🤖 AI Summary
This work addresses the inherent unreliability of large language models in safety-critical applications, where hallucinations remain a persistent challenge and existing external detection methods suffer from high latency, computational overhead, and dependence on external APIs. To overcome these limitations, the authors propose an intrinsic reliability monitoring framework that leverages linear probes to analyze hidden states during forward propagation. By jointly examining softmax outputs and latent intent representations, the method constructs a lightweight “cognitive dissonance gap” metric that quantifies in real time the inconsistency between semantic confidence and internal certainty. Crucially, this approach incurs no additional inference steps or external calls, enabling low-latency, low-overhead reliability assessment. Empirical evaluations demonstrate its strong statistical efficacy and robust out-of-distribution generalization across multiple model architectures.
📝 Abstract
As Large Language Models (LLMs) are increasingly deployed in mission-critical software systems, detecting hallucinations and ``faked truthfulness'' has become a paramount engineering challenge. Current reliability architectures rely heavily on post-generation, black-box mechanisms, such as Retrieval-Augmented Generation (RAG) cross-checking or LLM-as-a-judge evaluators. These extrinsic methods introduce unacceptable latency, high computational overhead, and reliance on secondary external API calls, frequently violating standard software engineering Service Level Agreements (SLAs). In this paper, we propose the Cognitive Circuit Breaker, a novel systems engineering framework that provides intrinsic reliability monitoring with minimal latency overhead. By extracting hidden states during a model's forward pass, we calculate the ``Cognitive Dissonance Delta'' -- the mathematical gap between an LLM's outward semantic confidence (softmax probabilities) and its internal latent certainty (derived via linear probes). We demonstrate statistically significant detection of cognitive dissonance, highlight architecture-dependent Out-of-Distribution (OOD) generalization, and show that this framework adds negligible computational overhead to the active inference pipeline.