{"id":"b0091a64-8634-43fb-af79-ae5610875900","arxiv_id":"2411.16073","paper_version":4,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"Per-task multiplicative masks on frozen transformer attention layers give strong class-incremental accuracy, but the claimed convergence proof and forgetting-free guarantee are not supported.","lead":"Soft-TransFormers learns a per-task real-valued mask on top of a frozen pre-trained transformer, claiming class-incremental learning with near-zero forgetting. The paper reports strong benchmark gains over prompts, adapters, and LoRA, but its theory is flawed and the headline numbers rely on test-time batch inference.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Appendix A.1's Lemma 1 assumes the conclusion: the telescoping identity in Eq. (16) is algebraically false, so the bounded-drift and harmless-task-error claims are unproven.","rationale":"The reader's weakest-assumption pinpoints the same flaw: Eq. (16) in Appendix A.1 assumes ||w(1)-w*||^2 = ||w(T+1)-w*||^2, which is not implied by the update rule. My reading of the manuscript confirms this is not a typo isolated to one line; the same false equality is reused in Eqs. (12), (16) and (20), and the strict ordering B_m < B_p < B is asserted rather than proven. These are the load-bearing steps for the abstract's claims of bounded drift, structural elimination of forgetting, and harmlessness of task-inference errors. The empirical sections still contain a strong method and extensive comparisons, but the central theoretical argument is currently unsupported. I additionally notice a separate, concrete internal inconsistency: the 'Upper-Bound of Soft-TF' entries (93.90 on 10-Split-CIFAR100 in Table 3; same value in Table 6) are lower than the reported Gradient-ID accuracies (97.87 and 98.05) for the same benchmark. If the upper bound is meant to be the best achievable accuracy with perfect task inference, the reported numbers contradict it; if it is not, the label is misleading. This makes the empirical claims harder to trust, although it is secondary to the proof failure. I therefore see no new concern that would move the reader's REJECT verdict in a different direction; the appropriate disposition is unchanged.","tokens_in":27718,"tokens_out":3831,"duration_ms":33087,"concrete_test":"Independently re-derive Lemma 1 for the update w(t+1)=w(t)-eta*v_t. Write the telescoping sum explicitly without the erroneous -||w(t+1)-eta*v_t||^2 term, and verify whether Eq. (16) holds or whether it should be ||w(1)-w*||^2 - ||w(T+1)-w*||^2. Then, using the released code on 10-Split-CIFAR100, compute B_m = ||w_m(T+1)-w*|| for the mask update and B_p = ||w_p(T+1)-w*|| for the prompt update at the end of training; if B_m >= B_p, Theorem 2's ordering fails. Finally, run the Soft-TF model with oracle (true) task IDs on 10-Split-CIFAR100 and compare the resulting accuracy to the printed 'Upper-Bound of Soft-TF' of 93.90; if oracle accuracy exceeds 93.90, the table is internally inconsistent and the reported numbers cannot all be correct.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central theoretical claim is that mask-only fine-tuning has bounded drift, which is used to justify 'forgetting is structurally eliminated' and 'task-inference errors are largely harmless.' This rests entirely on Lemma 1 and Theorem 2 in Appendix A.1, and that derivation is invalid. With the update rule w(t+1)=w(t)-eta*v_t, the completed-square identity in Eq. (14) should read D(w(t)-w*, v_t) = (1/(2eta))[-||w(t+1)-w*||^2 + ||w(t)-w*||^2] + (eta/2)||v_t||^2. The paper instead writes -||w(t+1)-eta*v_t||^2, which equals -||w(t)-2eta*v_t||^2, not the required squared distance to w*. Then Eq. (16) asserts that the telescoping sum collapses to ||w(1)-w*||^2 = ||w(T+1)-w*||^2. The telescoping sum actually equals ||w(1)-w*||^2 - ||w(T+1)-w*||^2, which is zero only if the conclusion is already assumed. A valid proof would drop the negative term to get an upper bound, not an equality, and the strict ordering B_m < B_p < B in Eq. (13) and (21) would still be needed; that ordering is asserted without derivation. Since the bounded-drift bound and the ordering are exactly what support the decoupling claim, the theory does not currently establish the abstract's central claims. There is also an internal numerical inconsistency: the 'Upper-Bound of Soft-TF' (93.90 on 10-Split-CIFAR100, Table 3 and Table 6) is lower than the reported Gradient-ID accuracy (97.87 in Table 3, 98.05 in Table 8), which is impossible if the upper bound is computed with perfect task identity.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Soft-TransFormers (Soft-TF), a continual learning method that keeps a pretrained transformer frozen and learns per-task real-valued multiplicative masks over selected self-attention layers, optionally combined with prompts. The authors claim that mask-only fine-tuning enjoys bounded parameter drift, that forgetting is structurally eliminated, and that wrong-task inference is largely harmless; they support these claims with a convex-Lipschitz convergence analysis in Appendix A.1. The empirical section reports strong results on Split-CIFAR100, Split-ImageNet-R, CUB-200, TinyImageNet, and a BERT language benchmark against prompt-, adapter-, and LoRA-based baselines.","tokens_in":28048,"tokens_out":7082,"duration_ms":56427,"significance":"If the claims were established, the paper would make a substantial contribution: a rehearsal-free, parameter-efficient CIL method with a frozen backbone, no task-ID oracle, and state-of-the-art accuracy would be practically valuable. The experimental program is broad, includes ablations over mask location and initialization, comparisons with WSN, and provides code links. However, the paper's distinctive theoretical contribution is the advertised proof of bounded drift and harmless task-inference errors, and that proof is not valid as written. Moreover, the paper's own tables contradict the decoupling claim and the 'Upper-Bound' label. The empirical results may still be of interest, but they do not, by themselves, establish the paper's central narrative.","major_comments":[{"comment":"The completed-square identity is algebraically incorrect. With the update w(t+1)=w(t)-eta*v_t, the correct identity is <w(t)-w*, v_t> = (1/(2*eta))[-||w(t+1)-w*||^2 + ||w(t)-w*||^2] + (eta/2)||v_t||^2. The paper instead writes -||w(t+1)-eta*v_t||^2, which equals -||w(t)-2*eta*v_t||^2 and is not the squared distance to w*. Consequently Eq. (16) is not a telescoping identity: the sum telescopes to ||w(1)-w*||^2 - ||w(T+1)-w*||^2, which equals zero only if the lemma's conclusion is assumed. The lemma also states the condition ||w(1)-w*||^2 = ||w(T+1)-w*||^2 as an assumption and later, in Eq. (17), invokes 'w(1)=0', contradicting the premise w(1) != 0 used in the Soft-TransFormer setting. Lemma 1 and Theorem 2 therefore do not establish bounded drift.","section":"Appendix A.1, Lemma 1, Eqs. (14)-(16)"},{"comment":"The strict ordering B_m < B_p < B is asserted, not derived. The text supports it only with 'since all m are learnable parameters', but no argument shows that the distance from the mask-adapted final iterate to w* is smaller than the corresponding prompt-only distance, nor that both are smaller than the full-parameter distance. Because this ordering is exactly what is used to conclude that wrong-task masks are harmless and that accuracy is decoupled from task-inference reliability, the main theoretical conclusion is unsupported.","section":"Appendix A.1, Eqs. (13), (20), (21)"},{"comment":"The reported 'Upper-Bound of Soft-TF' is not an upper bound. On 10-Split-CIFAR100, Tables 3 and 6 report an upper bound of 93.90, while the same tables report Soft-TF with Gradient ID at 97.87, and Table 8 reports 98.05 for the same configuration. An upper bound cannot be lower than the method's own accuracy. The paper does not define how the upper-bound row is computed; this internal inconsistency undermines the claim that Soft-TF reaches an upper bound.","section":"Tables 3, 6, and 8"},{"comment":"The claim that 'task-inference errors are largely harmless' and that accuracy is 'decoupled from task-inference reliability' is contradicted by the paper's own Prompt-ID versus Gradient-ID results. For example, Table 1 and Table 8 show 92.35 with Prompt ID versus 98.05 with Gradient ID on 10-Split-CIFAR100, and 76.62 versus 83.70 on 10-Split-ImageNet-R. A 5-7 point drop is a substantial dependence on task-inference quality, not a demonstration of harmlessness.","section":"Abstract and Tables 1, 3, 8"}],"minor_comments":[{"comment":"The statement 'if for all t we have that ||vt <= rho||' contains a typo; it should read ||v_t|| <= rho.","section":"Appendix A.1, Theorem 2"},{"comment":"The 'Upper-Bound of Soft-TF' row has missing entries for the pretrained-dataset and task-ID columns, and the same value 93.90 is repeated for both 10-Split-CIFAR100 and 5-Split-CUB200 in a way that is not explained.","section":"Table 3"},{"comment":"Lines 14-16 of Algorithm 1 are incomplete: 'Calculate the prompted feature by' is followed by an empty line, and the per-sample loss line has no expression.","section":"Algorithm 1"},{"comment":"There are numerous typographical issues, including 'Lipshitz' for 'Lipschitz', 'exsistances' for 'existence', and inconsistent spacing in 'T ransF ormers' and 'W ell-initialized'; these should be corrected.","section":"Throughout"}],"recommendation":"reject","confidential_remarks":"The paper's headline claims are tied to an invalid proof and are additionally contradicted by its own numerical tables. Even if the empirical results are reproducible, the manuscript would need a fundamental reframing to separate the empirical study from the theoretical claims, and the 'upper-bound' and decoupling statements would need to be withdrawn or replaced. I do not see a path to a sound version within a normal revision cycle."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing you should know: the empirical idea is the real payload. Masking the QKV/O projections of a frozen pretrained ViT with per-task real-valued masks, initialized at one, and coupling that with prompts and a gradient-based task-ID selector, gives 97-99% on 10/20-split CIFAR-100 with near-zero forgetting, beating the prompt/adapter/LoRA baselines they compare against. The method is simple, plug-and-play, and the code is linked. That part is a meaningful contribution to the continual learning literature.\n\nNow the problems. The theory in Appendix A.1 does not work. Lemma 1's completed-square identity is algebraically wrong: after the update w(t+1)=w(t)-ηv_t, the term should involve ||w(t+1)-w*||^2, but they write ||w(t+1)-ηv_t||^2, which is not the same. Then Eq. (16) claims the telescoping sum collapses to ||w(1)-w*||^2 = ||w(T+1)-w*||^2, which is exactly what the lemma needs to prove. So the bounded-drift and 'harmless task-inference' conclusions are not established. The B_m<B_p<B ordering is also asserted, not derived.\n\nThe data themselves also contradict the harmlessness claim. With Prompt ID, accuracy is 5-8 points lower than with Gradient ID on CIFAR, so task-inference errors are far from harmless. And the 'Upper-Bound of Soft-TF' is 93.90, while the reported Gradient-ID accuracy is 98.05—an upper bound below the observed result is impossible. That's a red flag, maybe a labeling mistake, but it needs an explanation. Also, all results are single-run, and the abstract promises HiDe-Prompt and NoRGa couplings that don't appear in the experiments.\n\nThis is a paper with a strong empirical nugget wrapped in an unsound theoretical story. The method might be genuinely useful, and the experiments are extensive. But as written, the central claims are not supported.\n\nMy recommendation: don't desk-reject it. Send it to peer review, but expect heavy revision—either fix the proof properly or drop it, reconcile the upper-bound arithmetic, and add error bars. A serious referee could help the authors turn this into a solid empirical paper. If they don't fix the numbers, it stays weak.","headline":"Real method, broken theory: the mask-based continual learner is worth a look, but the paper's proof and its own numbers don't back the strong claims.","tokens_in":28642,"tokens_out":4497,"would_cite":false,"duration_ms":40553,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T05","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"Soft-TF claims that a frozen pre-trained transformer can keep learning new classes through per-task real-valued masks, with a proof that mask fine-tuning stays near the pre-trained solution.","keywords":["continual learning","class incremental learning","soft masks","frozen pre-trained transformer","catastrophic forgetting","parameter-efficient fine-tuning","well-initialized lottery ticket hypothesis","prompt-based continual learning"],"falsifier":"The decisive check is to measure accuracy on task t when the selector is forced to pick a different task's mask and compare with the correct mask and with the unmasked frozen backbone; the paper's claim predicts little drop, so a large drop refutes it. Also, on any nontrivial instance of the update $w(t+1) = w(t) - \\eta v_t$, checking whether $||w(1)-w^*||^2$ equals $||w(T+1)-w^*||^2$ would show whether the lemma's key equality is a theorem.","tokens_in":27392,"feed_emoji":"🎭","tokens_out":12243,"duration_ms":91141,"temperature":0.7,"pith_summary":"Soft-TransFormers (Soft-TF) claims that a frozen pre-trained transformer can learn a stream of new tasks by attaching one real-valued multiplicative mask per task to the query, key, value, and output projections of selected self-attention layers, while the underlying weights stay untouched. Because every mask is initialized at one, training begins exactly at the pre-trained solution, and the paper proves that under convex-Lipschitz assumptions, mask-space gradient descent has convergence rate and parameter drift controlled by the distance from the pre-trained weights to the task-optimal configuration. Two properties are claimed to follow: masks are never overwritten, so forgetting is structurally eliminated; and every masked model stays near the shared backbone, so selecting the wrong task's mask still evaluates a near-generalist function and task-inference errors are largely harmless. The paper reports leading accuracy with very low forgetting on vision and language class-incremental benchmarks at trainable-parameter budgets comparable to prompt, adapter, and LoRA methods, positioning the frozen-backbone-plus-soft-mask recipe as a replacement for those approaches.","feed_headline":"Per-task soft masks give near-zero forgetting on frozen transformers","feed_subtitle":"Each mask starts at one, so a wrong task guess still works almost as well, and old tasks are never touched.","key_machinery":"The load-bearing mechanism is the soft mask: a real-valued multiplier $m$, initialized to exactly 1 on the selected projection weights $w_Q$, $w_K$, $w_V$, and $w_O$, so the effective weight is $w \\odot m$ and optimization starts at the pre-trained point. Alongside the shared G-prompt and task-specific E-prompt, each task owns a copy of these masks, and task selection picks the right mask at inference by task-key matching or by a one-shot gradient/entropy criterion. The analysis centers on the attention expression $m_Q \\cdot (\\text{prompt and patch outer products}) \\cdot m_K^T$, reduced to convex-Lipschitz online learning, and the proof's conclusion is the ordering of drift bounds $B_m < B_p < B$ for masked, prompted, and unmodified configurations. This ordering is what converts the masks' initialization-at-one into near-zero forgetting and wrong-mask harmlessness: the architecture is built so that every decision the model makes happens through small multiplicative perturbations of a shared well-initialized backbone.","core_discovery":"The central claim is that mask-only fine-tuning of a well-pre-trained transformer is not just another parameter-efficient trick: it is a way to keep the model's trajectory inside a small neighborhood of the pre-trained solution. For a frozen weight $w$ and a learnable real-valued mask $m$, the effective parameter is $w \\odot m$, and the analysis in the appendix treats the attention bilinear form built from $m_Q z_t z_t^T m_K$ under standard convex-Lipschitz assumptions, deriving an upper bound on the average regret that is proportional to $B_m \\rho / \\sqrt{T}$, where $B_m$ is the distance from the masked configuration to the task optimum. The bound ranks mask drift below prompt drift ($B_m < B_p < B$), so the masked network stays closest to the shared pre-trained weights. That bounded drift is the load-bearing reason for both advertised properties: no parameter is ever overwritten, so forgetting is structurally impossible, and each task subnetwork is so close to the generalist backbone that even a wrong task mask yields a near-generalist function, decoupling accuracy from task-inference reliability. On the reported CIFAR-100 splits this yields high-90s accuracy with forgetting below one point for the gradient-based task selector, and on the BERT language benchmark it averages 79.4 across four task orders.","pith_inferences":["The paper leaves implicit that wrong-mask harmlessness can be tested directly: random or deliberately wrong task assignment should degrade accuracy only mildly, and that experiment would separate the near-generalist claim from the quality of the task selector.","The distance-to-optimum bound suggests that pre-training progress, not the continual-learning update, dominates the reported gains; comparing the same architecture at successive pre-training checkpoints would show how quickly the mask advantage grows as initialization improves.","Because each task is an independent multiplier on the same backbone, averaging or interpolating masks across tasks is a natural next step for forward transfer, letting new tasks start from a blend of old masks rather than from the identity mask."],"forward_implications":["Forgetting becomes a structural guarantee rather than a managed risk: the frozen backbone and each task's mask are never overwritten, so no replay buffer or experience memory is needed to protect old tasks.","Task-inference errors stop being catastrophic: a wrong mask still leaves the model close to the pre-trained generalist, so class-incremental accuracy should remain high even when task identity is misidentified.","The method is a plug-in for existing prompt frameworks: soft masks can be trained alongside L2P, DualPrompt, PGP, and, per the abstract, HiDe-Prompt and NoRGa, with no change to inference cost beyond the unmodified backbone.","Initialization quality becomes a first-order factor: the proved bound ties convergence and drift to the distance between the pre-trained weights and the task-optimal solution, so better pre-training should translate directly into better continual learning.","At matched parameter budgets, mask fine-tuning outperforms the reported Adapter, LoRA, and prompt-only baselines on the tested vision and language CIL benchmarks, reaching high-90s accuracy with near-zero forgetting on CIFAR-100 splits."],"supporting_citations":[{"why":"Supplies the prompt-pool and task-key matching scheme (L2P) that Soft-TF builds on for Prompt ID and as a plug-in host.","marker":"Wang et al. 2022c"},{"why":"Supplies the G/E prompt framework and the DualPrompt baseline that Soft-TF extends and compares against.","marker":"Wang et al. 2022b"},{"why":"Supplies the PGP baseline and the experimental protocol (epochs, batch sizes, prompt lengths) Soft-TF adopts.","marker":"Qiao et al. 2024"},{"why":"Defines the Vision Transformer backbone whose frozen weights Soft-TF masks.","marker":"Dosovitskiy et al. 2020"},{"why":"Source of the convex-Lipschitz convergence analysis used in Appendix A.1 for the bounded-drift proof.","marker":"Shalev-Shwartz and Ben-David 2014"},{"why":"Provides the winning-subnetwork (WSN) continual-learning baseline that Soft-TF outperforms and serves as a comparison upper bound.","marker":"Kang et al. 2022"},{"why":"Introduces mask/supermask-based continual learning that soft real-valued masks extend to frozen transformers.","marker":"Wortsman et al. 2020"},{"why":"Defines the Adapter parameter-efficient fine-tuning baseline compared at matched trainable-parameter budgets.","marker":"Houlsby et al. 2019"},{"why":"Defines the LoRA low-rank fine-tuning baseline compared at matched trainable-parameter budgets.","marker":"Hu et al. 2021"}],"fun_headline_variants":["Soft masks make frozen transformers nearly forget-free","Mask-only fine-tuning keeps transformers from forgetting","Soft-TF: near-zero forgetting with frozen backbones","Task-specific soft masks stop transformer forgetting"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof's load-bearing premise is that mask-space updates satisfy $||w(1)-w^*||^2 = ||w(T+1)-w^*||^2$ and that the mask drift bound sits strictly below the prompt drift bound; the equality does not follow from the update rule, so the bounded-drift conclusion is being assumed rather than derived.","fun_headline_variants_meta":{"raw":{"variants":["Soft masks make frozen transformers nearly forget-free","Mask-only fine-tuning keeps transformers from forgetting","Soft-TF: near-zero forgetting with frozen backbones","Task-specific soft masks stop transformer forgetting"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000305,"raw_usage":{"total_tokens":1838,"prompt_tokens":1122,"completion_tokens":716,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":738,"completion_tokens_details":{"reasoning_tokens":659}},"tokens_in":738,"tokens_out":716,"duration_ms":7256,"temperature":1.0,"reasoning_tokens":659,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T13:34:46.572013+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"The decisive check is to measure accuracy on task t when the selector is forced to pick a different task's mask and compare with the correct mask and with the unmasked frozen backbone; the paper's claim predicts little drop, so a large drop refutes it. Also, on any nontrivial instance of the update $w(t+1) = w(t) - \\eta v_t$, checking whether $||w(1)-w^*||^2$ equals $||w(T+1)-w^*||^2$ would show whether the lemma's key equality is a theorem.","supporting_citations":[],"review_version":2}