Lagged Coupling: Internal Representations Become Readable Before They Become Causal
Across the full Pythia suite (160M-12B, eight checkpoints, four task families), a linear probe can read a target variable from the residual stream as early as step 1,000 at every scale, yet steering along that same reading direction remains null-equivalent in 43 of 48 model-checkpoint cells. Internal readability systematically outruns causal efficacy, and the lag does not shrink with scale.
The study decomposes lagged coupling into three dissociable tracks: internal readability, saturated (AUROC >= 0.990) from the first checkpoint everywhere; behavioral readability, which develops gradually and progressively later at larger scales (12B reaches 0.909 only at the final checkpoint); and causal efficacy, almost always null-equivalent, occasionally counterproductive early, with one isolated positive pulse (12B, step 8,000, z = +2.49) the grid cannot resolve. The ordering is dominantly read-before-write (11/11 units, no inversion). Representation headroom along the probe direction grows up to 57x with training and scale while causal write-in stays below 0.11% of headroom.
The evidence indicates a systematic dissociation between linear readability and causal steerability in transformer residual streams. A linear probe achieves near-perfect decoding of a target variable from early training, but interventions along the probe direction do not produce corresponding behavioral changes. This suggests that the model encodes information in directions that are not yet causally integrated into downstream computation. The observed growth in representation headroom (up to 57x) without proportional causal write-in implies that the model accumulates latent representational capacity before it learns to use it for behavior. The isolated positive pulse at 12B, step 8,000 (z = +2.49) is a potential signal of a phase transition or checkpoint-specific effect that warrants further investigation with finer-grained checkpoints and multiple seeds.
This finding challenges assumptions that interpretability tools like linear probes directly reveal causally relevant features. For AI safety and alignment, it implies that a model may internally represent concepts long before those representations influence outputs, complicating efforts to steer models by editing internal activations. For model developers, the lag between internal readability and behavioral readability suggests that capability emergence may be delayed relative to representational learning, which could inform training curricula or early stopping criteria. The lack of scale-dependent reduction in the lag indicates that simply increasing model size does not resolve the dissociation, pointing to a need for architectural or training innovations to align internal representations with causal pathways.
For AI companies, understanding lagged coupling could improve model steering and alignment techniques, reducing the risk of deploying models that internally represent harmful or unintended concepts without behavioral control. It may also inform more efficient training by identifying when representations become behaviorally relevant, potentially saving compute. For interpretability tooling vendors, there is an opportunity to develop causal probing methods that go beyond linear readability. The finding may also influence regulatory discussions about model transparency, as it shows that internal representations do not straightforwardly map to model behavior.
Observable next signals include: (1) replication of the lagged coupling effect in other model families (e.g., Llama, GPT) to test generality; (2) finer-grained checkpoint analysis around the 12B, step 8,000 positive pulse to determine if it is a reproducible phase transition; (3) development of intervention methods that target causal directions rather than probe directions; (4) investigation of whether lagged coupling correlates with downstream task performance or safety-relevant behaviors; (5) exploration of training objectives that explicitly encourage causal integration of early-readable representations.