WLA-0: Prototype for World-Language-Action Models

• The paper introduces the WLA-0 prototype that unifies world modeling, language reasoning, and action synthesis using an autoregressive Transformer backbone.
• It achieves state-of-the-art performance on benchmarks like RoboTwin2.0 Clean (92.94% success) and LIBERO (98.6%), with inference latency of ~40 ms on an RTX 5090.
• WLA-0 predicts both semantic subtasks and fine-grained physical dynamics, enabling long-horizon reasoning and effective cross-embodiment learning from multimodal inputs.

WLA-0 is a prototype world-language-action (WLA) model, introduced as an embodied foundation model that jointly performs world modeling, language reasoning, and action synthesis from multimodal robot inputs. It takes textual instructions, images, and robot states as inputs, and predicts textual subtasks, subgoal images, and robot actions, combining the world modeling interface associated with world-action models and the language reasoning capacities associated with vision-language-action models. Its core design choice is an autoregressive Transformer backbone that predicts the next state at both the semantic level and the level of fine-grained physical dynamics. The reported WLA-0 prototype has 2B active parameters, achieves 40 ms per inference on an NVIDIA RTX 5090, and is evaluated on both simulated and real-world robotic benchmarks, including 92.94\% success rate on RoboTwin2.0 Clean and 56.5\% success rate on RMBench (Yang et al., 4 Jun 2026).

1. Concept and scope

WLA-0 was proposed as the first instantiation of a new class of World-Language-Action models intended to unify three functions that had previously been emphasized separately: predicting physical dynamics, reasoning over instructions and subgoals, and synthesizing executable robot actions (Yang et al., 4 Jun 2026). The model is described as conjoining the world modeling interface used to learn from extensive egocentric videos with the language reasoning capacities needed for complex long-horizon tasks.

A central feature is that WLA-0 predicts the next state in two complementary forms. One is a semantic-level textual intention, expressed as textual subtasks. The other is a fine-grained physical dynamics representation that supports future-state prediction and action generation. This dual representation distinguishes the model from systems that only emit low-level controls or only perform high-level task decomposition.

The paper situates WLA-0 relative to WAMs and VLAs rather than treating it as a minor variant of either family. Its backbone is explicitly autoregressive, in contrast to the bidirectional diffusion Transformer used in WAMs. This design is intended to support both text generation and multimodal causal prediction in a single model (Yang et al., 4 Jun 2026).

A common misconception would be to view WLA-0 as merely a vision-language-action policy with an auxiliary prediction head. The reported formulation is stronger: the model is trained so that world modeling supervises a dedicated representation of physical dynamics, and that representation is then used to ease the characterization of the state-action correlation for the Action Expert (Yang et al., 4 Jun 2026).

2. Architectural organization

The architecture comprises three principal modules: an Autoregressive Transformer Backbone, a World Expert, and an Action Expert (Yang et al., 4 Jun 2026).

The Autoregressive Transformer Backbone is the core multimodal model. It is initialized from a large pretrained vision-LLM such as RynnBrain-2B, and processes images, text, robot states, and memory. According to the reported design, it sequentially predicts Textual Intention and Physical Dynamics. The backbone therefore serves both as a language reasoning engine and as a dynamics rollouter.

The World Expert is a lightweight diffusion Transformer, exemplified by SANA-600M, that receives the physical dynamics representation and the current state to predict a latent feature representation of the future observation. The model uses VAE features for this prediction. The paper states that the world modeling objective is auxiliary at train time and optional at inference time (Yang et al., 4 Jun 2026).

The Action Expert consumes the same latent dynamics and the current proprioceptive state to synthesize an action chunk. It is implemented as a flow-matching head. The reported decomposition implies that WLA-0 does not generate actions directly from raw observations alone; instead, it interposes a latent representation of physical dynamics shared with world modeling.

The model’s inputs and outputs are explicitly multimodal:

Component	Reported modalities
Inputs	Textual instructions, images, robot states, memory
Intermediate predictions	Textual intention, physical dynamics, meta-queries
Outputs	Textual subtasks, future images, action chunk

The architecture also introduces meta-queries, described as special tokens appended to the Transformer context. These tokens extract and aggregate relevant context through attention and generate the physical dynamics vector used by both the World Expert and the Action Expert (Yang et al., 4 Jun 2026).

3. Representation, objectives, and inference modes

The model predicts a window of subtasks according to

$$\mathcal{S}_t = f(\mathbf{o}_{t-h}, \mathbf{o}_t, \ell, \mathcal{M})$$

where $\ell$ is the original instruction and $\mathcal{M}$ is the memory buffer (Yang et al., 4 Jun 2026). The latent dynamics are then produced via meta-queries: $$\mathbf{h}_t = f(\mathbf{o}_{t-h}, \mathbf{o}_t, \ell, \mathcal{M}, \mathcal{S}_t, \mathbf{Q})$$ and this latent code is used by both experts.

The paper defines the action and world-model outputs as

$$\mathbf{a}_{t:t+n} = f_{\mathrm{act}}(\mathbf{h}_t, \mathbf{q}_t)$$

and

$$\mathbf{o}_{t+n} = f_{\mathrm{wm}}(\mathbf{h}_t, \mathbf{o}_t).$$

This organization makes the latent physical-dynamics representation a shared interface between semantic reasoning and control (Yang et al., 4 Jun 2026).

Training uses a three-term objective: $$\mathcal{L} = \mathcal{L}_\mathrm{act} + \alpha\,\mathcal{L}_\mathrm{wm} + \beta\,\mathcal{L}_\mathrm{lang}$$ with $\mathcal{L}_\mathrm{act}$ the flow-matching loss for the Action Expert, $\mathcal{L}_\mathrm{wm}$ the flow-matching loss for world modeling, and $\mathcal{L}_\mathrm{lang}$ the cross-entropy loss for subtask generation. The reported example values are $\ell$0 and $\ell$1 (Yang et al., 4 Jun 2026).

Two inference regimes are described. In Efficient Mode, the World Expert is disabled, leaving only the AR Backbone and Action Expert active; this is the mode associated with the reported ~40 ms per call on RTX 5090. In Test-Time Scaling (TTS) Mode, the model samples $\ell$2 candidate action chunks, uses the World Expert to imagine future images for each candidate, and scores them with a learned value model: $\ell$3 The best action chunk is then selected for execution. The paper further states that imagined frames can be fed back to extend the planning horizon (Yang et al., 4 Jun 2026).

A common misconception is that WLA-0 must always run explicit world prediction online. The reported design states the opposite: world prediction can be disabled during inference, while still allowing it to be activated for test-time scaling when improved control is desired (Yang et al., 4 Jun 2026).

4. Benchmark performance

The reported empirical results cover RoboTwin2.0, LIBERO, RMBench, and real-world robotic tasks (Yang et al., 4 Jun 2026).

On RoboTwin2.0 Clean, WLA-0 achieves 92.94\% average success. On RoboTwin Rand., the summary table reports 90.02\%. On LIBERO, WLA-0 attains 98.6\% average success, and with TTS this rises to 98.9\%. The paper describes these as state-of-the-art results for the reported settings (Yang et al., 4 Jun 2026).

On RMBench, which emphasizes long-horizon behavior and memory, WLA-0 achieves 56.5\% average success, compared with 28.5\% for the best prior Mem-0 baseline. The paper also reports that removing the language subtask loss reduces performance to 17.3\%, which is presented as evidence for the importance of language-decomposed subtasks in long-horizon execution (Yang et al., 4 Jun 2026).

The provided comparison table includes the following entries:

Method	RoboTwin Clean	LIBERO Avg
WLA-0	92.94	98.6
Motus	88.66	97.7
Fast-WAM	91.88	97.6

The real-world experiments emphasize dynamic and out-of-distribution tasks. On Dispose Trash, which is described as latency-sensitive and dynamic, WLA-0 is reported to outperform Motus and $\ell$4, with the explanation that those systems lag due to higher latency or lack of history dependence. The paper also states that WLA-0 reduces inference latency by ~40× vs. Motus (Yang et al., 4 Jun 2026).

These results support a specific interpretation of WLA-0’s contribution: the model is not only optimized for broad benchmark coverage, but also for the interaction between reasoning depth, history dependence, and control latency.

5. Long-horizon reasoning and memory

WLA-0 is explicitly designed for complex long-horizon tasks, and the reported mechanism for this capability is the joint use of textual subtasks, memory, and physical dynamics (Yang et al., 4 Jun 2026).

The backbone receives a memory buffer and outputs a window of intermediate subtasks aligned with the planning horizon. This means the model does not treat the instruction as a single immutable prompt; instead, it incrementally reformulates the task into local semantic intentions. The paper associates this with WLA-0’s strong performance on RMBench, where language-level memory and progress tracking are credited with the nearly twofold improvement over the best prior baseline (Yang et al., 4 Jun 2026).

The ablation in which the language subtask loss is removed is particularly important because it isolates the role of semantic decomposition. The drop from 56.5\% to 17.3\% is reported as highlighting the critical role of language-decomposed subtasks. This makes WLA-0 notable not simply for incorporating language tokens, but for making subtask prediction a supervised and quantitatively consequential component of the architecture.

A plausible implication is that WLA-0 treats language not merely as a conditioning signal but as an internal planning substrate. That interpretation is consistent with the model’s explicit prediction of Textual Intention before downstream action synthesis, though the paper’s concrete claim is limited to the reported module structure and ablation outcomes (Yang et al., 4 Jun 2026).

6. Learning from robot videos and cross-embodiment transfer

A prominent claim of WLA-0 is that it holds the promise to learn novel tasks directly from cross-embodiment robot videos without action annotations (Yang et al., 4 Jun 2026). The experiments addressing this point combine action-labeled seen tasks with action-free videos of unseen tasks, drawn from both the same and cross-embodiment robots.

For unseen tasks, adding same-embodiment action-free videos increases the baseline unseen-task success rate from ~13\% to 34.4\%. Using cross-embodiment action-free videos yields 28.8\%, which the paper characterizes as retaining most of the performance (Yang et al., 4 Jun 2026). These numbers are presented as evidence that WLA-0 can extract and transfer actionable knowledge from diverse robot videos even without explicit action supervision.

This claim should be interpreted carefully. The reported setup does not state that WLA-0 is trained without any action labels at all; rather, it is trained on action-labeled seen tasks and additionally on action-free videos of unseen tasks. The stronger statement is prospective: WLA-0 “holds the promise” of learning novel tasks directly from cross-embodiment robot videos without action annotations (Yang et al., 4 Jun 2026).

The broader significance of this result lies in the model’s intermediate representation. Because the AR backbone predicts semantic-level subtasks and fine-grained physical dynamics, and because the World Expert supervises future-state prediction while the Action Expert maps the same latent dynamics to control, the architecture is designed to separate task semantics from embodiment-specific motor trajectories. This suggests why cross-embodiment transfer is plausible, although the exact mechanism is described in the paper through the reported modules and results rather than through a formal invariance theorem.

7. Position within embodied model design

WLA-0 is positioned as a unification of two previously distinct emphases in embodied learning: the world modeling interface associated with WAMs and the language reasoning capacities associated with VLAs (Yang et al., 4 Jun 2026). Its defining architectural commitment is the use of an autoregressive Transformer backbone in place of the bidirectional diffusion Transformer used in WAMs.

This substitution matters because the AR backbone is described as capable of both sequential text modeling and latent distribution modeling, allowing a causal chain of the form “predict subtask $\ell$5 predict physical dynamics $\ell$6 predict action.” The World Expert remains diffusion-based, but it is moved into an expert head rather than used as the central sequence model (Yang et al., 4 Jun 2026).

The model also departs from approaches that bind world prediction to online control at every step. Through meta-queries, world prediction can implicitly influence action generation, even when the World Expert is disabled during inference. This yields a hybrid design: world modeling shapes the learned control interface during training, while low-latency action execution remains available at test time (Yang et al., 4 Jun 2026).

Several boundaries of the current evidence should be stated neutrally. The paper reports a WLA-0 prototype with 2B active parameters, not a universal or final WLA system. The strongest claims are supported on the listed benchmarks and real-world tasks. The statement regarding direct learning from cross-embodiment robot videos without action annotations is framed as a promise rather than as a fully closed experimental conclusion (Yang et al., 4 Jun 2026). Within those bounds, WLA-0 is presented as a unified embodied model in which semantic reasoning, future-state modeling, and low-level action synthesis are trained as a single coordinated system.