ReacTOD's bounded ReAct loop lifts dialogue state accuracy by 14 points

ReacTOD tackles a critical pain point in task-oriented dialogue systems: LLMs prone to hallucination and format errors that cascade into wrong actions (e.g., booking the wrong hotel date). The architecture reformulates NLU as discrete tool calls inside a bounded ReAct loop, where each dialogue state update is validated by a symbolic validator enforcing action compliance, schema conformance, and coreference consistency. This deterministic guardrail enables iterative self-correction, boosting accuracy by up to 9.3 percentage points over single-pass inference on MultiWOZ, with a 93.1% self-correction rate on intercepted errors.

Importantly, ReacTOD uses incremental state prediction and on-demand history retrieval to keep prompts compact, improving instruction adherence even in parameter-constrained models. On MultiWOZ 2.1, gpt-oss-20B achieves 52.71% joint goal accuracy (14-point gain over prior SOTA), and Qwen3-8B reaches 47.34%. On the Schema-Guided Dialogue benchmark, Claude-Opus-4.6 hits 80.68% JGA under end-to-end evaluation with predicted domains, while Qwen3-32B reaches 64.09%. The system requires no task-specific training data, making it a strong candidate for real-world, zero-shot deployment across reservation, transaction, and service request scenarios.