{"id":"0e47926b-90f0-4083-a56d-d7117456c410","arxiv_id":"2608.10804","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"BPG sizes per-domain adapters inversely to a feature-separability score and replaces hard domain selection with confidence-weighted logit fusion, setting state-of-the-art accuracy and near-zero forgetting on three domain incremental learning benchmarks.","lead":"This paper introduces BPG, a method for domain incremental learning that sizes per-domain adapters according to feature separability and softly mixes domain experts at test time. It reports consistent accuracy gains and near-zero forgetting on DomainNet, CDDB, and CORe50, offering a practical recipe for continual learning under shifting visual domains.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The capacity-allocation rule rests on an unmeasured monotone link between feature separability s_t and optimal adapter size; the derivation assumes κ_t ∝ s_t without evidence, and the empirical support is post-hoc and imperfect (Sketch vs Quickdraw).","rationale":"The reader's weakest assumption identifies the same load-bearing concern: the monotone separability-to-capacity relationship is asserted rather than established. My reading of Section III-E confirms that the 'derivation' hinges on an unmeasured κ_t, and the empirical evidence in Section IV-F is post-hoc and contains counterexamples to strict monotonicity (Sketch vs Quickdraw). This is not a fatal flaw, because the paper's empirical results could still hold even if the rule is suboptimal, but it directly undermines the claimed mechanism. The proposed test—inverting the allocation rule under a matched parameter budget—would settle whether the direction of the rule matters. If the inverted rule performs similarly, then the adaptive capacity allocation is not the source of BPG's gains, and the central contribution is weakened. The verdict remains CONDITIONAL because the concern is addressable with additional experiments and does not by itself invalidate the reported results.","tokens_in":24400,"tokens_out":9844,"duration_ms":103296,"concrete_test":"On DomainNet (ViT-B/16), replace the capacity rule with its inverse, r_t = (s_t / s_0) r_0, using the same total parameter budget as BPG-Adapter (Σ r_t ≈ 1495). Compare A_T and F_T against the 68.33/1.28 reported for BPG-Adapter in Table VII. If the inverted rule achieves accuracy within 1 point or lower forgetting, the direction of the separability-to-capacity mapping is not the source of the gain. As a secondary check, compute the rank correlation between the per-domain empirically optimal r from Fig. 4 and s_t from Table I; if it is not significantly negative, the monotone assumption is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central mechanism BPG-Adapter sets r_t = (s_0/s_t) r_0 (Eq. 14). Its justification in Section III-E derives r* ∝ 1/κ_t from a toy risk model, then asserts κ_t ∝ s_t without any measurement of κ_t. The only empirical support is Section IV-F, but that is the same data that motivated the rule, and the correlation is not clean: Sketch (s=0.156) peaks at r=256 while Quickdraw (s=0.151) peaks at r=1024, despite nearly identical separability. Moreover, the derivation solves a budget-constrained optimization, but Eq. (14) does not enforce any total budget and can produce r_t exceeding the embedding dimension d for low-s domains; the per-dataset hand-setting of r_0 (64 for DomainNet/CORe50, 1 for CDDB) is the only safeguard. If the monotone mapping fails on a new domain, BPG-Adapter misallocates capacity. This is load-bearing because the paper claims the capacity rule is the mechanism behind the state-of-the-art results, not merely a neutral heuristic.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes BPG, a parameter-isolation framework for domain incremental learning. BPG-Adapter sets each domain's adapter hidden dimension according to an inverse-proportionality rule based on a frozen-backbone feature separability score, while BPG-Inference replaces hard domain selection with a confidence-weighted soft mixture over domain-specific experts using k-means prototypes. Experiments on DomainNet, CDDB, and CORe50 with ViT-B/16 and CLIP backbones report state-of-the-art average accuracy and very low forgetting on several benchmarks, including A_T=72.19 and F_T=0.22 on DomainNet with ViT. The paper also shows that BPG-Inference can be applied as a plug-and-play module to existing prompt-based methods, and provides ablations, parameter-matched comparisons, and efficiency analyses.","tokens_in":24651,"tokens_out":16157,"duration_ms":137921,"significance":"If the results hold, BPG offers a simple and effective recipe for parameter-isolation DIL, and the soft-mixture inference module is a useful and broadly applicable contribution. The paper's experimental core is solid: ablations isolate the two components, the per-domain matrix in Table X directly illustrates the forgetting dynamics, and Section IV-G demonstrates gains over uniform scaling at matched parameter budgets. The paper also ships unusually detailed efficiency analyses (training time, latency, parameter counts) and makes the adaptive-capacity idea concrete via a reproducible separability score. The main weaknesses are the theoretical justification of the capacity rule, which relies on an unverified monotonicity assumption, and an apparent inconsistency between the printed per-domain matrix and the reported forgetting value.","major_comments":[{"comment":"The derivation of the capacity rule from Eq. (23) to Eq. (24) is not justified. The conclusion r* ∝ 1/κ_t ignores that A_t, N_t, and the logarithmic term in Eq. (23) all vary across domains, so the stated assumptions do not imply monotonicity of r* in κ_t. More importantly, the assumed monotone link κ_t ∝ s_t is load-bearing but only supported post-hoc by Fig. 4, which actually contains a counterexample: Sketch (s=0.156) peaks at r=256 while Quickdraw (s=0.151) peaks at r=1024, despite nearly identical separability scores. I recommend measuring κ_t directly (e.g., fitting per-domain risk curves) or demonstrating the monotone relationship on additional datasets before claiming that Eq. (14) approximates an optimal allocation.","section":"Section III-E and Section IV-F"},{"comment":"The per-domain accuracy matrix reported in Table X does not reproduce the stated F_T values under the metric defined in Eq. (25). Recomputing from the printed entries for 'BPG-Adapter only' gives approximately F_T=1.59, not 1.28; for the full BPG matrix the recomputed value is about 0.24, close to but not equal to the reported 0.22. Please clarify the exact computation (e.g., any rounding, seed averaging, or a different forgetting definition) and provide matrix entries with enough precision so that the headline forgetting numbers are reproducible.","section":"Table X"},{"comment":"The Lagrangian derivation in Eq. (21) assumes a fixed total budget R_tot, but Eq. (14) does not enforce any budget; the total capacity is an emergent function of s_0, r_0, and the sequence of s_t values. For low-separability domains, the rule can produce r_t exceeding the embedding dimension d, and the only safeguard in the paper is the per-dataset hand-setting of r_0 (64 for DomainNet/CORe50, 1 for CDDB). Please report the sensitivity of the final results to r_0 on CDDB and CORe50, and discuss how r_0 is selected without access to test performance, since this is a free parameter that directly controls the scale of all adapter widths.","section":"Section III-E.b and Section IV-G"}],"minor_comments":[{"comment":"The text states that BPG-Adapter is evaluated with r_0 = 4, 16, and 64, but Fig. 3 appears to show only one BPG-Adapter curve; please clarify whether the curve is a representative run, an average, or a different summary of the three settings.","section":"Section IV-D.a and Fig. 3"},{"comment":"The pruning step is described as using the uniform distribution over domains as a threshold; please state explicitly that this threshold is applied to the softmax-normalized weights w_t rather than to the raw distances a_t.","section":"Section III-D.b"},{"comment":"The CORe50 row contains a malformed entry ('91.87–92.14–'); please fix the formatting so the table is readable.","section":"Table IX"},{"comment":"The adapter parameter count is given as 2L d r_t + L d + L r_t, but the adapter structure in Section III-B does not define which parameters correspond to the L d and L r_t terms; please specify whether these are biases or other parameters.","section":"Section IV-G.a"},{"comment":"The t-SNE visualizations in Fig. 4 are described in the text, but the figure is not referenced before Section IV-F; please add the reference where the per-domain accuracy numbers are first discussed.","section":"Section IV-F"}],"recommendation":"major_revision","confidential_remarks":"The Table X discrepancy is the most concrete correctness issue and should be verified before acceptance; if the headline F_T values in Tables IV-VII were computed with the same unspecified metric, they may need correction. The capacity-rule justification is the weakest intellectual contribution; the authors should either strengthen it with direct evidence or present Eq. (14) more modestly as a heuristic. The empirical results are otherwise strong and the paper is well within the scope of the journal."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick summary: this is a solid, well-executed empirical paper on rehearsal-free domain incremental learning. The headline numbers are strong — 72.19% average accuracy and 0.22% forgetting on DomainNet with ViT-B/16, consistent gains on CDDB and CORe50 — and the per-domain accuracy matrix in Table X backs up the forgetting claim by showing early-domain accuracy actually improves under soft mixture rather than merely being averaged away. The training-free soft-mixture component (BPG-Inference) is the cleanest contribution: it consistently lifts several existing prompt-based methods (S-iPrompts, PINA, C-Prompt, KA-Prompt) by nontrivial margins, which makes it a genuinely useful plug-in independent of the adapter-allocation story. The paper also reports wall-clock times and parameter budgets, which is more than most DIL papers do.\n\nThe weak link is BPG-Adapter's capacity-allocation rule. The 'derivation' in Section III-E is a toy risk model that ends with an unmeasured assumption that the error-decay rate κ_t is proportional to the separability score s_t. The empirical support in Fig. 4 is the same data that motivated the rule, and it is not clean: Sketch (s=0.156) peaks at r=256 while Quickdraw (s=0.151) peaks at r=1024, despite nearly identical separability. To be fair, for Sketch the accuracy difference between r=256 and r=1024 is 0.02, so the practical damage is small on these benchmarks. The hand-setting of r0 per dataset (64 vs 1) is a real fragility with no sensitivity study on CDDB, and the rule has no built-in budget constraint — a future domain with tiny s_t could produce an absurd r_t. No code or data is provided, which makes exact reproduction and fair comparison hard. These are all addressable rather than fatal.\n\nThe central empirical claim — that soft mixture beats hard selection and adaptive capacity beats uniform capacity at matched parameter budgets — is well supported by the ablations. The exact form of the allocation rule is a heuristic, and the paper would be improved by presenting it as such, with a proper sensitivity study rather than a pseudo-derivation.\n\nWho this is for: anyone working on parameter-isolation DIL, especially prompt and adapter methods. BPG-Inference is worth trying on your own method. The paper deserves a serious referee; the concerns are revision-level, not desk-reject-level.","headline":"Solid empirical paper with a useful plug-in; the capacity-allocation story is a heuristic dressed up as theory, but the experiments largely support the central claim.","tokens_in":25174,"tokens_out":3360,"would_cite":true,"duration_ms":32554,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Domain incremental learning can reach near-zero forgetting by sizing each domain's adapter to its feature separability and softly mixing domain experts at test time, reaching 72.19% accuracy with 0.22% forgetting on DomainNet.","keywords":["domain incremental learning","catastrophic forgetting","parameter-efficient fine-tuning","adapters","feature separability","soft domain mixture","continual learning","capacity allocation"],"falsifier":"On a new multi-domain benchmark, compute each domain's separability score $s_t$ and then grid-search its optimal adapter dimension in isolation; if any high-separability domain peaks only at a large adapter dimension, or any low-separability domain peaks at a small one, the monotone inverse rule $r_t=(s_0/s_t)r_0$ fails.","tokens_in":1985,"feed_emoji":"🧩","tokens_out":2260,"duration_ms":93902,"temperature":0.7,"pith_summary":"The paper tries to show that two persistent weaknesses of parameter-isolation domain-incremental learning can be fixed together: uniform adapter capacity ignores domain difficulty, and hard domain-ID selection at test time is brittle. It claims that giving harder domains larger adapter hidden dimensions and easier domains smaller ones, sized by a simple separability score, improves plasticity without wasting parameters, and combining this with a confidence-weighted soft mixture of domain-specific models at inference nearly eliminates forgetting. If true, rehearsal-free continual learning on pre-trained transformers improves substantially: BPG reports 72.19% average accuracy with 0.22% average forgetting on DomainNet with ViT-B/16, close to the oracle-domain-ID upper bound, with similar gains on CDDB and CORe50. The soft mixture is also plug-and-play, lifting accuracy when added to existing prompt-based methods.","feed_headline":"Adapter size follows domain difficulty; forgetting hits 0.22%","feed_subtitle":"A separability score sizes each adapter, and soft domain mixing replaces brittle domain-ID selection at inference.","key_machinery":"The load-bearing object is the feature-separability score $s_t=R^{\\mathrm{bcs}}_t/(R^{\\mathrm{wcs}}_t+\\varepsilon)$, the ratio of between-class scatter to within-class scatter of frozen-backbone features for domain $t$. This feeds the capacity rule $r_t=(s_0/s_t)r_0$, which allocates a larger hidden dimension to domains whose classes are more entangled, with $s_0$ and $r_0$ fixed by a reference domain (ImageNet for separability; $r_0=64$ on DomainNet and CORe50, $r_0=1$ on CDDB). The supporting theoretical motivation is a qualitative risk decomposition $R_t(r_t)=A_t e^{-\\kappa_t r_t}+B\\sqrt{r_t/N_t}$, whose stationarity condition equates a marginal plasticity gain with a generalization and budget cost, yielding $r^*_t\\propto 1/\\kappa_t$ and, via a monotone map $\\kappa_t\\propto s_t$, the inverse rule. The second mechanism is BPG-Inference: $k=5$ k-means prototypes per domain, nearest-prototype distances $a_t=\\min_j\\|z-\\nu_{t,j}\\|_2$, softmax confidences $w_t$, a parameter-free pruning of domains below the uniform-confidence threshold, renormalization, and logit fusion $y_{\\mathrm{final}}=\\sum_t w'_t y_t$.","core_discovery":"The paper's central claim is that the plasticity–generalization tension in parameter-isolation DIL is removable in a rehearsal-free setting. BPG-Adapter sets each domain's adapter hidden dimension by the inverse-proportionality rule $r_t=(s_0/s_t)r_0$, where $s_t=R^{\\mathrm{bcs}}_t/(R^{\\mathrm{wcs}}_t+\\varepsilon)$ is a feature-separability ratio of between-class scatter to within-class scatter computed on frozen backbone features, and $s_0$ comes from ImageNet as the reference domain. BPG-Inference replaces the single-domain commitment with a soft mixture: per-domain k-means prototypes define a nearest-prototype affinity, a softmax turns distances into confidences, candidate domains below the uniform-confidence threshold are pruned, and the surviving per-domain classifiers are fused. On DomainNet with ViT-B/16 the framework reports 72.19% average accuracy and 0.22% average forgetting, within two points of the oracle-domain-ID upper bound; on CLIP it reports 75.72% with 0.59% forgetting, with analogous gains on CDDB and CORe50, and the soft mixture improves existing prompt methods when plugged into their inference.","pith_inferences":["If the monotone separability-to-capacity relationship holds generally, the hand-set reference capacity $r_0$ could itself be determined from the distribution of separability scores across the encountered domains, removing the last per-dataset hyperparameter; the paper does not explore this.","The large improvement from replacing hard selection with soft mixture suggests that much of what is reported as 'forgetting' in parameter-isolation methods may actually be domain-ID misclassification at test time rather than representation drift; reporting accuracy conditioned on correct domain assignment would disentangle the two sources.","Because BPG-Inference needs only frozen features and prototypes, it can be applied retrospectively to any already-trained parameter-isolation model; a natural stress test is to apply it to methods whose adapters were sized uniformly and measure how much of the gain comes from the soft mixture alone."],"forward_implications":["On DomainNet with ViT-B/16, BPG reaches 72.19% average accuracy and 0.22% average forgetting, surpassing the next-best rehearsal-free method by 4.24% in accuracy and closing the gap to the oracle-domain-ID upper bound to 2.02 points.","On the CLIP backbone, BPG reaches 75.72% with 0.59% forgetting on DomainNet, 93.91% with 0.12% forgetting on CDDB, and 92.46% accuracy on CORe50, ahead of the corresponding prompt-based baselines.","BPG-Inference is a plug-and-play module: added on top of S-Prompts, PINA, C-Prompt, and KA-Prompt, it improves average accuracy on all three benchmarks, with the largest gain being +5.42 points on DomainNet.","BPG-Adapter transfers to LoRA as well as adapters, showing the capacity rule is not tied to a specific parameter-efficient fine-tuning module.","With $r_0=64$, BPG's total parameter budget matches a uniform adapter at $r=256$ yet improves both accuracy and forgetting, indicating that the gain comes from reallocating capacity rather than adding parameters."],"supporting_citations":[{"why":"Defines the parameter-isolation prompt-learning baseline and the K-NN hard domain-selection protocol that BPG builds on and compares against.","marker":"[22]"},{"why":"The CLIP mixture-of-prompts method whose hard selection BPG-Inference improves on and that serves as a CLIP-backbone baseline.","marker":"[39]"},{"why":"A parameter-isolation method with patch-shuffle hard selection; BPG-Inference is tested on top of PINA and PINA is a main comparison baseline.","marker":"[40]"},{"why":"Compositional prompting baseline that BPG-Inference also upgrades, providing a soft-selection comparison point.","marker":"[41]"},{"why":"ImageNet supplies the reference separability score $s_0$ used in the capacity rule $r_t=(s_0/s_t)r_0$.","marker":"[73]"},{"why":"DomainNet is the primary benchmark and the source of the headline accuracy and forgetting numbers.","marker":"[76]"},{"why":"CDDB supplies the continual deepfake-detection benchmark used for the method's second set of results.","marker":"[77]"},{"why":"CORe50 supplies the non-overlapping 11-domain object-recognition benchmark for the method's generalization results.","marker":"[78]"}],"fun_headline_variants":["Domain difficulty sizes adapters; soft mixing cuts forgetting to 0.22%","Separability score sizes each adapter; soft domain fusion at test","Adapter dimension from feature scatter; soft mixing beats hard ID","BPG: separable adapters, soft test-time fusion, 0.22% forgetting","Per-domain adapter size by separability; soft mixture inference"],"cache_read_input_tokens":27392,"weakest_assumption_plain":"The paper assumes that a domain's feature-separability score reliably orders how much adapter capacity it needs, via an increasing map from separability to error-decay rate, and it hand-sets the reference capacity $r_0$ per dataset; if that ordering or the chosen reference is wrong, capacity is misallocated.","fun_headline_variants_meta":{"raw":{"variants":["Domain difficulty sizes adapters; soft mixing cuts forgetting to 0.22%","Separability score sizes each adapter; soft domain fusion at test","Adapter dimension from feature scatter; soft mixing beats hard ID","BPG: separable adapters, soft test-time fusion, 0.22% forgetting","Per-domain adapter size by separability; soft mixture inference"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000349,"raw_usage":{"total_tokens":1925,"prompt_tokens":983,"completion_tokens":942,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":599,"completion_tokens_details":{"reasoning_tokens":848}},"tokens_in":599,"tokens_out":942,"duration_ms":8283,"temperature":1.0,"reasoning_tokens":848,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T17:04:49.807599+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a new multi-domain benchmark, compute each domain's separability score $s_t$ and then grid-search its optimal adapter dimension in isolation; if any high-separability domain peaks only at a large adapter dimension, or any low-separability domain peaks at a small one, the monotone inverse rule $r_t=(s_0/s_t)r_0$ fails.","supporting_citations":[{"cited_title":"Mop-clip: A mixture of prompt-tuned clip models for domain incre- mental learning,","cited_arxiv_id":null,"evidence_quote":"The CLIP mixture-of-prompts method whose hard selection BPG-Inference improves on and that serves as a CLIP-backbone baseline."},{"cited_title":"Non- exemplar domain incremental learning via cross-domain concept in- tegration,","cited_arxiv_id":null,"evidence_quote":"A parameter-isolation method with patch-shuffle hard selection; BPG-Inference is tested on top of PINA and PINA is a main comparison baseline."},{"cited_title":"Compositional prompting for anti- forgetting in domain incremental learning,","cited_arxiv_id":null,"evidence_quote":"Compositional prompting baseline that BPG-Inference also upgrades, providing a soft-selection comparison point."},{"cited_title":"Moment matching for multi-source domain adaptation,","cited_arxiv_id":null,"evidence_quote":"DomainNet is the primary benchmark and the source of the headline accuracy and forgetting numbers."},{"cited_title":"A continual deepfake detection benchmark: Dataset, methods, and essentials,","cited_arxiv_id":null,"evidence_quote":"CDDB supplies the continual deepfake-detection benchmark used for the method's second set of results."},{"cited_title":"Core50: a new dataset and benchmark for continuous object recognition,","cited_arxiv_id":null,"evidence_quote":"CORe50 supplies the non-overlapping 11-domain object-recognition benchmark for the method's generalization results."}],"review_version":1}