Dynamic Abstention: Temporal Decay and Recovery of Concept Boundaries in LLMs

TL;DR: What if abstention and concept boundary maintenance in LLMs isn’t static, but changes over a conversation—sometimes decaying, sometimes recovering? A longitudinal probing study could reveal such dynamics and suggest training fixes.

Research Question: How stable are LLMs’ internal representations of concept boundaries (e.g., “dead vs. alive”) over multiple conversational turns, and can we model or intervene in their temporal decay/recovery?

Hypothesis: LLMs’ concept boundary encoding degrades over extended conversations, especially under sustained incongruence, but could be stabilized through targeted prompts or memory-augmented architectures.

Experiment Plan: Setup: Design multi-turn dialogues where concept-incongruent prompts are introduced and revisited over time (e.g., recurring references to a “dead” character).
Methodology: Use probing to track the internal representation of the relevant concept boundary across turns.
Interventions: Test “reminder” prompts or external memory mechanisms to reinforce boundaries.
Analysis: Measure abstention rates, accuracy, and representational drift over time.

References:

• Bai, X., Peng, I., Singh, A., & Tan, C. (2025). Concept Incongruence: An Exploration of Time and Death in Role Playing. arXiv.org.
• Sun, C.-E., Oikarinen, T. P., Ustun, B., & Weng, T.-W. (2024). Concept Bottleneck Large Language Models. International Conference on Learning Representations.