Selected work

Representative Papers

When RAG Meets Query Planning: Logical Query Trees for Resolving Exploratory Reasoning Problems

Jul 01, 2026

This work addresses the limitations of traditional retrieval-augmented generation (RAG) in handling highly uncertain exploratory reasoning problems (ERPs), where it is prone to retrieval noise and error propagation due to the absence of an end-to-end query planning mechanism. To overcome this, the authors propose PlanRAG, a novel framework that introduces database-inspired query planning into RAG for the first time. PlanRAG formalizes complex natural language queries as logical query trees (LQTs) and constructs high-quality LQTs using a multidimensional cost model combined with dynamic programming. It then executes a concurrent, iterative pipeline of retrieval, rewriting, aggregation, and generation, enabling parallel processing of subqueries. Evaluated on the newly introduced WikiWeb-ERP dataset, PlanRAG significantly outperforms existing iterative and graph-based RAG approaches, demonstrating its effectiveness and superiority in exploratory reasoning tasks.

0 citationsRead paper

ComLQ: Benchmarking Complex Logical Queries in Information Retrieval

Nov 14, 2025

Existing information retrieval (IR) benchmarks primarily focus on semantic matching for single- or multi-hop queries, failing to adequately evaluate models’ ability to handle complex logical queries involving first-order logic operations—such as conjunction, disjunction, and negation. Method: We introduce ComLQ, the first benchmark tailored for complex logical querying, comprising 2,909 structured logical queries and 11,251 candidate documents. We propose a subgraph-guided LLM-based data construction method to ensure logical structural alignment between queries and documents, and design LSNC@K—a novel metric quantifying retrieval consistency under negation. Data quality is ensured via GPT-4o generation, expert verification, and subgraph-informed prompting. Contribution/Results: Zero-shot evaluations reveal that state-of-the-art retrieval models exhibit significant performance degradation on negation-heavy queries, confirming ComLQ’s rigor and its value in exposing critical limitations in current IR systems.

0 citationsRead paper

Logical Consistency is Vital: Neural-Symbolic Information Retrieval for Negative-Constraint Queries

May 28, 2025

To address the low relevance of dense retrievers on complex queries involving negative constraints—caused by their neglect of semantic intent—this paper proposes Neural-Symbolic Integrated Retrieval (NS-IR). NS-IR is the first dense retrieval framework to incorporate First-Order Logic (FOL) modeling, enabling fine-grained re-ranking via logic consistency discrimination, logic alignment, and connective-aware constraint enforcement, while jointly optimizing logic-driven embedding representations. We introduce NegConstraint, the first benchmark dataset specifically designed for queries with negative constraints. Experiments demonstrate that NS-IR significantly outperforms state-of-the-art methods on negative-constraint queries; it also achieves superior performance in zero-shot web search and low-resource settings. The code and dataset are publicly released.

0 citationsRead paper

Recent publications

Latest Papers

When RAG Meets Query Planning: Logical Query Trees for Resolving Exploratory Reasoning Problems

Jul 01, 2026

This work addresses the limitations of traditional retrieval-augmented generation (RAG) in handling highly uncertain exploratory reasoning problems (ERPs), where it is prone to retrieval noise and error propagation due to the absence of an end-to-end query planning mechanism. To overcome this, the authors propose PlanRAG, a novel framework that introduces database-inspired query planning into RAG for the first time. PlanRAG formalizes complex natural language queries as logical query trees (LQTs) and constructs high-quality LQTs using a multidimensional cost model combined with dynamic programming. It then executes a concurrent, iterative pipeline of retrieval, rewriting, aggregation, and generation, enabling parallel processing of subqueries. Evaluated on the newly introduced WikiWeb-ERP dataset, PlanRAG significantly outperforms existing iterative and graph-based RAG approaches, demonstrating its effectiveness and superiority in exploratory reasoning tasks.

0 citationsRead paper

ComLQ: Benchmarking Complex Logical Queries in Information Retrieval

Nov 14, 2025

Existing information retrieval (IR) benchmarks primarily focus on semantic matching for single- or multi-hop queries, failing to adequately evaluate models’ ability to handle complex logical queries involving first-order logic operations—such as conjunction, disjunction, and negation. Method: We introduce ComLQ, the first benchmark tailored for complex logical querying, comprising 2,909 structured logical queries and 11,251 candidate documents. We propose a subgraph-guided LLM-based data construction method to ensure logical structural alignment between queries and documents, and design LSNC@K—a novel metric quantifying retrieval consistency under negation. Data quality is ensured via GPT-4o generation, expert verification, and subgraph-informed prompting. Contribution/Results: Zero-shot evaluations reveal that state-of-the-art retrieval models exhibit significant performance degradation on negation-heavy queries, confirming ComLQ’s rigor and its value in exposing critical limitations in current IR systems.

0 citationsRead paper

Logical Consistency is Vital: Neural-Symbolic Information Retrieval for Negative-Constraint Queries

May 28, 2025

To address the low relevance of dense retrievers on complex queries involving negative constraints—caused by their neglect of semantic intent—this paper proposes Neural-Symbolic Integrated Retrieval (NS-IR). NS-IR is the first dense retrieval framework to incorporate First-Order Logic (FOL) modeling, enabling fine-grained re-ranking via logic consistency discrimination, logic alignment, and connective-aware constraint enforcement, while jointly optimizing logic-driven embedding representations. We introduce NegConstraint, the first benchmark dataset specifically designed for queries with negative constraints. Experiments demonstrate that NS-IR significantly outperforms state-of-the-art methods on negative-constraint queries; it also achieves superior performance in zero-shot web search and low-resource settings. The code and dataset are publicly released.

0 citationsRead paper