Causal retention improves adaptive control in interactive AI agents
Causal Retention in Interactive Agents: Interface Factorization and Selective Adaptation
Machine LearningArtificial Intelligence
Summary
Sometimes AI agents learn from their experiences, but it's hard to know what exactly they keep that helps them adapt later. The authors study a concept called causal retention, which looks at whether learned internal states can still answer questions about how their actions cause effects even when some conditions change. They built a system called Causal Core that carefully controls how the agent updates and reads its knowledge to keep reliable cause-effect information. Their experiments show this approach helps agents recognize when things like delays change without harming their overall task performance.
What this means in practice
- •For robotic control engineers: Improve robot controllers to recognize and adapt to changes in action delays while maintaining task stability.
- •For autonomous vehicle developers: Enhance models to maintain accurate causal predictions under changing environmental conditions or sensor delays.
Authors
Shengjun Zhang, Tingyi Liu, Dong Xie, Yunlong Dong, Xiang Wang, Cheng Zeng
Abstract
Task performance need not determine which intervention mechanism an agent retains. We study causal retention: whether a frozen learned state answers a mechanism-probe map fixed independently of training, including action, context, direct target, value, and delay. For finite structural causal model classes, the optimal probe error is a Bayes decision risk. It vanishes exactly when every learning-interface fiber lies within one probe-answer fiber; any state obtained by post-processing that interface inherits the same lower bound. A posterior-coverage theorem characterizes budgeted retesting, while an exact edit decomposition shows that the shifted set is the unique support of an error-free target update. Causal Core implements these conditions through evidence-gated writing, readout filtering, temporal credit, hidden-context setup, and local diagnostic updates. Experiments cover finite causal systems, continuous simulators, an official TD-MPC2 world model, and Qwen2.5-7B-Instruct. A frozen Qwen last-layer probe reaches 0.958 balanced accuracy on source mechanisms but 0.583 on changed delays; the gated mechanism state reaches 1.000 and accepts only 0.056 of synchronized-readout candidates. In TD-MPC2, five target states per actuator recover effect-sign accuracy from 0.057 to 0.948 without degrading stable responses. Causal retention is therefore distinct from task sufficiency and source-domain decodability.