Why Action‑Centric World Models Outperform Generalist: The GigaWorld‑Policy Breakthrough

Background: Goal vs. Idea in World Models

Generalist’s recent long‑form article "Going Beyond World Models & VLAs" argues that goals are more important than tool labels, urging a return to the core purpose of enabling machines to act efficiently and accurately in the physical world rather than debating whether to build VLA (vision‑language‑action) or pure world models.

Limitations of Conventional World Models

Two fatal flaws are identified:

Goal misalignment : video generation is treated as the goal, while high‑frequency, precise action output is the real objective, leading to bloated architectures and mismatched compute resources.

Real‑world constraints : rendering high‑dimensional pixels incurs huge computational cost, causing intolerable inference latency and error propagation from video prediction to action sequences, ultimately breaking physical interaction.

The authors at 极佳视界 assert that any design requiring extensive computation unrelated to the final goal cannot be optimal; embodied intelligence needs a “practical” rather than a “fantasist” approach.

GigaWorld‑Policy: Action‑Centric World Model

极佳视界 introduced GigaWorld‑Policy, an open‑source world model that shifts video generation from a mandatory component to an optional one during inference.

Training “strict teacher” : The model receives dual supervision—action prediction and video generation—using massive internet video data. Video generation acts as a rigorous auxiliary task that forces the model to internalize dynamics consistent with real physics.

Inference “Action‑Only” mode : When deployed, the video generation module retreats completely, allowing the model to switch to a pure action‑output mode that directly issues high‑frequency control commands.

This redesign eliminates the compute burden of rendering pixels, aligning architecture with the true goal of action execution.

Alignment with Real‑World Constraints

GigaWorld‑Policy embodies the principle: if inference must perform large amounts of computation unrelated to the goal, the design is sub‑optimal. The model treats physical laws like an intuitive “subconscious” rather than rendering every visual detail.

Data Efficiency via Transfer Scaling Law

OpenAI’s Transfer Scaling Law (2023) shows that performance on a target task depends not only on model size but also on the alignment between pre‑training (source) data distribution and target‑task data distribution. GigaWorld‑Policy’s action‑centric design ensures that the representations learned during pre‑training are naturally aligned with downstream action‑output tasks, dramatically reducing transfer penalty.

Three‑Stage Training Pipeline

Physical Knowledge (source‑domain pre‑training) : Leverages massive internet videos to teach the base model broad physical commonsense and visual representations.

Spatiotemporal Adaptation (cross‑domain) : Incorporates first‑person, real‑robot, and simulation videos to narrow the distribution gap between source and target domains.

Precise Alignment (target‑domain fine‑tuning) : Requires only a small amount of labeled real‑robot action data to finalize the control policy.

Experimental Results

Using just 10 % of real‑robot data, GigaWorld‑Policy matches the performance of traditional VLA approaches that consume 100 % of the data, delivering a ten‑fold improvement in training efficiency.

On an A100 GPU, inference speed reaches 360 ms per step, with lower memory consumption and a ten‑fold speed advantage over Motus.

In real‑world task evaluations, GigaWorld‑Policy achieves an average success rate of 83 %, nine times faster than Motus and 7 percentage points higher in success.

Open‑Source Release and Impact

GigaWorld‑Policy, its code, and the accompanying paper are fully open‑sourced (project homepage, GitHub repository, and arXiv link). Earlier work, GigaWorld‑1, topped the WorldArena benchmark with a composite score exceeding 60, surpassing Google, NVIDIA, and Alibaba, and its code/dataset have been downloaded over 24 k times on HuggingFace.

The shift from conceptual debate to concrete architectural innovation signals that embodied intelligence has moved beyond proof‑of‑concept, and action‑centric world models may become a milestone on the path toward physical AGI.