FiSMiness: A Finite State Machine Based Paradigm for Emotional Support Conversations

Researchers propose FiSMiness, a framework integrating Finite State Machines with large language models to improve emotional support conversations by enabling models to systematically reason through emotional states, support strategies, and responses. The approach outperforms multiple baseline methods including chain-of-thought and fine-tuning approaches on ESC datasets, demonstrating that structured reasoning paradigms can enhance LLM performance on specialized dialogue tasks.

FiSMiness represents a meaningful advancement in applying structured computational models to conversational AI, specifically addressing limitations in how LLMs handle emotional support interactions. Rather than relying on simple inference or prompt engineering techniques, the framework uses finite state machines—a foundational computer science concept—to impose architectural discipline on the reasoning process during each conversational turn. This allows the model to explicitly track emotional states, deliberate on appropriate support strategies, and generate contextually relevant responses in a systematic manner.

The motivation stems from a recognized gap in existing ESC approaches: while LLMs demonstrate strong capabilities across many tasks, their application to long-term emotional support lacks the structured planning necessary for sustained effectiveness. FSMs provide a formal mechanism for modeling state transitions and decision logic, creating a more predictable and interpretable conversational flow. This bridges the gap between raw language understanding and purposeful intervention design.

The research demonstrates practical value by outperforming not only direct inference methods but also more sophisticated approaches like self-refinement and chain-of-thought prompting. Notably, FiSMiness achieves superior results even when compared against larger models, suggesting that architectural elegance and structural reasoning can offset raw parameter scaling. This finding challenges the prevailing assumption that bigger models automatically solve harder problems.

For the broader AI development community, this work validates the continued relevance of classical computer science concepts in the LLM era. Rather than abandoning formal methods in favor of purely neural approaches, integrating symbolic reasoning structures with neural inference offers a complementary path forward. Future applications might extend FSM-based frameworks to other specialized dialogue domains requiring structured reasoning and long-term coherence.

• →Finite State Machines integrated with LLMs improve emotional support conversations by systematizing emotional reasoning and response generation
• →FiSMiness outperforms larger models and multiple baseline techniques despite using equivalent or fewer parameters
• →The framework demonstrates that structured planning mechanisms enhance LLM performance on specialized conversational tasks requiring long-term consistency
• →Classical computer science concepts remain valuable for addressing limitations in neural language models on specialized applications
• →Emotional support conversation research advances toward more interpretable and systematic approaches rather than pure end-to-end learning