2026

CUDACoach: Guiding CUDA Kernel Optimization with Diverse Strategies Learned from Verified Code
CUDACoach: Guiding CUDA Kernel Optimization with Diverse Strategies Learned from Verified Code

Guoqiang Zhang, Junrui Shen, Shiyao Li, Yi Cai, Shiqiang Wu, Zhiheng Xi, Tao Gui, Qi Zhang, Xuanjing Huang, Hongzhong Zheng

Submitted to ICLR 2027 Under Review

CUDACoach learns repair and optimization strategies to guide a frozen coder through verified strategy mining and reinforcement learning with a two-level diversity reward. Evaluated on all 250 KernelBench tasks with three downstream coders, it achieves 60.4% correctness with Qwen3-Coder-Plus and produces correct kernels faster than PyTorch on 19.2% of tasks.

CUDACoach: Guiding CUDA Kernel Optimization with Diverse Strategies Learned from Verified Code

Guoqiang Zhang, Junrui Shen, Shiyao Li, Yi Cai, Shiqiang Wu, Zhiheng Xi, Tao Gui, Qi Zhang, Xuanjing Huang, Hongzhong Zheng

Submitted to ICLR 2027 Under Review

CUDACoach learns repair and optimization strategies to guide a frozen coder through verified strategy mining and reinforcement learning with a two-level diversity reward. Evaluated on all 250 KernelBench tasks with three downstream coders, it achieves 60.4% correctness with Qwen3-Coder-Plus and produces correct kernels faster than PyTorch on 19.2% of tasks.

NovGauge: A Fine-Grained Benchmark for Diagnosing LLMs' Capability in Paper Novelty Assessment
NovGauge: A Fine-Grained Benchmark for Diagnosing LLMs' Capability in Paper Novelty Assessment

Guoqiang Zhang, Kexin Tan, Ming Zhang, Li Ju, Wenqing Jing, Zhonghan Yue, Jiayi Chen, Shiqiang Wu, Shaofan Liu, Yue Zhang, Yuankai Ying, Yang Shi, Tao Gui, Qi Zhang, Xuanjing Huang

Submitted to EMNLP 2026 Under Review

NovGauge is a human-anchored benchmark for diagnosing LLM-based scientific novelty assessment across task, problem, and method dimensions. Evaluation of 18 LLMs on 619 paper pairs and 50 multi-paper sets shows that faithfulness verification reduces most models' raw F1 by more than half, exposing substantial hallucination and unsupported-evidence failures.

NovGauge: A Fine-Grained Benchmark for Diagnosing LLMs' Capability in Paper Novelty Assessment

Guoqiang Zhang, Kexin Tan, Ming Zhang, Li Ju, Wenqing Jing, Zhonghan Yue, Jiayi Chen, Shiqiang Wu, Shaofan Liu, Yue Zhang, Yuankai Ying, Yang Shi, Tao Gui, Qi Zhang, Xuanjing Huang

Submitted to EMNLP 2026 Under Review

NovGauge is a human-anchored benchmark for diagnosing LLM-based scientific novelty assessment across task, problem, and method dimensions. Evaluation of 18 LLMs on 619 paper pairs and 50 multi-paper sets shows that faithfulness verification reduces most models' raw F1 by more than half, exposing substantial hallucination and unsupported-evidence failures.