acceptodds
Withdrawn from ICLR 2027

ShareVLA: Share What Is Common, Specialize What Is Needed for Multi-Task Vision-Language-Action Learning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Vision-Language-Action (VLA) models adapt pretrained vision-language models to robot control and now solve individual manipulation tasks reliably, yet a single policy that executes many skills remains an open problem. Naive multi-task training suffers from interference among skills, and the prevailing alternative, merging independently trained experts, requires task-specific masks and must leave the deepest layers of the action expert unmerged. This fragility is traced to its source: independently trained action experts diverge from their very first block, so that no component of one expert transfers to another skill and every new skill requires a full training run. We therefore propose ShareVLA, which reverses the order of merging and training. A shared trunk covering of the action expert is trained once on demonstrations from all skills and frozen. Lightweight expert heads are then trained on top of it, one per skill. The shared frozen trunk enables all expert heads to be assembled into a single model with zero merge error by construction, without backbone masks and with only a single backbone forward pass at inference. A lightweight router selects the head. On LIBERO, ShareVLA reaches average success at the same parameter count as the strongest merging-based method while using of its training compute, and exceeds multi-task training of the same architecture by points. These results indicate that multi-skill VLA policies need not merge what was learned apart, and that deciding what to share before specializing is simpler and cheaper.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision-Language-Action Models, Multi-Task Learning, Robotic Manipulation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_Y2yrOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes ShareVLA, a share-then-specialize training strategy for multi-task VLA policies. It first trains a shared action-expert trunk on all skills, freezes it, and then trains lightweight skill-specific heads together with a router. The goal is to avoid the incompatibility of independently trained experts and the need for post-hoc model merging. On LIBERO, ShareVLA achieves 91.7% average success compared with 90.2% for MergeVLA-TIES, and the paper reports a claimed 44% training-compute cost relative to merge-based training. It also evaluates robustness on LIBERO-Plus.

Critical strengths
  1. The proposed share-then-specialize pipeline is simple and practically appealing. Sharing most of the action expert and specializing only lightweight heads produces a compact single policy and avoids explicit post-hoc merging of the action experts.

  2. The method achieves strong in-distribution performance on LIBERO. ShareVLA reaches 91.7% average success and performs competitively with the strongest reported merging baseline.

Critical weaknesses
  1. The claimed 44% training-compute advantage is based on asymmetric accounting. ShareVLA reports 88k gradient steps versus 200k for merge-based training, but this count only includes the steps performed by ShareVLA itself. Its Stage-1 backbone is constructed by TIES-merging publicly released single-task backbones, whose training cost is excluded.
    This comparison is not symmetric. If existing single-task checkpoints are treated as free for ShareVLA, then a merging method should also be allowed to start from existing experts without paying their training cost. Conversely, if the comparison is intended to measure the end-to-end cost of obtaining the final multi-task policy, then the cost of producing the single-task checkpoints used by ShareVLA should also be included. This is particularly important because the paper’s own ablation shows that the adapted/merged backbone is essential: using the original pretrained backbone gives 65.8%, compared with 89.3% using the merged backbone. Therefore, the excluded component is not incidental to the final performance.

  2. The experimental validation is too narrow, and the robustness result substantially favors the baselines. The main evaluation is only on LIBERO. ShareVLA obtains only a 1.5-point improvement. On LIBERO-Plus, however, ShareVLA drops a lot, substantially below both the same-architecture multi-task baseline MergeVLA-TIES.
    There is no second benchmark, no different embodiment, and no real-robot evaluation. The paper itself acknowledges that evaluation is limited to simulation, a single embodiment, and a single 0.5B backbone. Given the small in-distribution gain and the much worse OOD robustness, I do not think the evidence is sufficient to support the broad conclusion that share-then-specialize is generally preferable to expert merging.

Questions for authors
  1. Under a symmetric compute-accounting convention, what is the actual training-cost comparison between ShareVLA and merging? Please report either end-to-end cost from the same pretrained checkpoint, or incremental cost assuming the same single-task checkpoints are already available to both methods.

  2. Can the authors provide results on an additional benchmark, embodiment, or real robot? Given the small gain on LIBERO and the substantially worse LIBERO-Plus result, this would materially affect my assessment.

Reviewer_374POpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper diagnoses merging of independently trained VLA experts fails because experts diverge from the first block of the action expert. And this paper proposes ShareVLA, which train and freeze a shared trunk on all skills, train one lightweight head per skill, and assemble them with zero merge error by construction plus a lightweight router. On LIBERO, ShareVLA reaches 91.7% average success at the same 0.70B parameters as MergeVLA while using 44% of its training compute.

Critical strengths
  1. By learning and freezing a shared trunk before specializing heads, the merge error becomes zero by construction. Adding a skill costs only a 4% parameter head instead of a full expert.

  2. ShareVLA uses 44% of the training compute while achieving 91.7% on LIBERO at the same parameter count. The ablations (head steps, head capacity, routing) are thorough and support the key design decisions.

Critical weaknesses
  1. All experiments were simulated on LIBERO / LIBERO-Plus and no real-robot verification was conducted. But there are huge gap between simulation and real-world environments. The results presented in this paper on LIBERO-Plus have already demonstrated vulnerabilities under camera perspectives, sensor noise, and initial state perturbations, and all of them are common in real-world deployments.

  2. Low performance compared with joint-training SOTA. OpenVLA-OFT reaches 97.1% on the four suites via mixed fine-tuning, above the 91.7% reported here. I doubt whether the Merging method is suitable for VLA training. Although MergeVLA has a lower accuracy in joint training, it is hard to determine whether it is a problem with the training strategy/model hyperparameters or the mixed dataset itself.

Questions for authors
  1. What is the difference between this mechanism of selecting different heads through a router and MoE? Some works are already using MoE architecture for VLA training [1,2].

  2. How to divide different skills? Do you think it is more appropriate to divide based on different datasets from LIBERO, or based on robot actions such as pick and place, turning on and off switches, pushing and pulling drawers, etc.?

  3. The model architecture in this paper is limited to MergeVLA. For more general VLA architectures such as PI-series and OFT, do you think the strategy in this paper needs to be adjusted?

[1] Zhao, Han, et al. "More: Unlocking scalability in reinforcement learning for quadruped vision-language-action models." 2025 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2025.

[2] Du, Zhiying, et al. "Himoe-vla: Hierarchical mixture-of-experts for generalist vision-language-action policies." arXiv preprint arXiv:2512.05693 (2025).

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

In this paper, the authors look at model merging as the way to create a multi-task policy by combining single-task policies. They demonstrate that MergeVLA's action experts diverge from each other making merging brittle. To fix this, they create SHareVLA where the action experts share a trunk and just have one last block that differs between them. Training proceeds in three stages. The trunk is trained once on mixed multi-skill data and frozen. One head per skill is then trained on top of the frozen trunk. Finally, the heads are assembled around a single copy of the trunk together with a lightweight router that infers the active skill from one backbone forward pass. The authors evaluate on the LIBERO benchmark and show that ShareVLA achieves high success rates.

Critical strengths
  • The paper is clear and easy to follow, diagnoses a clear issue in model merging, and fixes it successfully with an intuitive idea to share and then specialize.

  • The results on LIBERO show the capabilities of the method. The ablations on trunk sharing, head capacity, and misrouting are very helpful.

Critical weaknesses
  • It would be great to get real world results on robots or further results on more than a single benchmark to quantify the method's broad effectiveness. It would also be great to see standard deviations to show statistical significance of the results. This is my main concern.

  • Given the solution is effectively a form of multi-task training, the reviewer believes it would be useful to compare to wider Libero results including popular VLA models like pi and WAM models like dreamzero. These comparisons are missing in the paper and in general the comparisons seem very limited when only focused on MergeVLA (and VLA-Adapter).

Questions for authors

Please see Weaknesses

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

Related papers

Discussion (0)

Sign in to comment.