Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Hierarchical Sequence Iteration for Heterogeneous Question Answering

Ruiyi Yang, Hao Xue, Imran Razzak, Hakim Hacid, Flora D. Salim

TL;DR

This work tackles the brittleness of retrieval-augmented generation for multi-hop, heterogeneous QA by introducing Hierarchical Sequence Iteration for Heterogeneous Question Answering (HSEQ). HSEQ unifies text, tables, and knowledge graphs into a reversible hierarchical sequence and employs a structure-aware iteration policy guided by a short plan to collect just-enough evidence before answer synthesis, with an optional contradiction-driven refinement. Key contributions include a modality-agnostic adapter, a budget-aware iterative controller, and a provenance-preserving canonical evidence package that supports auditable, reliable QA. Across HotpotQA, HybridQA, and MetaQA, HSEQ achieves consistent EM/F1 gains over strong baselines while delivering predictable latency under token/tool budgets, demonstrating robust performance and interpretability in heterogeneous QA settings.

Abstract

Retrieval-augmented generation (RAG) remains brittle on multi-step questions and heterogeneous evidence sources, trading accuracy against latency and token/tool budgets. This paper introducesHierarchical Sequence (HSEQ) Iteration for Heterogeneous Question Answering, a unified framework that (i) linearize documents, tables, and knowledge graphs into a reversible hierarchical sequence with lightweight structural tags, and (ii) perform structure-aware iteration to collect just-enough evidence before answer synthesis. A Head Agent provides guidance that leads retrieval, while an Iteration Agent selects and expands HSeq via structure-respecting actions (e.g., parent/child hops, table row/column neighbors, KG relations); Finally the head agent composes canonicalized evidence to genearte the final answer, with an optional refinement loop to resolve detected contradictions. Experiments on HotpotQA (text), HybridQA/TAT-QA (table+text), and MetaQA (KG) show consistent EM/F1 gains over strong single-pass, multi-hop, and agentic RAG baselines with high efficiency. Besides, HSEQ exhibits three key advantages: (1) a format-agnostic unification that enables a single policy to operate across text, tables, and KGs without per-dataset specialization; (2) guided, budget-aware iteration that reduces unnecessary hops, tool calls, and tokens while preserving accuracy; and (3) evidence canonicalization for reliable QA, improving answers consistency and auditability.

Hierarchical Sequence Iteration for Heterogeneous Question Answering

TL;DR

This work tackles the brittleness of retrieval-augmented generation for multi-hop, heterogeneous QA by introducing Hierarchical Sequence Iteration for Heterogeneous Question Answering (HSEQ). HSEQ unifies text, tables, and knowledge graphs into a reversible hierarchical sequence and employs a structure-aware iteration policy guided by a short plan to collect just-enough evidence before answer synthesis, with an optional contradiction-driven refinement. Key contributions include a modality-agnostic adapter, a budget-aware iterative controller, and a provenance-preserving canonical evidence package that supports auditable, reliable QA. Across HotpotQA, HybridQA, and MetaQA, HSEQ achieves consistent EM/F1 gains over strong baselines while delivering predictable latency under token/tool budgets, demonstrating robust performance and interpretability in heterogeneous QA settings.

Abstract

Retrieval-augmented generation (RAG) remains brittle on multi-step questions and heterogeneous evidence sources, trading accuracy against latency and token/tool budgets. This paper introducesHierarchical Sequence (HSEQ) Iteration for Heterogeneous Question Answering, a unified framework that (i) linearize documents, tables, and knowledge graphs into a reversible hierarchical sequence with lightweight structural tags, and (ii) perform structure-aware iteration to collect just-enough evidence before answer synthesis. A Head Agent provides guidance that leads retrieval, while an Iteration Agent selects and expands HSeq via structure-respecting actions (e.g., parent/child hops, table row/column neighbors, KG relations); Finally the head agent composes canonicalized evidence to genearte the final answer, with an optional refinement loop to resolve detected contradictions. Experiments on HotpotQA (text), HybridQA/TAT-QA (table+text), and MetaQA (KG) show consistent EM/F1 gains over strong single-pass, multi-hop, and agentic RAG baselines with high efficiency. Besides, HSEQ exhibits three key advantages: (1) a format-agnostic unification that enables a single policy to operate across text, tables, and KGs without per-dataset specialization; (2) guided, budget-aware iteration that reduces unnecessary hops, tool calls, and tokens while preserving accuracy; and (3) evidence canonicalization for reliable QA, improving answers consistency and auditability.
Paper Structure (92 sections, 6 theorems, 15 equations, 3 figures, 10 tables, 1 algorithm)

This paper contains 92 sections, 6 theorems, 15 equations, 3 figures, 10 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. HSEQ: A Multi-Agent Heterogeneous Question Answering Framework
  3. Background and Setup
  4. Heterogeneous QA with budgets.
  5. From retrieval to guided iteration.
  6. HSEQ Architecture
  7. HSEQ-Adapter(HSEQ-A)
  8. HSEQ-Iterator(HSEQ-I)
  9. Guidance prior.
  10. Iteration control.
  11. HSEQ-Head (HSEQ-H).
  12. Guidance Generation.
  13. Answer synthesis and optional refinement.
  14. Learning to Use HSEQ with Open-Source LLMs
  15. Fine-tuning HSEQ-I
  16. ...and 77 more sections

Key Result

Theorem 1

For any finite corpus $X$, under (T1)--(T3), the encoder $\Phi$ is injective up to $\equiv$, i.e., $\Psi(\Phi(X))\equiv X$.

Figures (3)

  • Figure 1: HSEQ overview. (i) HSEQ-A linearizes heterogeneous sources into $S_h$ with level tags, parent pointers, and standardized metadata; (ii) HSEQ-I iterates over a windowed stream of segments under budgets, guided by $g$, and queries $\Phi$ for sufficiency; (iii) $\kappa$ compacts $M_t$ into provenance-preserving evidence; (iv) HSEQ-H produces the final answer and optionally triggers a brief refinement if inconsistencies are detected.
  • Figure 2: HSEQ $S_h$ construction: Different modalities of data are transformed into unified sequence by HSEQ-A
  • Figure 3: HSEQ_I is build by training questions from multiple datasets. After guidance sets are generated, LoRA is applied for finetuning

Theorems & Definitions (14)

  • Theorem 1: Faithful linearization
  • proof
  • Lemma 1: Prefix coverage under $k$-selection
  • proof
  • Proposition 1: Guaranteed halt
  • proof
  • Theorem 2: Budgeted selection complexity
  • proof
  • Definition 1: Canonicalizer
  • Proposition 2: Soundness and auditability
  • ...and 4 more