Regime-Conditional Retrieval: Theory and a Transferable Router for Two-Hop QA
A new AI router uses text features to decide how to answer complex, two-step questions.
Researcher Andre Bacellar has published a paper introducing a new theoretical framework and a practical tool for improving two-hop question answering (QA). The work formalizes a key challenge in multi-hop retrieval: questions fall into two distinct regimes. In the 'Q-dominant' regime, the final answer entity is named in the question itself, while in the 'B-dominant' regime, it's only found in an intermediate 'bridge' passage. The paper proves three theorems establishing the mathematical basis for this split, showing that performance is tied to specific text features and that the crucial element for B-dominant questions is the relation-bearing sentence, not just an entity name.
Building on this theory, Bacellar created RegimeRouter, a lightweight model that acts as a decision engine. It analyzes a question using five simple text features derived from the theoretical predicates to instantly choose the optimal retrieval strategy: a direct question-only search or a more complex search that includes a key relation sentence. Trained on just 881 examples from the 2WikiMultiHopQA dataset, RegimeRouter was then tested zero-shot on two other popular benchmarks, MuSiQue and HotpotQA. It delivered significant performance gains, improving Recall@5 by +5.6 and +5.3 percentage points, respectively, demonstrating its ability to generalize without additional training.
- Theoretical framework splits two-hop QA into Q-dominant and B-dominant regimes, proven by three theorems with strong statistical backing (R² ≥ 0.90, p < 0.001).
- RegimeRouter uses only five text features to route queries, trained on a small set of 881 examples with 5-fold cross-validation.
- Achieved zero-shot improvements of +5.6 pp and +5.3 pp in Recall@5 on MuSiQue and HotpotQA, proving transferability across datasets.
Why It Matters
Provides a more efficient, explainable way to improve complex reasoning in RAG systems and AI assistants without massive retraining.