mae. Thus, the adaptive alpha may simply act as a stable learning-rate schedule rather than a forgetting safeguard. The ablation in Table 4 shows NKA outperforms fixed momentum, but this could be because the alpha trajectory converges quickly to a near-optimal constant (see Fig. 4), not because the mae signal carries diagnostic information about retention. Without a direct comparison between mae and actual forgetting, the mechanism's causal contribution to the reported SOTA numbers remains unsubstantiated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes PEARL, a rehearsal-free class-incremental learning method built on a frozen pretrained ViT. It removes the input-dependent \"query-select\" prompt choice and instead maintains an input-agnostic global prompt generated by a small prompt encoder over a prompt pool; a momentum update (Eq. 3) mixes the previous and current prompt tokens, and the momentum weight is adapted by the NKA mechanism (Eq. 11) using the MAE between current and previous logits on old-class outputs (Eq. 10). Experiments on six benchmarks and several ablations report average/final accuracy exceeding prior prompt-based and representation-based methods, including a 2.24% average-accuracy gain over RanPAC, the second-best method.","tokens_in":22164,"tokens_out":6439,"duration_ms":60525,"significance":"If the empirical claims hold, PEARL is a conceptually simple and potentially impactful result: it shows that an input-agnostic prompt with adaptive momentum can outperform prior prompt-based, representation-based, and model-mixture methods in PTM-based CIL, challenging the recent survey conclusion that prompt-based methods underperform. The paper ships code, uses standard benchmarks and backbones, reports multiple ablations including hyperparameter sensitivity, and includes an OVOR comparison and forgetting-rate tables; these are genuine strengths. The main weaknesses are that the forgetting-proxy signal in Eq. 10 is not validated against actual forgetting, hyperparameters are selected on the same benchmarks without a validation split, and Table 1 lacks variance estimates. These issues do not invalidate the method but weaken the causal-explanatory and state-of-the-art claims as currently stated.","major_comments":[{"comment":"The mae signal is computed on current-session images, using the old-class slice of the current logits and the previous logits for the same images. Since these images come from the new task, the old-class logits are responses to out-of-distribution inputs; stability of these responses under a prompt update does not by itself measure retention of previously learned classes. True forgetting is the drop in accuracy on old-task test data after learning a new task, and the paper never measures or correlates mae with such forgetting. The ablation in Table 4 and the convergence curves in Fig. 4 are consistent with NKA acting as a robust learning-rate schedule: alpha converges to a near-constant value regardless of initialization. Please add per-session measurements of actual forgetting on old-class test data, report a direct correlation between mae and forgetting, or include an ablation in which the adaptive signal is replaced by a fixed schedule to show that the adaptive component, not merely the converged schedule, is responsible for the reported gains.","section":"Negative-feedback Knowledge Accumulation, Eq. (10); Table 4"},{"comment":"With the prompt-pool length M=100 and the ImageNet-R 40-task setting, M/N = 2.5, so the slice P[1 + M/N * (t-1) : M/N * t] is not an integer range. It is unclear whether the reported results use floor/ceil boundaries, an adjusted pool length, or some other allocation of prompts to tasks. Please specify the exact slicing used for each benchmark and ensure that the pseudo-code, Eq. (7), and the segmented positional encoding are consistent for all six benchmarks, since this directly affects reproducibility.","section":"Sequential Prompt Adaptation, Eq. (7); Implementation Details"},{"comment":"The values lambda=12500, gamma=0.99, theta_min/theta_max, and the prompt-pool length are selected experimentally on the same benchmarks used for the final comparison, with no held-out validation split or nested model-selection protocol. The supplementary text states that gamma and lambda are \"determined experimentally\" and reports tuning curves on VTAB and ImageNet-A. Tuning on the test benchmark can inflate the reported SOTA numbers. Please either use a validation split for hyperparameter selection or report the complete sensitivity grid and justify that the final choices do not materially change the comparison with RanPAC and other baselines.","section":"Supplementary Section 5; Implementation Details"},{"comment":"The text states that results are the average of three random runs, but Table 1 reports no standard deviations or confidence intervals. Given that the gain over RanPAC is only 0.13% in average accuracy on CIFAR-100 and PEARL's final accuracy is lower than RanPAC there (89.02 vs. 89.30), the \"state-of-the-art\" claim needs error bars and, ideally, a paired comparison across the three runs to establish that the reported improvements are not within run-to-run variation.","section":"Table 1 and Comparison with State-of-the-art Methods"}],"minor_comments":[{"comment":"The objective is written as arg max E_{(x,y)~D_t} I(y != l_t), which maximizes the error; it should be I(y = l_t) or an equivalent loss-minimization form.","section":"Eq. (4)"},{"comment":"The text refers to \"CIFAR00\" in the datasets paragraph; this should be CIFAR-100.","section":"Datasets, Experiments"},{"comment":"Please specify the shapes of l_t and l_{t-1} and clarify where lambda is applied; as written it is ambiguous whether lambda multiplies both terms and how the different lengths of the two logit vectors are handled.","section":"Eq. (10)"},{"comment":"The method name \"RacPAC\" is a typo; it should be RanPAC.","section":"Supplementary Table 4"},{"comment":"Line 10 contains the typo \"Initilize\"; also, the pseudo-code resets alpha_tau to 0.99 at the start of every session, which should be stated explicitly in the main text alongside Eq. (11).","section":"Algorithm 1"}],"recommendation":"major_revision","confidential_remarks":"This is a standard empirical CIL paper with a clear scope fit. The requested experiments, particularly the validation of the Eq. (10) forgetting proxy against actual per-session forgetting, are feasible within the manuscript's scope. If those experiments confirm the proxy's diagnostic value, the paper could become a solid contribution; if not, the NKA mechanism's claimed causal role would need to be substantially reframed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Honest take: this is a real contribution. The combination of an input-agnostic global prompt—a small prompt encoder over a per-task partitioned pool with segmented positional encoding—and an adaptive momentum update is not something I've seen in L2P, DualPrompt, CODA-Prompt, OVOR, or HiDe-Prompt. The reported gains, about 2.2 average points over RanPAC across six benchmarks, are substantial if they hold. The ablations are mostly convincing: both the encoder and momentum are needed, and the adaptive momentum is robust to the initial α where fixed momentum collapses at low α. Code is available; the protocol is standard; the supplementary includes hyperparameter sweeps and forgetting rates.\n\nThe main soft spot is the NKA feedback signal. Eq. (10) computes MAE between current and previous logits on the current session's new-task images, restricted to old-class outputs. The paper calls that a measure of knowledge retention, but it never validates the proxy against actual old-task accuracy. On new-task images, old-class logits are out-of-distribution responses; their stability under a prompt update does not obviously track whether old-class accuracy has dropped. So the attractive story—negative feedback regulates forgetting—is not established. The ablation shows NKA beats fixed momentum, but that's consistent with the adaptive rule simply finding a good learning-rate schedule; Fig. 4 shows α converges to a constant for all initializations, which suggests the signal may be doing less diagnostic work than advertised.\n\nSecond, the SOTA numbers rest on hyperparameters (λ, γ, θ bounds, α0) selected from experiments on the same benchmarks used in the final comparison, with no proper validation split. That's common in this subfield, but it does weaken the significance claim. Also, Table 1 reports averages over three runs without error bars, so I can't tell how much of the 2.24% gap is noise.\n\nFinally, the 'knowledge container theory' is a metaphor, not a theory, and 'first use of negative feedback in incremental learning' is the kind of claim that won't survive contact with the literature on control-theoretic continual learning. Those overclaims should be trimmed, but they don't undermine the method.\n\nThis deserves peer review. A referee should push for a direct test of the forgetting proxy: correlate mae with actual old-task accuracy drops, and add error bars. If the proxy validation fails, the paper still stands as an adaptive-momentum prompt method with strong numbers, just with a less exciting story.","headline":"A genuinely new prompt-based CIL method with strong numbers, but the negative-feedback story is under-validated and needs referee pressure.","tokens_in":22688,"tokens_out":4155,"would_cite":true,"duration_ms":37746,"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":"PEARL claims input-agnostic prompts beat prior class-incremental methods by 2.24% average accuracy.","keywords":["class-incremental learning","pre-trained models","prompt learning","catastrophic forgetting","input-agnostic prompt","negative feedback regulation","momentum update","continual learning"],"falsifier":"Take a stream whose new tasks come from a visually different domain and hold out a test set of previous-task classes. If PEARL's $\\alpha_\\tau$ is pushed up by the MAE on new-task inputs while true accuracy on the held-out old classes drops, or if PEARL with adaptive $\\alpha$ does not beat a fixed $\\alpha$ tuned on the same stream, the central claim is refuted. A direct version: measure the correlation between the NKA MAE and held-out old-class accuracy across sessions; a correlation near zero on any benchmark would show the feedback signal is not a forgetting proxy.","tokens_in":21680,"feed_emoji":"🧠","tokens_out":6579,"duration_ms":50777,"temperature":0.7,"pith_summary":"This paper argues that prompt-based class-incremental learning has been held back by its reliance on input-dependent prompt selection, the \"query-select\" mechanism that mixes knowledge from different tasks in a disorganized way. It proposes PEARL, a method that replaces selection with an input-agnostic global prompt generated by a small prompt encoder, updated by a momentum rule whose weight is set by negative feedback from the model's own outputs. The claim is that this removes the model's dependence on data distribution, regulates forgetting without any distillation loss, and lets the prompt keep accumulating knowledge across sessions. If the claim holds, prompt-based methods move from lagging other pre-trained-model approaches to leading them, with state-of-the-art average accuracy on six benchmarks and a small memory footprint.","feed_headline":"Input-agnostic prompts lift class-incremental accuracy 2.24%","feed_subtitle":"Negative feedback sets the momentum weight, preserving old classes while learning new ones across six benchmarks.","key_machinery":"Two coupled components carry the argument. Sequential Prompt Adaptation (SPA) treats the prompt pool as a sequence, assigns each task a frozen segment, and passes the concatenation through a session-sharing Transformer encoder to produce a global prompt token $[PT]^t_i$ with segmented positional encoding that marks which task each segment belongs to. Negative-feedback Knowledge Accumulation (NKA) then blends the previous session's prompt token with the current one via $[PT]^{mem}_i = \\alpha_\\tau [PT]^{t-1}_i + (1-\\alpha_\\tau)[PT]^t_i$, where $\\alpha_\\tau = \\gamma \\alpha_{\\tau-1} + (1-\\gamma)\\sigma(\\mathrm{MAE}(l_t[0:K(t-1)]\\cdot\\lambda, l_{t-1}\\cdot\\lambda))$. The MAE on the old-class slice is the feedback signal: it is the paper's proxy for how much old knowledge has been lost, and it makes the momentum weight adaptive rather than hand-tuned.","core_discovery":"PEARL's central claim is that catastrophic forgetting in pre-trained-model class-incremental learning can be controlled by a single global prompt that is the same for every input in a session, updated with an adaptive momentum weight. The global prompt is produced by a frozen-backbone prompt encoder over a task-partitioned prompt pool, so no per-input query or selection is needed. The momentum weight $\\alpha_\\tau$ is driven by the mean absolute error between current and previous logits restricted to old classes; high divergence raises $\\alpha_\\tau$ to preserve the old prompt, low divergence lowers it to admit new knowledge. The paper reports that this mechanism produces state-of-the-art results, beating the second-best method by 2.24% average accuracy and 1.65% final accuracy across six benchmarks, and that it converges to a task-dependent value independent of the initialization of $\\alpha_0$.","pith_inferences":["The NKA feedback signal is computed on current-task data, so if new-task inputs are far out-of-distribution for old classes, the MAE may stop tracking true forgetting; a testable extension is to replace the old-class slice with a small stored exemplar set and compare.","Because the paper shows $\\alpha_\\tau$ converges to a value that varies across tasks, the converged weight could be used as a measure of inter-task correlation to decide how many prompts to allocate to future tasks.","The same negative-feedback momentum recipe could transfer to domain-incremental or task-agnostic settings, where the old-class slice would be replaced by a domain-discrimination or novelty score."],"forward_implications":["Prompt-based class-incremental learning no longer needs input-dependent selection; inference on a new sample uses one global prompt token, cutting both compute and the risk of mismatched prompts.","The method is insensitive to the initial momentum weight: with fixed $\\alpha$, accuracy collapses when $\\alpha_0$ is low, while the NKA update maintains high accuracy across initial values from 0.6 to 0.999.","Knowledge retention is handled by the adaptive weight instead of a distillation loss, simplifying the training objective to plain cross-entropy.","The approach works across streams of different lengths and domains: six benchmarks from 5-task to 40-task streams all show average and final accuracy above the compared methods."],"supporting_citations":[{"why":"Defines the L2P query-select baseline that PEARL diagnoses as the source of disorganized knowledge mixing and replaces with an input-agnostic prompt.","marker":"Wang et al. 2022b"},{"why":"Provides the DualPrompt baseline and the general-expert prompt design used as a comparison and ablation target.","marker":"Wang et al. 2022a"},{"why":"Supplies CODA-Prompt, the prompt-pool decomposition and attention combination that PEARL builds on and compares against.","marker":"Smith et al. 2023"},{"why":"RanPAC is the second-best baseline and the source of the no-trainable-parameter classification head update used in PEARL.","marker":"McDonnell et al. 2024"},{"why":"Survey that supplies the claim that prompt-based methods underperform other PTM-based approaches, which PEARL sets out to overturn.","marker":"Zhou et al. 2024a"},{"why":"Prefix-tuning defines how the prompt tokens are attached as key/value prefixes in the attention blocks of the frozen ViT.","marker":"Li and Liang 2021"},{"why":"The Transformer encoder and positional encoding that SPA adapts into a session-sharing prompt encoder with segmented positions.","marker":"Vaswani et al. 2017"},{"why":"Provides the six-benchmark protocol and the average/final accuracy metrics used in the experiments.","marker":"Zhou et al. 2024c"}],"fun_headline_variants":["PEARL's input-agnostic global prompt lifts CIL accuracy 2.24%","Adaptive momentum global prompt: CIL forgetting tamed, SOTA","One prompt, no input queries: PEARL's CIL edge with negative feedback","Negative feedback tunes prompt, class-incremental accuracy up 2.24%","Global prompt with adaptive momentum: PEARL beats CIL second best"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole adaptive schedule rests on the assumption that the mean absolute error between current and previous logits, measured only on old-class outputs for inputs of the current session, actually tracks how much old knowledge the model has forgotten; if new-task inputs are not informative about old classes, the feedback weight will not regulate forgetting correctly.","fun_headline_variants_meta":{"raw":{"variants":["PEARL's input-agnostic global prompt lifts CIL accuracy 2.24%","Adaptive momentum global prompt: CIL forgetting tamed, SOTA","One prompt, no input queries: PEARL's CIL edge with negative feedback","Negative feedback tunes prompt, class-incremental accuracy up 2.24%","Global prompt with adaptive momentum: PEARL beats CIL second best"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00041,"raw_usage":{"total_tokens":2125,"prompt_tokens":947,"completion_tokens":1178,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":563,"completion_tokens_details":{"reasoning_tokens":1072}},"tokens_in":563,"tokens_out":1178,"duration_ms":8704,"temperature":1.0,"reasoning_tokens":1072,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T15:29:51.272292+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a stream whose new tasks come from a visually different domain and hold out a test set of previous-task classes. If PEARL's $\\alpha_\\tau$ is pushed up by the MAE on new-task inputs while true accuracy on the held-out old classes drops, or if PEARL with adaptive $\\alpha$ does not beat a fixed $\\alpha$ tuned on the same stream, the central claim is refuted. A direct version: measure the correlation between the NKA MAE and held-out old-class accuracy across sessions; a correlation near zero on any benchmark would show the feedback signal is not a forgetting proxy.","supporting_citations":[],"review_version":1}