Does monitoring help more by choosing what to correct than when to intervene?

The paper studies "behavioral drift, a progressive decline in persona consistency" in LLM-simulated ADHD student personas, across 1,200 28-turn conversations, four LLMs and two persona intensities. It compares five prompt-level mechanisms that vary when to intervene (static versus adaptive) and what to inject (reinjection versus reflective reminder), plus a novel adaptive condition where a monitor generates behavior-specific instructions. Relative to no intervention, reinjection reduced modeled LLM-rated drift by 35–38% and reflective reminders by 22–27%. Behavior-specific instruction reduced it by 87%. The authors report "no evidence that adaptive timing outperformed static scheduling," and conclude that monitoring "appears more useful for deciding what to correct than when to intervene."

The discussion frames this as two different uses of the same monitor. When monitoring "only determined when to repeat fixed intervention content," the overall trajectory was no better than a fixed schedule. When "behavioral measurements shaped what the intervention addressed," the result was the largest reduction, a flatter trajectory than either of the other adaptive conditions. That advantage "appeared from checkpoint 3 onward" and held in both intensity strata. In the authors' words, monitoring "was therefore not uniformly beneficial; its value depended on how the measurement informed the intervention." The effect is also bounded: drift continued to decline in every condition, including the best one.

This sits beside Can training user simulators reduce persona drift in dialogue? as a prompt-level counterpart. That work changes the simulator's weights with consistency rewards, while this one leaves the model alone and edits the context mid-conversation. The two share the premise that drift is a measurable, correctable failure rather than a fixed property of simulation, and the earlier note's split into three drift types matches the idea of correcting specific behaviors. It also echoes Why does supervised learning fail to enforce persona consistency?: in both, the corrective signal pays off when it names the inconsistency. The same weights-free logic appears in Why do language models lose performance in longer conversations?, where a separate component recovers multi-turn performance without weight updates. The 55% and 87% figures come from different metrics and setups, so they should not be ranked against each other.

The excerpt leaves several things open. It does not define "post-knot decline rate," the drift rater, the monitor's design, or what the four models were. Nor does it say how the behavior-specific instructions were worded, and the authors themselves note that this condition "requires component-level testing." That means the 87% figure cannot yet be attributed to the monitor's current-state information rather than to instruction specificity, since the behavior-specific condition differs from the other conditions in both respects. Drift was rated by an LLM, and the excerpt does not report validation against how real students with ADHD behave. What the evidence supports is a design lean: when building a drift monitor for simulated personas, spend effort on what it tells the persona, not on scheduling when it speaks. It also supports treating any single intervention as a delay of drift rather than a fix.