Open Models Almanac
ModelsX-VLA

X-VLA authors

X-VLA

What X-VLA’s own documents say it is built from — every method with the sentence that describes it.

Parameters
0.9B
Attention
Florence-2-large backbone; soft prompts per embodiment
Input
text+image -> action
Licence
Apache-2.0
Released
2025-11-04

Against the consensus recipe

Its documents state 0 of the 18 methods the field agrees on.

The recipe from the overview: the methods the most labs adopt, per stage. A missing one is something its documents do not say, not something the model lacks.
  1. Data curation

  2. Training objective

  3. Model architecture

  4. Optimization

  5. Post-training

  6. Inference & serving

Checked in its code

Its code shows 0 of the 19 architecture features checked.

Sources say what a lab chose to describe; the modeling code says what the checkpoint runs. Each feature below was put to a code verifier, and only the config values and source lines it cited that were found in the files are shown, linked to the line.

Read from the model repository’s own code (modeling_xvla.py, modeling_florence2.py, configuration_florence2.py, transformer.py, action_hub.py, configuration_xvla.py) and config.json at revision c1c4a64a7e, by anthropic/claude-sonnet-5 on 2026-11-07.

FeatureCodeIts documentsVerdict
Grouped-query attention not in its codenot stated

The Florence2 text encoder/decoder attention and the SoftPromptedTransformer attention both use standard multi-head attention with equal…

show evidence

The Florence2 text encoder/decoder attention and the SoftPromptedTransformer attention both use standard multi-head attention with equal numbers of query, key and value heads, with no separate kv head count.

self.k_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
        self.v_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
        self.q_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
        self.out_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
modeling_florence2.py · L763–766
Multi-head latent attention not in its codenot stated

No low-rank KV latent projection exists anywhere in the attention implementations; keys/values are projected directly to full embed_dim as…

show the whole verdict

No low-rank KV latent projection exists anywhere in the attention implementations; keys/values are projected directly to full embed_dim as in standard attention.

Sliding-window attention not in its codenot stated

No sliding-window masking mechanism is implemented in the Florence2 text attention or the action transformer attention; attention is…

show the whole verdict

No sliding-window masking mechanism is implemented in the Florence2 text attention or the action transformer attention; attention is either full bidirectional (encoder) or full causal (decoder).

Interleaved sliding-window and global attention not in its codenot stated

There is no interleaving of sliding and full-attention layers; all layers use the same uniform full/causal attention pattern.

Indexer-selected sparse attention (DeepSeek Sparse Attention) not in its codenot stated

No indexer-based top-k token selection mechanism is present in any attention module.

Linear-attention or state-space layers alongside full attention not in its codenot stated

All token-mixing layers (Florence2 encoder/decoder, action transformer) use standard softmax attention; no linear-attention or state-space…

show the whole verdict

All token-mixing layers (Florence2 encoder/decoder, action transformer) use standard softmax attention; no linear-attention or state-space mixers are implemented.

Gated DeltaNet layers not in its codenot stated

No gated delta-rule recurrent mixer is implemented anywhere in the provided files.

Mamba-2 layers not in its codenot stated

No Mamba-2 or SSM-style selective state-space block is implemented in the given modeling files.

Learnable attention sink not in its codenot stated

None of the attention modules (Florence2Attention, Florence2SdpaAttention, Florence2FlashAttention2, or the transformer.py Attention) add…

show the whole verdict

None of the attention modules (Florence2Attention, Florence2SdpaAttention, Florence2FlashAttention2, or the transformer.py Attention) add a learned per-head sink logit to the softmax denominator.

Gated attention output not in its codenot stated

No sigmoid gate multiplies the attention output before the output projection in any attention class; outputs go straight from attn@v to…

show the whole verdict

No sigmoid gate multiplies the attention output before the output projection in any attention class; outputs go straight from attn@v to out_proj/proj.

QK normalization not in its codenot stated

The language-model attention paths (Florence2Attention family) have no q/k normalization; the only q_norm/k_norm code exists in…

show evidence

The language-model attention paths (Florence2Attention family) have no q/k normalization; the only q_norm/k_norm code exists in transformer.py's Attention class guarded by qk_norm flag, which is never set true or exposed via XVLAConfig/SoftPromptedTransformer, so it defaults off for this model.

self.q_norm = norm_layer(self.head_dim) if qk_norm else nn.Identity()
        self.k_norm = norm_layer(self.head_dim) if qk_norm else nn.Identity()
transformer.py · L117–118
Partial RoPE not in its codenot stated

No rotary position embedding is applied at all (text model uses learned absolute positional embeddings), so there is no partial RoPE…

show evidence

No rotary position embedding is applied at all (text model uses learned absolute positional embeddings), so there is no partial RoPE mechanism.

self.embed_positions = Florence2LearnedPositionalEmbedding(
            config.max_position_embeddings,
            embed_dim,
        )
modeling_florence2.py · L1486–1489
YaRN RoPE scaling not in its codenot stated

There is no rope_scaling key in the config and no RoPE-scaling code (YaRN or otherwise) anywhere in the provided source, consistent with…

show the whole verdict

There is no rope_scaling key in the config and no RoPE-scaling code (YaRN or otherwise) anywhere in the provided source, consistent with the model using learned absolute positions instead of RoPE.

Layers without positional encoding (NoPE) not in its codenot stated

All attention layers in the text model use the same learned absolute positional embedding scheme uniformly; there is no mixture of layers…

show the whole verdict

All attention layers in the text model use the same learned absolute positional embedding scheme uniformly; there is no mixture of layers with and without positional encoding.

Mixture of experts not in its codenot stated

Feed-forward blocks (fc1/fc2 in Florence2EncoderLayer/DecoderLayer, and Mlp in transformer.py) are single dense MLPs with no router or…

show evidence

Feed-forward blocks (fc1/fc2 in Florence2EncoderLayer/DecoderLayer, and Mlp in transformer.py) are single dense MLPs with no router or expert selection.

self.fc1 = nn.Linear(self.embed_dim, config.encoder_ffn_dim)
        self.fc2 = nn.Linear(config.encoder_ffn_dim, self.embed_dim)
modeling_florence2.py · L1239–1240
Shared expert not in its codenot stated

There is no MoE structure at all, so no shared-expert mechanism exists in this model.

Auxiliary-loss-free load balancing (selection bias) not in its codenot stated

No router or expert-selection bias mechanism exists since the model has no mixture-of-experts layers.

Multi-token prediction layers not in its codenot stated

No config key or code declares extra next-next-token prediction modules; the model predicts a single action chunk via diffusion-style…

show the whole verdict

No config key or code declares extra next-next-token prediction modules; the model predicts a single action chunk via diffusion-style denoising, not multi-token text prediction.

SwiGLU feed-forward not in its codenot stated

Feed-forward blocks use GELU activation (Florence2 text config activation_function is gelu, and transformer.py's Mlp uses nn.GELU), not a…

show evidence

Feed-forward blocks use GELU activation (Florence2 text config activation_function is gelu, and transformer.py's Mlp uses nn.GELU), not a SiLU-gated SwiGLU structure.

florence_config.text_config.activation_function = "gelu"
self.activation_fn = ACT2FN[config.activation_function]
modeling_florence2.py · L1237
self.act = nn.GELU(approximate="tanh")
transformer.py · L70

filed at the root 1

Flow-based vision-language-action model core

“a clean, flow-matching-based VLA design relying exclusively on soft-prompted standard Transformers”

multimodal architecture 7

Per-hardware-configuration learnable prompt embeddings core

“By associating each hardware configuration with learnable prompt embeddings, the model can flexibly capture domain-specific knowledge”

Streamlined encoding pipeline with soft prompts core

“Our design introduces a streamlined encoding pipeline that integrates soft prompts and explicitly disentangles high- and low-dimensional…”

Cross-attention-based resampler used

“we implement a cross-attention–based resampler that maps domain-specific observations into a shared representation space before feeding…”

Domain-specific resamplers and action heads used

“Each domain is assigned its own resampler and a dedicated action head, while the core Transformer backbone remains shared across all…”

Embodiment-aware language prompt concatenation used

“These descriptions are concatenated with the task instruction and processed by the Florence-Base encoder”

Pretrained vision-language encoder used

“A pretrained VLM encoder (Florence-Large (Xiao et al., 2024) in X-VLA) is used for the main vision-language stream”

Shared vision backbone used

“auxiliary views such as wrist-views are processed with a shared vision backbone”

6 more mentioned or evaluated
Action-modality isolation evaluated

multimodal architecture “We isolate the action modality from visual–language inputs.”

HPT-style domain-specific input projection evaluated

multimodal architecture “domain-specific projection layers are also applied on top of multi-modal inputs to align them before being fed into the backbone”

Multimodal Diffusion Transformer decoder evaluated

multimodal architecture “A multimodal variant of DiT that allocates separate parameters for different input modalities and integrates them through attention”

Natural-language hardware-configuration prompts evaluated

multimodal architecture “natural language descriptions of hardware configurations are provided as additional inputs”

Parallel MLP-Mixer-based action decoder evaluated

channel mixer “this design employs a parallel MLP-Mixer (Tolstikhin et al., 2021)–based action module alongside a pretrained VLM for vision–language…”

Separate embodiment-specific action projection heads evaluated

prediction head “This strategy addresses heterogeneity by assigning separate projection heads at the model output to map action tokens into…”

Not used 2

Heterogeneity-guided MoE routing not used

expert routing “we designed a heterogeneity-guided routing strategy that aimed to activate experts based on embodiment-specific cues.”

Load-balancing regularization not used

expert load balancing “To mitigate this, we give another try to introduce load-balancing regularization (Wang et al., 2024a), but the resulting rapid switching…”

filed at the root 1

Optimal-transport path for flow matching used

“To train the velocity field, we use the OT (optimal transport) path”

learning-rate schedule 1

Reduced learning rates for soft prompts and vision-language modules used

“use of a reduced learning rate for the soft prompts as well as for the vision–language modules”

training stability 1

Two-step prompt-and-head warm-up adaptation default

“All models are adapted using our proposed two-step procedure: during the first 1,000 iterations, only the soft prompts and action heads…”

data sourcing 1

DAgger-style data collection used

“we find it essential to adopt a DAgger-style data collection strategy (Ross et al., 2011)”

data mixture & curriculum 2

Cross-domain and cross-trajectory shuffling core

“this includes cross-domain and cross-trajectory shuffling, ensuring that the model is consistently exposed to a diverse and balanced…”

Weighted sampling of heterogeneous datasets core

“we adopt a weighted sampling strategy combined with a carefully designed shuffling pipeline.”

1 more mentioned or evaluated
Privacy-sensitive content filtering and bias mitigation mentioned

data filtering “we encourage future researchers to exercise caution when curating data for training large-scale robotics models, particularly by filtering…”

supervised fine-tuning 4

Embodiment-specific fine-tuning used

“Currently, deployment still relies on embodiment-specific adaptation, typically involving the collection of a small number of…”

Joint policy adaptation used

“Then, we jointly optimize both the backbone and the warmed-up prompt”

Periodic policy training with failure-targeted demonstration collection used

“we train ACT (Zhao et al., 2023) after every 100 collected episodes, identify its failure modes, and then collect targeted demonstrations…”

Prompt warm-up with the pretrained backbone frozen used

“The prompt is first warmed up while keeping the pretrained weights frozen.”

2 more mentioned or evaluated
Low-Rank Adaptation evaluated

supervised fine-tuning “we adopt PEFT techniques such as Low-Rank Adaptation (LoRA)”

Joint fine-tuning across multiple embodiments evaluated

supervised fine-tuning “we conduct joint finetuning experiments on a mixture of downstream datasets, including Libero, BridgeData, and Calvin-ABC”

Not used 1

Heterogeneous Low-Rank Adaptation not used

supervised fine-tuning “we experimented with Low-Rank Adaptation (LoRA)-style modules (Hu et al., 2022), where domain-specific adapters were introduced in…”

agentic scaffolding 1

Embodiment-aware language prompts used

“we provide embodiment-aware textual descriptions that encode hardware configurations and camera setups for each domain.”

filed at the root 1

Open-loop evaluation used

“We conduct open-loop validation experiments to monitor pretraining convergence and ensure fair comparisons across different architectures…”

benchmark 1

Held-out validation set construction used

“we carefully construct a dedicated validation set. Specifically, we sample trajectories from AGIBOT-beta (Bu et al., 2025) that are…”

filed at the root 2

Swinging motions for cloth smoothing used

“At that point, we employ swinging motions to complete the smoothing stage and then transition to stage II.”

Two-stage cloth-folding demonstrations used

“we decompose the folding task into two stages: (1) smoothing out the cloth from a highly disordered state, and (2) folding the smoothed…”

unfiled 15

Embodiment-specific learnable soft prompts core

“introducing separate sets of learnable embeddings for each distinct embodiment”

Embodiment-specific soft prompts core

“We introduce a soft-prompt mechanism that explicitly absorbs embodiment-specific variability while preserving a shared backbone for…”

Soft prompting with learnable embodiment-specific data-source embeddings core

“we assign a set of learnable embeddings to each data source as Soft Prompts”

Soft-prompted cross-embodiment robot learning core

“X-VLA introduces a Soft Prompt approach with minimally added parameters”

Standard Transformers with soft prompts core

“relying exclusively on soft-prompted standard Transformers”

Concatenation of proprioceptive states, action tokens, and time embeddings used

“we concatenate them along with their corresponding time embeddings within the flow-matching pipeline”

Repeated stage-I demonstrations until cloth keypoints emerge used

“we collect demonstrations for stage I in a repetitive manner until meaningful keypoints, such as the two corners or two ends of the cloth…”

Represent the next four seconds of intended trajectory with 30 anchor points used

“generate a sequence of 30 anchor points that summarize the intended trajectory over the next 4 seconds”

Rotate6D encoding for absolute end-effector rotation used

“the absolute EEF rotation encoded using the Rotate6D representation”

Soft prompt learning used

“In this work, we adopt the soft prompt paradigm for robotics”

Standardize actions as end-effector pose and discretized gripper state used

“we standardize the action space into end-effector (EEF) pose representation”

Temporal downsampling of actions to abstract action intentions used

“a simple temporal downsampling strategy can help abstract action intentions and thereby facilitate more efficient pretraining.”

Temporal downsampling of demonstrations into abstract action-intention representations used

“we temporarily downsample demonstrations to construct abstract representations of action intentions”

1 more mentioned or evaluated
Standard Diffusion Transformer decoder conditioned on multimodal features evaluated

“A direct application of the Diffusion Transformer (DiT) decoder (Peebles & Xie, 2023) that generates actions conditioned on multimodal…”

Not used 1

Domain-specific parameter learning with a shared backbone not used

“we initially approached this problem from the perspective of heterogeneous parameter learning. Concretely, we assigned distinct parameter…”

Sources

The curated source list for this model. "not fetched" means the URL is recorded but its text was not read in the current run.

SourceKindPublisherStatus
huggingface.co/2toINF/X-VLA-Ptmodel cardX-VLA authorsread
arxiv.org/abs/2510.10274technical reportX-VLA authorsread
↑