Causal mechanisms of online-update failure in delayed-feedback CI prediction

Determine the causal mechanisms responsible for the failure of retained online LARC updates in delayed-feedback continuous-integration prediction, including the respective contributions of update magnitude, parameter history, feedback delay, and class composition.

Background

The delayed-feedback continuous-integration replay found that retaining real-feedback updates increased half-Brier loss relative to resetting the numerical policy state, despite real feedback remaining less harmful than permuted feedback. A fixed-prefix intervention further showed same-batch non-descent for the original update and no consistent future-risk improvement from shrinking the update across seeds.

The paper identifies update size, accumulated parameter history, feedback delay, and label composition as plausible contributors, but does not isolate their individual causal effects. Resolving these mechanisms would require additional interventions that separately manipulate these factors while preserving the other aspects of the replay protocol.

References

It makes update size and parameter history concrete design questions, while leaving the causes of failure unresolved. Rank alone controls neither local descent nor usefulness on later predictions.

— LARC: Low-Rank Adaptive Residual Connections for Learning in Frozen Models  (2609.40063 - Zou et al., 30 Sep 2026) in Discussion section; see also Section 6, “Separating Update Magnitude from Parameter History”