Adaptation, Not Priming: A Repulsive In-Context Aftereffect Carried by a Sparse Set of Attention Heads

In-context learning is assumed to assimilate: repeated exposure pulls the next prediction toward the repeated pattern, as induction and copy mechanisms attest. We report the opposite sign for graded scalar attributes, and establish three things about it. The phenomenon. Across five attribute constructs (formality, sentiment, warmth, arousal, politeness) and five open-weight models, an ambiguous probe's judgment shifts away from the adaptor pole on all twenty-five construct-by-model cells, bootstrap intervals excluding zero throughout and effect sizes reaching |d| = 4.6: a repulsive aftereffect opposite in sign to priming. The carrier. Two-stage activation patching isolates eight attention heads whose mean-ablation abolishes the recoil on two independently pretrained full-attention substrates (91%, 7.66 to 0.72; 99.8%, 8.10 to -0.02) and attenuates it on a grouped-query substrate (37%), replicates on held-out probes, and exceeds all 100 random head-set controls on every substrate — with the norm matching fully satisfied (100 of 100 draws) on OLMo-2-7B and only approximated on the other two. The behavioral law. The causal contribution grows monotonically with exposure; the recoil builds up and saturates, relaxes along a fitted exponential as neutral material intervenes (R² ≥ 0.92), acts far more strongly on the automatic route than on deliberate comparison (removing 88 to 99.8% of the recoil costs explicit comparison at most 10 points), and is weaker for a pole the model already leans toward. That profile is consistent with adaptation accounts from psychophysics — divisive normalization in particular, which we adopt as an organizing hypothesis rather than an identified computation. Two pre-registered extrapolations came out against us: the same ablation does not remove batch-order drift in a judge pipeline, and head-set overlap does not predict cross-construct transfer, which stimulus coupling dominates instead.