Announcement_paper_token_audit

We’ve released a new preprint! In today’s LLM-as-a-service ecosystem, users are vulnerable to providers overcharging them by misreporting the number of tokens required to encode the generated string. We develop a sequential auditing framework that enables a trusted third-party auditor to detect whether a provider is misreporting tokens. Our framework offers strong theoretical guarantees, and in experiments, it detects a misreporting provider after at most 70 queries.