Researchers often employ truth probes to test whether a layer’s activations encode factual information. When a probe’s prediction labels coincide with the task’s prescribed action labels, the probe can only rely on the labels themselves and cannot distinguish truth from action—this phenomenon is termed perfect aliasing.
In a controlled binary‑reporting game, truth and prescribed‑action probes fitted on compliant contexts (where the model’s output matches the ground‑truth) solve the same optimization problem. When the same probes are evaluated on rival contexts (where the model’s output opposes the ground‑truth), their labels become complements, forcing their AUROCs to satisfy $$\text{AUROC}_{\text{truth}} + \text{AUROC}_{\text{action}} = 1$$. This identity holds across 751 cell‑layer pairs to floating‑point precision.
To disentangle output symbols from semantic actions, we introduce randomized codebooks and then fit probes on a mixture of compliant and rival contexts. Evaluating a reward‑trained Gemma‑2‑9B policy, the conventional probe scores $0.006 \pm 0.005$ AUROC on all rival trials (averaged over three seeds), whereas the mixed‑fit probe achieves a perfect 1.000 AUROC on the same held‑out activations.
It is important to note that mixed fitting uses more training examples and has access to labeled rival contexts; thus the comparison demonstrates linear recoverability rather than a pure decorrelation benefit.
Further, two probes that are perfect on‑distribution can diverge dramatically on rival activations, scoring $0.080$ and $0.986$ respectively.
These findings highlight that probes primarily measure label correlation and do not guarantee preserved functional belief, causal use of the recovered direction, or a deployable deception detector. The accompanying code and aggregated results are released with the paper.
Review: The work convincingly exposes the perfect aliasing issue inherent to probes trained on compliant contexts and proposes a mixed‑context fitting remedy. However, establishing a causal link between probe signals and model intent remains an open challenge, calling for deeper investigations beyond correlation metrics.