Speculative Uncertainty recovers a failure signal from a black-box coding agent using only its output tokens

arXiv·medium signalXBlueskyLinkedInCopy link

arXiv 2609.05274 (submitted 2026-09-04) attacks the problem that LLM software-engineering agents act confidently wrong and the mistake is only visible after an expensive execute-and-retry cycle. The method inverts speculative decoding to build a draft-model gate that produces a predictive failure signal without access to logits, weights, activations, or repeated sampling. For anyone driving a hosted agent through a CLI, this is the class of technique that could gate a destructive action before it runs rather than reverting afterwards.