Priors Persist Through Suppression: A Stroop Paradigm for Lexical Override (arxiv.org)
arXiv:2606.07555v1 Announce Type: cross
Abstract: Glossaries, technical specifications, and system prompts routinely ask language models to use familiar words in unfamiliar ways. When this works, the lexical prior persists through override rather than being replaced: it continues to operate after the local rule applies, with the rule lowering its logit rather than installing the new meaning on top. We test this with a Stroop-style paradigm: a remapping rule ("doctor" means "forest") pitted against the query word's lexical-prior distractor ("hospital"), with matched neutral controls. Across 11 open-weight models spanning four families and 1B--9B parameters, lexical-prior strength predicts interference even after item-level controls for answer prior, frequency, tokenization, and prompt wording. Activation patching on five aligned models locates a source-position triplet (definition subject, definition target, query word) that nearly fully recovers the conflict effect (aggregate $R \in [0.92, 1.06]$). A definition-target swap shows the triplet performs binding rather than identity matching. Dissociation experiments isolate target preservation as the binding-specific signature: distractor suppression occurs under matched, swap, and item-mismatched conditions alike, whereas target logit collapse occurs only when the definition-target position is corrupted. Behavior and mechanism converge on the same channel: the lexical prior is where both interference originates and where override leaves its mark.
Abstract: Glossaries, technical specifications, and system prompts routinely ask language models to use familiar words in unfamiliar ways. When this works, the lexical prior persists through override rather than being replaced: it continues to operate after the local rule applies, with the rule lowering its logit rather than installing the new meaning on top. We test this with a Stroop-style paradigm: a remapping rule ("doctor" means "forest") pitted against the query word's lexical-prior distractor ("hospital"), with matched neutral controls. Across 11 open-weight models spanning four families and 1B--9B parameters, lexical-prior strength predicts interference even after item-level controls for answer prior, frequency, tokenization, and prompt wording. Activation patching on five aligned models locates a source-position triplet (definition subject, definition target, query word) that nearly fully recovers the conflict effect (aggregate $R \in [0.92, 1.06]$). A definition-target swap shows the triplet performs binding rather than identity matching. Dissociation experiments isolate target preservation as the binding-specific signature: distractor suppression occurs under matched, swap, and item-mismatched conditions alike, whereas target logit collapse occurs only when the definition-target position is corrupted. Behavior and mechanism converge on the same channel: the lexical prior is where both interference originates and where override leaves its mark.
Comments