Yikai Zhu

Shenggui Li, Chao Wang, Yikai Zhu et al.

Large language models incur high inference latency due to sequential autoregressive decoding. Speculative decoding alleviates this bottleneck by using a lightweight draft model to propose multiple tokens for batched verification. However, its adoption has been limited by the lack of high-quality draft models and scalable training infrastructure. We introduce SpecForge, an open-source, production-oriented framework for training speculative decoding models with full support for EAGLE-3. SpecForge incorporates target-draft decoupling, hybrid parallelism, optimized training kernels, and integration with production-grade inference engines, enabling up to 9.9x faster EAGLE-3 training for Qwen3-235B-A22B. In addition, we release SpecBundle, a suite of production-grade EAGLE-3 draft models trained with SpecForge for mainstream open-source LLMs. Through a systematic study of speculative decoding training recipes, SpecBundle addresses the scarcity of high-quality drafts in the community, and our draft models achieve up to 4.48x end-to-end inference speedup on SGLang, establishing SpecForge as a practical foundation for real-world speculative decoding deployment.

Yikai Zhu, Kunfeng Chen, Qihuang Zhong et al.

Retrieval-augmented generation (RAG) has emerged as a promising paradigm for enhancing large language models (LLMs) on multi-hop question answering (QA), which requires reasoning over evidence from multiple documents. Current multi-hop RAG methods generally focus on either query-side task decomposition or corpus-side knowledge graph construction. Despite their progress, these methods still struggle to achieve satisfactory performance on complex multi-hop QA tasks. To this end, we propose ConRAG, a consensus-driven multi-view RAG framework that effectively boosts LLMs on complex multi-hop QA. The core of ConRAG is to systematically optimize both the query and corpus sides and to leverage multi-view evidence (relation, entity, and text signals) for more accurate retrieval. Extensive experiments on three multi-hop QA benchmarks show that ConRAG consistently outperforms all baselines by a clear margin, e.g., up to +26.9% average performance gains over vanilla RAG, and enables Gemma-4-31B to achieve a new state-of-the-art record on the challenging MuSiQue benchmark.

Chenkai Guo, Yikai Zhu, Jing Yangum et al.

By incorporating spatial location information, spatial-transcriptomics clustering yields more comprehensive insights into cell subpopulation identification. Despite recent progress, existing methods have at least two limitations: (i) topological learning typically considers only representations of individual cells or their interaction graphs; however, spatial transcriptomic profiles are often noisy, making these approaches vulnerable to low-quality topological signals, and (ii) insufficient modeling of spatial neighborhood information leads to low-quality spatial embeddings. To address these limitations, we propose SPHENIC, a novel Spatial Persistent Homology Enhanced Neighborhood Integrative Clustering method. Specifically, SPHENIC incorporates invariant topological features into the clustering network to achieve stable representation learning. Additionally, to construct high-quality spatial embeddings that reflect the true cellular distribution, we design the Spatial Constraint and Distribution Optimization Module (SCDOM). This module increases the similarity between a cell's embedding and those of its spatial neighbors, decreases similarity with non-neighboring cells, and thereby produces clustering-friendly spatial embeddings. Extensive experiments on 14 benchmark spatial transcriptomic slices demonstrate that SPHENIC achieves superior performance on the spatial clustering task, outperforming existing state-of-the-art methods by 3.31%-6.54% over the best alternative.

Jianbo Dong, Bin Luo, Jun Zhang et al.

The emergence of Large Language Models (LLMs) has necessitated the adoption of distributed training techniques, involving the deployment of thousands of GPUs to train a single model. Unfortunately, the efficiency of large-scale distributed training systems is often suboptimal due to the increased likelihood of hardware errors in high-end GPU products and the heightened risk of network traffic collisions. Moreover, any local hardware failure can disrupt training tasks, and the inability to swiftly identify faulty components leads to a significant waste of GPU resources. And, prolonged communication due to traffic collisions can substantially increase GPU waiting times. To address these challenges, we propose a communication-driven solution, namely the C4. The key insights of C4 are twofold. First, the load in distributed training exhibits homogeneous characteristics and is divided into iterations through periodic synchronization, therefore hardware anomalies would incur certain syndrome in collective communication. By leveraging this feature, C4 can rapidly identify the faulty components, swiftly isolate the anomaly, and restart the task, thereby avoiding resource wastage caused by delays in anomaly detection. Second, the predictable communication model of collective communication, involving a limited number of long-lived flows, allows C4 to efficiently execute traffic planning, substantially reducing bandwidth competition among these flows. The C4 has been extensively deployed across real-world production systems in a hyperscale cloud provider, yielding a significant improvement in system efficiency, from 30% to 45%. This enhancement is attributed to a 30% reduction in error-induced overhead and a 15% reduction in communication costs.

Yikai Zhu, Jianhao Shen, Ming Zhang

In the task of factoid question answering over knowledge base, many questions have more than one plausible interpretation. Previous works on SimpleQuestions assume only one interpretation as the ground truth for each question, so they lack the ability to answer ambiguous questions correctly. In this paper, we present a new way to utilize the dataset that takes into account the existence of ambiguous questions. Then we introduce a simple and effective model which combines local knowledge subgraph with attention mechanism. Our experimental results show that our approach achieves outstanding performance in this task.