Interpretability and Selective Unlearning in Evolving Contexts: Towards Explainable Agentic Adaptation

TL;DR: What if ACE could not only grow smarter playbooks, but also explain exactly why it made each adaptation—and let you “unlearn” specific strategies or facts on demand?

Research Question: How can we enhance the transparency, auditability, and controllability of ACE’s evolving playbooks, enabling users to trace, explain, and selectively remove or “unlearn” specific knowledge, strategies, or biases?

Hypothesis: Incorporating structured provenance tracking (e.g., metadata for each context edit, decision rationale, source attribution) and interactive unlearning tools will empower users to both audit the adaptation process and surgically remove problematic or outdated knowledge, improving trust and governance.

Experiment Plan: Design a provenance-aware ACE variant where each context update is tagged with source, rationale, timestamp, and impact. Develop an interface for visualizing the evolution of playbooks and for specifying patterns/facts to “unlearn.” Test the system on scenarios requiring unlearning of outdated/biased knowledge (e.g., regulatory/privacy requests, evolving domain norms). Measure ease and reliability of selective unlearning, impact on downstream task performance, and user interpretability via human studies.

References:

• Wang, S., Feng, J., Liu, T., Pei, D., & Li, Y. (2025). Mitigating Geospatial Knowledge Hallucination in Large Language Models: Benchmarking and Dynamic Factuality Aligning. Conference on Empirical Methods in Natural Language Processing.
• Chen, S., Gu, J., Han, Z., Ma, Y., Torr, P. H. S., & Tresp, V. (2023). Benchmarking Robustness of Adaptation Methods on Pre-trained Vision-Language Models. Neural Information Processing Systems.