acceptodds
Under review as a conference paper at ICLR 2027

Socialized Continual Learning: Recursive Weight Space Editing for Scalable Multi-Expert Collaboration

Abstract

Intelligent systems increasingly rely on collaboration among specialized agents to exchange complementary expertise. However, existing collaborative paradigms such as Federated Learning (FL) and Socialized Learning (SL) typically assume one-shot or fixed-stage interaction, limiting their ability to adapt to open-ended environments where new experts and capabilities emerge continuously. To address this gap, we introduce Socialized Continual Learning (SCL), a continual collaborative learning paradigm that enables previously socialized experts to interact with newly emerging experts while retaining their specialized capabilities. The central challenge is to reconcile newly acquired expertise with knowledge accumulated through previous interactions. To this end, we propose Recursive Weight Space Editing (RWSE), which formulates SCL as a constrained functional alignment problem and recursively integrates new expertise into an evolving shared model. RWSE derives recursive sufficient statistics and an exact closed-form residual editing rule, enabling efficient knowledge integration without revisiting previously incorporated experts. Furthermore, a nonlinear spectrally anisotropic regularization mechanism imposes stronger protection on historically sensitive directions while preserving flexibility for new expertise, thereby mitigating inter-expert interference and maintaining prior specialization. Extensive experiments across diverse continual collaborative learning settings demonstrate the effectiveness of RWSE for continual expert integration.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Socialized Continual Learning, Multi-Expert Collaboration, Weight-Space Editing

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_Q1EfOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces Socialized Continual Learning, in which previously socialized experts integrate newly arriving expertise while retaining specialization. Recursive Weight-Space Editing (RWSE) expresses integration for an editable linear mapping as a quadratic functional-alignment problem. It accumulates historical feature Gram matrices and derives a closed-form residual update, augmented by a nonlinear spectral penalty that protects directions with large historical eigenvalues. Experiments on CIFAR-100 and ImageNet-100 compare classification performance under 5- and 10-expert partitions, including a synthetic-feature setting and a regularization ablation.

Critical strengths
  1. The quadratic formulation makes the trade-off between fitting incoming expert outputs and preserving the previous mapping explicit. Compressing historical feature information into a Gram matrix is a potentially useful way to avoid storing all prior feature batches.

  2. The reported accuracy improvements span both datasets and both expert partitions, and the synthetic-feature experiments examine a relevant limitation of relying on original examples during collaboration.

Critical weaknesses
  1. There is a substantial gap between the multi-expert problem being claimed and the algorithm that is actually specified. Section 3 describes a collection of separately specialized experts that should all expand their capabilities, whereas Section 4 specifies one evolving shared linear mapping W. It is not clear whether a separate RWSE trajectory is maintained for every expert, how those trajectories differ, or how the updated collection realizes directionality for all experts. More fundamentally, Eq. (2) defines V_t = K_t W_t for the incoming expert, and the update then evaluates the same K_t through the previous shared mapping. This requires compatible input coordinates and output semantics at the edited layer. The paper does not adequately specify where these correspondences are extracted in independently trained ResNet-18 experts, how feature-coordinate mismatch is handled, how new class outputs are aligned, or which layers are edited. 
  2. The evaluation does not reliably establish the claimed specialization, retention, and scalability benefits. First, the stated BWT formula is the mean of final minus initial task accuracy, for which larger values indicate better retention. However, Figure 4 labels BWT as lower-is-better, displays positive values, and the text treats the lowest BWT as best. This could be an omitted sign reversal, but the actual metric must be resolved before accepting the retention conclusion. Second, Table 2's Avg values appear to average expert and general accuracy equally: for RWSE on CIFAR-100, (69.06 + 65.43)/2 = 67.245, matching 67.25. If those columns cover 10 expert classes and all 90 remaining classes as the 10/10 description suggests, the class-balanced overall accuracy would instead be 65.793. If General is a single equally sized group or an intermediate stage, that protocol needs to be specified. The table therefore does not currently provide an unambiguous measure of the all-class versatility claim. These combined methodological and reporting gaps materially weaken the conclusions rather than amounting only to presentation issues.
Questions for authors
  1. Please specify the complete ResNet-18 integration procedure: edited layers, sources of K_t and V_t, cross-expert coordinate and class alignment, handling of upstream feature changes, and the ordering of SGD and closed-form edits. Does every original expert retain a distinct updated model?

  2. What quantity is actually plotted on the BWT axis? Please provide the signed per-task accuracy matrix and define the exact class groups, collaboration stage, and weighting used in Tables 1 and 2.

  3. Can you reconcile the ablation arithmetic and report matched isotropic, linear-spectral, and nonlinear-spectral penalties with trial variability?

Reviewer_pbkqOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes Socialized continual learning, which combines federated learning, socialized learning, and continual learning. The paper claims the central challenge of this setting is to reconcile all acquired knowledge. I find the setting sensible, but the experiments section is basically continual learning with a model-merging setting. Overall, the paper presents limited novelty.

Critical strengths
  1. The idea of combining federated learning, socialized learning, and continual learning is interesting.
  2. The paper reports consistent performance improvement.
Critical weaknesses

1.The paper motivates socialized continual learning by combining federated learning and socialized learning, as in Fig 1, and Section 3 defines different experts E1, E2, and so on. However, the method only maintains one shared Wt, which aggregates all experts. The main motivation of peer-to-peer exchange, as in Fig 1 never happens, and it becomes one merged expert. 2.The main technical contribution is updating the Gram matrix in one batch. It is very similar to sequential merging methods. Thus, I find the overall novelty weak. 3.Overall, the method merges different experts with closed-form updates; however, it's not clear what experts mean in this context. Are experts the whole backbone model or some layers? Also, it’s not clear what is being trained. 4.There are many formatting issues. Specifically in Section 3 lines “Let E = {E1,E2, . . . , t} denotes” and line numbers 192-195 where Versatility and Directionality start the sentence but seems like author intended to use dot?

Questions for authors

See weaknesses

1 comment

I have checked the discussion and found no response from the author. So my concerns remain unaddressed. I keep my original evaluation and rating.

Reviewer_AEraOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

The paper extends socialized learning to a continual learning setup with new experts joining over time. The continual learning dynamics emphasize on versatility (incorporation of others' knowledge) and directionality (superior performance on the personal task).

Critical strengths
  • The paper is easy to follow.
  • The methodology is reasonable.
  • The experiments are extensive.
Critical weaknesses

Significance

My major concern of the paper is its significance. The proposed quadratic objective is very straightforward. Incorporating eigendirections to preserve historic knowledge is pretty standard and direct. And this quadratic objective is the only methodological contribution of the work. Additionally, this method has many limitations:

  • It only applies to linearly editable models.
  • It's formulation, especially anisotropic regularization term, is very heuristic, without any theoretical explanation or motivation.
  • It does not have any theoretical guarantee. Especially for the directionality, the method does not seem to guarantee this proposed metric.
  • Its hyperparameters such as the regularization strength is hard to design.

Some of these limitations are old and some are new, but I don't see much effort devoted to address them. It's also unclear to me what is the motivation is care about directionality.

Presentation

The presentation of the paper could also be greatly improved. Some examples:

  • Lines 175 and 192, typos in EE.
  • Line 192, "Versatility" and "Directionality" should have a period after them.
  • Line 205, if the arrival has an order, better to use other notation than curly brackets.
  • The presentation of Eq. (5) is unclear.
Questions for authors
  • What is "manifold-aware" in the contributions?
  • Is the method maintaining a single shared model? Then how does it provide personalized directionality for each experts? And does it need a central server to maintain this shared model?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.