Yijun Shen (沈逸骏)

Master student @ECNUER. Intern @BAAIBeijing.
Working on World Model, Embodied Agent.

[Yijun Shen (沈逸骏) @bui_thanh64689]

CausGround turns interaction trajectories into reusable action preconditions and effects.

As more data accumulates, it can ground richer causal knowledge in the current world state, helping a frozen planner reason backward over preconditions and forward over effects.

[Yijun Shen (沈逸骏) @bui_thanh64689]

Embodied planning needs to know when actions are executable and what they change.

Introducing 𝗖𝗮𝘂𝘀𝗚𝗿𝗼𝘂𝗻𝗱, our work at BAAI on causal interaction grounding for adaptive planning without model updates.

Paper: https://t.co/V5aW9z70dO
Project: https://t.co/XTYSqvSfbp

[Yijun Shen (沈逸骏) @bui_thanh64689]

4. We demonstrate a promising path toward scaling configurable perception across diverse real-world video data and four large visual backbones.

[Yijun Shen (沈逸骏) @bui_thanh64689]

Our latest paper, “World Models with Configurable Perception”, explores configurable states as information bottlenecks for robust planning in noisy worlds.
⬇️
Paper: https://t.co/YqvjITUeSp
Web: https://t.co/piXqdM2nhK
Code: https://t.co/6VnlnRnNJe
Data: https://t.co/7J8f74Qsbt

[Yijun Shen (沈逸骏) @bui_thanh64689]

3. Physical-quantity probing and distracted-to-clean retrieval show that ConfigWM better preserves task-relevant information.

[Yijun Shen (沈逸骏) @bui_thanh64689]

2. ConfigWM achieves more robust and stable planning than task-agnostic baselines in noisy worlds.

[Yijun Shen (沈逸骏) @bui_thanh64689]

Experimental Results:
1. Task-agnostic representations may retain both task-relevant and irrelevant dynamics, while ConfigWM preserves the task-relevant states according to language queries.

[Yijun Shen (沈逸骏) @bui_thanh64689]

@DavidSunok47 Good question! In term of this, I think the best way is to enrich the data, especially the causal data.

[Yijun Shen (沈逸骏) @bui_thanh64689]

Inspired by the configurator view of autonomous intelligence, we introduce ConfigWM: World Models with Configurable Perception! By using language to configure perception, ConfigWM constructs task-relevant predictive states and learns action-conditioned dynamics for planning.

[Yijun Shen (沈逸骏) @bui_thanh64689]

World modeling needs configurable perception.
A good state should act as an information bottleneck: preserving just enough task-relevant information for planning while filtering out irrelevant distractions.