ReacTOD: Bounded Neuro-Symbolic NLU for Zero-Shot Dialogue State Tracking
ReacTOD is a constrained neuro-symbolic framework designed for task-oriented dialogue systems, which reinterprets natural language understanding as distinct tool invocations within a self-correcting ReAct loop that relies on deterministic validation. This system tackles issues like hallucinations and formatting mistakes found in moderately-sized LLMs, which may lead to erroneous actions, such as reserving a hotel on an incorrect date. By utilizing a bounded ReAct loop, it enhances accuracy by as much as 9.3 percentage points compared to single-pass inference on the MultiWOZ benchmark. A symbolic validator ensures that every dialogue state update adheres to action compliance, schema adherence, and coreference consistency, achieving a 93.1% self-correction rate on identified errors and generating structured execution traces. The paper can be accessed on arXiv under reference 2605.19077.
Key facts
- ReacTOD is a bounded neuro-symbolic architecture for task-oriented dialogue systems.
- It reformulates NLU as discrete tool calls within a self-correcting ReAct loop.
- The system improves accuracy by up to 9.3 percentage points over single-pass inference on MultiWOZ.
- A symbolic validator enforces action compliance, schema conformance, and coreference consistency.
- Achieves a 93.1% self-correction rate on intercepted errors.
- Incremental state prediction and on-demand history retrieval keep prompts compact.
- The paper is available on arXiv under reference 2605.19077.
Entities
Institutions
- arXiv