Attention Fragmentation: Reward-Biased Over-Promotion of Speculative Branches
by Nick Clark | Published March 27, 2026
When the reward signal biases promotion thresholds downward, speculative branches are promoted too readily. The agent pursues many partially developed ideas simultaneously, unable to sustain focus on any single execution path long enough to complete it. This attention fragmentation pattern is the architectural analog of ADHD-like behavior, arising from a structural promotion bias rather than a character flaw.
What It Is
Attention fragmentation occurs when the affective reward signal reduces promotion thresholds to the point where speculative branches are advanced before they are adequately evaluated. The agent begins many execution paths but completes few because each new speculative branch is promoted before the previous one reaches conclusion.
The fragmentation is self-reinforcing: the novelty of each new branch provides a reward signal that further reduces promotion thresholds, creating a cycle of increasingly rapid branch switching.
Why It Matters
Attention fragmentation is one of the most common disruption patterns in cognitive architectures with reward-modulated promotion. It produces agents that appear active and engaged but accomplish little because no execution path reaches completion. The agent is not lazy or unmotivated; its promotion thresholds are miscalibrated.
How It Works
The reward signal from novel branch initiation feeds back into the promotion threshold calibration. In healthy operation, this feedback is bounded. In attention fragmentation, the feedback amplifies: novelty reward reduces thresholds, which produces more branch switches, which produces more novelty reward.
Detection involves monitoring branch completion rates, average branch lifetime, and promotion threshold trajectory. A sustained decrease in branch lifetime with increasing branch count signals attention fragmentation.
What It Enables
Understanding attention fragmentation as a promotion threshold miscalibration enables precise intervention: recalibrating the reward-to-threshold feedback loop, introducing minimum branch lifetime requirements, or temporarily increasing promotion thresholds to force existing branches to completion before new ones are initiated.
