OwlSight investigates whether one reasoning monitor can look for different kinds of behaviour without being retrained for each objective. The paper calls the requested monitoring condition a Factor: a natural-language instruction describing what to inspect in a chain-of-thought trace.
The distinction is practical. A broad request to detect privacy leakage and a specific request to detect disclosure of a particular email address are related, but they are not the same monitoring task.
Specify the behaviour at the right level
The framework distinguishes fine-grained and coarse-grained factors. OwlSight is trained on traces annotated at both levels, across capability, safety and trustworthiness objectives. The paper describes 11 categories in its training data.
Changing the factor changes the question asked of the trace. This makes the monitor's instruction an important part of the evaluation, rather than an incidental prompt wrapped around a fixed classifier. A broad factor may expose a class of failures, while a narrow one can test a concrete requirement.
Read the results as monitoring measures
The evaluation uses 32,222 factor-conditioned queries over 15,971 reasoning trajectories. The paper reports macro-average G-Mean squared scores of 69.57 at fine granularity and 66.80 at coarse granularity. Those are monitoring metrics, not the monitored agent's task-success rate.
Against the evaluated general-purpose models, the reported improvements range from 1.32 to 30.57 points at fine granularity and 11.71 to 27.81 points at coarse granularity. The authors also report testing on unseen models' reasoning traces, rephrased factors, compressed traces and obfuscation.
These comparisons support the paper's monitor-focused approach within the tested conditions. They do not establish that every harmful intention appears in the text available to a monitor, or that a successful detection automatically prevents an action.
Connect a detection to an operational decision
Our interpretation is that a factor-based monitor is most useful when its output has a defined consequence. An evaluator should know whether a positive finding triggers an alert, a pause, a review or a narrower follow-up check.
For a reproduction, use the same trace with both a broad and a specific factor, then inspect the evidence behind each result. Include harmless traces that resemble the target behaviour to understand false alarms. Also test what happens when the trace is missing relevant information.
OwlSight contributes a way to vary the oversight question without rebuilding the monitor. The larger oversight system still needs reliable trace access, evidence review and a clear response to a detection.
Comments