Wang, X., Hu, Y., Du, W., Cheng, R., Wang, B. & Zou, D. (2025). Towards understanding fine-tuning mechanisms of LLMs via circuit analysis. ICML 2025. arXiv:2502.11812.
Read in full on 2026-10-26 (NOTE-1183). One account was drawn from it: THEORY-997. The identification from T24 was checked against the reading and holds (see the history).
Why this is here
Cited at T24 for the point that the relations among a network’s components can reorganize while the components stay the same, which the reply reads as the network analogue of members reorganizing into new perspectives. It bears on QUESTION-275, and on CLAIM-367 (which unit a training-time condition belongs to depends on how units are individuated).