![[2602.14643] Arbor: A Framework for Reliable Navigation of Critical Conversation Flows](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.14643] Arbor: A Framework for Reliable Navigation of Critical Conversation Flows
Summary
The paper presents Arbor, a framework designed to enhance the navigation of critical conversation flows in high-stakes environments like healthcare, improving accuracy and efficiency.
Why It Matters
Arbor addresses the limitations of traditional monolithic approaches in AI, particularly in healthcare, where accurate decision-making is crucial. By breaking down decision processes into manageable tasks, it allows for better performance even with smaller models, making advanced AI more accessible and reliable in critical applications.
Key Takeaways
- Arbor improves mean turn accuracy by 29.4 percentage points.
- It reduces per-turn latency by 57.1%, enhancing efficiency.
- The framework allows smaller models to perform comparably to larger ones.
- Decision trees are standardized for dynamic retrieval, improving navigation.
- Arbor's architecture reduces dependence on model size and capability.
Computer Science > Artificial Intelligence arXiv:2602.14643 (cs) [Submitted on 16 Feb 2026] Title:Arbor: A Framework for Reliable Navigation of Critical Conversation Flows Authors:Luís Silva, Diogo Gonçalves, Catarina Farinha, Clara Matos, Luís Ungaro View a PDF of the paper titled Arbor: A Framework for Reliable Navigation of Critical Conversation Flows, by Lu\'is Silva and 3 other authors View PDF HTML (experimental) Abstract:Large language models struggle to maintain strict adherence to structured workflows in high-stakes domains such as healthcare triage. Monolithic approaches that encode entire decision structures within a single prompt are prone to instruction-following degradation as prompt length increases, including lost-in-the-middle effects and context window overflow. To address this gap, we present Arbor, a framework that decomposes decision tree navigation into specialized, node-level tasks. Decision trees are standardized into an edge-list representation and stored for dynamic retrieval. At runtime, a directed acyclic graph (DAG)-based orchestration mechanism iteratively retrieves only the outgoing edges of the current node, evaluates valid transitions via a dedicated LLM call, and delegates response generation to a separate inference step. The framework is agnostic to the underlying decision logic and model provider. Evaluated against single-prompt baselines across 10 foundation models using annotated turns from real clinical triage conversations. Arbor imp...