Selected work

Representative Papers

KAP: Bridging the Knowledge Selection-Runtime Consumption Gap in LLM Systems

Jul 27, 2026

This work addresses the inefficiency in large language model inference caused by flattening structured knowledge selection signals into ordinary token sequences, which leads to redundant key-value (KV) cache accesses, increased latency, and reduced throughput. To overcome this limitation, the paper introduces Knowledge Access Planning (KAP), a novel paradigm that elevates structured knowledge selection to a first-class runtime primitive. KAP bridges knowledge selection and KV cache consumption through a unified intermediate representation and leverages the GraphSpec compiler-executor framework to enable plan-driven, efficient KV cache access. Experimental results demonstrate that, on long-context question answering tasks ranging from 4K to 128K tokens, KAP achieves comparable answer quality while accessing only 5.5% of the original KV states, substantially reducing latency and redefining the scaling efficiency of long-context generation.

0 citationsRead paper

When Good OCR Is Not Enough: Benchmarking OCR Robustness for Retrieval-Augmented Generation

Apr 29, 2026

This work addresses a critical gap in optical character recognition (OCR) evaluation by demonstrating that conventional character-level metrics fail to capture OCR performance in industrial retrieval-augmented generation (RAG) systems. The authors introduce a robust benchmark tailored to OCR-first RAG pipelines, encompassing 11 challenging document types. Their analysis reveals a significant disconnect between high OCR accuracy and effective RAG performance: structural and semantic errors—whose impact varies across document categories—frequently cause retrieval failures. Through a comprehensive assessment integrating character error rate (CER), word error rate (WER), and downstream task effectiveness, the study shows that state-of-the-art OCR models exhibit substantial performance degradation on real-world industrial documents. These findings underscore that traditional OCR metrics are insufficient predictors of RAG utility. The benchmark is publicly released to support future research.

0 citationsRead paper

$Agent^2$: An Agent-Generates-Agent Framework for Reinforcement Learning Automation

Sep 15, 2025

Reinforcement learning (RL) agent development has long suffered from high expertise barriers and inefficient manual iteration. This paper proposes an LLM-driven end-to-end automation framework that, for the first time, synthesizes high-performance RL agents directly from natural language task specifications and environment code. Methodologically, it introduces a dual-agent architecture that decouples MDP modeling from algorithmic optimization; an agent self-generation mechanism enabling “agents building agents”—a novel paradigm in automated RL design; and a model-context protocol integrating NLP understanding, RL formalization, adaptive training, and feedback-driven analysis. Evaluated across standard benchmarks—including MuJoCo, MetaDrive, Multi-Agent Particle Environment (MPE), and StarCraft Multi-Agent Challenge (SMAC)—the framework consistently outperforms human-designed agents, achieving substantial average performance gains, with peak improvements reaching 55%.

0 citationsRead paper

PhaseNAS: Language-Model Driven Architecture Search with Dynamic Phase Adaptation

Jul 28, 2025

To address the challenge of balancing exploration breadth and search efficiency in Neural Architecture Search (NAS), this paper proposes a Large Language Model (LLM)-based dynamic phase-adaptive NAS framework. Methodologically, it introduces: (1) a dynamic phase-transition mechanism that adaptively regulates the search process based on real-time performance feedback; (2) a structured template language ensuring precise architectural specification and consistent code generation; and (3) tight integration of LLM reasoning, finite-state machine control, and template-driven coding to enhance search stability and generalizability. Evaluated on NAS-Bench-Macro, the framework achieves significantly improved ranking accuracy. On image classification tasks, it reduces search time by 86%. In object detection, it automatically evolves a lightweight, high-efficiency YOLOv8 variant—achieving higher mAP while consuming fewer computational resources.

0 citationsRead paper

XeMap: Contextual Referring in Large-Scale Remote Sensing Environments

Apr 30, 2025

Locating mid-scale semantic entities (e.g., residential blocks, farmland clusters, industrial zones) in remote sensing imagery remains challenging due to their contextual dependence and lack of precise pixel-level supervision. Method: This paper proposes a novel cross-modal task—Contextual Reference Mapping (XeMap)—for text-driven, context-aware pixel-level referring localization. We design XeMap-Net, featuring a Hierarchical Multi-Scale Semantic Alignment (HMSA) module that jointly leverages self-attention and cross-modal attention, trained via zero-shot cross-modal alignment to eliminate reliance on pixel-level annotations. Contribution/Results: Evaluated on our newly constructed XeMap-Set benchmark under zero-shot settings, our approach significantly outperforms existing state-of-the-art methods. It achieves the first text-to-pixel fine-grained contextual mapping for remote sensing scenes, establishing a new paradigm for large-scale Earth surface semantic understanding.

0 citationsRead paper

Recent publications

Latest Papers

KAP: Bridging the Knowledge Selection-Runtime Consumption Gap in LLM Systems

Jul 27, 2026

This work addresses the inefficiency in large language model inference caused by flattening structured knowledge selection signals into ordinary token sequences, which leads to redundant key-value (KV) cache accesses, increased latency, and reduced throughput. To overcome this limitation, the paper introduces Knowledge Access Planning (KAP), a novel paradigm that elevates structured knowledge selection to a first-class runtime primitive. KAP bridges knowledge selection and KV cache consumption through a unified intermediate representation and leverages the GraphSpec compiler-executor framework to enable plan-driven, efficient KV cache access. Experimental results demonstrate that, on long-context question answering tasks ranging from 4K to 128K tokens, KAP achieves comparable answer quality while accessing only 5.5% of the original KV states, substantially reducing latency and redefining the scaling efficiency of long-context generation.

0 citationsRead paper

When Good OCR Is Not Enough: Benchmarking OCR Robustness for Retrieval-Augmented Generation

Apr 29, 2026

This work addresses a critical gap in optical character recognition (OCR) evaluation by demonstrating that conventional character-level metrics fail to capture OCR performance in industrial retrieval-augmented generation (RAG) systems. The authors introduce a robust benchmark tailored to OCR-first RAG pipelines, encompassing 11 challenging document types. Their analysis reveals a significant disconnect between high OCR accuracy and effective RAG performance: structural and semantic errors—whose impact varies across document categories—frequently cause retrieval failures. Through a comprehensive assessment integrating character error rate (CER), word error rate (WER), and downstream task effectiveness, the study shows that state-of-the-art OCR models exhibit substantial performance degradation on real-world industrial documents. These findings underscore that traditional OCR metrics are insufficient predictors of RAG utility. The benchmark is publicly released to support future research.

0 citationsRead paper

$Agent^2$: An Agent-Generates-Agent Framework for Reinforcement Learning Automation

Sep 15, 2025

Reinforcement learning (RL) agent development has long suffered from high expertise barriers and inefficient manual iteration. This paper proposes an LLM-driven end-to-end automation framework that, for the first time, synthesizes high-performance RL agents directly from natural language task specifications and environment code. Methodologically, it introduces a dual-agent architecture that decouples MDP modeling from algorithmic optimization; an agent self-generation mechanism enabling “agents building agents”—a novel paradigm in automated RL design; and a model-context protocol integrating NLP understanding, RL formalization, adaptive training, and feedback-driven analysis. Evaluated across standard benchmarks—including MuJoCo, MetaDrive, Multi-Agent Particle Environment (MPE), and StarCraft Multi-Agent Challenge (SMAC)—the framework consistently outperforms human-designed agents, achieving substantial average performance gains, with peak improvements reaching 55%.

0 citationsRead paper

PhaseNAS: Language-Model Driven Architecture Search with Dynamic Phase Adaptation

Jul 28, 2025

To address the challenge of balancing exploration breadth and search efficiency in Neural Architecture Search (NAS), this paper proposes a Large Language Model (LLM)-based dynamic phase-adaptive NAS framework. Methodologically, it introduces: (1) a dynamic phase-transition mechanism that adaptively regulates the search process based on real-time performance feedback; (2) a structured template language ensuring precise architectural specification and consistent code generation; and (3) tight integration of LLM reasoning, finite-state machine control, and template-driven coding to enhance search stability and generalizability. Evaluated on NAS-Bench-Macro, the framework achieves significantly improved ranking accuracy. On image classification tasks, it reduces search time by 86%. In object detection, it automatically evolves a lightweight, high-efficiency YOLOv8 variant—achieving higher mAP while consuming fewer computational resources.

0 citationsRead paper

XeMap: Contextual Referring in Large-Scale Remote Sensing Environments

Apr 30, 2025

Locating mid-scale semantic entities (e.g., residential blocks, farmland clusters, industrial zones) in remote sensing imagery remains challenging due to their contextual dependence and lack of precise pixel-level supervision. Method: This paper proposes a novel cross-modal task—Contextual Reference Mapping (XeMap)—for text-driven, context-aware pixel-level referring localization. We design XeMap-Net, featuring a Hierarchical Multi-Scale Semantic Alignment (HMSA) module that jointly leverages self-attention and cross-modal attention, trained via zero-shot cross-modal alignment to eliminate reliance on pixel-level annotations. Contribution/Results: Evaluated on our newly constructed XeMap-Set benchmark under zero-shot settings, our approach significantly outperforms existing state-of-the-art methods. It achieves the first text-to-pixel fine-grained contextual mapping for remote sensing scenes, establishing a new paradigm for large-scale Earth surface semantic understanding.

0 citationsRead paper