Search papers, labs, and topics across Lattice.
This paper introduces Dependency-Aware Intermediate QA Supervision (DAIS), a novel framework that enhances chain-of-thought supervision by generating stage-level QA records that incorporate dependencies from previous reasoning steps. By conditioning local answers on prior states, DAIS significantly improves final-answer accuracy across multiple benchmarks, including GDPR, AIACT, MedQA, and FOLIO, outperforming traditional methods such as flat chain-of-thought and independent-QA approaches. The findings indicate that this dependency-aware method not only boosts accuracy but also provides a lightweight auxiliary supervision signal that enriches the reasoning process without the need for extensive additional data.
DAIS boosts final-answer accuracy by up to 5.6% by leveraging dependency-aware intermediate QA records that enhance reasoning coherence.
Chain-of-thought (CoT) supervision exposes intermediate rationales, but flat rationale targets usually optimize a single reasoning sequence and provide limited supervision on how local conclusions should support later decisions. We introduce Dependency-Aware Intermediate QA Supervision (DAIS), a training-time framework that converts filtered teacher rationales into stage-level QA records. Each intermediate record predicts a local answer conditioned on the previous states needed for that decision, while the final-answer record keeps the original task format; evaluation therefore uses only the original input and optional context. Across GDPR, AIACT, MedQA, and FOLIO with multiple Qwen backbones, DAIS improves average final-answer accuracy over answer-only, flat chain-of-thought, and independent-QA baselines. On policy-compliance benchmarks, it achieves a largest gain of 5.6% and an average gain of 4.2% over the strongest non-DAIS baseline. Controlled ablations show that valid previous-state conditioning contributes beyond longer targets or additional intermediate text, supporting dependency-conditioned intermediate QA as a lightweight auxiliary supervision signal for standard final-answer inference.
