Online Conformal Prediction with Adversarial Semi-bandit Feedback via Regret Minimization

The key technical move here is the 'semi-bandit' framing: the system receives feedback on some actions but not all, and the feedback source is adversarial rather than cooperative or random. That combination is what makes this harder than standard online conformal prediction, and it's the part the summary underweights.

Conformal prediction has been appearing across several recent threads in our coverage. The April 16 paper on LLM judge reliability ('Diagnosing LLM Judge Reliability') used conformal prediction sets to produce per-instance confidence estimates when aggregate metrics looked deceptively clean. That application assumed a relatively cooperative feedback environment. This new work addresses what happens when that assumption breaks, which is a meaningful extension toward production conditions where labels are delayed, missing, or strategically withheld. The MADE benchmark paper from the same date also flagged uncertainty quantification as critical for high-stakes settings, reinforcing that the demand for robust UQ methods is coming from multiple applied directions simultaneously.

The practical test is whether this method holds coverage guarantees on a real safety-critical deployment, such as a medical triage or content moderation pipeline, where adversarial label withholding is documented rather than simulated. If a follow-up empirical study appears within the next six months using real partial-feedback logs, the theoretical contribution will have cleared its most important hurdle.