🤖 AI Summary
This study investigates the precise representational locus of behaviorally relevant states in large language models, moving beyond conventional approaches that rely solely on prompt-based interventions followed by successful post-training. Through controlled routing tasks—augmented with support data selection, held-out query evaluation, and necessity/sufficiency ablation experiments—the work distinguishes, for the first time under comparable conditions, between fixed interface reuse and prompt-position shifting, demonstrating that the former provides stronger evidence for state reuse. Methodologically, it introduces interface-matching controls, zero-retraining compiled transfer, trainable prompt-slot optimization, and a generation–reasoning branch analysis. The approach achieves precise state transfer on GPT-2 in Triop and arithmetic tasks, with fixed interfaces recovering most routing accuracy, and further validates cross-architectural consistency in Qwen models.
📝 Abstract
Prompt-based interventions can change model behavior, but trained success alone does not identify where the behaviorally relevant state is represented. We study this question in controlled routing tasks using interfaces chosen on support data, held-out query evaluation, and matched necessity, sufficiency, and wrong-interface controls. On GPT-2 triop, an early interface supports exact transfer under these tests. On GPT-2 add/sub, zero-retrain compiled transfer at the fixed interface recovers most of donor routing accuracy, while trainable prompt slots can relearn the same behavior at several other positions only after additional support examples and optimization. These results distinguish fixed-interface reuse from prompt relocation in a setting where the two can be tested directly. Qwen routing provides a cross-architecture consistency check for the same matched-interface pattern at the operator token, although donor-specific identity on the local V-path remains unresolved. Generation and reasoning branches are used to map scope: they show broader transport or weaker controller identifiability once control depends on longer trajectories or harder selection. In controlled routing, fixed-interface transfer is therefore stronger evidence of reuse than trained prompt success alone.