Extending consensus sampling to multiple or agentic interactions

Develop an extension of Algorithm 1 (Consensus sampling from k distributions, risk-competitive with the safest s) to settings involving multiple interactions, including agentic interactions, where the system engages in more than one interaction rather than a single-shot prompt–response scenario.

Background

The paper proposes a consensus sampling algorithm that aggregates k generative models to produce outputs with risk bounded relative to the safest s models, and analyzes per-prompt robustness, abstention, and information leakage.

While Section “Varying prompts” treats conditioning on a single prompt, the work’s guarantees are primarily single-shot. The authors explicitly note the need to go beyond this setting, identifying multi-step or agentic interaction scenarios as an open direction. Such settings raise additional challenges, including composition across rounds and potential cumulative leakage, that are not addressed by the current framework.