3 papers
cs.LG2026
Reading the Gate, Not the Interference: Output-Side Interference Measurement Does Not Track Merge Collapse
Chencheng Zhu
Task-arithmetic merging works until it doesn't, and the field diagnoses why by measuring interference inside the merged model. We take the most direct such measure, the exact layer…
cs.LG2026
When Do Task Vectors Interfere? Mapping the Validity Boundaries of Weight-Space Composition
Chencheng Zhu, Xiaoyang Li, Taotao Cai
Task arithmetic composes skills by adding weight displacements, and merged models are then judged on benchmark suites. We measure when that composition is functionally additive, an…
cs.LG2026
A Shared Learning Rate Is Not a Neutral Control in Selective On-Policy Distillation
Chencheng Zhu
Selective on-policy distillation trains a student only at the token positions a selector scores highest, and the literature compares selectors under a single shared learning rate--…