Isolate the causal contributions of structured clarification components

Determine the separate effects of ambiguity pre-checking, category-conditioned clarification, and interpretation-state-gated answering in Category-Aware Clarification ReAct by evaluating a generic structured-clarification baseline that omits named categories and by conducting per-component ablations.

Background

Category-Aware Clarification ReAct combines three components: an ambiguity pre-check, category-conditioned clarification questions, and interpretation-state-gated answering that prevents a final answer until the relevant interpretation fields are resolved. In a GPT-5 pilot, the combined policy improves accuracy over standard ReAct.

Because all three components are introduced simultaneously, the observed improvement cannot be attributed specifically to category-conditioned questioning. The paper therefore identifies a generic structured-clarification comparison without named categories and separate ablations of the three components as the controls required to determine which mechanism produces the gain.

References

A generic-structured clarification baseline that keeps the structure but drops the named categories, together with a per-component ablation, is the clean control and remains future work.

— FinInteract: Benchmarking Clarification and Intent Integration in Ambiguous Financial Question Answering  (2609.24002 - Wang et al., 21 Sep 2026) in Appendix, Section 6.5.3, “Inference-Time Intervention: Category-Aware Clarification ReAct,” paragraph “Pilot results”