Cascaded Language Models for Cost-effective Human-AI Decision-Making

TL;DR

This work tackles cost-aware human–AI decision-making by proposing a cascaded LLM framework that routes prompts through a cheap base model, a more expensive large model, and, if needed, a human expert. It introduces two principled policies—deferral based on confidence and abstention based on uncertainty—together with Bayesian calibration and online learning to adapt thresholds using human feedback. The method is evaluated across ARC, MMLU, and medical QA benchmarks, showing improved accuracy-cost trade-offs and reduced abstention with online refinement, though results on MedQA are mixed. The proposed approach offers a practical, scalable pathway to integrate AI assistance with expert oversight in high-stakes domains.

Abstract

A challenge in human-AI decision-making is to balance three factors: the correctness of predictions, the cost of knowledge and reasoning complexity, and the confidence about whether to abstain from automated answers or escalate to human experts. In this work, we present a cascaded LLM decision framework that adaptively delegates tasks across multiple tiers of expertise -- a base model for initial candidate answers, a more capable and knowledgeable (but costlier) large model, and a human expert for when the model cascade abstains. Our method proceeds in two stages. First, a deferral policy determines whether to accept the base model's answer or regenerate it with the large model based on the confidence score. Second, an abstention policy decides whether the cascade model response is sufficiently certain or requires human intervention. Moreover, to overcome static policies and accommodate changing task difficulty, we incorporate an online learning mechanism which uses human feedback. We demonstrate this approach to general question-answering (ARC-Easy, ARC-Challenge, and MMLU) and medical question-answering (MedQA and MedMCQA). Our results demonstrate that our cascaded strategy outperforms single-model baselines in most cases, achieving higher accuracy while reducing costs and providing a principled approach to handling abstentions.

Figures (16)

• Figure 1: Cascaded LLM Human-AI Decision-Making Framework Examples. Given a decision-making problem, the system (1) generates an initial response with a base model, (2) verifies correctness probability, (2.5) defers to a larger model if needed, (3) assesses response uncertainty, and (3.5) abstains to a human expert if necessary. If feedback is available, deferral and abstention modules are adjusted over time. For this system to work efficiently, the modules should uphold three desiderata: the deferral policy regenerates responses only when necessary, the abstention policy escalates to humans only when uncertainty is high, the system continuously improves with feedback.
• Figure 2: Decision flow of the two-tiered cascaded LLM system. The base model first evaluates each query. Confident, low-uncertainty responses are accepted; uncertain ones are passed to the large model or, if still uncertain, deferred to a human expert. Online feedback progressively improves these policies.
• Figure 3: Cost-Accuracy Trade-off for Calibrated Verification Methods (Qwen-2.5 1.5B$\rightarrow$7B). Accuracy versus cost per sample is shown for the cascaded model using various verification methods. Performance above the linear interpolation line between base and large model baselines indicates a positive cost-benefit. Error bars represent standard error.
• Figure 4: Cumulative Regret in Online Setting (Qwen-2.5 1.5B $\rightarrow$ 7B). Cumulative system risk over time. Training data is collected only when abstentions occur. The cascaded system consistently achieves lower regret.
• Figure 5: Imperfect Experts. We increase the percentage of flipped labels during system calibration, simulating imperfect experts, which in turn increases the system's risk of error.