{"id":"30a85d71-a205-46a2-8163-99f6ee1b35d5","arxiv_id":"2504.13822","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A survey and comparison of Parameter-Efficient Continual Fine-Tuning methods for adapting large pre-trained models to sequential tasks without catastrophic forgetting.","lead":"This paper surveys methods that combine continual learning with parameter-efficient fine-tuning, such as LoRA, adapters, and prompts, so large AI models can learn new tasks without forgetting old ones. It also reports a comparison of ten methods on three image benchmarks, though the comparison lacks the experimental details needed to trust the ordering.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Tables 4 and 5 lack any experimental protocol, so the accuracy and forgetting rankings that support the paper's PECFT performance claims are unverified; Table 3 unit errors reinforce this concern.","rationale":"The reader's weakest assumption identifies exactly the same load-bearing concern: the Table 4/5 comparison assumes a consistent, described protocol that the paper never provides. My reading agrees that this is the most consequential flaw, because the survey uses these tables to rank methods and derive design principles in Section 5.7, not merely to illustrate the literature. The Table 3 unit error is a separate but related red flag that the quantitative sections were not carefully checked. However, the concern does not require rejection: the paper's qualitative synthesis, taxonomy, and references do support the existence and utility of the PECFT area, and the claimed trade-offs could be restored by adding a full protocol, fixing units, and relabeling any copied results as 'as reported in original papers.' A conditional acceptance with mandatory corrections is therefore appropriate, and the verdict stays unchanged from the reader's assessment.","tokens_in":35573,"tokens_out":3545,"duration_ms":35766,"concrete_test":"Request from the authors the complete experimental configuration and code for the Section 5.6 comparison, then rerun the CIFAR-100 10-task rows of Table 4 for APER (adapter) and L2P under that protocol with a shared backbone and fixed hyperparameters. If the reported values (86.31 and 80.01) or the relative ordering of the ten methods cannot be reproduced, the accuracy and forgetting trade-offs that support the survey's recommendations are not established; additionally, recompute Table 3 percentages after correcting the model-size units to confirm the parameter-efficiency claims.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 5.6 presents the paper's main quantitative contribution: a comparative evaluation of ten PECFT methods in Table 4 (average accuracy) and Table 5 (average forgetting), followed by concrete recommendations such as 'adapter methods for peak performance' and 'LoRA methods for consistent multi-domain deployment.' These conclusions directly support the central claim that PECFT is a mature paradigm whose trade-offs are known. However, Section 5.6 specifies no experimental protocol: no backbone model, no task-splitting rule, no hyperparameters (LoRA rank, prompt length, learning rate, epochs), no seeds, no memory/rehearsal settings, and no statement of whether the numbers were produced by the authors under a unified setup or copied from the original papers. The section even acknowledges that 'individual methods often rely on different experimental settings,' yet the table treats all rows as comparable. This is not a cosmetic omission: the selection of backbone and protocol materially changes accuracy and forgetting, so the rankings and the Section 5.7 design principles drawn from them are not established. The concern is compounded by Table 3, which labels BERT Base as 109B and ViT B16 as 85B, implying units off by roughly 1000x; the reported percentages only make sense if the intended values are in millions. The central qualitative thesis that a PECFT area exists is well supported by the cited literature, but the quantitative evidence intended to characterize state-of-the-art trade-offs is not.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This survey defines and reviews the emerging area of Parameter-Efficient Continual Fine-Tuning (PECFT), which combines Continual Learning (CL) with Parameter-Efficient Fine-Tuning (PEFT) to adapt large pre-trained models to sequential tasks while limiting forgetting and computational cost. The paper first presents background on CL scenarios and classical CL algorithms, then reviews PEFT methods (adapters, prompts, reparameterization, selective methods), and then proposes a taxonomy of PECFT methods organized into adapter-based, LoRA-based, prompt-based, and unified frameworks. It defines evaluation metrics (average accuracy, average forgetting, forward/backward transfer, model size efficiency), reports a comparative performance evaluation in Tables 4 and 5, distills design principles (parameter separation, interference prevention), and discusses future directions including multi-modality, model merging, reasoning tasks, and realistic settings.","tokens_in":35857,"tokens_out":2526,"duration_ms":24893,"significance":"If the survey's central claims are accepted, it provides a useful systematization of a rapidly growing intersection between CL and PEFT, with a clear taxonomy (Figure 2) and a structured catalog of representative methods. The qualitative coverage is broad and largely recognizable to a reader familiar with the literature, and the synthesis of design principles in Section 5.7 is a genuine contribution that goes beyond a simple enumeration of methods. However, the paper's quantitative contribution is load-bearing for its claim to characterize the state of the art and to derive practical recommendations; those numbers are currently not supported by any described experimental protocol, and Table 3 contains unit and consistency errors. The paper also includes self-acknowledged limitations in Section 5.6 that are not resolved in the presentation. The qualitative survey value is real, but the quantitative evidence must be repaired or explicitly downgraded before the paper can be accepted.","major_comments":[{"comment":"The comparative evaluation that drives the recommendations in Section 5.6 and the design principles in Section 5.7 is not reproducible or verifiable as reported. No experimental protocol is given: there is no specification of the backbone model, the task-splitting rule, the optimizer, learning rate, epochs, LoRA rank, prompt length, batch size, number of seeds, or whether rehearsal/memory buffers were used. The section itself concedes that 'individual methods often rely on different experimental settings,' yet the rows of Tables 4 and 5 are treated as directly comparable. Because accuracy and forgetting values depend strongly on these choices, the rankings and the conclusions drawn from them (e.g., 'adapter methods for peak performance,' 'LoRA methods for consistent multi-domain deployment') are not established. Please either provide the full protocol and state whether the numbers were produced by the authors under a unified setup, or reposition Tables 4 and 5 as a literature-reported comparison with clear per-source attribution and explicit caveats that no cross-method ranking should be inferred.","section":"Section 5.6, Tables 4 and 5"},{"comment":"Table 3 contains unit errors and internal inconsistencies that undermine its quantitative content. BERT Base Uncased is labeled (109B), ALBERT Base V2 (11B), RoBERTa Base (124B), ViT B16 (85B), and CLIP ViT B16 (149B); these are roughly three orders of magnitude too large and should be in millions. The reported percentages are also not internally consistent with the listed total and trainable parameter counts: for example, ViT B16 lists QLoRA trainable parameters as 43,906 with 0.51%, while LoRA lists 442,368 with 0.52% on the same model, and CLIP ViT B16 lists QLoRA as citation [88] rather than [22] or [23]. Please correct the units, recompute all percentages from a single consistent source, and verify the citations in each row.","section":"Table 3"},{"comment":"The forgetting analysis in Section 5.6.2 makes strong comparative claims on the basis of a single scenario (CIFAR-100, 10 tasks) with no protocol or error-bar derivation described. Statements such as 'L2P suffers from severe forgetting (10.61%), nearly doubling the forgetting of any other competing method' and the conclusion that 'subspace-constrained approaches like InfLoRA offer the best stability-performance balance' depend on the comparability of the table entries. Please clarify how the standard deviations were computed, how many runs were performed, and whether the same backbone, optimizer, and task order were used for all ten methods; otherwise these claims should be substantially softened.","section":"Section 5.6.2, Table 5"},{"comment":"References [22] and [23] are the same QLoRA paper (Dettmers et al.), which creates confusion in the text and in Table 3, where some rows cite [22] and one row cites [88] for QLoRA. Please deduplicate the citation and ensure every QLoRA row points to the single correct reference.","section":"References [22] and [23]"}],"minor_comments":[{"comment":"Equation (1) uses the summation index t for both the task index and the inner sum over samples, which is notationally confusing; please replace the inner index with i or n.","section":"Section 2.1, Eq. (1)"},{"comment":"The definition of Model Size Efficiency (MS) is unclear: the formula 'min(1, (sum_i Mem(theta1)/Mem(theta_i))/N)' appears to average ratios but the min operator and the placement of N are ambiguous, and no reference is given for this exact formulation. Please clarify the definition or cite a source.","section":"Section 5.5, Eq. (18)"},{"comment":"The text describes QLoRA with reference [22] while the reference list contains an identical duplicate [23]; after deduplication, please also check the in-text citation for QLoRA in Table 3, which currently cites [88] in the CLIP block.","section":"Section 4.2"},{"comment":"The description of Mix-and-Match adapters states that they 'match full fine-tuning performance on some benchmarks while fine-tuning only about 6.7% of parameters,' but the cited source [43] is listed as a parallel conference version of [42]; please reconcile the two citations and clarify which paper is being described.","section":"Section 4.1.1"},{"comment":"Several sentences contain grammatical errors or missing words, e.g., 'On one hand, CL enables models to continually adapt whiles PEFT ensures that efficiency whiles the model adapts' in Section 1 and 'The organization of this paper as follows' in the same section. A thorough language edit is needed.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper is best viewed as a qualitative survey with an under-supported quantitative appendix. The central qualitative thesis is defensible and the taxonomy is useful, but the comparative tables and Table 3 need substantial repair. I would ask the authors to either provide a full experimental protocol and correct the tables, or explicitly reframe the numerical sections as literature-reported values with clear caveats. The duplicate QLoRA reference and the self-admitted lack of a unified protocol should also be addressed in the revision. The paper is likely within the scope of the journal once these issues are resolved."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere is my read of arXiv:2504.13822. The genuinely useful part is the synthesis: the paper organizes PECFT into adapter-based, LoRA-based, prompt-based, and unified-framework families, and the qualitative descriptions of the main methods are recognizable and mostly accurate. The background on CL scenarios and PEFT is standard but solid. Section 5.7 on design principles is a reasonable distillation of how frozen backbones, orthogonal subspaces, modular routing, and compositional prompts trade off. For a practitioner wanting a map of this emerging subfield, this survey has real value.\n\nThe soft spot is exactly where the stress-test note lands: Tables 4 and 5 are presented as a comparative evaluation but no protocol is given. No backbone, task split, hyperparameters (LoRA rank, prompt length, epochs, learning rate), seeds, or statement about whether the numbers come from unified runs or from the original papers. The text even acknowledges that methods rely on different experimental settings, then treats the rows as comparable and draws recommendations like \"adapter methods for peak performance.\" That is not a cosmetic omission; with different backbones and protocols, both accuracy and forgetting rankings could shift. The Section 5.7 principles are partly built on those rankings, so they inherit the problem.\n\nThe quantitative troubles extend to Table 3, which labels BERT Base as 109B and ViT B16 as 85B—clearly millions, not billions. There are duplicated QLoRA references ([22] and [23] are the same paper), a suspicious 43,906 trainable-parameter entry for QLoRA on ViT B16 next to 442,368 for LoRA, and minor issues like the ambiguous VPT acronym and treating Rainbow Memory as a scenario when it is a method.\n\nThe good news: the central qualitative claim—that PECFT is a distinct and growing paradigm—is well supported by the cited literature. The taxonomy is genuinely new as an aggregation. None of the flaws make the survey incoherent; they make the empirical section untrustworthy as it stands.\n\nMy recommendation: send it to peer review, but expect major revision. The authors should either remove the comparative tables and the ranking-based recommendations, or redo the evaluation under a clearly described protocol with code and data. The rest of the survey deserves referee time.","headline":"A useful qualitative map of PECFT that is undermined by a comparative evaluation with no experimental protocol and a parameter table with clear unit errors; worth revising, not rejecting.","tokens_in":36341,"tokens_out":3344,"would_cite":false,"duration_ms":31375,"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":"Large pre-trained models can keep adapting to new tasks with tiny parameter updates while retaining old knowledge.","keywords":["continual learning","parameter-efficient fine-tuning","catastrophic forgetting","low-rank adaptation","prompt tuning","adapters","pre-trained models","survey"],"falsifier":"Running the ten methods compared in Tables 4 and 5 on CIFAR-100 with 10 tasks under a single, fixed protocol (identical epochs, learning rate, LoRA rank, prompt length, and backbone) and showing that the accuracy ranking or the forgetting ranking changes materially would overturn the survey's comparative conclusions. Likewise, correcting Table 3 so that all parameter counts are in the same units would settle whether the efficiency comparison is meaningful.","tokens_in":35370,"feed_emoji":"🧠","tokens_out":6840,"duration_ms":58379,"temperature":0.7,"pith_summary":"This survey tries to establish that parameter-efficient continual fine-tuning (PECFT) is a coherent and useful paradigm: large pre-trained models can be adapted to an endless stream of tasks by updating only a small set of parameters, without sacrificing old knowledge. It argues that continual learning supplies the sequential-learning machinery while parameter-efficient fine-tuning supplies the cheap adaptation, and that their combination is what makes lifelong deployment of large models practical. The survey organizes existing methods into adapter-based, LoRA-based, prompt-based, and unified frameworks, and compares them on accuracy and forgetting across three benchmarks. A sympathetic reader would care because the claim, if right, means a single large model can keep evolving in the real world without prohibitive retraining costs.","feed_headline":"Large models can keep learning new tasks with tiny parameter updates","feed_subtitle":"Survey maps adapter, LoRA, and prompt methods that fight forgetting while adapting to task streams.","key_machinery":"The organizing machinery is a taxonomy of PECFT methods plus the design principles that separate them: parameter separation (modular separation, subspace separation, and input space separation) and interference prevention (orthogonality constraints, architectural modularity, and compositional isolation). The taxonomy takes each PEFT family—adapters, LoRA, prompts, and unified frameworks—and shows how continual-learning mechanisms such as routing, merging, masking, prompt pools, and orthogonal projection attach to it. The comparative tables use Average Accuracy and Average Forgetting as the load-bearing metrics that make the paradigm's trade-offs visible.","core_discovery":"The paper's central claim is that PECFT methods enable large pre-trained models to continuously evolve and adapt to new tasks while preserving past knowledge and maintaining computational efficiency. Concretely, the survey reports that adapter-based methods achieve the highest average accuracy, LoRA-based methods offer the most consistent cross-dataset performance, prompt-based methods are the lightest but lag in accuracy, and subspace-constrained methods such as InfLoRA yield the lowest forgetting (4.61% on CIFAR-100 with 10 tasks). It also finds that high accuracy and low forgetting are decoupled: no method tops both tables, and the choice of parameter-separation and interference-prevention strategy matters more than the choice of PEFT module alone.","pith_inferences":["If the reported trade-offs persist under a truly unified protocol, a natural next step is hybrid designs that combine orthogonal-subspace updates (low forgetting) with adapter expansion (high accuracy), a combination no single method in the survey currently offers.","The survey's future-direction section points toward model merging; a testable extension is to treat each task's LoRA or adapter weights as a task vector and merge them sequentially with interference-reducing operators, which would make PECFT methods memory-constant rather than growing with tasks.","Because the comparison only covers image classification, the claimed rankings may not transfer to language or reasoning tasks; a natural extension is to rerun the same methods on sequential question answering or instruction-tuning streams, where representation drift behaves differently.","Before relying on the efficiency comparison, a reader should re-derive the parameter counts in Table 3 from the original papers, since entries like 'BERT Base Uncased (109B)' appear to use inconsistent units."],"forward_implications":["Practitioners who need peak accuracy on a specific domain should prefer adapter-based PECFT methods, which top the averaged accuracy comparison at the cost of more trainable parameters.","For multi-domain deployment, LoRA-based methods such as LoRAC-IPC and SD-LoRA give the most consistent accuracy across CIFAR-100, CUB-200, and Tiny ImageNet, with moderate, predictable degradation when the task count doubles.","Stability-critical applications can use subspace-constrained methods like InfLoRA, whose 4.61% average forgetting on CIFAR-100 is the best in the comparison.","Prompt-based methods, despite being the most parameter-light, underperform by 3-6% and scale poorly, so they are a reasonable choice only when memory is the dominant constraint.","Unified frameworks such as LAE and APER decouple the continual-learning strategy from the underlying PEFT module, so any adapter, LoRA, or prompt method can be turned into a continual learner; LAE loses only 1.32% accuracy when moving from 10 to 20 CIFAR-100 tasks."],"supporting_citations":[{"why":"Supplies the bottleneck adapter module that adapter-based PECFT methods extend with routing, fusion, and expansion.","marker":"[50]"},{"why":"Supplies the low-rank decomposition that defines LoRA-based PECFT and its orthogonal-subspace variants.","marker":"[52]"},{"why":"Introduces trainable prompt tokens that prompt-based PECFT methods turn into prompt pools and compositional prompts.","marker":"[77]"},{"why":"Early demonstration that a prompt pool on a frozen pre-trained network can handle class-, domain-, and task-agnostic continual learning.","marker":"[153]"},{"why":"Extends prompt-pool prompting with general and expert prompts, anchoring the prompt-based branch of the taxonomy.","marker":"[152]"},{"why":"Introduces decomposed attention-based prompts and is a key comparison point in the accuracy and forgetting tables.","marker":"[133]"},{"why":"Provides the subspace-constrained LoRA method whose 4.61% forgetting is the best stability result in Table 5.","marker":"[82]"},{"why":"Supplies the self-expanding mixture-of-adapters method that posts the highest CIFAR-100 accuracy among adapter methods.","marker":"[146]"},{"why":"Frames adapt-once-and-merge as a unified PECFT strategy and achieves the top CUB-200 accuracy in Table 4.","marker":"[171]"},{"why":"Unified framework that converts arbitrary PEFT modules into continual learners, underpinning the unified-framework category.","marker":"[36]"}],"fun_headline_variants":["Tiny tweaks, lifelong learning: PEFT survey","Forgetting less with small updates: PECFT","Adapters, LoRA, prompts: which best for continual AI?","Keep AI learning forever with tiny parameter changes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparative rankings in Tables 4 and 5 assume that every method was evaluated under a consistent experimental protocol with the same backbone, hyperparameters, and evaluation conditions; if the numbers came from heterogeneous original setups, the reported ordering is not reliable.","fun_headline_variants_meta":{"raw":{"variants":["Tiny tweaks, lifelong learning: PEFT survey","Forgetting less with small updates: PECFT","Adapters, LoRA, prompts: which best for continual AI?","Keep AI learning forever with tiny parameter changes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000332,"raw_usage":{"total_tokens":1838,"prompt_tokens":930,"completion_tokens":908,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":546,"completion_tokens_details":{"reasoning_tokens":842}},"tokens_in":546,"tokens_out":908,"duration_ms":9031,"temperature":1.0,"reasoning_tokens":842,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:58:31.770489+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Running the ten methods compared in Tables 4 and 5 on CIFAR-100 with 10 tasks under a single, fixed protocol (identical epochs, learning rate, LoRA rank, prompt length, and backbone) and showing that the accuracy ranking or the forgetting ranking changes materially would overturn the survey's comparative conclusions. Likewise, correcting Table 3 so that all parameter counts are in the same units would settle whether the efficiency comparison is meaningful.","supporting_citations":[{"cited_title":"Learn- ing to prompt for continual learning","cited_arxiv_id":null,"evidence_quote":"Early demonstration that a prompt pool on a frozen pre-trained network can handle class-, domain-, and task-agnostic continual learning."},{"cited_title":"Dualprompt: Complementary prompting for rehearsal-free contin- ual learning","cited_arxiv_id":null,"evidence_quote":"Extends prompt-pool prompting with general and expert prompts, anchoring the prompt-based branch of the taxonomy."},{"cited_title":"Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning","cited_arxiv_id":null,"evidence_quote":"Introduces decomposed attention-based prompts and is a key comparison point in the accuracy and forgetting tables."},{"cited_title":"Self-expansion of pre-trained models with mixture of adapters for continual learning, 2024","cited_arxiv_id":null,"evidence_quote":"Supplies the self-expanding mixture-of-adapters method that posts the highest CIFAR-100 accuracy among adapter methods."},{"cited_title":"Revisiting class-incremental learning with pre-trained models: Gener- alizability and adaptivity are all you need","cited_arxiv_id":null,"evidence_quote":"Frames adapt-once-and-merge as a unified PECFT strategy and achieves the top CUB-200 accuracy in Table 4."}],"review_version":1}