acceptodds
Under review as a conference paper at ICLR 2027

MorphAct: Contextual Parameterization of Vision–Language–Action Models

16 views

Abstract

Vision–language–action (VLA) models seek to support diverse robot behaviors within a shared policy. Conventional fine-tuning, however, fits a single static update to operations whose perceptual and control demands change even within one task. When these operations favor competing parameter adjustments, fitting their mixture can leave individual operations insufficiently adapted. Our atomic-task diagnostics reveal competing update directions in both vision–language and action parameters, with source-improving updates increasing loss on other tasks. This motivates MorphAct, a contextual parameterization framework that shares an adaptation rule rather than a fixed update, allowing adaptation to specialize without splitting the policy into operation-specific experts. Two coupled generators produce context-dependent low-rank updates for the vision–language and action modules. Their coupling follows the policy's computational dependency: vision–language updates shape the representations that condition action-side parameter generation. Both generators are trained jointly through the native action objective, without operation labels or expert-adapter targets. MorphAct improves average success rates across three VLA backbones on LIBERO and Meta-World, with gains of up to 15.13%. In real-world evaluation, it achieves gains of up to 17% over π₀. Ablations support the benefits of contextual parameter generation and representation coupling.

Primary area:
foundation or frontier models, including LLMs
Keywords:
vision-language-action models, parameter generation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_zFvrOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes MorphAct, a contextual parameterization framework that shares an adaptation rule rather than a fixed update, allowing adaptation to specialize without splitting the policy into operation-specific experts. It improves average success rates across three VLA backbones on LIBERO, Meta-World, and a real-world robot.

Critical strengths
  1. The motivation is reasonable and supported by experiments (Fig 2).
  2. The method design is simple yet effective. Then VLM side generates based on the context, and then the action side generates based on the updated VLM parameters. And this is a dynamic update process across the episode, which aligns the problem they found in robocasa365 atomic tasks.
  3. The experiment results show its advantages on three different VLA backbones and two benchmarks, especially on MetaWorld.
Critical weaknesses
  1. The paper mentioned other VLA parameter-generation methods like Hyper-LoRA and PhaseLoRA. But there lacks a direct benchmarking comparison between MorphAct and these method.
  2. Some important figures seem AI-generated (Fig 3), the readability is not good. The font and the too much text comment in the figure affect the understanding of the whole architecture.
Questions for authors
  1. Can you provide some benchmarking comparison between MorphAct and Hyper-LoRA/PhaseLoRA, or provide the reasons if you can not.
3 comments
AuthorsOpenReview ↗

Rebuttal to Reviewer zFvr

Weakness 1 & Question 1 — Comparison with Hyper-LoRA and PhaseLoRA

The paper mentioned other VLA parameter-generation methods like Hyper-LoRA and PhaseLoRA. But there lacks a direct benchmarking comparison between MorphAct and these method.

Can you provide some benchmarking comparison between MorphAct and Hyper-LoRA/PhaseLoRA, or provide the reasons if you can not.

Thank you for this suggestion. We ran PhaseLoRA [1] on the same backbone and benchmark as MorphAct (X-VLA, LIBERO, all four suites). We also trained a static LoRA whose total trainable-parameter budget is comparable to MorphAct's, under the same data and schedule, to check whether the gain comes from extra parameters:

X-VLA, LIBEROSuccess (%)
Static LoRA, comparable trainable-parameter budget94.00
PhaseLoRA [1]94.70
MorphAct, budget-matched run96.50
MorphAct, Table 396.60

MorphAct exceeds PhaseLoRA by 1.90 points and the budget-matched static LoRA by 2.50 points. PhaseLoRA modulates action-side LoRA per action chunk. MorphAct differs in two respects: it also generates VLM-side updates, and its action-side generator reads the adapted VLM representation. Table 3 tests each of these: without VLM-side generation (Action-only) success is 81.00%, and when the action generator reads unadapted VLM features (Unadapted) it is 86.00%, against 96.60% for MorphAct.

We were not able to complete a reproduction of Hyper-LoRA [2] on X-VLA within the discussion period, so we do not report a number for it. Its central design choice is, however, tested in Table 3. Hyper-LoRA generates the LoRA update from the language instruction, so the update depends only on the task and stays fixed throughout an episode. The Task-ID variant uses this kind of condition inside MorphAct's own architecture and training: one fixed generation condition per task, while the policy still receives the current image, instruction, and robot state. Each LIBERO task has its own instruction, so task identity carries the same task information as the instruction. Task-ID reaches 88.25%, compared with 96.60% for MorphAct (Section 5.1). This comparison isolates task-level from context-level conditioning with identical generators and training; it does not reproduce Hyper-LoRA's generator architecture or training recipe, as Section 5.1 of the revision now states.

The PhaseLoRA and static-LoRA comparisons are in Section 5.1 and Appendix D.6 of the revision, and Section 5.1 now also states the relation between Hyper-LoRA and the Task-ID variant.

Weakness 2 — Readability of Fig. 3

Some important figures seem AI-generated (Fig 3), the readability is not good. The font and the too much text comment in the figure affect the understanding of the whole architecture.

Thank you for this comment. We agree that the original Fig. 3 contained too much text at too small a font size, which made the overall architecture harder to follow. We have redrawn it in the revision:

  • All labels, module names, and symbols are set in a larger font that is readable at print size.
  • Long in-figure notes are removed, such as “no feature detachment”, “no operation labels”, “separate generator weights”, and “no test-time optimization”, so the figure keeps only module names, the symbols used in the equations, and short labels.
  • Panel and box titles are shortened, for example “(a) Joint learning through adapted features” and “(b) Contextual generation and chunk-level reuse”, while the layout keeps solid arrows for data, dashed arrows for parameter updates, and one color per module.

Thank you again for your kind support. We hope these results address your concerns. If you have further questions, we would be glad to answer them, and we would be very grateful if you would consider raising your score.

References

[1] PhaseLoRA: Control-Regime-Conditioned Low-Rank Adaptation for Continuous-Action Vision-Language-Action Policies.

[2] Efficient Domain Adaptation of Robotic Foundation Models via Hypernetwork-Generated LoRA. NeurIPS 2024 Workshop.

Thanks for your rebuttal. I will keep my recommendation as a weak acceptance.

Thank you for your careful consideration of our submission and rebuttal. We appreciate your time and constructive feedback.

Reviewer_DFMJOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper argues that common VLA tasks usually consist of a mixture of multiple subtasks and atomic actions within one trajectory. The paper suggests that sharing parameter updates among these tasks could lead to competing updates from different actions. The authors diagnose this on RoboCasa atomic tasks by showing that cross-task gradients are more likely to conflict than those within. Based on this argument, the paper proposes MorphAct, which uses a hypernetwork-style generator to generate per-chunk LoRA weights, conditioning on observations, enabling context-dependent parameters without operation labels.

Critical strengths

Strengths

  1. The problem is well motivated. As VLA develops, more tasks focus on high-level or compound instructions and trajectories, and studying these mixtures of actions through weights is an interesting direction.
  2. Evaluation is broad. The paper evaluates on three backbones and three different benchmarks (LIBERO, Meta-World, real-world), with consistent gains, including large improvements on Meta-World.
  3. The paper reports computational cost trade-offs clearly (about 18–21% added latency and about 1.7 GiB added memory).
  4. Conditioning action-side generation on the adapted VLM features is a reasonable design, with gradients shown to be routed into generator training (Eq. 4)
Critical weaknesses
  1. The diagnostic experiment does not show strong evidence. The mean cosine is at 1.05×10−31.05 \times 10^{-3} for cross task gradient alignments, which could be a noise-level sign change. Since the gradient is calculated from sampled batches, it could also be introduced by per-sample perturbation and randomness, as commonly seen in SGD-based training, no confidence interval, repeated experiments on different seeds, or random-gradient reference are reported.
  2. The diagnosis and experiments do not well support the motivation claim. The motivation primarily focuses on the potential mixture of actions and conflict within a trajectory or high-level instruction, but the diagnosis focuses on separate trajectories with different atomic tasks. Whether a high-level single trajectory contains conflicting signals within is not well shown. The link between them is relatively weak.
  3. The success-rate comparisons (Tables 1 and 2) are against originally reported baseline numbers. The only shared-LoRA comparison is in the cost table (Table A8), where active factors are matched (about 50M) but MorphAct adds about 300M trainable parameters and about 20% latency. The improvements may therefore come from capacity or extra training effort rather than the proposed method.
  4. The motivation is loosely connected to the proposed method. While the proposed method can generate per-chunk state-dependent weights. It is not shown whether the generated weights truly resolve or mitigate the conflicts within the trajectory, for example, it is unknown whether generated weights actually vary by much across time within the same trajectory. The Section 2 diagnostics are computed on full-parameter gradients of a RoboCasa-pretrained π0.5\pi0.5 , whereas MorphAct trains only low-rank generators on a frozen backbone and is evaluated on different models and benchmarks. It is therefore unclear whether the measured conflict holds or reduces after training MorphAct.
  5. Key training and architecture hyperparameters are not shown (like LoRA rank, learning rate, etc.) for better reproducibility.
  6. Ablations are only on X-VLA/LIBERO. Constant (87.50) and Task-ID (88.25) are lower than the unmodified X-VLA baseline (93.05), and VLM-only collapses (43.50), which suggests the generator pathway is hard to optimize and makes the source of the gains unclear.
Questions for authors
  1. Do the results hold on more capacity-comparable setups, like LoRA of similar number of trainable parameters, or MoE LoRAs with same number of trainable and active parameters?
  2. Does the proposed method produce truly state-dependent LoRA that differs across atomic tasks within the same trajectory?
  3. Does the gradient conflict persist when gradients are measured using LoRA parameters, and is it reduced in a trained MorphAct compared with a shared LoRA?
7 comments
AuthorsOpenReview ↗

Rebuttal to Reviewer DFMJ (Part 1)

First, thank you for your careful review and constructive feedback. Below are our point-by-point responses to your concerns and questions.

Weakness 1 — Statistical reliability of the gradient diagnostic: whether a mean cross-task cosine of 1.05×10⁻³ reflects systematic conflict rather than sampling noise.

Thank you for the interesting comment on the diagnostic statistics. A mean cosine near zero is what our claim predicts. If different atomic operations needed compatible updates, their gradients would align positively. A mean close to zero says that a step taken for one operation brings, on average, no benefit to another. The question is whether the negative cases are systematic or noise. They also explain why the mean is small: within each cross-task comparison, positive and negative parameter-group contributions coexist in 99.45% (VLM) and 95.93% (action) of the pairs and cancel 36.13% and 30.90% of the absolute contribution (Table A4).

The diagnostic is computed on one fixed pretrained checkpoint, so the relevant randomness is batch sampling. Appendix A.1 repeats the measurement over 16 independent batch-resampling rounds. We now add confidence intervals and a reference in which task labels are shuffled:

StatisticTrue task labelsShuffled task labels
Mean cross-task cosine, 95% CI over 16 rounds1.05 [0.93, 1.17] ×10⁻³0.21 [0.05, 0.37] ×10⁻³
Negative-alignment rate, cross-task vs. within-taskVLM 47.80% vs. 38.94%; action 48.28% vs. 42.69%49.60% vs. 49.40%

With shuffled labels, cross-task and within-task pairs become indistinguishable, as noise would produce. With the true labels, negative alignment is 8.9 points more frequent across tasks than within tasks for the VLM and 5.6 points for the action module. After averaging over the 16 rounds, 890 of 2,080 VLM task pairs and 918 of 2,080 action task pairs remain negatively aligned (Table A2).

The one-step intervention tests the consequence directly and is not exposed to batch noise in the same way, because target observations, action labels, flow times, and noise are fixed before and after each update. Across four rounds, all 260 updates that lower the loss on their source batch raise the loss on batches from 22–46 other tasks (median 32).

The 16 rounds measure variation over batches, not over independently trained models, as Appendix A.1 states. The revision adds the confidence intervals and the shuffled-label reference to Appendix A.2.

Weakness 2 — Link between the diagnostic and the motivation: why a comparison across atomic tasks supports a motivation about conflicting demands within one trajectory.

Thank you for this interesting point. Our motivation is stated at the level of atomic operations, and conflict within a trajectory follows from it. A compound task is a sequence of atomic operations, such as opening, grasping, transporting, and placing, and one high-level trajectory passes through several of them. If an update that helps one operation hurts another, no single static update can serve the whole trajectory. The diagnostic tests this premise directly.

We use RoboCasa365 [1] because it defines 65 atomic tasks and builds its composite tasks from them, so its task IDs separate operations with boundaries given by the benchmark rather than by our own segmentation. In this comparison, every one of the 260 source-improving updates raises the loss on 22–46 other operations (median 32).

Whether the generated parameters actually follow these changes within a single trajectory, and whether that matters for execution, is the subject of Question 2; we answer it with within-episode measurements below.

Weakness 3 & Question 1 — Capacity and training effort: whether the gains come from the extra generator parameters or training effort rather than from the method.

We thank the reviewer for the interesting question on capacity. Following common practice, Tables 1 and 2 report each baseline's official results: π0 and π0.5 from OpenPI's official LIBERO fine-tuning at 30k steps [2], and X-VLA from its Ours-LoRA configuration [3]. Appendix D.3 of the revision states each source and setting. To control for capacity and training effort, we retrained the comparisons below with the same backbone, data, and schedule:

X-VLA, LIBEROSuccess (%)
Static LoRA, comparable trainable-parameter budget94.00
PhaseLoRA [4]94.70
MorphAct, budget-matched run96.50

Matching the trainable-parameter budget with a static LoRA does not close the gap, and a dynamic adapter on the same backbone does not either.

Table 3 also isolates the method at fixed capacity. Constant, Task-ID, and Unadapted have the same generators, parameter count, and training as MorphAct and differ only in what the generators read. They reach 87.50%, 88.25%, and 86.00%, against 96.60% for MorphAct, so the gain depends on the contextual input rather than on the extra parameters.

AuthorsOpenReview ↗

Rebuttal to Reviewer DFMJ (Part 2)

On the parameter counts in Table A8: the 300.09M generator parameters are a function that outputs LoRA factors once per replanning step, and the policy's forward pass uses only the 49.99M generated factor entries, the same size as a shared LoRA of the same rank. The generation step accounts for the measured 18–21% latency increase. Table A8 in the revision now reports generator parameters, active factors, and total trainable parameters side by side, and Appendix D.3 explains how they relate.

Weakness 4 & Question 2 — Within-trajectory variation and the diagnostic setting: whether the generated LoRA actually changes within one trajectory, and why the diagnostic uses a different model and benchmark from the trained method.

Thank you for the interesting question. Yes, and the change matters for execution. We traced the effective update ΔWt=(α/r)AtBt\Delta W_t=(\alpha/r)A_tB_t through a successful X-VLA rollout of the LIBERO-Long task "put both the alphabet soup and the tomato sauce in the basket", which grasps, carries, and places two objects in sequence. The instruction and scene layout are fixed for the whole episode. Across 27 consecutive replanning steps, the update changes by 24.1% on average, measured as ∥ΔWt−ΔWt−1∥F/∥ΔWt−1∥F\lVert \Delta W_t-\Delta W_{t-1}\rVert_F/\lVert \Delta W_{t-1}\rVert_F over all adapted projections (29.2% for the VLM, 23.5% for the action module).

To test whether this variation is used, we keep the same checkpoint and generate both updates once from the first observation, then hold them fixed for the rest of the episode. Images, robot state, VLM features, and action-conditioning inputs are still refreshed at every replanning step, and initial states, seeds, and execution settings are matched. On the same LIBERO cases, success falls from 96.5% to 0%: fresh observations do not compensate for frozen parameters. Since this is an inference-time intervention, its trained counterpart is the Task-ID variant, whose update is fixed within each episode by construction; it reaches 88.25% against 96.60%.

The revision adds the within-episode trace of ΔWt\Delta W_t and the episode-fixed control as Appendix E (Tables A12 and A13).

The two settings serve different purposes. Section 2 needs a benchmark with an explicit atomic-operation taxonomy, which RoboCasa365 provides. The main experiments need standard benchmarks with established baselines across several backbones, which LIBERO and Meta-World provide. Testing the method on data other than the data that motivated it also keeps the design from being tuned to the diagnostic.

Question 3 — Gradient conflict in LoRA parameters: whether the conflict persists when measured in LoRA parameters, and whether it is reduced in a trained MorphAct compared with a shared LoRA.

We appreciate this interesting question. To connect the two, we repeated the diagnostic in the parameters that are actually trained downstream. On X-VLA/LIBERO-Spatial, we take gradients of the native action loss with respect to the factors of a trained static LoRA. Across repeated task-batch draws, 41.39% of cross-task gradient pairs have negative cosine, so the conflict persists in the low-rank adaptation parameters. Using the same task batches, observations, action targets, noise, and timesteps, the corresponding rate for MorphAct's shared generator weights is 19.44%.

The two rates are computed in different parameter spaces, LoRA factors and generator weights, so we read the comparison as how often each method's shared trainable parameters receive opposing task gradients, not as a like-for-like cosine comparison. MorphAct's generator produces different updates ΔW(c)\Delta W(c) for different contexts cc, so tasks with opposing demands are less often forced to pull the same parameters in opposite directions. The revision summarizes this analysis at the end of Section 2 and gives the full protocol and results in Appendix F (Table A14).

A comparison that does not depend on the parameter space is the loss change produced by an actual update. We repeated the one-step protocol of Section 2.2 (Fig. 2(b)) in the same X-VLA/LIBERO-Spatial setting. For each source-task batch, we take one gradient step from the trained checkpoint, restarting from that checkpoint for every source, and evaluate the loss on fixed batches of the other nine tasks; observations, action targets, noise, and timesteps are held fixed before and after the step. For static LoRA the step updates the LoRA factors, and for MorphAct it updates the shared generator weights. The step size of each method is set so that both reduce the source-batch loss by the same mean relative amount (2.40%), so the two updates make equal progress on their own task. Over four rounds (40 source updates per method):

AuthorsOpenReview ↗

Rebuttal to Reviewer DFMJ (Part 3)

X-VLA, LIBERO-Spatial, one-step updateSource loss reducedOther tasks with higher loss (median, of 9)Mean relative loss change on other tasks
Static LoRA40 / 406 (range 4–8)+0.71%
MorphAct40 / 402 (range 0–4)+0.06%

With the same progress on the source task, a static-LoRA update raises the loss of a median of six of the other nine tasks, whereas a MorphAct update raises the loss of two, and the mean relative loss change on the other tasks falls from +0.71% to +0.06%. The interference found in Section 2 therefore also appears in the low-rank parameters used downstream, and MorphAct reduces it in terms of the task loss itself.

Weakness 6 — Ablations: whether the low scores of Constant, Task-ID, and VLM-only indicate that the generator pathway is hard to optimize, and whether ablations on X-VLA/LIBERO suffice.

Thank you for the interesting observation on the ablations. The low VLM-only score comes from what the variant freezes, not from the generator. Each variant is trained separately with the same initialization, data, and training budget, and a module whose generated updates are disabled is frozen with no static LoRA in its place. VLM-only therefore keeps the whole action module, which maps features and robot state to actions, at its pretrained weights. A static LoRA on the VLM alone, with the action module frozen in the same way, reaches 41.80%, close to the 43.50% of generated VLM-only updates. The 93.05% reference also adapts action-side components [3] and is not a like-for-like comparison.

Constant and Task-ID keep the generator and remove its contextual input. Without an informative input they fall below a directly optimized static LoRA (94.00%); they are included to isolate the conditioning signal within a fixed architecture, where current-context conditioning adds 9.10 and 8.35 points. The coupling comparison is the cleanest test of optimization difficulty: Unadapted and MorphAct have identical architectures, parameter counts, and training, and differ only in whether the action generator reads unadapted or adapted VLM features. The 10.60-point gap therefore reflects the coupling rather than how hard the pathway is to optimize. The full table is in our response to Reviewer Qa7R (Question 3).

These component comparisons are run on X-VLA/LIBERO, while the complete method is evaluated across three backbones and two benchmarks.

We hope these results address your concerns, and we would be grateful if you would reconsider your rating in light of this evidence. If you have further questions, we would be glad to answer them!

References

[1] RoboCasa365: A Large-Scale Simulation Framework for Training and Benchmarking Generalist Robots. ICLR 2026.

[2] Physical Intelligence. LIBERO Benchmark Results. OpenPI GitHub repository. https://github.com/Physical-Intelligence/openpi/

[3] X-VLA: Soft-Prompted Transformer as Scalable Cross-Embodiment Vision-Language-Action Model. ICLR 2026.

[4] PhaseLoRA: Control-Regime-Conditioned Low-Rank Adaptation for Continuous-Action Vision-Language-Action Policies.

Reviewer_Qa7ROpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

MorphAct adapts vision-language-action models by generating context-dependent low-rank updates for both the vision-language and action modules. The action-side generator is conditioned on the adapted vision-language representation, and both generators are trained jointly through the backbone's native action objective, without operation labels or expert-adapter targets. Updates are generated at each replanning context and reused within an action chunk. The paper motivates this design through gradient and one-step transfer diagnostics on 65 RoboCasa tasks, and evaluates it on three VLA backbones using LIBERO and Meta-World, as well as four physical manipulation tasks.

Critical strengths
  1. The coupling between representation and action adaptation is technically coherent. Conditioning action-side parameter generation on the adapted VLM representation follows the policy's computational dependency. The formulation preserves the backbone's native action objective and does not require operation annotations. Generating updates once per replanning context and reusing them within a chunk also makes the design practical for existing action-generation procedures.

  2. The closed-loop results provide useful evidence of practical effectiveness. Aggregate performance improves across three backbone configurations on both simulation benchmarks, with consistent improvements on LIBERO-Long. The physical experiments further report gains across four tasks on two platforms.

Critical weaknesses
  1. The experiments do not directly establish the motivating claim about adaptation demands across stages of the same task. The introduction and Fig.1 emphasize changing requirements within a trajectory, such as grasping, transport, and placement. However, the diagnostics compare native RoboCasa task IDs; the within-task reference compares batches from different episodes without grouping them by execution stage. The Task-ID ablation and LIBERO-Long results support contextual generation, but do not distinguish stage-dependent adaptation from responses to changing layouts, object positions, or robot states. The physical tasks primarily involve grasping, placement, and retrieval, and report only overall completion. They provide limited coverage of different interaction mechanisms and no direct analysis of stage-dependent adaptation.

  2. The comparisons do not sufficiently isolate the proposed design from additional learned capacity or related dynamic adaptation methods. For π0\pi_0 and π0.5\pi_{0.5}, Tab.A8 reports 300.09M generator parameters, while shared LoRA uses 49.99M factor entries. The matched active-factor counts account for the per-context update budget, but do not match the learned adaptation capacity. The main-table baseline configurations and total trainable parameter budgets are also insufficiently specified. A capacity-matched static adaptation baseline and a relevant dynamic adapter, such as the cited PhaseLoRA, would help establish the contribution of dual-module generation and representation coupling. This matters because the structured synthesis primitive is adopted from HY-WU. Moreover, the LIBERO improvement on π0.5\pi_{0.5} is modest, making the additional capacity and execution cost important to the practical assessment.

Questions for authors
  1. Evidence for the central motivation. Can the authors provide direct evidence of different adaptation demands across stages of the same task, and show how the generated updates address them? Stage-level analysis on existing simulation or physical tasks would help. A controlled comparison between per-context generation and episode-fixed generated updates, while preserving current observation inputs, would also help assess the benefit of updating parameters within a trajectory.

  2. Baseline and capacity comparisons. What adaptation configuration does each main-table Baseline use? Please specify the trainable modules, total trainable parameter counts, and training budgets. Is there evidence that the advantage remains against a static baseline with comparable learned capacity or a closely related dynamic adapter?

  3. Interpretation of the ablations. All Tab.3 variants fall below the ordinary X-VLA baseline of 93.05%; VLM-only reaches only 43.50%. Were the variants trained separately with comparable optimization and tuning budgets? When generated updates are disabled in one module, does that module retain conventional static adaptation? Please explain how these results separate the benefit of contextual coupling from difficulties optimizing the ablated configurations.

  4. Scope of adaptation and generalization. Please clarify which objects, layouts, and visual conditions differ between training and evaluation, particularly in the physical experiments. Are results available under controlled environment changes or with fewer demonstrations? These would help determine whether the generated updates remain useful outside the training contexts; the current results do not establish this aspect of the method.

8 comments
AuthorsOpenReview ↗

Rebuttal to Reviewer Qa7R (Part 1)

Thank you for the detailed review. Below are our point-by-point responses to your concerns and questions.

Weakness 1 & Question 1 — Stage-dependent adaptation: whether the operation-level diagnostic and the current results show different adaptation demands across stages of the same task, and whether refreshing the generated update within a trajectory helps.

We thank the reviewer for this interesting question. Our motivation is stated at the level of atomic operations, and the stage structure in Fig. 1 follows from it. A compound task is carried out as a sequence of atomic operations, such as opening, grasping, transporting, and placing, and each stage of a trajectory is one of these operations. Different stages call for different adaptations because different operations do. The diagnostic therefore tests the underlying question directly: does an update that helps one atomic operation hurt the others?

RoboCasa365 [1] suits this test because it defines 65 atomic tasks and builds its composite tasks as sequences of them. Comparing its native task IDs is a comparison between atomic operations, with the operation boundaries given by the benchmark rather than by our own segmentation of long trajectories. In this comparison, every one of the 260 updates that lowers the loss on its source operation also raises the loss on 22–46 other operations (median 32). A shared static update thus conflicts at exactly the level at which a compound task changes from one stage to the next.

For direct within-trajectory evidence, we traced the generated update through a successful LIBERO-Long rollout of "put both the alphabet soup and the tomato sauce in the basket". The instruction and the scene layout stay fixed for the whole episode; what changes is the progress of the task, that is, which object is being grasped, carried, or released. Across 27 consecutive replanning steps, the effective update ΔWt=(α/r)AtBt\Delta W_t=(\alpha/r)A_tB_t changes by 24.1% on average, measured as ∥ΔWt−ΔWt−1∥F/∥ΔWt−1∥F\lVert\Delta W_t-\Delta W_{t-1}\rVert_F/\lVert\Delta W_{t-1}\rVert_F over all adapted projections (29.2% on the VLM side and 23.5% on the action side). Object positions and robot state do change during the episode, but these changes are how the stages appear in the observation. MorphAct conditions on the current context and needs no stage labels, so it responds to a new stage through exactly these signals.

We also ran the controlled comparison the reviewer suggested. On the same trained checkpoint, we generate the VLM-side and action-side updates once from the first observation of each episode and keep them fixed until the episode ends. Everything else is unchanged: images, robot state, VLM features, and action-conditioning inputs are refreshed at every replanning step, and initial states, seeds, and execution settings match the per-context evaluation. On the same LIBERO cases, success falls from 96.5% with per-context generation to 0% with episode-fixed updates. Current observations alone cannot replace the refreshed parameters. Because this control is applied at inference time, it shows what the trained policy relies on; the trained counterpart is the Task-ID variant in Table 3, whose update depends only on the task and is therefore fixed within each episode while the policy still receives current observations. Trained this way, it reaches 88.25%, compared with 96.60% for MorphAct.

The revision adds the within-episode trace and the episode-fixed control as Appendix E (Tables A12 and A13), referenced from Section 5.1. The introduction now also state the atomic-operation basis of the motivation.

Weakness 2 & Question 2 — Baselines, capacity, and related adapters: how the main-table baselines are configured, and whether the gains remain against a capacity-matched static adapter and a dynamic adapter such as PhaseLoRA.

We thank the reviewer for this interesting comment. MorphAct improves all three backbones on both benchmarks, with the largest gains where the baselines leave room. On Meta-World it adds 15.13 points to π0.5; on LIBERO it adds 3.55 points to X-VLA, and LIBERO-Long gives the largest suite-level gain for every backbone (+9.0, +7.2, and +2.2 points for X-VLA, π0, and π0.5). π0.5 already reaches 96.85% on LIBERO, and raising it to 97.35% removes 15.9% of its remaining failures.

The main-table baseline numbers are taken from the original reports: π0 and π0.5 from OpenPI's official LIBERO fine-tuning results at 30k steps [2], and X-VLA from the Ours-LoRA configuration of the X-VLA paper, which trains 9M parameters [3]. Appendix D.3 of the revision lists these sources together with the trainable parameters of each configuration, and Appendix D.4 gives our training budget for each backbone. Because these are reported numbers, we also ran the retrained comparisons below, in which everything except the adaptation method is matched.

AuthorsOpenReview ↗

Rebuttal to Reviewer Qa7R (Part 2)

Weakness 2 & Question 2 — Baselines, capacity, and related adapters (continued).

The 300.09M generator parameters and the 49.99M active factor entries (π0/π0.5) measure different things. The generator is a function that outputs LoRA factors. It runs once per replanning step, and the policy's forward pass then uses only the generated factors, which have the same size as a shared LoRA of the same rank. To compare at equal learned capacity, we additionally trained a static LoRA with a comparable total trainable-parameter budget under the same backbone, data, and schedule, and compared against PhaseLoRA [4] on the same backbone:

X-VLA, LIBEROSuccess (%)
Static LoRA, comparable trainable-parameter budget94.00
PhaseLoRA94.70
MorphAct96.60

Extra static capacity does not close the gap. The ablations in Table 3 make the same point at fixed capacity: Constant, Task-ID, and Unadapted have exactly the same generators and parameter count as MorphAct and differ only in what the generators read, yet they reach 87.50%, 88.25%, and 86.00% against 96.60%.

Our contribution lies in generating updates for both the VLM and the action module from the current context, and in conditioning the action-side generator on the adapted VLM representation; HY-WU's structured synthesis primitive [5] serves as the component that produces the factors. The Constant and Task-ID variants isolate the first, and the Unadapted variant isolates the second.

The Transformer projection weights stay frozen. MorphAct trains the two generators, the vision encoder and visual projections, and the native action interfaces, plus soft prompts for X-VLA. Generation adds 18–21% end-to-end latency (Table A8). The revision adds this accounting in Appendix D.3, and Table A8 now lists the total trainable parameters.

Question 3 — Interpretation of the ablations: whether the Table 3 variants were trained comparably, whether a disabled module keeps static adaptation, and how the results separate contextual coupling from optimization difficulty.

Thank you for the interesting question about the ablations. Each variant in Table 3 is trained separately, from the same initialization and with the same data and training budget as the full model. When generated updates are disabled in a module, that module is frozen; it receives no static LoRA in their place.

VLM-only therefore keeps the entire action module at its pretrained weights. The action module maps VLM features and robot state to actions, so in this variant the representation adapts to the tasks while the mapping to actions cannot, which explains its 43.50%. The 93.05% X-VLA reference adapts action-side components as well [3], so comparing it with VLM-only mixes this restriction with the method. To separate the two, we trained a static LoRA on the VLM alone, with the action module frozen in the same way. It reaches 41.80%, close to the 43.50% of generated VLM-only updates: the drop comes from freezing the action module, not from the generator.

Variant (X-VLA, LIBERO)What differs from MorphActSuccess (%)
MorphAct–96.60
Static LoRA (new)static update, comparable trainable-parameter budget94.00
Task-IDgenerators read the task identity88.25
Constantgenerators read a constant87.50
Unadaptedaction generator reads unadapted VLM features86.00
Action-onlyVLM frozen81.00
VLM-only static LoRA (new)action module frozen, static VLM LoRA41.80
VLM-onlyaction module frozen43.50

The remaining variants keep both adaptation paths and change one design choice at a time. Constant and Task-ID are not static baselines: they keep the generator and remove its contextual input. With no informative input they fall below a directly optimized static LoRA (94.00%), and we report them to isolate the conditioning signal within a fixed architecture. Within that architecture, conditioning on the current context adds 9.10 and 8.35 points. Unadapted and MorphAct share the architecture, parameter count, and training, and differ only in whether the action generator reads unadapted or adapted VLM features. Optimization difficulty from architecture or capacity is the same for both, so the 10.60-point gap measures the coupling itself.

The revision reports the two new rows in Table A10 (Appendix D.6) and refers to them in Sections 5.1 and 5.2. The Table 3 caption now also states the training and freezing protocol.

AuthorsOpenReview ↗

Rebuttal to Reviewer Qa7R (Part 3)

Question 4 — Training and evaluation conditions: which conditions differ between training and evaluation, especially on the physical robots, and whether the generated updates remain useful under controlled environment changes.

We appreciate this interesting question. In simulation we follow the standard protocol of OpenVLA [6], OpenVLA-OFT [7], π0 [8], and X-VLA [3]. One policy is trained jointly on the four LIBERO suites and evaluated with 50 closed-loop rollouts per task (500 per suite) from the benchmark's initial states; on Meta-World MT50, one policy is trained on all 50 tasks and evaluated with 20 rollouts per task. The LIBERO suites are built to vary spatial layout (Spatial), object identity (Object), task goal (Goal), and long-horizon composition (Long), so the four-suite score already covers these variation types.

In the physical experiments on SO-101 and xArm6, each task uses 100 demonstrations, and each method is evaluated over 100 trials. The initial object placement is randomized for every trial. The baseline and MorphAct are run on the same sequence of placements, so the i-th trial of both methods starts from the same configuration; scene difficulty and demonstration budget are thus matched between methods while the placement varies across trials. Under these conditions, MorphAct improves success by 13–17 points on all four tasks.

For controlled environment changes, we evaluated on LIBERO-Plus [9], which perturbs LIBERO along seven dimensions: object layout, camera viewpoint, robot initial state, language instruction, lighting, background, and sensor noise.

LIBERO-PlusSpatialObjectGoalLongOverall
X-VLA75.6567.7167.0067.1769.29
MorphAct82.7275.8174.0374.6776.72

MorphAct improves every suite by 7.0–8.1 points and the episode-weighted overall score by 7.43 points, so the generated updates remain useful under conditions not seen in training.

We hope these results address your concerns, and we would be grateful if you would reconsider your rating in light of this evidence. If you have further questions, we would be glad to answer them!

References

[1] RoboCasa365: A Large-Scale Simulation Framework for Training and Benchmarking Generalist Robots. ICLR 2026.

[2] Physical Intelligence. LIBERO Benchmark Results. OpenPI GitHub repository. https://github.com/Physical-Intelligence/openpi/

[3] X-VLA: Soft-Prompted Transformer as Scalable Cross-Embodiment Vision-Language-Action Model. ICLR 2026.

[4] PhaseLoRA: Control-Regime-Conditioned Low-Rank Adaptation for Continuous-Action Vision-Language-Action Policies.

[5] HY-WU (Part I): An Extensible Functional Neural Memory Framework and An Instantiation in Text-Guided Image Editing.

[6] OpenVLA: An Open-Source Vision-Language-Action Model.

[7] Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success.

[8] π0: A Vision-Language-Action Flow Model for General Robot Control.

[9] LIBERO-Plus: In-depth Robustness Analysis of Vision-Language-Action Models.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Thanks and Follow-up as the Author Discussion Period Nears Its End

Dear Program Chairs, Senior Area Chairs, Area Chairs, and Reviewers,

As the author-participation period draws to a close, we would sincerely appreciate any further comments or questions while we can still respond.

We have responded to all reviewers’ comments point by point and incorporated additional experiments and clarifications into the revised manuscript. We especially thank Reviewers zFvr and DFMJ for their positive follow-up feedback: zFvr reaffirmed their initial support, and DFMJ raised their score from 2 to 3, finding the new results convincing and helpful.

We would be particularly grateful to know whether our responses and additional experiments have addressed Reviewer Qa7R’s concerns, and would welcome the opportunity to clarify any remaining points. We sincerely hope our additional experiments and clarifications have addressed most of your concerns. Several of your concerns overlap with those raised by DFMJ, so our responses in that discussion may also be helpful. We would welcome any remaining questions or suggestions and would be happy to discuss them further.

We sincerely thank everyone in the review process for their time, effort, and thoughtful consideration.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.