Selected work

Representative Papers

FinanceComplexQA: Benchmarking Agentic Reasoning on Industrial-grade Financial Documents

Jul 21, 2026

Existing AI agents exhibit inconsistent performance in understanding and reasoning over real-world, complex financial documents, largely due to the absence of evaluation benchmarks that reflect industrial settings. This work proposes FinanceComplexQA—the first open-ended, generative benchmark specifically designed for complex-layout financial documents—encompassing six realistic scenarios, seven task types, and 2,026 challenging questions. The benchmark leverages a novel Finance-LaTeX SKILL pipeline to synthetically generate 2,000 professional documents and 6,000 bilingual question-answer pairs. Integrated with RAG, multi-hop reasoning, and an Agent-as-a-Judge evaluation framework, FinanceComplexQA enables systematic assessment of mainstream agents across critical dimensions such as numerical computation, summarization, and domain-specific analysis, thereby uncovering their strengths and limitations in practical financial contexts.

0 citationsRead paper

Atomic-to-Compositional Generalization for Mobile Agents with A New Benchmark and Scheduling System

Jun 10, 2025

Existing mobile agents exhibit limited generalization to atomic tasks, struggling with real-world composite tasks that involve multi-step reasoning, cross-interface navigation, and context-dependent execution. Method: We introduce UI-NEXUS—the first benchmark explicitly designed for generalization from atomic to composite tasks—and propose AGENT-NEXUS, a lightweight scheduling system. It features dynamic subtask decomposition and state-aware scheduling, enabling seamless composition of atomic capabilities into composite reasoning without model retraining. The system integrates a multimodal large language model–driven architecture, a controllable local UI simulation environment, and an online evaluation framework supporting both English- and Chinese-language mobile applications. Contribution/Results: We systematically define and evaluate three composite operation types: simple concatenation, context transfer, and deep exploration. Experiments show AGENT-NEXUS improves task success rates by 24%–40% on UI-NEXUS, significantly mitigating under-execution, over-execution, and attention drift, while incurring negligible inference overhead.

0 citationsRead paper

LightRetriever: A LLM-based Hybrid Retrieval Architecture with 1000x Faster Query Inference

May 18, 2025

To address high query encoding latency and substantial resource overhead in LLM-driven hybrid retrieval, this paper proposes a heterogeneous, decoupled lightweight query encoding architecture: documents retain full LLM-based encoding capability, while queries bypass real-time LLM inference entirely and instead leverage GPU-accelerated embedding lookup. This is the first design achieving complete decoupling of query and document encoding, drastically reducing computational load. On an H800 GPU, it achieves over 1000× query inference speedup; even without GPU acceleration, it attains 20× speedup. Crucially, it maintains 95% of the full-LLM retrieval accuracy on large-scale benchmarks. The core contribution lies in reformulating query encoding in hybrid retrieval as an efficient embedding lookup task—enabling an order-of-magnitude improvement in inference efficiency with negligible accuracy degradation.

0 citationsRead paper

Recent publications

Latest Papers

FinanceComplexQA: Benchmarking Agentic Reasoning on Industrial-grade Financial Documents

Jul 21, 2026

Existing AI agents exhibit inconsistent performance in understanding and reasoning over real-world, complex financial documents, largely due to the absence of evaluation benchmarks that reflect industrial settings. This work proposes FinanceComplexQA—the first open-ended, generative benchmark specifically designed for complex-layout financial documents—encompassing six realistic scenarios, seven task types, and 2,026 challenging questions. The benchmark leverages a novel Finance-LaTeX SKILL pipeline to synthetically generate 2,000 professional documents and 6,000 bilingual question-answer pairs. Integrated with RAG, multi-hop reasoning, and an Agent-as-a-Judge evaluation framework, FinanceComplexQA enables systematic assessment of mainstream agents across critical dimensions such as numerical computation, summarization, and domain-specific analysis, thereby uncovering their strengths and limitations in practical financial contexts.

0 citationsRead paper

Atomic-to-Compositional Generalization for Mobile Agents with A New Benchmark and Scheduling System

Jun 10, 2025

Existing mobile agents exhibit limited generalization to atomic tasks, struggling with real-world composite tasks that involve multi-step reasoning, cross-interface navigation, and context-dependent execution. Method: We introduce UI-NEXUS—the first benchmark explicitly designed for generalization from atomic to composite tasks—and propose AGENT-NEXUS, a lightweight scheduling system. It features dynamic subtask decomposition and state-aware scheduling, enabling seamless composition of atomic capabilities into composite reasoning without model retraining. The system integrates a multimodal large language model–driven architecture, a controllable local UI simulation environment, and an online evaluation framework supporting both English- and Chinese-language mobile applications. Contribution/Results: We systematically define and evaluate three composite operation types: simple concatenation, context transfer, and deep exploration. Experiments show AGENT-NEXUS improves task success rates by 24%–40% on UI-NEXUS, significantly mitigating under-execution, over-execution, and attention drift, while incurring negligible inference overhead.

0 citationsRead paper

LightRetriever: A LLM-based Hybrid Retrieval Architecture with 1000x Faster Query Inference

May 18, 2025

To address high query encoding latency and substantial resource overhead in LLM-driven hybrid retrieval, this paper proposes a heterogeneous, decoupled lightweight query encoding architecture: documents retain full LLM-based encoding capability, while queries bypass real-time LLM inference entirely and instead leverage GPU-accelerated embedding lookup. This is the first design achieving complete decoupling of query and document encoding, drastically reducing computational load. On an H800 GPU, it achieves over 1000× query inference speedup; even without GPU acceleration, it attains 20× speedup. Crucially, it maintains 95% of the full-LLM retrieval accuracy on large-scale benchmarks. The core contribution lies in reformulating query encoding in hybrid retrieval as an efficient embedding lookup task—enabling an order-of-magnitude improvement in inference efficiency with negligible accuracy degradation.

0 citationsRead paper