Safety as Computation: Certified Answer Reuse via Capability Closure in Task-Oriented Dialogue
arXiv cs.AI / 2026/3/24
💬 オピニオンIdeas & Deep AnalysisModels & Research
要点
- The paper proposes a new paradigm for task-oriented dialogue where “safety certification” is treated as a computational primitive that enables answer reuse across turns rather than re-deriving responses each time.
- It argues that, in capability-based systems, the certification step effectively computes a fixed-point closure (cl(At)) that contains all answers reachable from the current configuration.
- The authors implement this idea with a Certified Answer Store (CAS) augmented by Pre-Answer Blocks (PAB), which materializes all derivable follow-up answers along with minimal provenance witnesses at each certified turn.
- By answering later queries through formal containment checks, the approach aims to reduce response latency to sub-millisecond times and eliminate redundant retrieval/generation.
- Overall, the work reframes safety verification as both a correctness mechanism and an efficiency enabler for dialogue systems through formal computation of reachable answers.
