Causal Retention in Interactive Agents: Interface Factorization and Selective Adaptation
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.
Related
- ParetoChat: An LLM-Based Interface for Interactive Multi-Objective Optimization
- Certified Behavioral Retention for Continual Adaptation with Held-Out Anchors
- A Domain-specific adaptation network for elliptic interface problems
- Sharp Probabilistic Normwise Backward Error Bounds for LU Factorization
- A Categorical Realization of the (2-)Category of Monoids via Sch{ü}tzenberger Categories and Strict Factorization Systems
BibTeX
@misc{zhang2026causal,
title = {{Causal Retention in Interactive Agents: Interface Factorization and Selective Adaptation}},
author = {Shengjun Zhang and Tingyi Liu and Dong Xie and Yunlong Dong and Xiang Wang and Cheng Zeng},
howpublished = {arXiv},
year = {2026},
month = sep,
doi = {10.48550/arxiv.2609.30650},
eprint = {2609.30650},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2609.30650},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Causal Retention in Interactive Agents: Interface Factorization and Selective Adaptation
AU - Shengjun Zhang
AU - Tingyi Liu
AU - Dong Xie
AU - Yunlong Dong
AU - Xiang Wang
AU - Cheng Zeng
PY - 2026
DA - 2026/09/25
PB - arXiv
DO - 10.48550/arxiv.2609.30650
UR - https://arxiv.org/abs/2609.30650
AB - 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.
ER -