acceptodds
Under review as a conference paper at ICLR 2027

DualMem-Evo: Dual-Memory Evolution for Replay-Free Continual Adaptation of Vision-Language Models

Abstract

Vision-language models (VLMs) have demonstrated remarkable transferability, yet their deployment in dynamic environments requires continuous adaptation to emerging tasks without access to previous training data. This setting poses a fundamental challenge: how to acquire new knowledge while preserving previously learned semantic representations. Existing continual adaptation approaches typically rely on either shared adaptation modules, which suffer from severe interference, or isolated task experts, which limit knowledge transfer across tasks. In this work, we propose DualMem-Evo, a dual-memory continual adaptation framework that views lifelong learning as a process of knowledge evolution rather than static knowledge preservation. Our framework introduces two complementary memories: a semantic evolution memory that continuously accumulates transferable cross-task knowledge, and an episodic expert memory that preserves task-specific expertise. To enable effective memory reuse under a strict no-replay constraint, we develop a Prototype-Guided Expert Evolution mechanism that dynamically retrieves, interacts with, and consolidates historical expert memory modules through cross-modal semantic prototypes and low-rank memory fusion. Furthermore, we introduce functional memory separation and stability-aware semantic regulation to maintain representation diversity while preventing semantic drift during incremental adaptation. The empirical results demonstrate that our framework achieves the state-of-the-art performance.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
continual learning, vision-language models, parameter-efficient fine-tuning, low-rank adaptation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_xFdHOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a replay-free, persistently adaptive framework for visual language models called DualMem-Evo, which combines evolutionary semantic memory with task-specific contextual experts. The authors introduce prototype-guided expert retrieval and low-rank integration to reuse historical knowledge, and incorporate memory separation and semantic stability regularization.

Critical strengths
  1. The evaluation covers heterogeneous datasets, different task orders, limited supervision, unidentified module selection, and class-incremental learning.

  2. Architecture ablation experiments use comparable LoRA overall rankings, which helps distinguish the impact of memory organization from the impact of increased adapter capacity.

Critical weaknesses
  1. Robustness of the principal gains has not been confirmed. Under Order-I conditions, the reported improvements compared to DPW were approximately 0.3 and 0.7 percentage points in Average and Last, respectively; improvements were also generally small under other MTIL settings. Similarly, adding PGEEM to controlled dual memory only improved Average from 79.36 to 79.59 and Last from 88.78 to 88.95, equivalent to only 0.23 and 0.17 percentage points of gain, respectively. The paper does not report the training seed mean or variability for these principal comparisons, making it difficult to determine whether the observed differences reflect consistent gains or normal fluctuations during training runs. This is particularly important in component ablation experiments, as small incremental gains are used to support the contribution of PGEEM. The authors should replicate principal baseline comparisons and related ablation experiments under matched experimental conditions using multiple independent training seeds and report the mean and standard deviation or confidence intervals. Reporting variability in performance differences will help further determine whether the proposed method consistently outperforms baseline methods and whether PGEEM offers reproducible additional advantages.

  2. Appendix A.3 describes scoring candidate objects using normalized prediction confidence and image-text consistency, followed by candidate selection and temperature-weighted fusion. However, this description does not fully clarify the specific scoring functions, normalization operations, number of candidate objects, temperature, or parameter selection procedures. For example, it is unclear how these two scoring terms are calculated and combined, on which candidate objects or sample sets the normalization is performed, and how many candidate objects are evaluated and retained. The mathematical relationship between candidate object scores and final fusion weights also needs to be clarified. These choices directly determine which candidate objects contribute to the prediction and their weights, so different reasonable implementations may produce different results. Authors should provide complete formulas or pseudocode, along with the numerical settings used in the experiments. They should also explain how these settings were chosen, what data were used for selection, and whether the same settings were shared across different datasets and task sequences. A focused sensitivity analysis of the number of candidates and fusion temperature helps assess the dependence of the reported gains on these choices.

Questions for authors
  1. Could you provide the complete equations or pseudocode for the inference process in Appendix A.3, including all candidate scores, selection rules, fusion parameters, and their tuning schemes? Could you compare the complete model and the model containing only episodic memories under the same inference rules to differentiate between training gains and candidate fusion gains?

  2. After completing T tasks, are all T integrated episodic memory experts retained? If so, the ranking of each expert is bounded, but the total storage capacity still increases with T. What storage object does the compression ratio in Appendix B refer to? Please specify the total storage capacity and the number of uppropagation steps or latency during inference.

16 comments
AuthorsOpenReview ↗

1-Response to Reviewer xFdH

Response to Reviewer xFdH

We sincerely thank Reviewer xFdH for recognizing our comprehensive evaluation across multiple heterogeneous datasets, different task orders, few-shot regimes, identification-free module selection (IFMS-MTIL), and class-incremental learning (CIL), as well as our rank-matched (Comparable Overall Ranks) architectural ablation design (CS1, CS2). Your precise and incisive questions regarding the multi-seed statistical stability of our main and ablation gains, complete inference fusion equations and rigorous hyperparameter selection criteria, controlled architectural comparison under unified inference rules, and the exact scope of SVD compression and deployment overhead go straight to the heart of experimental rigor. Below, we address each of your points in detail.


1. Magnitude Analysis of Main & PGEEM Ablation Gains and Multi-Seed Verification (W1)

(1) Contextualizing the 0.3%–0.7% Gains on MTIL: Empirical Performance Ceiling and Cross-Protocol Consistency
On MTIL benchmarks, recent CLIP-based continual learning baselines (such as DPW and KeepLoRA+) already operate in a high-accuracy regime. To objectively quantify the remaining room for improvement, we computed an offline Empirical Ceiling reference on the Order-I accuracy matrix: for each unseen task, we take the highest transfer accuracy observed at any stage prior to learning that task; for each seen task, we replace its entry with the score of an independent per-task LoRA fine-tuning (Oracle LoRA Fine-tune) model. This yields a theoretical combined reference of 71.1% (Transfer) / 80.4% (Average) / 89.4% (Last).
Consequently, the strongest baseline DPW (70.4% / 79.3% / 88.3%) is only 0.7% / 1.1% / 1.1% away from this empirical ceiling. Within this highly saturated regime, DualMem-Evo's 70.4% / 79.6% / 89.0% effectively eliminates roughly 30%–64% of the remaining gap between the strongest baseline and the empirical ceiling. More importantly, this advantage does not occur by chance in a single setting, but holds consistently across Order-II (+0.1% / +0.4% / +0.5%), 5-shot (+0.0% / +0.2% / +0.5%), 16-shot (Average +0.4%, Last +0.5%), IFMS-MTIL (+0.3% / +0.4% / +0.3%), and all 6 CIL benchmarks (up to +2.27% in Average; see Appendix C.1; and +2.54% / +4.09% / +4.21% over CL-LoRA under matched settings; see our Response to Reviewer ELGv, Section 1).

(2) Multi-Seed Statistical Evaluation Under Strictly Matched Protocols (Multiple Independent Training Seeds)
Following your recommendation, under strictly matched experimental conditions, we repeated our experiments across multiple independent training random seeds for the strongest baseline DPW, our controlled dual-memory baseline (Regulated Dual Memory w/o PGEEM), and the full DualMem-Evo (w/ PGEEM), reporting means, standard deviations, and paired differences:

Model Configuration (MTIL Order-I, Multi-Seed)Transfer (Mean ± Std)Average (Mean ± Std)Last (Mean ± Std)
Strong Baseline DPW70.38 ± 0.0579.31 ± 0.0688.32 ± 0.10
Regulated Dual Memory (w/o PGEEM)70.14 ± 0.0379.35 ± 0.0488.76 ± 0.11
DualMem-Evo (Full Model, w/ PGEEM)70.39 ± 0.0279.58 ± 0.0188.95 ± 0.05
Net Gain vs. DPW (Δvs. DPW\Delta _ {\text{vs. DPW}})+0.01+0.27+0.63
Paired Gain from PGEEM (ΔPGEEM\Delta _ {\text{PGEEM}})+0.24 ± 0.04+0.23 ± 0.04+0.19 ± 0.15

The multi-seed statistics clearly demonstrate that: (i) Full DualMem-Evo exhibits minimal sensitivity to training seed randomness (with standard deviations of only 0.02%, 0.01%, and 0.05% on Transfer, Average, and Last—substantially smaller than DPW's variance), reliably outperforming the strong baseline DPW; and (ii) incorporating PGEEM delivers stable, reproducible positive paired gains across all three metrics (+0.24% / +0.23% / +0.19%) while significantly reducing the standard deviation of final retention (Last) from 0.11% to 0.05%. Furthermore, in the newly added Appendix D.5 and our controlled retrieval ablation, we confirm that our Centered signed collaboration—designed in accordance with the inevitable positive/negative polarization of layer-wise weights—not only significantly accelerates early-to-mid training convergence and lowers cross-entropy loss during new-memory formation, but also consistently outperforms Nearest and Random retrieval across all final metrics (see our Response to Reviewer Ttpt, Section 3 for detailed analysis and comparison tables).

AuthorsOpenReview ↗

2-1Response to Reviewer xFdH

2. Complete Inference Equations in Appendix A.3, Leak-Free Selection Protocol, and Hyperparameter Sensitivity Analysis (W2, Q1a)

2.1 Complete Inference Scoring, Normalization, and Temperature-Scaled Fusion Equations

Under standard MTIL (Order-I, Order-II, 5-shot, 16-shot), seen tasks are evaluated directly using the current semantic memory GtG _ {t} paired with the corresponding episodic expert Ej\mathcal{E} _ {j}; for unseen future tasks (as well as unknown/ambiguous samples in IFMS-MTIL), the model performs label-free selection and fusion over the candidate pool Ccand\mathcal{C} _ {\mathrm{cand}} formed by the learned expert set {Ee}e=0t−1\lbrace \mathcal{E} _ {e} \rbrace _ {e=0}^{t-1} (each combined with the current GtG _ {t}) plus frozen CLIP.

For an individual test sample xx and target class set Ctarget\mathcal{C} _ {\mathrm{target}} (of size C=∣Ctarget∣C = |\mathcal{C} _ {\mathrm{target}}|) specified by the evaluation protocol, each candidate model e∈Ccande \in \mathcal{C} _ {\mathrm{cand}} extracts normalized visual features f^I,e(x)\widehat{f} _ {I,e}(x) and target class text features f^T,e(c)\widehat{f} _ {T,e}(c) to compute cosine similarity logits and class probability distributions: Le,c(x)=100f^I,e(x)⊤f^T,e(c),pe,c(x)=softmax⁡c∈Ctarget(Le,c(x))L _ {e,c}(x) = 100\widehat{f} _ {I,e}(x)^\top \widehat{f} _ {T,e}(c), \qquad p _ {e,c}(x) = \operatorname{softmax} _ {c\in\mathcal{C} _ {\mathrm{target}}}\bigl(L _ {e,c}(x)\bigr) Note: All normalizations are performed independently within each single test sample xx across target classes c∈Ctargetc \in \mathcal{C} _ {\mathrm{target}} or across candidate experts ee, with zero reliance on statistics from other samples in the test batch. We evaluate the reliability of candidate ee on sample xx along two complementary dimensions:

  1. Normalized Prediction Confidence: Derived from the normalized Shannon entropy Ue(x)∈[0,1]U _ {e}(x) \in [0, 1] of the class distribution as the confidence term 1−Ue(x)1 - U _ {e}(x): Ue(x)=−∑c∈Ctargetpe,c(x)log⁡(pe,c(x)+10−8)log⁡CU _ {e}(x) = \frac{-\sum _ {c\in\mathcal{C} _ {\mathrm{target}}} p _ {e,c}(x)\log\bigl(p _ {e,c}(x)+10^{-8}\bigr)}{\log C}
  2. Cross-Modal Image–Text Consistency: Defined as the maximum cosine similarity between the sample's visual feature and the best-matching class text feature: Me(x)=max⁡c∈CtargetLe,c(x)100M _ {e}(x) = \max _ {c\in\mathcal{C} _ {\mathrm{target}}} \frac{L _ {e,c}(x)}{100}
  3. Joint Reliability Scoring and Top-KevalK _ {\mathrm{eval}} Temperature-Scaled Fusion: Multiplying the two terms yields the candidate reliability score ρe(x)\rho _ {e}(x), requiring a high-scoring candidate to simultaneously exhibit "low predictive uncertainty (sharp distribution)" and "high cross-modal semantic alignment": ρe(x)=(1−Ue(x))Me(x)\rho _ {e}(x) = \bigl(1 - U _ {e}(x)\bigr) M _ {e}(x) After computing ρe(x)\rho _ {e}(x) for all t+1t+1 candidates, we retain at most the Top-Keval=min⁡(5,t+1)K _ {\mathrm{eval}} = \min(5, t+1) highest-scoring candidates to form the set V(x)\mathcal{V}(x), and map them via fusion temperature τfuse=0.1\tau _ {\mathrm{fuse}} = 0.1 into final fusion weights πe(x)\pi _ {e}(x): πe(x)=softmax⁡e∈V(x)(ρe(x)τfuse)=exp⁡(ρe(x)/0.1)∑e′∈V(x)exp⁡(ρe′(x)/0.1),y^=arg⁡max⁡c∈Ctarget∑e∈V(x)πe(x)Le,c(x)\pi _ {e}(x) = \operatorname{softmax} _ {e\in\mathcal{V}(x)}\left(\frac{\rho _ {e}(x)}{\tau _ {\mathrm{fuse}}}\right) = \frac{\exp\bigl(\rho _ {e}(x)/0.1\bigr)}{\sum _ {e'\in\mathcal{V}(x)}\exp\bigl(\rho _ {e'}(x)/0.1\bigr)}, \qquad \hat{y} = \arg\max _ {c\in\mathcal{C} _ {\mathrm{target}}} \sum _ {e\in\mathcal{V}(x)} \pi _ {e}(x) L _ {e,c}(x) (Note: For the complete equations of IFMS-MTIL empirical quantile compatibility routing and CIL hierarchical owned-class margin inference, as well as the full list of training/evaluation hyperparameters across all experimental tables, please see our Response to Reviewer ELGv, Sections 3.2 and 3.3.)
AuthorsOpenReview ↗

2-2&3Response to Reviewer xFdH

2. Complete Inference Equations in Appendix A.3, Leak-Free Selection Protocol, and Hyperparameter Sensitivity Analysis (W2, Q1a)

2.2 Strict Selection Protocol for Parameters and Inference Configurations (Data Provenance & Selection Protocol)

In response to your concern regarding "how these hyperparameters were selected, what data were used, and whether they are shared across datasets," we explicitly clarify: the selection of all model parameters and inference configurations strictly follows the six-stage train/validation isolation protocol below, without ever using any test-set samples or test results for tuning:

StageData UsedAction & Strict Criterion
1. Data SplitOfficial training set of currently accessible tasksPartition into a Train split and a Validation split, and save sample indices
2. Model TrainingTrain split onlyOptimize dual-memory LoRA parameters, layer-wise collaboration coefficients wkiw _ {k}^i, and trainable weights
3. Config SelectionValidation split onlyEvaluate on the validation split to select fusion temperature τfuse\tau _ {\mathrm{fuse}}, candidate count KevalK _ {\mathrm{eval}}, routing thresholds, etc.
4. Config FreezingZero reference to any test resultsFreeze the final selected global configuration and save the configuration alongside its validation selection basis
5. Final EvaluationOfficial test setCompute the full accuracy matrix and Transfer / Average / Last metrics once under the frozen configuration
6. Multi-Seed VerificationFixed protocol & frozen configurationVary only the training random seed, reporting mean, standard deviation, and paired gains

Cross-Task and Cross-Sequence Sharing: Once determined on the validation split, Keval=5K _ {\mathrm{eval}}=5 and τfuse=0.1\tau _ {\mathrm{fuse}}=0.1 serve as globally frozen hyperparameters shared identically across all datasets and all sessions in Order-I, Order-II, 5-shot, 16-shot, and IFMS-MTIL, without any per-dataset or per-order adjustment.

2.3 Sensitivity Analysis on Retained Candidate Count KevalK _ {\mathrm{eval}} and Fusion Temperature τfuse\tau _ {\mathrm{fuse}}

Following your suggestion, we conducted dedicated sensitivity sweeps over the retained candidate count KevalK _ {\mathrm{eval}} (on full Order-I) and the fusion temperature τfuse\tau _ {\mathrm{fuse}} (across the first 5 sessions):

Retained Candidates After Scoring KevalK _ {\mathrm{eval}}1 (Top-1 Hard Routing)35 (Default)79
Transfer (%) (Order-I)70.0370.2670.3770.4670.49
Fusion Temperature τfuse\tau _ {\mathrm{fuse}}0.010.050.1 (Default)0.20.5
Transfer (%) (5-Session)64.6864.4064.4564.3464.31

The experiments demonstrate that: (i) compared to hard single-expert selection (Keval=1K _ {\mathrm{eval}}=1, 70.03%), multi-candidate soft fusion (Keval≥3K _ {\mathrm{eval}}\ge 3) effectively mitigates the risk of single-expert misselection and maintains smooth gains across Keval∈[3,9]K _ {\mathrm{eval}}\in [3, 9] (with default Keval=5K _ {\mathrm{eval}}=5 balancing accuracy and computational efficiency); and (ii) as the fusion temperature τfuse\tau _ {\mathrm{fuse}} varies across a 50×50\times range [0.01,0.5][0.01, 0.5], Transfer fluctuates by only 0.37% (and by only 0.14% within [0.05,0.5][0.05, 0.5]), proving that our method is highly robust to inference hyperparameters and that its gains do not rely on delicate hyperparameter tuning.

Reviewer_TtptOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

DualMem-Evo proposes a dual-memory continual learning framework for vision-language models, combining a shared memory that evolves across tasks with frozen per-task expert memories retrieved via prototype-based routing. It reports small but consistent gains over prior methods on several continual learning benchmarks.

Critical strengths
  1. The experiments cover a wide range of settings (full-shot, few-shot, identifier-free, and class-incremental) rather than a single benchmark, and the method stays competitive or best across nearly all of them, suggesting the gains are not overfit to one evaluation setup.

  2. The method achieves its results with notably fewer trainable parameters and no auxiliary data compared to several strong baselines, which is a meaningful practical advantage for a parameter-efficient continual learning method.

Critical weaknesses
  1. The shared memory plus per task experts framing overlaps with complementary memory continual learning methods already cited in Related Work. The paper's main claimed novelty, replacing nearest neighbor expert retrieval with centered, deviation-based routing, is not directly ablated against nearest-neighbor retrieval.

2.The method adds several components on top of a dual-LoRA base (prototype routing, SVD consolidation, CKA separation, KL stability regularization), but the resulting gains over the strongest baselines are mostly under 1 point and inconsistent across settings, making it unclear whether this added complexity is well justified.

Questions for authors
  1. Could you clarify which components of the framework are already established in prior continual learning work (e.g., dual/complementary memory designs, prototype-based retrieval) versus which parts are genuinely new, and what new insight the new parts offer?

  2. Could you report the training and inference overhead (time/memory) introduced by PGEM (prototype construction, routing, SVD consolidation) relative to the baselines, especially as the number of historical experts grows?

  3. The centered routing assumes experts with negative affinity provide "complementary or contrastive" knowledge, but this is only validated by comparing centered routing against random routing as a whole. Is there an ablation isolating the specific contribution of the negatively-routed experts to verify this assumption directly?

  4. Does error in the semantic memory from a poorly-learned session propagate to later sessions, since G_{t-1} is used as the teacher for stability regularization?

  5. PGEM computes one task-level prototype per session. How robust is this if a session's data isn't from a single clean distribution?

  6. Expert count grows linearly with the number of tasks. How does routing cost/scalability hold up over much longer task sequences (e.g., hundreds of tasks)?

8 comments
AuthorsOpenReview ↗

1-Response to Reviewer Ttpt

Response to Reviewer Ttpt

We sincerely thank Reviewer Ttpt for the positive assessment and for recognizing our method's broad generalization across four distinct continual learning protocols, as well as its practical value in outperforming strong baselines with fewer trainable parameters and zero historical replay or auxiliary datasets (CS1, CS2). Your insightful questions regarding our novelty boundary, complementary knowledge motivation and retrieval ablation, module cost–benefit trade-offs, training/inference overhead, potential error propagation, and long-horizon scalability have helped us articulate the core insights of "dynamic memory evolution" much more clearly. Below, we address each concern in detail.


1. Boundary Between Existing Building Blocks and Our Core Novelty & New Insights (W1, Q1)

We fully agree that dual-memory (fast/slow learning) concepts, prototype retrieval, knowledge distillation, CKA, and truncated SVD exist individually as mathematical tools in prior literature. However, representative prior methods such as CL-LoRA (CVPR 2025) treat task-specific adapters as static isolated parameter snapshots that are frozen and sealed immediately after training. Our core novelty lies in constructing a closed-loop "Memory Evolution & Absorption" mechanism in which frozen historical experts dynamically participate in forming and consolidating new memories:

  1. Cross-Modal Evolving Memory Bank and Layer-Wise Co-Evolution of Frozen Experts (PGEEM): While historical episodic experts remain strictly frozen in subsequent sessions to prevent catastrophic forgetting, they are not passive bystanders—they actively participate in current-task residual forward and gradient propagation via learnable layer-wise signed collaboration coefficients wkiw _ {k}^i (240 scalar parameters in total), and their multi-expert evolutionary trajectory is absorbed into a single fixed-rank expert Et\mathcal{E} _ {t} via truncated SVD at the end of the session. As shown in Appendix D.1, adaptive co-evolution yields an intra-session functional contribution redistribution total variation of DTV=0.1733D _ {\mathrm{TV}} = 0.1733—32×32\times higher than static fixed fusion (DTV=0.0054D _ {\mathrm{TV}} = 0.0054)—confirming that historical memories are substantively and dynamically reorganized during optimization.
  2. Conditional Marginal CKA Decoupling and Single-Path Local Stability Rule: On the visual encoder, we minimize the CKA similarity between the conditional marginal contributions of the semantic and episodic pathways (Ldec\mathcal{L} _ {\mathrm{dec}}, driving inter-branch redundancy down from 0.8470.847 to 0.0450.045 while preserving non-zero active norms on both branches; see Appendix D.2), and apply intra-session stability distillation (Lstability\mathcal{L} _ {\mathrm{stability}}) exclusively along the shared semantic pathway GtG _ {t}, fully unlocking the task-specific specialization potential of TtT _ {t}.

Direct Comparison with CL-LoRA Under Identical Configurations: Adapting CL-LoRA to the exact same CLIP ViT-B/16 backbone and MTIL protocol across the first 5 sessions of Order-I, DualMem-Evo (64.45% / 71.79% / 84.71%) substantially outperforms CL-LoRA (61.91% / 67.70% / 80.50%) by +2.54%, +4.09%, and +4.21% in Transfer, Average, and Last, respectively (see our Response to Reviewer ELGv, Section 1 for the full comparison table and architectural details), strongly validating the advantage of dynamic memory co-evolution over traditional static dual-branch designs.

AuthorsOpenReview ↗

2-Response to Reviewer Ttpt

2. Justifying Architectural Necessity and Value Amidst Compact Headline Gains (W2)

Regarding the observation that headline gains on some MTIL benchmarks appear within 1%, we explain the substantive value of our design from three perspectives: benchmark saturation ceilings, multi-seed statistical significance under a leak-free protocol, and synergistic module ablations (see also our Response to Reviewer xFdH, Sections 1 and 2 for complete tables):

  1. MTIL Empirical Ceiling Dictates That 0.3%–0.7% Is Already a Substantial Breakthrough: On the Order-I matrix, combining the best historically observed transfer scores with independent per-task LoRA fine-tuning (Oracle) scores yields an offline empirical ceiling of 71.1% / 80.4% / 89.4%. Because the strongest baseline DPW (70.4% / 79.3% / 88.3%) already approaches this ceiling, DualMem-Evo (70.4% / 79.6% / 89.0%) actually closes roughly 30%–64% of the remaining gap between the strongest baseline and the empirical ceiling; meanwhile, on the less saturated CIL benchmarks (Appendix C.1), our Average accuracy gains reach +1.39% to +2.27%.
  2. Six-Stage Leak-Free Selection Protocol and Multi-Seed Paired Significance: We strictly enforce a six-stage protocol ("Train/Val split →\to Train split parameter optimization →\to Val split hyperparameter selection →\to Config freezing with zero access to test results →\to Official test set evaluation →\to Multi-seed verification"; see our Response to Reviewer xFdH, Section 2.2). Across multiple independent training random seeds, DualMem-Evo achieves 70.39±0.02 / 79.58±0.01 / 88.95±0.05, not only consistently outperforming the strong baseline DPW (70.38±0.05 / 79.31±0.06 / 88.32±0.10), but also delivering stable positive paired gains (+0.24±0.04 / +0.23±0.04 / +0.19±0.15) over our regulated dual-memory baseline (w/o PGEEM: 70.14±0.03 / 79.35±0.04 / 88.76±0.11) while reducing the standard deviation on Last by ~54%.
  3. Complementary Component Roles: As shown in our module ablation (Section 4.2), removing Lstability\mathcal{L} _ {\mathrm{stability}} causes Last to drop sharply by 2.84%, removing Ldec\mathcal{L} _ {\mathrm{dec}} reduces Transfer by 0.48%, and incorporating PGEEM synergistically elevates all three metrics.
AuthorsOpenReview ↗

3-Response to Reviewer Ttpt

3. Deep Motivation, Process-Level Verification, and Centered / Nearest / Random Retrieval Ablation for "Complementary Knowledge" (W1, Q3)

We thank the reviewer for raising this pivotal question. We address your concern from three levels: our empirical discovery during early development (design motivation), optimization dynamics during new-memory formation (process-level verification), and controlled Centered / Nearest / Random comparisons (outcome-level verification):

  1. Why Choose Centered (Mean-Centered Signed Weights)? — Motivated by the "Inevitable Positive/Negative Polarization of Layer Weights" to Accelerate Memory Formation
    During the early development of our framework, an in-depth analysis of freely optimized layer-wise collaboration coefficients wkiw _ {k}^i without prior constraints revealed a key empirical regularity: during cross-task adaptation, the emergence of both positive and negative weights across layers is inevitable—because constructing representations for a new task naturally requires both positively reusing shared historical features (additive synergy) and negatively suppressing historical feature directions that cause confusion or interference (subtractive contrast). Motivated by this finding, rather than adopting traditional nearest-neighbor weighting with purely positive initial values, we specifically designed our mean-centered bidirectional initialization (Centered, s~k=sk−sˉ\tilde{s} _ {k} = s _ {k} - \bar{s}): by explicitly supplying both positive (shared transfer) and negative (contrastive repulsion) priors at training step 00 in alignment with the natural evolution of layer-wise weights, Centered directly accelerates the formation of new episodic memories.
  2. Process-Level Empirical Analysis (Newly Added in Appendix D.5) To directly verify how positive and negative historical memories accelerate memory formation, we densely tracked current-task accuracy and cross-entropy loss every 100 optimization steps during new-memory formation under three configurations: Signed (both positive and negative weights), Positive-only, and Negative-only. Experiments on DTD and CIFAR100 clearly demonstrate that the Signed strategy exhibits noticeably faster accuracy convergence and lower cross-entropy loss during early-to-mid training, confirming from an optimization-dynamics perspective that the Centered design fulfills its goal of accelerating memory formation.
  3. Controlled Retrieval Ablation: Centered vs. Nearest vs. Random
    Following your suggestion, we conducted a strictly controlled comparison on a streamlined 5-task continual learning sequence (Caltech101 →\to CIFAR100 →\to DTD →\to EuroSAT →\to MNIST) comparing our mean-centered signed retrieval (Centered, positive + negative weights), standard nearest-neighbor retrieval (Nearest, positive weights only), and random retrieval (Random):
Retrieval & Initialization StrategyTransfer (%)Average (%)Last (%)
Nearest (Standard Nearest-Neighbor, Positive Weights Only)59.2678.7892.68
Random (Random Historical Expert Retrieval)59.4978.8792.56
Centered (Mean-Centered Signed Weights, Ours)59.5178.9492.90

The empirical data strongly support our design motivation: Centered achieves the best performance across all three metrics (Transfer, Average, and Last), outperforming standard Nearest by +0.25%, +0.16%, and +0.22%, and outperforming Random by +0.02%, +0.07%, and +0.34%, respectively (notably, positive-only Nearest even slightly trails Random on Transfer because purely positive nearest-neighbor priors can over-constrain early optimization within similar historical subspaces, whereas Centered balances positive transfer with contrastive regularization to simultaneously achieve the highest Transfer, Average, and Last).

Reviewer_ELGvOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies replay-free continual adaptation of vision-language models. On a frozen CLIP it maintains a continually updated semantic LoRA and per-task episodic LoRAs. Cross-modal prototypes are used to retrieve historical experts, signed layer-wise merging weights are learned, and truncated SVD compresses the merged result into the current expert; marginal-representation CKA and semantic-branch distillation are used to regulate stability. The experiments cover two full-shot task orders, 5-shot, 16-shot, identification-free module selection (IFMS-MTIL), and CIFAR-100/TinyImageNet class-incremental learning. The paper reports competitive system-level performance and analyzes the individual modules through capacity-matched and component-wise ablations.

Critical strengths
  1. Broad experimental coverage and a competitive full system. Beyond standard MTIL, the paper also evaluates few-shot, IFMS and class-incremental settings, and provides complete task-time accuracy matrices. Table 1 (PDF p. 7) reports Order-I Average/Last of 79.6/89.0, against DPW's 79.3/88.3; the three incremental settings on TinyImageNet also show positive gains. These results make the approach worth attention and prevent it from being reduced to a single setting that relies on known task identities.
  2. The architectural and component ablations provide a basis for analyzing the dual-branch design. Table 2 compares a single branch of rank 128 against two 64+64 branches, controlling the total rank of the base architecture; Table 3 further separates the effects of the regularization terms and of PGEEM. Distinguishing representational diversity from parameter orthogonality is also reasonably motivated. This evidence supports the feasibility of the system, which is why a clear-reject score of 1 was not given.
Critical weaknesses
  1. The boundary of the core novelty and the standalone benefit of the added mechanisms are not sufficiently established. CL-LoRA (https://arxiv.org/abs/2505.24816), which this paper already cites, has task-shared/task-specific LoRA and shared-knowledge distillation; this paper mainly uses it as a contrast on prototype statistics and lacks a direct technical comparison of the dual-adapter architecture. More importantly, in Table 3 (p. 9) PGEEM adds only 0.23 points of Average and 0.17 points of Last on top of the already-regularized dual memory, and the main results are mostly narrow margins, yet no statistical evidence across training seeds is provided. Centered low similarity does not automatically imply complementary knowledge, and the current analysis does not adequately separate the contributions of centered retrieval, random experts, negative weights and merge training. It therefore remains unclear whether the extra complexity yields a stable contribution that is distinct from prior work.
  2. The core routing rule and the actual resource cost lack sufficient definition, which affects reproducibility and efficiency assessment. Appendix A.3 (p. 15) does not fully give the formulas and values for confidence normalization in identification-free inference, the image–text consistency combination, the number of candidates, the temperature, and the CIL owned-class margins. The optimizer, learning rate, LoRA scaling, and key regularization/distillation settings are also incompletely specified. Top-K prototype retrieval at training time cannot substitute for these test-time rules. In addition, truncated SVD only compresses one current merged result, while the historical expert pool still grows with the number of tasks; Table 15 (p. 26) counts only trainable parameters and does not account for total storage and multi-expert inference cost. Since the test-time selection mechanism directly affects accuracy, these gaps have a material impact on the overall judgment.
Questions for authors
  1. Can you provide the complete expert-selection, score-normalization, temperature-fusion and class-margin formulas for standard MTIL, IFMS and CIL, together with all corresponding training and inference configurations? Which configuration set is used in each table?
  2. After truncated SVD, are all historical experts still retained? Please report separately the total number of saved parameters, the actual number of candidate evaluations, inference latency and memory as the number of tasks grows, distinguishing current trainable parameters from full deployment cost.
  3. In a compatible setting with the same backbone, parameter count and inference budget, how large is the gain over CL-LoRA? Can paired results over multiple seeds support the 0.23/0.17 point gains of PGEEM, and separate the contributions of centered retrieval, random retrieval and positive-weights-only?
16 comments
AuthorsOpenReview ↗

1-Response to Reviewer ELGv

We sincerely thank Reviewer ELGv for the thorough and constructive review, and for recognizing our comprehensive experimental evaluation across diverse continual learning protocols and our rank-matched architectural ablations (CS1, CS2). Your insightful questions regarding the technical boundary with CL-LoRA, the multi-seed statistical significance of PGEEM, the motivation behind complementary knowledge, and the exact inference routing equations and deployment overhead have helped us substantially sharpen our contributions. Below, we address each of your points in detail.


1. Technical Boundary and Controlled Experimental Comparison with CL-LoRA (W1, Q3)

(1) Fundamental Differences in Core Mechanism and Architecture
We fully agree that shared/task-specific adapters and knowledge distillation are foundational concepts in continual learning. However, DualMem-Evo differs fundamentally from CL-LoRA (CVPR 2025) in both its knowledge reuse mechanism and cross-modal architecture:

  • From "Static Isolation" to "Memory Evolution & Absorption with Frozen Historical Experts": In CL-LoRA, task-specific branches merely isolate current-task features into standalone static snapshots. In contrast, the core of our work lies in enabling frozen historical episodic experts to dynamically participate in current-task gradient optimization via learnable layer-wise signed collaboration coefficients wkiw _ {k}^i without modifying historical parameters, and subsequently absorbing this collaborative evolutionary trajectory into a single, fixed-rank new episodic expert Et\mathcal{E} _ {t} via truncated SVD at the end of each session.
  • Bimodal MLP Topology and Single-Path Local Stability Distillation: CL-LoRA places shared and specific LoRAs solely in the shallow and deep attention Q/V projections of a unimodal ViT. DualMem-Evo deploys evolving semantic memory GtG _ {t} and episodic memory TtT _ {t} in parallel across the MLP layers of both visual and textual encoders in frozen CLIP, enforces functional decoupling via conditional marginal CKA (Ldec\mathcal{L} _ {\mathrm{dec}}, which reduces inter-branch redundancy from 0.8470.847 to 0.0450.045; see Appendix D.2), and applies stability distillation (Lstability\mathcal{L} _ {\mathrm{stability}}) exclusively along the semantic memory path GtG _ {t}, preventing historical semantic structure from constraining the task-specific plasticity of the current episodic expert.

(2) Direct Experimental Comparison Under Matched Backbone, Parameter Budget, and Protocol
We adapted CL-LoRA from its native unimodal ViT/CIL setting to the exact same CLIP ViT-B/16 backbone and MTIL protocol, and conducted a controlled comparison across the first 5 heterogeneous sessions of Order-I (Aircraft →\to Caltech101 →\to CIFAR100 →\to DTD →\to EuroSAT) against ZSCL, KeepLoRA+, and our method:

MethodTransfer (%)Average (%)Last (%)
ZSCL62.3067.2379.14
CL-LoRA (Adapted to CLIP / MTIL)61.9167.7080.50
KeepLoRA+63.5371.0083.74
DualMem-Evo (Ours)64.4571.7984.71

Under the exact same backbone and evaluation budget, DualMem-Evo outperforms CL-LoRA by +2.54% in Transfer, +4.09% in Average, and +4.21% in Last. This demonstrates that static dual-branch adapters and standard distillation alone are insufficient for cross-modal multi-domain continual adaptation, validating the distinct and independent contribution of our historical expert co-evolution and functional decoupling mechanisms.


AuthorsOpenReview ↗

2-Response to Reviewer ELGv

2. Multi-Seed Paired Verification of PGEEM and Motivation of Centered Complementary Knowledge (W1, Q3)

(1) Empirical Ceiling Reference and Multi-Seed Paired Statistics
On MTIL Order-I, combining the best historical transfer score for each unseen task with the independent per-task LoRA fine-tuning (Oracle) score for each seen task yields an offline empirical ceiling of 71.1% / 80.4% / 89.4%. DualMem-Evo (70.4% / 79.6% / 89.0%) effectively closes roughly 30%–64% of the remaining gap between the strongest baseline DPW and this empirical ceiling. We further performed multi-seed evaluations across independent training random seeds under strictly matched conditions (see also our Response to Reviewer xFdH, Section 1 for full details):

Model Configuration (MTIL Order-I, Multi-Seed)Transfer (Mean ± Std)Average (Mean ± Std)Last (Mean ± Std)
Strong Baseline DPW70.38 ± 0.0579.31 ± 0.0688.32 ± 0.10
Regulated Dual Memory (w/o PGEEM)70.14 ± 0.0379.35 ± 0.0488.76 ± 0.11
DualMem-Evo (Full Model, w/ PGEEM)70.39 ± 0.0279.58 ± 0.0188.95 ± 0.05
Paired Gain from PGEEM (Δ\Delta)+0.24 ± 0.04+0.23 ± 0.04+0.19 ± 0.15

The results show that DualMem-Evo exhibits remarkably small cross-seed standard deviations (only 0.01%–0.05%, substantially lower than DPW). Incorporating PGEEM yields consistently positive paired gains across all metrics while cutting the standard deviation on Last by more than half (from ±0.11\pm 0.11 to ±0.05\pm 0.05).

(2) Core Motivation for Negative Weights (Centered Complementary Knowledge) and Controlled Retrieval Ablation: Accelerating New Memory Formation
We thank the reviewer for scrutinizing the motivation behind complementary knowledge. During the early development of our framework, analyzing freely optimized layer-wise collaboration coefficients wkiw _ {k}^i without prior constraints revealed a fundamental empirical phenomenon: during cross-task adaptation, the emergence of both positive and negative weights across layers is inevitable (as adapting to a new task requires both positively reusing shared features and negatively suppressing confusable or interfering historical features). Therefore, our core motivation for designing the mean-centered (Centered) bidirectional initialization is to align with this natural polarization to accelerate the formation of new episodic memories:

  • Process-Level Acceleration Verification (Appendix D.5): Tracking new memory formation every 100 optimization steps, we compared Signed (both positive and negative weights), Positive-only, and Negative-only variants. On tasks such as DTD and CIFAR100, the Signed variant exhibits noticeably faster accuracy ramp-up and lower cross-entropy loss during early-to-mid training (complementing the 32×32\times higher intra-session contribution redistribution observed in Appendix D.1, DTV=0.1733D _ {\mathrm{TV}}=0.1733 vs. 0.00540.0054).
  • Empirical Comparison of Centered vs. Nearest vs. Random Retrieval: On the streamlined 5-task sequence (Caltech101 → CIFAR100 → DTD → EuroSAT → MNIST), our controlled ablation demonstrates that Centered (59.51% / 78.94% / 92.90%) consistently outperforms both positive-only Nearest (59.26% / 78.78% / 92.68%) and Random (59.49% / 78.87% / 92.56%) across Transfer, Average, and Last. For the complete comparison table and dynamics analysis, please refer to our Response to Reviewer Ttpt, Section 3.
AuthorsOpenReview ↗

3-1Response to Reviewer ELGv

3. Inference Routing Formulations, Selection Protocol, and Complete Hyperparameter Configurations (W2, Q1)

3.1 Strict Leak-Free Six-Stage Hyperparameter Selection Protocol

To strictly prevent any test-set tuning leakage, we adhere to a rigorous six-stage protocol: (1) Split (partition the official training set of currently accessible tasks into a Train split and a Val split, saving sample indices) →\to (2) Model Training (optimize LoRA weights and collaboration coefficients wkiw _ {k}^i exclusively on the Train split) →\to (3) Configuration Selection (select fusion temperature, candidate count, and routing thresholds solely on the Val split) →\to (4) Configuration Freezing (globally freeze the chosen configuration without ever referencing test results) →\to (5) Final Evaluation (compute the accuracy matrix once on the official test set) →\to (6) Multi-Seed Verification (vary training random seeds under the frozen configuration to report mean and standard deviation). For the detailed protocol table and sensitivity sweeps over candidate count and fusion temperature, please see our Response to Reviewer xFdH, Sections 2.2 and 2.3.

Comments on the paper

3 comments
AuthorsOpenReview ↗

2-5General Response to the Area Chair and Reviewers

2. Overview of New Experiments and Key Responses to Reviewers

  1. Clarifying the Scope of SVD Trajectory Consolidation, Reporting Runtime/Memory Overhead, and Verifying Lightweight Scaling Solutions (Addressing Reviewers ELGv, Ttpt, xFdH)
    We clarify that truncated SVD performs intra-session consolidation of the multi-expert evolutionary trajectory rather than collapsing the cross-session expert bank (eliminating any recursive calls to parent experts at test time; Appendix B), discuss semantic evolution stability across sessions as well as the design scope and limitations of our class-level cross-modal prototypes (Appendix D.6), report measured runtime and peak memory across training (including PyTorch allocator peak VRAM) and inference stages (Appendix E.1), and validate two practical lightweight scaling avenues—low-rank storage compression (reducing total adapter footprint by 75% while still outperforming prior baselines) and fixed first-3 expert constant-time unseen-task inference (less than 0.1% metric difference, requiring at most 4 forward passes with no pre-screening) (see Response to Reviewer ELGv, Section 4; Response to Reviewer Ttpt, Sections 4–6; Response to Reviewer xFdH, Section 4).
AuthorsOpenReview ↗

2-1 2-4General Response to the Area Chair and Reviewers

2. Overview of New Experiments and Key Responses to Reviewers

Addressing the five core themes raised across the three reviews, we have supplemented the rebuttal and revised manuscript with systematic controlled experiments and theoretical clarifications (detailed empirical tables and equations are provided in our individual responses to each reviewer):

  1. Clarifying the Fundamental Boundary with Recent Dual-Branch Methods and Adding Controlled Comparisons (Addressing Reviewers ELGv, Ttpt)
    We clearly delineate the mechanistic distinctions between DualMem-Evo and CL-LoRA (CVPR 2025) along two axes: "dynamic co-evolution of frozen experts followed by SVD trajectory absorption vs. static branch isolation" and "bimodal MLP dual-memory topology with conditional marginal CKA decoupling and single-path stability distillation." Furthermore, we adapted CL-LoRA to the exact same CLIP ViT-B/16 backbone, parameter budget, and MTIL protocol on the 5-session Order-I sequence for a controlled comparison, confirming the substantial performance advantage of our dynamic memory evolution mechanism (Appendix C.3; see Response to Reviewer ELGv, Section 1; Response to Reviewer Ttpt, Section 1).

  2. Offline Empirical Oracle Reference, Multi-Seed Paired Significance, and Controlled Disentanglement Under Unified Inference (Addressing Reviewers ELGv, Ttpt, xFdH)
    By constructing an offline empirical oracle composite reference (combining best observed pre-learning transfer scores and independent single-task Oracle LoRA scores as an empirical reference scale rather than a strict mathematical upper bound), we contextualize the saturated MTIL Order-I regime: DualMem-Evo matches the strongest baseline DPW on Transfer while closing roughly 27%–64% of its remaining Average/Last headroom relative to this reference, and achieves even larger gains on CIL benchmarks (Appendix C.1 and C.3). We also added multi-seed evaluations across 3 independent random seeds (including a direct multi-seed comparison against DPW and paired PGEEM deltas) and a rank-matched controlled comparison against Episodic Memory Only (r=128) under strictly identical inference fusion rules, confirming both cross-seed stability and the independent net contribution of our training mechanism (see Response to Reviewer xFdH, Sections 1 and 3).

  3. Articulating the Motivation of Centered Complementary Knowledge and Adding Process-Level Tracking & Retrieval Comparisons (Addressing Reviewers Ttpt, ELGv, xFdH)
    Grounded in our early development observation that unconstrained layer-wise collaboration weights consistently evolved into both positive and negative values across the analyzed heterogeneous cross-domain sessions and adapted layers, we clarify that our Centered bidirectional initialization is designed to align the initial state with this signed structure—enabling positive feature reuse alongside an additional degree of subtractive freedom relative to non-negative priors—to facilitate new-memory formation. We added dense step-by-step tracking during new-memory formation (Appendix D.5), showing task-dependent convergence acceleration from bidirectional sign flexibility (Signed vs. single-sign weighting, clearest on DTD and smaller on CIFAR100), alongside a 5-session comparison of coupled retrieval-and-initialization strategies (Centered vs. Nearest and Random, Appendix D.4), where Centered achieves the highest single-run point estimates across Transfer, Average, and Last (see Response to Reviewer Ttpt, Section 3).

  4. Providing Complete Inference Routing Equations, a Staged Selection Protocol, and Hyperparameter Robustness Analyses (Addressing Reviewers ELGv, xFdH)
    We provide the complete mathematical formulations for all three evaluation protocols (Standard MTIL, IFMS-MTIL, and CIL) and the complete hyperparameter configurations across all experimental benchmarks (Appendix A.3). We also explicitly document our six-stage development selection and cross-protocol transfer protocol—"(1) Partition the primary full-shot Order-I official training set into Train/Val splits →\to (2) Optimize model parameters on the Train split →\to (3) Select shared global defaults on the Order-I Val split →\to (4) Globally freeze shared defaults with zero access to test results →\to (5) Single-pass evaluation on the official test sets →\to (6) Multi-seed verification"—clarifying that Order-II, 5-shot/16-shot MTIL-FS, and IFMS-MTIL skip Stages 1–3 (Train/Val re-splitting and re-tuning) and inherit those shared defaults under their protocol-specific schedules (keeping all 5/16 shots for training with no extra labels accessed for few-shot tuning), while sensitivity sweeps confirm flat plateaus around candidate counts and fusion temperatures (Appendix E.2; see Response to Reviewer ELGv, Section 3; Response to Reviewer xFdH, Section 2).

AuthorsOpenReview ↗

1-General Response to the Area Chair and Reviewers

General Response to the Area Chair and Reviewers

We sincerely thank all three reviewers (ELGv, Ttpt, and xFdH) for their dedicated time and constructive feedback. We are deeply encouraged that the reviewers unanimously recognized the core strengths of our work: (1) under the strict and realistic constraints of zero historical sample replay (replay-free) and zero external auxiliary datasets, our framework achieves consistent state-of-the-art performance across four distinct continual learning protocols—standard Multi-Domain Task-Incremental Learning (MTIL Order-I/II), Few-Shot continual adaptation (5-shot/16-shot), Identification-Free Module Selection (IFMS-MTIL), and Class-Incremental Learning over a unified label space (CIL) (CS1); and (2) under fair, controlled comparisons with comparable overall trainable ranks, our ablations systematically validate the effectiveness of the dual-memory architecture and each regularization objective, outperforming strong full-model and recent adapter baselines (e.g., MoE-Adapters and DPW) with fewer trainable parameters (CS2).


1. Summary of Core Contributions and Technical Paradigm

In replay-free continual adaptation of vision-language models (CLIP), existing parameter-tuning and expert-based methods face a fundamental dilemma: compressing all tasks into a single shared module inevitably induces cross-modal catastrophic forgetting and zero-shot degradation, whereas freezing and sealing each learned task into an isolated static parameter snapshot not only severs the division of labor between shared semantics and task-specific specialization, but also reduces valuable historical episodic knowledge to a passive bystander during subsequent task learning.

Inspired by complementary memory systems, we propose DualMem-Evo (Dual-Memory Evolution), shifting the continual adaptation paradigm from static isolated snapshots to memory evolution and absorption, where frozen historical experts dynamically participate in forming and consolidating new memories:

  1. Bimodal Parallel Dual-Memory Architecture: Across the MLP blocks of both the visual and textual encoders in frozen CLIP, we deploy in parallel a continuously evolving shared semantic memory GtG _ {t} (for slow cross-task semantic integration) and a task-specific episodic memory TtT _ {t} (for rapid current-task specialization), enabling cross-modal continual adaptation without disrupting pretrained attention routing.
  2. Conditional Marginal Contribution Decoupling and Single-Path Local Stability Rule: By minimizing the CKA similarity between the conditional marginal contributions of the semantic and episodic branches (Ldec\mathcal{L} _ {\mathrm{dec}}), we encourage complementary functional specialization between the two pathways; simultaneously, we apply intra-session stability distillation (Lstability\mathcal{L} _ {\mathrm{stability}}) exclusively along the shared semantic pathway GtG _ {t}, preserving global cross-task semantic geometry while fully unlocking the task-specific plasticity of the episodic branch TtT _ {t}.
  3. Prototype-Guided Episodic Memory Evolution and SVD Absorption (PGEEM): While historical episodic expert parameters remain strictly frozen to prevent forgetting, relevant historical experts are dynamically awakened via fine-grained class-level cross-modal prototype retrieval and mean-centered signed initialization (Centered). Through lightweight learnable layer-wise collaboration coefficients wkiw _ {k}^i, these frozen experts actively participate in current-task residual forward propagation and gradient optimization; at the end of the session, the multi-expert joint evolutionary trajectory is consolidated via truncated SVD into a standalone, fixed-rank new episodic expert Et\mathcal{E} _ {t}, forming an evolving memory loop.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.