Characterize the effect of dependence among LLM calls on endpoint forecasting

Characterize how dependence among successive LLM calls affects the calibration and accuracy of state-conditioned endpoint forecasts for fixed-budget plurality voting, beyond the iid response-law assumption.

Background

The paper computes endpoint recovery probabilities from the realized vote state and remaining budget under an iid response law. Prospective experiments show that these state-conditioned forecasts are informative, but their calibration depends on estimating the future response distribution and on the validity of the iid assumption.

The authors explicitly identify dependence among calls as an unresolved issue. This leaves open whether the exact finite-horizon forecasting framework remains calibrated, or how it should be modified, when model responses are correlated across calls because of shared prompts, sampling procedures, adaptive policies, or other dependencies.

References

Dependence among calls remains unresolved.

— From Discovery to Decision: Finite-Budget Recoverability in LLM Voting  (2610.01014 - Li et al., 1 Oct 2026) in Section 3.2, subsection “State-Conditioned Forecasting”

The recoding uses gold and post-processes responses, so it tests only the transfer calculation and leaves the mechanism of natural permutation open.

— From Discovery to Decision: Finite-Budget Recoverability in LLM Voting  (2610.01014 - Li et al., 1 Oct 2026) in Appendix B, Section “Prospective Forecasting from the Vote State,” subsection “Oracle recoding of two wrong identities”