acceptodds
Under review as a conference paper at ICLR 2027

PreMerge for VLA: Building a Stronger Model by Merging Models with a Shared Base

Abstract

Vision-language-action (VLA) models commonly initialize from a single pretrained vision-language model (VLM), although models from different training stages may retain complementary knowledge. We propose PreMerge for VLA to combine VLMs with a shared base before target-task action training. First, we align their vision-language parameters and assign source weights from the magnitude and direction changes of the original weights relative to the base. A sign-consistency mask removes conflicting parameter updates. Second, we use a small set of target images and instructions, without action annotations, to calibrate the merged model through multilayer feature distillation. Source parameters remain frozen, and only tensor-level merging coefficients are optimized. Using the aligned VLM components of Qwen3-VL, Cosmos-Reason2, and GR00T-N1.7, PreMerge achieves average success rates of 98.9% on LIBERO and 92.7% on RoboTwin 2.0, exceeding the best single-source initialization by 1.0 and 3.2 percentage points, respectively. On four real-world manipulation tasks, it achieves 90.0%, compared with 82.5% for the strongest evaluated external baseline. The merged parameters are folded into one VLM, so subsequent action training and inference require no additional source branches. These results demonstrate the value of reusing models from different training stages to construct VLA initializations.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision-Language-Action Models, Model Merging

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

This paper proposes PreMerge, a framework for constructing VLA initializations by merging aligned VLM components from models sharing a common pretrained base. The method first combines adaptive parameter weighting with sign-conflict filtering, then learns tensor-level merging coefficients through feature distillation. Experiments on two simulation benchmarks and four real-world manipulation tasks show the improvements. However, actually, I do not really understand the motivation behind it. It seems the method works, but I do not understand why it works. The writing should be improved.

Critical strengths
  1. The paper addresses a practically relevant question: whether models from different training stages can be reused to construct a better initialization for downstream robot learning.
  2. Calibration requires no action annotations, and the merged parameters are folded into a single VLM.
  3. Experiments are overall good.
Critical weaknesses
  1. The most concern I have is that the connection between feature-change consistency and downstream task quality is insufficiently justified. The calibration target averages source features when their changes relative to the base have nonnegative cosine similarity; otherwise, it preserves the initial merged representation. However, agreement between feature changes does not necessarily imply that those changes are useful for robot control.

For example, both sources could increasingly encode the same background or camera-specific shortcut. Their feature changes would be aligned, yet distilling their average could reinforce a spurious correlation and hurt performance when the background changes, right? On the other hand, the paper does not sufficiently explain when consistency indicates useful shared knowledge rather than shared error. I mean, it depends on how good the models are? And also the variance between the training datasets.

  1. The motivation and technical novelty are not clearly articulated. Several central components are inherited from prior work: magnitude-, direction-, and rank-based weighting from WIDEN, sign-conflict erasure from SCE, and feature-distillation-based merging from ProDistill. It is therefore difficult to identify the specific methodological advance. Is the principal novelty the consistency-conditioned teacher-selection rule, the joint multilayer calibration, or their application before target-task action training? The abstract and introduction should state this distinction explicitly and explain the limitation of existing methods that necessitates the proposed modification.

  2. Several figures are difficult to read. The framework diagram and some multi-panel analysis figures contain very small text, labels, and equations. This makes an already complex pipeline harder to follow. The symbols are also too numerous; consider adding a symbol table. It's very hard to follow.

Questions for authors
  1. With the initial merged model, coefficient parameterization, and calibration budget fixed, how does the proposed teacher-selection rule compare with always averaging source features?
  2. Which layers or feature groups best distinguish the source backbones, and does removing or replacing those groups affect the gains from merging?
  3. I am curious which layers, token positions, or feature dimensions most clearly distinguish the sources? Do these differences reflect object identity, spatial relationships, physical interactions, or task-irrelevant appearance?
Reviewer_87vBOpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper merges aligned Qwen3-VL, Cosmos-Reason2 and GR00T-N1.7 VLM components before target-task action training. WIDEN-style weighting and SCE-style sign filtering produce fixed residuals; action-free target images and instructions then calibrate 986 tensor-level coefficients through feature distillation. The resulting single VLM reaches 98.9% on LIBERO and 92.7% on RoboTwin, exceeding the strongest single-source initialization by 1.0 and 3.2 percentage points.

Critical strengths
  1. Combining different fine-tuned versions of models with the same architecture to improve VLA initialization is a simple and reasonable approach. The reported results and component ablation experiments confirm its practical value.
  2. The evaluation covered two simulated benchmarks and four real-world tasks. The merged VLM was also evaluated using the GR00T, PI, and OFT action heads, demonstrating its compatibility with a variety of commonly used action interfaces.
Critical weaknesses
  1. Although PreMerge incorporates adaptive weighting and conflict filtering mechanisms, the experiments primarily compared it with equally weighted averaging and variants of its own components. No direct comparisons were made with mature model-merging approaches in the community, such as Task Arithmetic and TIES. Therefore, it remains unclear whether the proposed merging process offers advantages over existing, mature alternatives under the same source models and downstream training protocols.
  2. The method appears to contribute primarily through incremental improvements. Section 4.1 explicitly builds upon WIDEN and SCE, while feature-based coefficient learning via distillation has already been explored in ProDistill, including its end-to-end variant, DistillMerge. CapVector also investigated parameter space enhancement in VLA initialization. Existing component ablation studies support the effectiveness of the method but are insufficient to demonstrate the necessity of this combination or to quantify its advantages over these original methods.
  3. On the LIBERO dataset, performance improved by 1.0 percentage points compared to the strongest single data source, with calibration contributing 0.6 percentage points. In real-world evaluation, the comparison between 90.0% and 82.5% translates to an additional 6 successes out of 80 trials. Due to the lack of results across independent training seeds and an estimate of the uncertainty in performance differences, it is difficult to determine to what extent these performance improvements are attributable to the method rather than evaluation randomness.
Questions for authors
  1. Given the same source model and downstream training protocol, how does PreMerge perform compared to appropriately tuned approaches such as Task Arithmetic and TIES?
  2. Can you report retraining results on the LIBERO dataset and, where feasible, results for real-world tasks, including the number of training seeds, the mean, standard deviation, and confidence intervals of performance differences? Please distinguish between variability across different training runs and uncertainty arising from a limited number of evaluation trials.
  3. Could the performance gains be primarily attributed to the extensive action training of source models such as Cosmos-Reason and GR00T N1.7? Would incorporating other Qwen3-VL-based models trained only through SFT or GRPO, such as the spatial intelligence models PhysBrain1.5 and RoboBrain2.5, yield similar improvements? Further analysis of this question would help clarify the scope and limitations of the proposed method.
Reviewer_Kw2rOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes PreMerge, which combines aligned VLM components from Qwen3-VL, Cosmos-Reason2, and GR00T-N1.7 before downstream action training. It first constructs a merged initialization through adaptive source weighting and sign-based conflict filtering, then calibrates tensor-level merging coefficients using multilayer feature distillation on target images and instructions without action labels. The resulting weights are folded into a single VLM. The paper evaluates this initialization on LIBERO, RoboTwin 2.0, and four real-world manipulation tasks.

Critical strengths
  1. Component ablations. The paper separately evaluates conflict filtering, adaptive weighting, and task-level calibration. The reported results show incremental benefits from these components on both simulation benchmarks, providing some evidence that the complete pipeline improves over its simpler variants.
  2. Practical calibration and deployment. Calibration optimizes only 986 scalar coefficients without requiring action annotations. The calibrated weights are folded into one backbone, preserving the deployment architecture and inference cost of the downstream VLA.
Critical weaknesses
  1. Limited novelty. PreMerge appears to combine established ideas from Task Arithmetic (ICLR 2023), TIES-Merging (NeurIPS 2023), and ProDistill (ICML 2025) for VLA initialization. The paper does not sufficiently establish a substantive advance from its teacher-selection rule or calibration constraints.
  2. Missing real-world demonstrations. The paper reports 90% real-world success but provides only tables and static images, without complete autonomous rollout videos. Successful and failed demonstrations for each task, linked to evaluation trials, would help assess task difficulty, execution quality, and failure modes.
  3. Unverified complementary knowledge transfer. Source coefficients, CKA, and PCA do not show which semantic or physical capabilities are transferred or retained after action training. The reported gains support a beneficial initialization effect more directly than complementary knowledge reuse.
  4. Irreversible conflict filtering. The unweighted residual sum in Equation (9) can select a direction inconsistent with the weighted combination. It may also discard useful corrections introduced by later robot training. Since calibration only rescales fixed residuals with nonnegative coefficients (Equations (10)–(11)), it cannot recover discarded updates. The paper does not establish why these decisions should remain fixed rather than be revised using target data.
  5. Missing implementation attribution. The supplementary code builds on StarVLA’s training infrastructure and Qwen-based action frameworks, but the manuscript neither cites StarVLA nor acknowledges it as the implementation foundation.
  6. Excessive formalism and fragmented presentation. Equations for routine operations, scattered citations, and frequent cross-references obscure the core algorithm and its novel components. A concise, self-contained method description would improve clarity.
Questions for authors

See Weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.