Selected work

Representative Papers

GraphInfer-Bench: Benchmarking LLM's Inference Capability on Graphs

Jun 09, 2026

This work addresses the lack of effective evaluation of open-ended reasoning capabilities of large language models (LLMs) on graph-structured data, particularly for tasks requiring integration of node features with their neighborhood context. To this end, we introduce GraphInfer-Bench, a novel benchmark that systematically defines and evaluates graph reasoning through five types of descriptive and comparative tasks whose answers cannot be derived from a single node or path alone. A rigorous four-stage quality control pipeline ensures high dataset fidelity. Experiments on 42,000 samples reveal that current LLMs fail to fully solve these tasks, while standard graph neural networks (GNNs) match or outperform even the strongest closed-source LLMs in zero-shot settings across most categories. We further compare diverse approaches—including graph-text alignment models, LLM zero-shot inference, Graph2Text fine-tuning, and baseline GNNs—highlighting key challenges and promising directions for advancing graph-based reasoning.

0 citationsRead paper

Beyond the Golden Teacher: Enhancing Graph Learning through LLM-GNN Co-teaching

Jun 09, 2026

This work addresses the challenges of few-shot learning on textual attributed graphs (TAGs), where graph neural networks (GNNs) struggle with cold-start nodes, large language models (LLMs) are susceptible to textual ambiguity, and existing approaches relying on a single “gold teacher” propagate blind spots. To overcome these limitations, the authors propose a bidirectional LLM-GNN collaborative teaching framework that abandons the fixed-teacher assumption and instead enables joint training through iterative cross-model pseudo-label exchange and dynamic preference optimization. A key innovation is round-level pseudo-label preference optimization (RPL-PO), which constructs DPO-style preference pairs by tracking prediction trajectories from inter-model disagreement to consensus. Evaluated on six benchmark datasets, the method achieves significant gains over state-of-the-art approaches, improving 3-shot accuracy by 7.86% on Cora and 7.73% on ogbn-arxiv, while also demonstrating strong generalization to 5-shot and zero-shot cross-dataset settings.

0 citationsRead paper

Quantifying and Defending against the Privacy Risk in Logit-based Federated Learning

Jun 06, 2026

In logit-based federated learning, a semi-honest server can exploit shared logits to steal clients’ private models, posing a severe privacy threat. This work is the first to theoretically and empirically quantify such model leakage risks, demonstrating that effective attacks remain feasible even using only unrelated public data. To address this vulnerability, we propose AdaMSA, an adaptive model stealing attack, and introduce a targeted logits perturbation defense mechanism that significantly mitigates privacy leakage with negligible impact on training performance. Our approach achieves a synergistic balance between privacy preservation and model utility, thereby filling a critical gap in understanding and mitigating model-level privacy risks in logit-based federated learning.

0 citationsRead paper

HexAGenT: Efficient Agentic LLM Serving via Workflow- and Heterogeneity-Aware Scheduling

May 15, 2026

This work addresses the challenge of end-to-end latency guarantees for online agent workflows in heterogeneous LLM serving clusters that decouple prefill and decode stages. It proposes the first scheduling framework tailored to such workflows, modeling each request as a runtime-revealed DAG. The framework predicts per-workflow independent completion times and prioritizes ready nodes based on their risk of missing service-level objectives (SLOs). It jointly optimizes placement across heterogeneous GPUs and queue priorities for both prefill and decode phases, while explicitly accounting for KV cache capacity constraints and cross-stage communication overhead. Experiments on A100/H100/H200 clusters demonstrate that, compared to baselines, the proposed approach reduces the average resource requirements to meet 95% and 99% workflow SLOs by 20.1% and 33.0%, respectively, with peak reductions reaching 45.0% and 80.5%.

0 citationsRead paper

Does Synthetic Layered Design Data Benefit Layered Design Decomposition?

May 14, 2026

Existing image generation methods produce flat outputs that are difficult to edit, while real-world layered data is scarce and non-scalable, hindering the practical deployment of layer decomposition techniques. This work proposes a data-driven approach based on SynLayers, a purely synthetic dataset, operating within the CLD framework by integrating vision-language models (VLMs) to generate textual supervision alongside bounding box inputs. It demonstrates for the first time that purely synthetic data can effectively substitute real data for layered design decomposition. The method overcomes data scalability limitations, enables balanced control over the distribution of layer counts, and outperforms non-scalable alternatives such as PrismLayersPro. Experiments show that model performance saturates at around 50,000 samples, significantly alleviating the layer count imbalance problem.

0 citationsRead paper

Recent publications

Latest Papers

GraphInfer-Bench: Benchmarking LLM's Inference Capability on Graphs

Jun 09, 2026

This work addresses the lack of effective evaluation of open-ended reasoning capabilities of large language models (LLMs) on graph-structured data, particularly for tasks requiring integration of node features with their neighborhood context. To this end, we introduce GraphInfer-Bench, a novel benchmark that systematically defines and evaluates graph reasoning through five types of descriptive and comparative tasks whose answers cannot be derived from a single node or path alone. A rigorous four-stage quality control pipeline ensures high dataset fidelity. Experiments on 42,000 samples reveal that current LLMs fail to fully solve these tasks, while standard graph neural networks (GNNs) match or outperform even the strongest closed-source LLMs in zero-shot settings across most categories. We further compare diverse approaches—including graph-text alignment models, LLM zero-shot inference, Graph2Text fine-tuning, and baseline GNNs—highlighting key challenges and promising directions for advancing graph-based reasoning.

0 citationsRead paper

Beyond the Golden Teacher: Enhancing Graph Learning through LLM-GNN Co-teaching

Jun 09, 2026

This work addresses the challenges of few-shot learning on textual attributed graphs (TAGs), where graph neural networks (GNNs) struggle with cold-start nodes, large language models (LLMs) are susceptible to textual ambiguity, and existing approaches relying on a single “gold teacher” propagate blind spots. To overcome these limitations, the authors propose a bidirectional LLM-GNN collaborative teaching framework that abandons the fixed-teacher assumption and instead enables joint training through iterative cross-model pseudo-label exchange and dynamic preference optimization. A key innovation is round-level pseudo-label preference optimization (RPL-PO), which constructs DPO-style preference pairs by tracking prediction trajectories from inter-model disagreement to consensus. Evaluated on six benchmark datasets, the method achieves significant gains over state-of-the-art approaches, improving 3-shot accuracy by 7.86% on Cora and 7.73% on ogbn-arxiv, while also demonstrating strong generalization to 5-shot and zero-shot cross-dataset settings.

0 citationsRead paper

Quantifying and Defending against the Privacy Risk in Logit-based Federated Learning

Jun 06, 2026

In logit-based federated learning, a semi-honest server can exploit shared logits to steal clients’ private models, posing a severe privacy threat. This work is the first to theoretically and empirically quantify such model leakage risks, demonstrating that effective attacks remain feasible even using only unrelated public data. To address this vulnerability, we propose AdaMSA, an adaptive model stealing attack, and introduce a targeted logits perturbation defense mechanism that significantly mitigates privacy leakage with negligible impact on training performance. Our approach achieves a synergistic balance between privacy preservation and model utility, thereby filling a critical gap in understanding and mitigating model-level privacy risks in logit-based federated learning.

0 citationsRead paper

HexAGenT: Efficient Agentic LLM Serving via Workflow- and Heterogeneity-Aware Scheduling

May 15, 2026

This work addresses the challenge of end-to-end latency guarantees for online agent workflows in heterogeneous LLM serving clusters that decouple prefill and decode stages. It proposes the first scheduling framework tailored to such workflows, modeling each request as a runtime-revealed DAG. The framework predicts per-workflow independent completion times and prioritizes ready nodes based on their risk of missing service-level objectives (SLOs). It jointly optimizes placement across heterogeneous GPUs and queue priorities for both prefill and decode phases, while explicitly accounting for KV cache capacity constraints and cross-stage communication overhead. Experiments on A100/H100/H200 clusters demonstrate that, compared to baselines, the proposed approach reduces the average resource requirements to meet 95% and 99% workflow SLOs by 20.1% and 33.0%, respectively, with peak reductions reaching 45.0% and 80.5%.

0 citationsRead paper

Does Synthetic Layered Design Data Benefit Layered Design Decomposition?

May 14, 2026

Existing image generation methods produce flat outputs that are difficult to edit, while real-world layered data is scarce and non-scalable, hindering the practical deployment of layer decomposition techniques. This work proposes a data-driven approach based on SynLayers, a purely synthetic dataset, operating within the CLD framework by integrating vision-language models (VLMs) to generate textual supervision alongside bounding box inputs. It demonstrates for the first time that purely synthetic data can effectively substitute real data for layered design decomposition. The method overcomes data scalability limitations, enables balanced control over the distribution of layer counts, and outperforms non-scalable alternatives such as PrismLayersPro. Experiments show that model performance saturates at around 50,000 samples, significantly alleviating the layer count imbalance problem.

0 citationsRead paper