Habituation & Sensitization: Decaying Focus on Repeated Stimuli
If you sit in a room with a ticking clock, you will eventually stop hearing it. This is habituation—a neurological process where response to a repetitive, non-threatening stimulus decreases over time. Conversely, a sudden loud sound will cause sensitization, sharpening your focus.
Spector applies habituation modeling to filter out repetitive, redundant prompts, preventing them from clogging the agent's short-term context.
The Habituation Mathematical Model
Each time a prompt template, keyword, or semantic tag is repeated in rapid succession, its habituation factor \(H\) decays:
\[H_{t} = H_{t-1} \times e^{-\eta \Delta t} + (1.0 - e^{-\eta \Delta t})\]
When a stimulus occurs, we subtract a habituation penalty:
\[H_{t} \leftarrow H_{t} - \Delta H\]
Where:
- \(\eta\) is the recovery rate (causing attention to recover during silent periods).
- \(\Delta H\) is the habituation impact per occurrence.
- \(H_t\) is bounded between \(0.1\) and \(1.0\).
Sensitization and Alert Re-activation
If a completely new, high-arousal stimulus occurs (such as an error or exception signature), the sensitization mechanism triggers. This resets the habituation vector for that semantic pathway:
\[H_{t} \leftarrow \min(1.0, H_{t} + S_{arousal})\]
Where \(S_{arousal}\) is the intensity of the novel stimulus.
This ensures that while routine logs and repetitive prompt wrappers are suppressed, critical system errors instantly bypass the filters and receive maximum attention.