View-GRPO: Multi-View GRPO Methods

• View-GRPO is a family of methods that redefine groups in GRPO by aggregating semantically or geometrically distinct views, thereby providing denser and more informative reward signals.
• MV-GRPO augments text-to-image models by reusing sampled images with multiple semantically adjacent captions, enhancing reward variance without costly sample regeneration.
• VISTA applies view-based GRPO to GUI grounding by constructing groups from target-preserving crops and adding a self-verified cross-view anchor, resulting in robust model performance.

Searching arXiv for the cited works and closely related "View-GRPO" papers. arXiv search query: "View-GRPO OR MV-GRPO OR VISTA GRPO GUI grounding OR augmented condition space GRPO". View-GRPO denotes a family of Group Relative Policy Optimization formulations in which the comparison group is no longer defined solely by multiple rollouts under one fixed condition. In current arXiv usage, the term encompasses at least two closely related constructions. Multi-View GRPO (MV-GRPO) for flow-based text-to-image models evaluates the same sampled images under multiple semantically adjacent captions, thereby creating a dense multi-view reward mapping without regenerating samples (Bu et al., 13 Mar 2026). VISTA, for GUI grounding, constructs each GRPO group from multiple target-preserving views of the same GUI instance and adds a self-verified cross-view anchor that is excluded from the group baseline and activated only when the model has produced a maximum-reward rollout (Qiu et al., 12 Jun 2026). In both lines of work, the central objective is to make the relative reward structure inside each group more informative than in the single-view regime.

1. GRPO background and the redefinition of the group

Standard GRPO is a PPO-style RL method in which, for a given condition, one samples a group of trajectories, computes scalar rewards, and normalizes those rewards inside the group. Using the notation employed in the recent View-GRPO literature, the group statistics are

$$\mu_G = \frac{1}{G}\sum_{i=1}^G r_i,\qquad \sigma_G = \sqrt{\frac{1}{G}\sum_{i=1}^G (r_i-\mu_G)^2},$$

with normalized advantages

$A_i = \frac{r_i-\mu_G}{\sigma_G+\epsilon}.$

These advantages are then used inside a clipped surrogate objective, so GRPO replaces an explicit critic with group-internal relative normalization (Qiu et al., 12 Jun 2026).

What View-GRPO changes is not the existence of group-relative optimization, but the definition of what counts as a group. In single-view GRPO, the group is typically formed by several outputs conditioned on one prompt or one screenshot. In view-based variants, the group is constructed or rescored across multiple views of the same semantic task: multiple semantically adjacent captions for one prompt in MV-GRPO, or multiple target-preserving crops of one GUI instance in VISTA (Bu et al., 13 Mar 2026). This alters the reward distribution that enters GRPO, and therefore alters the effective advantage signal.

This move is compatible with broader theoretical descriptions of GRPO. Recent theory treats the GRPO policy gradient as a U-statistic and derives a universal scaling law for optimal group size, emphasizing that the behavior of GRPO depends critically on how information is distributed inside the group (Zhou et al., 1 Mar 2026). View-GRPO can therefore be understood as changing the internal geometry of the group rather than abandoning GRPO’s group-relative principle.

2. Why single-view GRPO becomes uninformative

The immediate motivation for View-GRPO is that the single-view regime often produces degenerate groups. In GUI grounding, when GRPO is applied to one fixed screenshot, hard instances often produce groups in which all rollouts fail, so all rewards are $0$; easy instances often produce groups in which all rollouts succeed, so all rewards are $1$. In both cases, $\sigma_G = 0$ and the resulting advantages collapse toward zero, yielding no useful relative signal (Qiu et al., 12 Jun 2026). VISTA reports that fewer than 5% of fixed-view groups are informative in this sense.

An analogous problem appears in flow-based text-to-image GRPO. There, the standard setup evaluates a group of generated samples only under a single prompt. MV-GRPO characterizes this as a sparse one-to-many condition-to-data reward mapping that insufficiently explores inter-sample relationships. Images that differ in lighting, composition, style, or other attributes may be ranked very differently under nearby captions, yet single-view GRPO uses only one textual condition to score them (Bu et al., 13 Mar 2026).

This suggests a direct connection between view construction and reward variance. Rather than increasing group size mechanically, View-GRPO seeks to increase the informativeness of the group by introducing structured diversity. That reading is consistent with the more general diagnosis, elsewhere in the GRPO literature, that useful optimization depends on informative within-group variation rather than raw sample count; Pro-GRPO, for example, explicitly diagnoses reward clustering and argues that high-variance subsets can outperform larger unfiltered groups (Ge et al., 17 Dec 2025).

3. MV-GRPO for flow-based text-to-image models

MV-GRPO, introduced for flow-based text-to-image alignment, augments the condition space rather than the sample space. For a fixed anchor prompt $\mathbf{c}$, one first samples a group of images $\mathbf{X}_G = \{\mathbf{x}_0^i\}_{i=1}^G$ once under the old policy. One then constructs an augmented set of captions

$$\{\mathbf{c}\}\cup \mathcal{V}_K,\qquad \mathcal{V}_K=\{\mathbf{c}_k\}_{k=1}^K,$$

where the additional captions are semantically adjacent yet diverse (Bu et al., 13 Mar 2026).

The paper formalizes the caption generator as a Condition Enhancer

$$\mathcal{E}:\mathcal{C}\times \mathcal{X}^G \to 2^\mathcal{C}.$$

Two instantiations are given. The online VLM enhancer uses a pretrained vision-LLM to generate posterior captions conditioned on individual sampled images and multi-perspective prompts. The offline LLM enhancer rewrites the anchor prompt with text-only operations such as ADD, DELETE, and PARAPHRASE, while preserving core semantics (Bu et al., 13 Mar 2026).

For each augmented caption $\mathbf{c}_k$, MV-GRPO recomputes group-normalized advantages

$A_i = \frac{r_i-\mu_G}{\sigma_G+\epsilon}.$0

and defines per-view importance ratios

$A_i = \frac{r_i-\mu_G}{\sigma_G+\epsilon}.$1

The final objective aggregates the original GRPO term for $A_i = \frac{r_i-\mu_G}{\sigma_G+\epsilon}.$2 with additional clipped GRPO terms over all $A_i = \frac{r_i-\mu_G}{\sigma_G+\epsilon}.$3 (Bu et al., 13 Mar 2026).

A distinctive feature is that MV-GRPO does not regenerate images for the new views. It reuses the original sampled trajectories and re-evaluates their transition probabilities under nearby captions. The justification given is empirical as well as probabilistic: because the augmented captions are semantically close to the anchor condition, the drift in log transition probabilities remains concentrated near zero (Bu et al., 13 Mar 2026). The method therefore turns one expensive batch of images into a multi-view reward matrix.

4. VISTA and view-based GRPO for GUI grounding

VISTA instantiates View-GRPO in GUI grounding by redefining the group across image views rather than across textual conditions. The input is a screenshot $A_i = \frac{r_i-\mu_G}{\sigma_G+\epsilon}.$4 and instruction $A_i = \frac{r_i-\mu_G}{\sigma_G+\epsilon}.$5, and the output is a coordinate string parsed into a click point $A_i = \frac{r_i-\mu_G}{\sigma_G+\epsilon}.$6. The reward is sparse and verifiable: $A_i = \frac{r_i-\mu_G}{\sigma_G+\epsilon}.$7 The key change is that the group is built from $A_i = \frac{r_i-\mu_G}{\sigma_G+\epsilon}.$8 target-preserving crops of the same GUI instance, not from repeated rollouts on one fixed screenshot (Qiu et al., 12 Jun 2026).

For each crop window $A_i = \frac{r_i-\mu_G}{\sigma_G+\epsilon}.$9, the original target box is remapped into the cropped coordinate frame by

$0$0

or equivalently by the pixel-space form given in the main text when $0$1 are already in pixels (Qiu et al., 12 Jun 2026). The semantics of the task are unchanged, but the geometry is altered: the target appears at different positions and scales in different views.

If $0$2, VISTA samples $0$3 completions per view,

$0$4

then computes model-only group statistics across all views,

$0$5

with advantages

$0$6

These enter the usual clipped GRPO term (Qiu et al., 12 Jun 2026).

VISTA then adds a self-verified cross-view anchor. The oracle answer is the center of the remapped target box,

$0$7

Its advantage is

$0$8

where $0$9 is the set of model rollouts with reward $1$0. The indicator enforces the self-verification gate: if no model rollout has maximum reward, the anchor contributes nothing. The anchor is excluded from the baseline, and the full objective is

$1$1

with $1$2 in the main experiments (Qiu et al., 12 Jun 2026).

5. Empirical profile

The two principal View-GRPO formulations improve different aspects of multimodal behavior, but both show gains over their single-view baselines.

Formulation	Representative result	Operational note
MV-GRPO (Bu et al., 13 Mar 2026)	Under HPS-v3 training, HPS-v3 rises from 0.150 to 0.155	Iteration time 191.95s vs 156.26s for Flow-GRPO-Fast and 1931.15s for image-level augmentation
VISTA (Qiu et al., 12 Jun 2026)	On ScreenSpot-Pro, Qwen3-VL 4B/8B/30B-A3B rises from 55.5/52.7/53.7 to 63.4/65.8/67.0	Crop diagnostic worst-view accuracy 92.42% and flip rate 5.80% for VISTA

For MV-GRPO, the reported improvements are not limited to one reward model. Under HPS-v3 training, MV-GRPO (VLM) reaches HPS-v3 $1$3, UR-v2-C $1$4, UR-v2-S $1$5, and ImageReward $1$6, exceeding the best single-view baselines listed in the paper. Under multi-reward training with HPS-v3+CLIP, it reaches HPS-v3 $1$7, UR-v2-C $1$8, UR-v2-S $1$9, and ImageReward $\sigma_G = 0$0. The number of views matters but saturates: without $\sigma_G = 0$1, HPS-v3 is $\sigma_G = 0$2; with $\sigma_G = 0$3, $\sigma_G = 0$4; with $\sigma_G = 0$5, $\sigma_G = 0$6; and with $\sigma_G = 0$7, $\sigma_G = 0$8 (Bu et al., 13 Mar 2026).

For VISTA, the most distinctive evidence concerns robustness across perturbed views. On the crop-perturbation diagnostic for Qwen3-VL-8B, the base model has Orig $\sigma_G = 0$9, Crop $\mathbf{c}$0, Worst $\mathbf{c}$1, VCR $\mathbf{c}$2, and Flip $\mathbf{c}$3. GRPO improves these to Orig $\mathbf{c}$4, Crop $\mathbf{c}$5, Worst $\mathbf{c}$6, VCR $\mathbf{c}$7, and Flip $\mathbf{c}$8. VISTA further improves them to Orig $\mathbf{c}$9, Crop $\mathbf{X}_G = \{\mathbf{x}_0^i\}_{i=1}^G$0, Worst $\mathbf{X}_G = \{\mathbf{x}_0^i\}_{i=1}^G$1, VCR $\mathbf{X}_G = \{\mathbf{x}_0^i\}_{i=1}^G$2, and Flip $\mathbf{X}_G = \{\mathbf{x}_0^i\}_{i=1}^G$3 (Qiu et al., 12 Jun 2026).

The VISTA ablations also show that its components are not interchangeable. On Qwen3-VL-8B, GRPO + crop reaches ScreenSpot-Pro $\mathbf{X}_G = \{\mathbf{x}_0^i\}_{i=1}^G$4, GRPO + anchor reaches $\mathbf{X}_G = \{\mathbf{x}_0^i\}_{i=1}^G$5, and full VISTA reaches $\mathbf{X}_G = \{\mathbf{x}_0^i\}_{i=1}^G$6. For anchor supervision, “Normalized” without gating drops to $\mathbf{X}_G = \{\mathbf{x}_0^i\}_{i=1}^G$7, “Const. SFT” reaches $\mathbf{X}_G = \{\mathbf{x}_0^i\}_{i=1}^G$8, and only “Normalized + gate” reaches $\mathbf{X}_G = \{\mathbf{x}_0^i\}_{i=1}^G$9. Increasing the number of anchors harms performance: $$\{\mathbf{c}\}\cup \mathcal{V}_K,\qquad \mathcal{V}_K=\{\mathbf{c}_k\}_{k=1}^K,$$0 anchor gives $$\{\mathbf{c}\}\cup \mathcal{V}_K,\qquad \mathcal{V}_K=\{\mathbf{c}_k\}_{k=1}^K,$$1, $$\{\mathbf{c}\}\cup \mathcal{V}_K,\qquad \mathcal{V}_K=\{\mathbf{c}_k\}_{k=1}^K,$$2 gives $$\{\mathbf{c}\}\cup \mathcal{V}_K,\qquad \mathcal{V}_K=\{\mathbf{c}_k\}_{k=1}^K,$$3, and $$\{\mathbf{c}\}\cup \mathcal{V}_K,\qquad \mathcal{V}_K=\{\mathbf{c}_k\}_{k=1}^K,$$4 gives $$\{\mathbf{c}\}\cup \mathcal{V}_K,\qquad \mathcal{V}_K=\{\mathbf{c}_k\}_{k=1}^K,$$5 (Qiu et al., 12 Jun 2026).

6. Interpretation, scope, and common misconceptions

One common misconception is that View-GRPO is a single canonical algorithm. The literature does not support that reading. In MV-GRPO, a view is a semantically adjacent caption generated by a Condition Enhancer, and the same trajectories are rescored under that augmented condition set (Bu et al., 13 Mar 2026). In VISTA, a view is a target-preserving crop of the same GUI instance, with exact coordinate remapping and cross-view comparison inside the group (Qiu et al., 12 Jun 2026). The shared principle is not a single implementation detail, but the replacement of single-view reward aggregation by multi-view reward aggregation.

A second misconception is that View-GRPO is merely data augmentation. The papers are explicit that the goal is not arbitrary augmentation. VISTA reports that a multi-image resize baseline severely degrades ScreenSpot-Pro from $$\{\mathbf{c}\}\cup \mathcal{V}_K,\qquad \mathcal{V}_K=\{\mathbf{c}_k\}_{k=1}^K,$$6 to $$\{\mathbf{c}\}\cup \mathcal{V}_K,\qquad \mathcal{V}_K=\{\mathbf{c}_k\}_{k=1}^K,$$7, because resizing changes scale but not the correct coordinates, so the views are not semantically and geometrically coupled in the required way (Qiu et al., 12 Jun 2026). MV-GRPO, conversely, augments the condition space rather than the image space, and its central claim is precisely that one can obtain denser supervision without costly sample regeneration (Bu et al., 13 Mar 2026).

A broader reading is that View-GRPO belongs to a shift in the GRPO literature from raw group size toward group informativeness. That is an inference rather than a named doctrine, but it is consistent with work diagnosing reward clustering and advocating high-variance or selectively pruned groups under fixed compute budgets (Ge et al., 17 Dec 2025). View-GRPO instantiates that idea in multimodal settings: it modifies what the model is compared against, so that the relative reward signal inside each group remains useful.

In that sense, View-GRPO is best understood not as an alternative to GRPO, but as a reparameterization of the group itself. The group ceases to be a bag of rollouts from one static input and becomes a structured set of semantically equivalent but geometrically or conditionally distinct views. That redefinition is the distinctive contribution of the current View-GRPO literature (Bu et al., 13 Mar 2026, Qiu et al., 12 Jun 2026).