USTC's Abductive World Model Gives Robots Causal Reasoning, Runs on Edge at 2-bit

Abductive World Model: Predict Then Reason Backwards

Traditional world models only predict the next frame. Baize Tongjing's AWM (Abductive World Model) adds a second step: after predicting, it performs abductive inference to explain why the predicted change occurs. The model decomposes the world into three questions:

Who is present? (Entity) – which objects appear in the scene.

How do they move? (Dynamic) – how each object evolves over time.

What will collide? (Relation) – which objects interact.

This yields a structured "world manual" rather than a blurry future frame, enabling robots to decide when to reach, which box arrives first, and whether a collision will occur.

Learning from Unlabeled Video with High Data Efficiency

AWM learns object, motion, and interaction patterns directly from massive unlabeled video, avoiding costly frame-by-frame annotation. On the Push-T robot manipulation benchmark:

Reaching 75% success rate requires only 2,000 trajectories for AWM, versus 14,728 for Causal-JEPA (7.3× fewer).

When both are trained on 14,728 trajectories, AWM achieves ≈92% success , significantly higher.

Occlusion tests show model predictions change 3.9× more when key objects are masked versus irrelevant objects, confirming predictions are grounded in recognized entities.

The team has published the paper "Abductive World Modeling via Causal Representation Learning" (arXiv:2609.36985) and open-sourced code at https://github.com/baiz-tech/AWM.

Comparison with JEPA-line World Models

AWM shares the latent-space prediction paradigm of Yann LeCun's JEPA family (e.g., Meta's V-JEPA 2 ), but differs in three key aspects:

Abductive step – explicit entity/dynamic/relation decomposition for explainable predictions.

Higher data efficiency – 7× fewer trajectories for equivalent Push-T performance.

Edge-first design – coupled with BAIZ-RUN for 2-bit quantization and multi-chip deployment (NVIDIA, AMD, Huawei Ascend).

Experiments show AWM outperforms V-JEPA 2 on physical prediction, event reasoning, and action understanding; action-understanding Top-1 accuracy improves by 68% .

BAIZ-RUN: 2-bit Quantization Engine for Edge Deployment

Running world models on robot-edge chips demands extreme compression. BAIZ-RUN achieves:

Weights quantized to 2-bit , activations to 4-bit via error-aware quantization.

Operator fusion and low-level compute optimizations to cut redundant ops and memory moves.

Multi-backend support: single artifact runs on NVIDIA, AMD, and Huawei Ascend, enabling domestic-hardware compatibility.

Model-agnostic: also targets V-JEPA, DINO-WM, VLA, and other multimodal models, aspiring to be a vLLM-like universal inference substrate.

On AMD Radeon integrated edge GPU with SSV2 dataset:

Memory footprint reduced 2.56× .

Inference frequency increased 94.6% .

Accuracy loss <1% .

Edge Compute Terminal: Plug-and-Play Embodied Brain

Baize plans a hardware terminal integrating AWM + BAIZ-RUN on domestic compute platforms, featuring:

Low cost – 2-bit quantization keeps per-unit cost controllable for mass deployment.

Edge reliability – fully offline inference, no cloud dependency.

Data never leaves factory – local closed-loop for security-sensitive sites.

Plug-and-play – standardized interfaces to reuse the same brain across robot form factors.

The company's positioning: "Robot OEMs provide the body; Baize provides the brain."

Real-World Validation in Three Industrial Scenarios

Within months of founding, Baize has signed and validated in:

Industrial loading/unloading – 5-second anomaly detection and replenishment on moving conveyor lines using dynamic prediction and fast decision-making.

Industrial quality inspection – generalized grasping and inspection across multiple SKUs, testing recognition and spatial localization.

Factory patrol – real-time obstacle avoidance for unknown obstacles via motion prediction and path planning.

Ecosystem Partnerships and Data Flywheel

Scenario partners: Wanyu Technology (supplier to Mercedes-Benz, Huawei, BYD), Weigang (top-20 China dairy), and a Luxshare Precision affiliate.

Robot-body partners: Zerith (Tsinghua-affiliated), Xiaobu Intelligence (USTC-affiliated), LimX Dynamics (pre-IPO).

Data collection base: Leveraging the National Advanced Manufacturing Center (国先中心) to join a municipal-scale data acquisition hub, creating a bidirectional loop where scene data refines the model.

Founder Bai Yinqi argues embodied AI cannot follow the LLM "general-first, deploy-later" path; physical-world variance demands scenario-driven entry then gradual generalization.

Funding and Recognition

Seed round of tens of millions RMB closed within 3 months of incorporation; lead investor Zhongxin Juyuan (semiconductor-focused) plus market-oriented funds with USTC Silicon Valley LPs.

Follow-on round in term-sheet stage; post-money valuation expected to exceed 400 million RMB .

Selected for Great Wall Strategy Consultants' "2026 China Tech Future Stars Report" (97 companies nationwide, 6 from Hefei).

Business model spans model teams (data pipeline), existing-model companies (BAIZ-RUN only), robot OEMs (full brain), and end-users (co-development with OEMs).

Outlook

Open challenges remain: AWM robustness in novel environments, BAIZ-RUN performance across more chipsets, and transition from pilot to volume deployment. However, the direction is clear: world models must leave papers and labs to become deployable embodied brains.