BAAI 等机构发布 CausGround：免微调提升具身规划能力
BAAI 带几家高校搞了个 CausGround，不用微调就能让 Qwen3.5 这些冻结模型做具身规划性能翻倍，搞智能体的可以看看。
BAAI 联合华东师大、西北工业大学、国科大、CASIA 推出 CausGround，一个即插即用的因果交互接地框架。它通过轻量符号层把原始交互蒸馏成可复用的 Action Preconditions & Effects，让智能体显式推理动作前置条件与后果，绕开端到端模型如 Astra 的高延迟和黑盒问题。该框架零参数更新、零微调，可直接挂接在冻结的 Qwen3.5、InternVL3.5 和 Gemma 4 骨干模型上，实验中使具身规划性能翻倍。
Facing the high latency and black-box bottlenecks of end-to-end models like Astra, @BAAIBeijing, ECNU, NPU, UCAS & CASIA officially introduce CausGround — a plug-and-play causal interaction grounding framework that doubles embodied planning with ZERO parameter updates!
What remains truly timeless in the embodied harness? An external causal add-on that continuously evolves with data. CausGround introduces a lightweight symbolic layer—distilling raw interactions into reusable Action Preconditions & Effects so agents explicitly reason about "what is required before" and "how the world changes after" without trial-and-error.
With ZERO fine-tuning or parameter updates, it acts as a plug-and-play causal engine for frozen Qwen3.5,InternVL3.5, and Gemma 4 backbones: 🌐 Project:https://t.co/20ynYqmSgs
