Abstract
Generative models may produce outputs that appear correct but lack verifiable reasoning, and methods to generate such reasoning sequentially can introduce significant latency. The disclosed technology describes an adaptive verification framework where a prompt triage model can assess a query's potential risk level. Low-risk queries may bypass verification, while higher-risk queries can trigger a parallel process where a primary model generates an answer as a causal proxy model concurrently constructs a structured reasoning graph. A heuristic verifier can then programmatically audit this graph for aspects such as logical coherence and factual grounding. This system provides a method for generating auditable reasoning for model outputs in a manner that can dynamically balance verification with performance, potentially reducing the latency and computational overhead associated with certain non-adaptive verification processes.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Mohbe, Neel and Piyush, "Adaptive Verification of Generative Models via Parallel Reasoning Generation", Technical Disclosure Commons, (November 17, 2025)
https://www.tdcommons.org/dpubs_series/8880