acceptodds
Withdrawn from ICLR 2027

Continual VLA Model Merging with Historical Subspace Constraints

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Vision-language-action (VLA) models must continually acquire new manipulation skills while preserving previously learned capabilities. Maintaining separate policies for individual skills incurs growing deployment overhead. Model merging provides a natural mechanism for consolidating skill-specific adaptations into a shared model. However, repeated adaptation and consolidation in rehearsal-free settings introduce a key challenge: current-stage supervision provides no direct signal about how incoming updates affect previously acquired skills, allowing interference to propagate and potentially accumulate across learning stages. We study rehearsal-free continual VLA model merging and propose history-aware subspace-constrained adaptation. Our key idea is to retain compact activation subspaces that summarize dominant directions in historical layer-input activations and use them as historical constraints during subsequent adaptation. The method constrains incoming backbone adaptation at two complementary levels: a soft orthogonal gradient constraint attenuates gradient components aligned with a cumulative historical subspace, while Subspace-L2 regularization penalizes the response induced by the effective backbone update along retained stage-specific historical input subspaces. After adaptation, the learned update is consolidated into the shared vision–language backbone, stage-specific action modules are retained, and historical subspaces are updated from stage-end rollouts, without retaining or replaying previous demonstrations. Experiments on within-suite and cross-suite LIBERO task streams show that our method achieves the highest final average success rate among the evaluated rehearsal-free approaches, reaching on the ten-task LIBERO-Goal stream and on the SpatialObjectGoal stream.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Vision-Language-Action Models, Continual Model Merging, Rehearsal-Free Learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_9ebqOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies rehearsal-free continual VLA model merging, where new manipulation skills are sequentially adapted and consolidated into a shared vision-language backbone without replaying historical demonstrations. The proposed method retains historical activation subspaces and uses them to constrain subsequent LoRA adaptation through a soft orthogonal gradient constraint and a Subspace-L2 regularizer, while stage-specific action modules are preserved. The method achieves the best reported rehearsal-free performance among the compared methods.

Critical strengths
  • Continual consolidation of VLA skills without retaining historical demonstrations is practical.
  • The proposed method substantially improves both final performance and trajectory-level retention.
  • Detailed storage accounting is reported rather than relying only on the final performance.
Critical weaknesses
  • Rehearsal-free continual learning, LoRA-based continual adaptation, and orthogonal-gradient protection are all well-established ideas. The main novelty lies in adapting and combining these mechanisms for continual VLA model merging rather than introducing a fundamentally new continual-learning principle.

  • Stage identity is required to select the corresponding stage-specific action module, while these modules are retained separately throughout continual learning. This limits realism for deployment where the relevant historical stage may be unknown.

  • The paper provides storage accounting but does not report training time, GPU hours, FLOPs, or the additional runtime.

  • The paper analyzes the merge coefficient, but does not systematically study important parameters such as the historical rank and the retained-energy threshold.

  • The cross-suite evaluation covers only three of the standard LIBERO suites (Spatial→Object→Goal), excluding LIBERO-Long. It is therefore unclear whether the observed retention advantage generalizes to longer-horizon and compositionally more complex tasks.

Questions for authors

See above.

Reviewer_y963OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies rehearsal-free continual adaptation and consolidation of modular vision–language–action policies. The method retains compact activation subspaces that summarize dominant directions in historical layer-input activations and use them as historical constraints during subsequent adaptation. Experiments on a ten-task LIBERO-Goal stream and a Spatial→Object→Goal stream report improvements over the evaluated sequential baselines.

Critical strengths

The paper propose historical activation subspaces which provide a reasonable mechanism for constraining this interference without replaying demonstrations. The result is encouraging.

Critical weaknesses
  1. Some result are confusing and required for more explanation. E.g, in Table 12, the number of evaluations is not quite consistent with the success rate of the report. Goal-10 (even including the caption) clearly states that if each task is evaluated 20 times, the single success rate would only be a multiple of 5%. However, it appears to be a multiple of 2%, which can match the 2% accuracy of 50 rollouts. In contrast, the task-by-task matrices of other methods are all integer multiples of 5%. Here, the author needs to explain. Moreover, there is an inconsistency between the stored equation and the reported values. According to Eq.21, Ssep = 3* (2389.18+306.90) = 8088.24, but table15 reports it as 9461.45, which requires the author to explain.

  2. The related work has not fully discussed the existing continuous learning methods based on historical activation subspaces, such as GPM and Scaled Gradient Projection. Please clarify the specific differences of this paper in terms of LoRA-A gradient constraints, cumulative memory construction, and effective-update regularization, and explain what contributions this paper has made compared to directly applying the existing subspace projection methods to this VLA architecture. And the comparison experimental results are encouraging.

Questions for authors
  1. The cross-suite experiment lacks complete component ablation. Currently, ablation has only been conducted on Goal-10. Personally, I think it's worth doing. This can show whether the method is effective across scenarios (and the experiment lacks fixed-merge (included in the ablation as a baseline)).

  2. Fig1 is too blurry and looks odd. It is recommended to change it to a vector graphic.

Reviewer_VQa9OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies rehearsal-free continual VLA model merging, where a policy sequentially learns new manipulation tasks without access to previous demonstrations and consolidates each adaptation into a shared model. The policy consists of a shared vision–language backbone and stage-specific action modules: at each stage, LoRA adapters and a new action module are trained on current-stage data, after which the LoRA update is merged into the backbone with a fixed coefficient and the action module is frozen and retained. To reduce interference, the method stores compact activation subspaces from stage-end rollouts and uses them to constrain subsequent adaptation at two levels: a soft orthogonal gradient constraint that attenuates LoRA gradients aligned with a cumulative historical subspace, and a Subspace-L2 regularizer that penalizes the update’s response along stage-specific historical subspaces. As I understand it, a new action module is added for each stage and selected by the stage identity at inference, so the setting is task-incremental. The method is evaluated on a ten-task LIBERO-Goal stream and a three-stage Spatial→Object→Goal stream, achieving final average success rates of 45.5% and 87.6%, respectively, the highest among the evaluated rehearsal-free baselines.

Critical strengths

Important and timely problem. As VLA policies are increasingly expected to acquire new manipulation skills after deployment, consolidating these skills into a shared model while controlling the interference that accumulates over repeated adaptation and merging is an important and underexplored problem. Recent work such as MergeVLA and RETAIN has established the feasibility of merging VLA policies, and this paper takes a step toward explicitly addressing interference during sequential consolidation.

Critical weaknesses

W1. The practical benefit of the proposed consolidation setting is insufficiently justified.
The paper assumes that historical demonstrations are unavailable but provides limited discussion of the concrete data-access and deployment constraints motivating this assumption. Moreover, the setting retains a separate action module for each stage and uses stage identity for module selection, making task-isolated LoRA deployment a particularly relevant alternative that also requires no historical replay. Since each action module (306.9 MiB) is already larger than a per-task LoRA adapter (198.2 MiB on average), retaining task-specific adapters adds relatively little storage. The paper’s own comparison shows that consolidation reduces deployment storage by 15.2% relative to task-isolated LoRA, while lowering cross-suite final average success from 97.2% to 87.6% (Table 17). Action modules and stage-specific subspace memories also accumulate as the stream grows. In the reported implementation, dense projector caching adds 8,027.35 MiB of training state; although this cache is unnecessary for inference and potentially avoidable, it complicates the efficiency argument for on-device continual adaptation. These trade-offs leave unclear under what practical constraints the proposed consolidation approach would be preferable to retaining task-specific adapters.

W2. The methodological novelty and empirical advantages over prior methods are insufficiently established.
Historical activation-subspace protection is closely related to GPM, while constrained low-rank adaptation is central to methods such as InfLoRA and O-LoRA. The paper neither discusses these connections nor establishes what its combination adds beyond them. The empirical comparisons also contain architectural differences and unexplained failures: RETAIN shares and interpolates action modules, whereas the proposed method retains stage-specific modules; OPCM obtains 0.0% on Object immediately after Object-stage adaptation and consolidation, without separating pre-merge acquisition from post-merge degradation. Additionally, OPCM’s task-level results in Table 12 are inconsistent with the stated 20-rollout binary-success evaluation, since several entries are not multiples of 5%. The cross-suite evaluation lacks a Fixed-Merge comparison, and the merge coefficient is selected based on reported cross-suite performance without a clearly specified independent validation protocol. Results are reported from single training runs, with 20 evaluation rollouts per task for Goal-10 and 50 for cross-suite experiments. Although the Goal-10 ablation supports a benefit from combining the constraints, substantial forgetting remains (BWT −50.0pp, with four tasks ending at 0–5% success). Consequently, the contribution beyond existing subspace methods and the source and robustness of the large cross-suite gains remain unclear.

Questions for authors

Q1. What concrete data-access and deployment constraints motivate the rehearsal-free consolidation setting? For example, do privacy or data-retention restrictions prevent access to historical robot demonstrations? Given the large gap to joint training (96.5% vs. 45.5% on Goal-10), what makes retaining demonstrations and retraining jointly infeasible or impractical in the intended scenario? Even when historical demonstrations are unavailable, when would consolidation be preferable to retaining task-specific LoRA adapters, given the 15.2% storage reduction and 9.6pp success-rate decrease reported in Table 17?

Q2. How well do the retained subspaces actually capture historical activations? The rank cap of 16 is small relative to the layer input dimensions, and the energy target η is not reported. What fraction of activation energy is retained per layer, i.e., how small is the residual r in Eq. 4 in practice?

Q3. How do architectural and implementation differences affect the baseline comparisons? In particular, can RETAIN be evaluated with the same stage-specific action-module retention as the proposed method? For OPCM, what are the Object success rates before and after consolidation, and what explains the post-consolidation score of 0.0%? Also, how were the entries in Table 12 computed, given that several are not multiples of 5% despite the stated 20-rollout evaluation?

Q4. Could the authors report Fixed-Merge on the cross-suite stream and results across multiple training seeds to isolate the contribution and assess the robustness of the proposed constraints? How was the merge coefficient selected, and was a separate validation protocol used?

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

Related papers

Discussion (0)

Sign in to comment.