Selected work

Representative Papers

AVA-Encoder: Towards Agent-Native Video Representation Learning

Aug 12, 2026

This work addresses the challenge that existing creative agents struggle to effectively learn from high-quality films, primarily due to the absence of structured video representations that are both content-faithful and amenable to agent-based reasoning and manipulation. To bridge this gap, the authors propose the AVA-Encoder framework, which introduces the first native video knowledge graph representation tailored for agent operation. This framework enables bidirectional conversion between video and knowledge graph through a self-encoding mechanism and incorporates a text-gradient optimization strategy guided by natural language update directions. The contributions include the first cinematic-scale knowledge graph dataset, a reconstruction benchmark, and a technical pipeline integrating multimodal asset linking, typed edge relations, and data-agnostic encoding. Experiments demonstrate that the method outperforms the strongest baseline by 20.7 percentage points in video reconstruction and achieves superior performance over hand-tuned strategies using 74.3% fewer prompts under a policy-only setting.

0 citationsRead paper

ForestBench: A Unified Graph Framework for Evaluating Multi-Agent Collaboration

Aug 09, 2026

This work addresses the lack of a unified and reproducible benchmark for evaluating collaborative processes in large language model (LLM)-based multi-agent systems, where reliance solely on outcomes or LLM-based judgments often leads to inconsistent assessments. The authors propose the first general evaluation paradigm grounded in collaboration graphs and reference forests: execution trajectories are mapped into collaboration graphs and structurally compared against task-specific reference forests, enabling efficient scoring without additional LLM inference. This approach preserves collaboration diversity and supports cross-method, reusable, and low-overhead structural evaluation. Built upon this framework, ForestBench comprises 844 tasks with ten successful reference graphs per task, delivers millisecond-level scoring, and demonstrates robust stability and validity across six mainstream multi-agent frameworks.

0 citationsRead paper

Scalable High-Fidelity Macromolecular Docking for GPU-Accelerated Supercomputers

Aug 07, 2026

This work addresses the challenges of scaling flexible macromolecular docking on GPU-based supercomputers, which are hindered by irregular computation, low parallelism, and load imbalance. The authors propose SparkleDock, a novel framework that restructures the Glowworm Swarm Optimization (GSO) algorithm to expose fine-grained parallelism at the individual level and reformulates energy scoring as matrix operations amenable to Tensor Cores. Coupled with a performance-model-driven scheduling strategy, SparkleDock achieves cross-GPU load balancing and out-of-core scalability. For the first time, this approach enables near-real-time GSO-based flexible docking on GPU supercomputers: achieving 9.7× and 18.9× speedups on a single A100 and H100 GPU, respectively, and reducing docking time from hours to seconds at a scale of 512 GPUs—delivering over two orders of magnitude overall acceleration.

0 citationsRead paper

MicroEvo: Knowledge-Guided LLM Sampling for Efficient Microarchitecture Design Space Exploration

Aug 06, 2026

This work addresses the challenges of microarchitectural design space exploration, which is hindered by an enormous search space and high costs of power, performance, and area (PPA) evaluation. Traditional approaches struggle to model complex microarchitectural dependencies and suffer from low learning efficiency. To overcome these limitations, this study introduces, for the first time, a knowledge-guided large language model (LLM) integrated with Monte Carlo Tree Search (MCTS). The proposed framework features LLM-driven evolutionary operators, a tree policy that balances Pareto contribution and diversity, and mechanisms for state-aware prompting and active knowledge accumulation, enabling efficient multi-objective optimization. Experimental results on industrial-grade cores demonstrate that the method improves Pareto front quality by up to 36.2% and achieves a 10.6× speedup in search efficiency, exhibiting strong scalability.

0 citationsRead paper

Research Paradigm of Materials Science Tetrahedra with Artificial Intelligence

Mar 14, 2026

This work addresses the methodological and ontological disparities between artificial intelligence and materials science by proposing a dual-paradigm integration framework to advance data-driven materials research. It extends the classical materials tetrahedron into a quintet—comprising matter, data, models, potential, and agents—to enhance AI-enabled materials discovery. Concurrently, it establishes a systematic AI research pathway structured around data, architecture, encoding, optimization, and reasoning, thereby formalizing the scientific logic inherent to AI itself. By synergistically integrating materials informatics, foundational AI architectures, data encoding strategies, and agent-based collaborative reasoning, this study constructs a structured theoretical framework for AI for Science in the domain of materials, fostering co-evolution between the formulation of scientific questions and their technological realization.

0 citationsRead paper

Recent publications

Latest Papers

AVA-Encoder: Towards Agent-Native Video Representation Learning

Aug 12, 2026

This work addresses the challenge that existing creative agents struggle to effectively learn from high-quality films, primarily due to the absence of structured video representations that are both content-faithful and amenable to agent-based reasoning and manipulation. To bridge this gap, the authors propose the AVA-Encoder framework, which introduces the first native video knowledge graph representation tailored for agent operation. This framework enables bidirectional conversion between video and knowledge graph through a self-encoding mechanism and incorporates a text-gradient optimization strategy guided by natural language update directions. The contributions include the first cinematic-scale knowledge graph dataset, a reconstruction benchmark, and a technical pipeline integrating multimodal asset linking, typed edge relations, and data-agnostic encoding. Experiments demonstrate that the method outperforms the strongest baseline by 20.7 percentage points in video reconstruction and achieves superior performance over hand-tuned strategies using 74.3% fewer prompts under a policy-only setting.

0 citationsRead paper

ForestBench: A Unified Graph Framework for Evaluating Multi-Agent Collaboration

Aug 09, 2026

This work addresses the lack of a unified and reproducible benchmark for evaluating collaborative processes in large language model (LLM)-based multi-agent systems, where reliance solely on outcomes or LLM-based judgments often leads to inconsistent assessments. The authors propose the first general evaluation paradigm grounded in collaboration graphs and reference forests: execution trajectories are mapped into collaboration graphs and structurally compared against task-specific reference forests, enabling efficient scoring without additional LLM inference. This approach preserves collaboration diversity and supports cross-method, reusable, and low-overhead structural evaluation. Built upon this framework, ForestBench comprises 844 tasks with ten successful reference graphs per task, delivers millisecond-level scoring, and demonstrates robust stability and validity across six mainstream multi-agent frameworks.

0 citationsRead paper

Scalable High-Fidelity Macromolecular Docking for GPU-Accelerated Supercomputers

Aug 07, 2026

This work addresses the challenges of scaling flexible macromolecular docking on GPU-based supercomputers, which are hindered by irregular computation, low parallelism, and load imbalance. The authors propose SparkleDock, a novel framework that restructures the Glowworm Swarm Optimization (GSO) algorithm to expose fine-grained parallelism at the individual level and reformulates energy scoring as matrix operations amenable to Tensor Cores. Coupled with a performance-model-driven scheduling strategy, SparkleDock achieves cross-GPU load balancing and out-of-core scalability. For the first time, this approach enables near-real-time GSO-based flexible docking on GPU supercomputers: achieving 9.7× and 18.9× speedups on a single A100 and H100 GPU, respectively, and reducing docking time from hours to seconds at a scale of 512 GPUs—delivering over two orders of magnitude overall acceleration.

0 citationsRead paper

MicroEvo: Knowledge-Guided LLM Sampling for Efficient Microarchitecture Design Space Exploration

Aug 06, 2026

This work addresses the challenges of microarchitectural design space exploration, which is hindered by an enormous search space and high costs of power, performance, and area (PPA) evaluation. Traditional approaches struggle to model complex microarchitectural dependencies and suffer from low learning efficiency. To overcome these limitations, this study introduces, for the first time, a knowledge-guided large language model (LLM) integrated with Monte Carlo Tree Search (MCTS). The proposed framework features LLM-driven evolutionary operators, a tree policy that balances Pareto contribution and diversity, and mechanisms for state-aware prompting and active knowledge accumulation, enabling efficient multi-objective optimization. Experimental results on industrial-grade cores demonstrate that the method improves Pareto front quality by up to 36.2% and achieves a 10.6× speedup in search efficiency, exhibiting strong scalability.

0 citationsRead paper

Research Paradigm of Materials Science Tetrahedra with Artificial Intelligence

Mar 14, 2026

This work addresses the methodological and ontological disparities between artificial intelligence and materials science by proposing a dual-paradigm integration framework to advance data-driven materials research. It extends the classical materials tetrahedron into a quintet—comprising matter, data, models, potential, and agents—to enhance AI-enabled materials discovery. Concurrently, it establishes a systematic AI research pathway structured around data, architecture, encoding, optimization, and reasoning, thereby formalizing the scientific logic inherent to AI itself. By synergistically integrating materials informatics, foundational AI architectures, data encoding strategies, and agent-based collaborative reasoning, this study constructs a structured theoretical framework for AI for Science in the domain of materials, fostering co-evolution between the formulation of scientific questions and their technological realization.

0 citationsRead paper