Sezgin
Saygili

Published the first comprehensive benchmark framework evaluating frontier LLMs as dental clinic phone agents across 483 standardized scenarios. DentesBench represents a critical step forward in assessing the real-world utility of conversational AI in medical administration, bridging the gap between general-purpose benchmarks and domain-specific clinic requirements.

The framework categorizes patient interactions into complex multi-turn dialogs involving scheduling, triage, insurance verification, and billing. By pitting models like GPT-4, Claude 3, and fine-tuned open-weight models against identical patient persona scripts, we mapped out the capability frontiers—and the systematic failure modes—of modern AI.

Our methodology highlighted that while most LLMs can handle standard appointment booking, they rapidly degrade in performance when navigating convoluted insurance contexts, multi-party calls, or nuanced medical triage constraints. The suite establishes a new standardized grading rubric focused on empathetic alignment, semantic accuracy, and hallucination resistance in high-stakes healthcare environments.

Surveys play a central role in economics, particularly in climate and public economics where researchers measure expectations, policy preferences, and willingness to pay. However, collecting survey data is expensive, slow, and constrained. This thesis investigates a cutting-edge methodological question: can large language models (LLMs) act as synthetic respondents to reliably reproduce economically meaningful response variation?

Using a private Finnish survey dataset on climate policy attitudes, transport emissions, and willingness to pay, I structured prompt sequences to test whether LLMs can accurately predict individual human responses from demographic and attitudinal parameters. This economically relevant setting examines support for carbon pricing—an area where heterogeneity in preferences is non-trivial and vital to policy making.

The findings offer a nuanced perspective on AI in economics research. While LLMs showed a capacity to recover high-level response patterns well—especially for binary or ternary categorical attitude questions—they significantly struggled when tasked with fine-grained classification or deeper subgroup differentiation. The results suggest that LLMs are currently best suited as a low-cost complement for pretesting questionnaires or exploring survey counterfactuals, rather than acting as a full replacement for field surveys measuring complex economic heterogeneity.