Step-TP is a dataset providing grounded, atomic step-level IR transitions and CoT supervision to enable reliable multi-step LLM-guided tensor program optimization instead of end-to-end imitation.
TritonRL: Training LLMs to think and code triton without cheating
7 Pith papers cite this work. Polarity classification is still indexing.
7
Pith papers citing it
years
2026 7representative citing papers
KernelBenchX benchmark shows task category explains nearly three times more variance in LLM kernel correctness than method choice, iterative refinement boosts correctness but reduces performance, and quantization remains unsolved.
Hawk raises NPU kernel generation accuracy from 49.4% to 80% and yields up to 2.2× speedups by retrieving and distilling structured hardware-aware knowledge without any model training.
EGG is a staged multi-agent system that encodes expert optimization principles to generate correct high-performance GPU kernels, reporting 2.13x average speedup over PyTorch on KernelBench and real workloads.
SpecGen introduces speculative generation to fork non-reasoning kernel candidates during LLM reasoning traces, enabling early termination and parallel profiling to reduce end-to-end optimization time on H200 GPUs.
daVinci-kernel trains one LLM to select, use, and summarize reusable GPU-kernel optimization skills in a single RL loop, reaching 37.2%, 70.6%, and 32.2% Fast-1 pass rates on KernelBench Levels 1-3 at 14B.
AscendOptimizer combines kernel rewinding for reusable experience with evolutionary search on hardware feedback to optimize Ascend NPU operators, delivering 1.21x geometric-mean speedup and faster performance on 53.47% of 101 tested operators versus baseline.
citing papers explorer
-
Step-TP: A Grounded, Step-Level Dataset with Chain-of-Thought Reasoning for LLM-Guided Tensor Program Optimization
Step-TP is a dataset providing grounded, atomic step-level IR transitions and CoT supervision to enable reliable multi-step LLM-guided tensor program optimization instead of end-to-end imitation.
-
KernelBenchX: A Comprehensive Benchmark for Evaluating LLM-Generated GPU Kernels
KernelBenchX benchmark shows task category explains nearly three times more variance in LLM kernel correctness than method choice, iterative refinement boosts correctness but reduces performance, and quantization remains unsolved.
-
Hawk: Harnessing Hardware-Aware Knowledge for High-Performance NPU Kernel Generation
Hawk raises NPU kernel generation accuracy from 49.4% to 80% and yields up to 2.2× speedups by retrieving and distilling structured hardware-aware knowledge without any model training.
-
EGG: An Expert-Guided Agent Framework for Kernel Generation
EGG is a staged multi-agent system that encodes expert optimization principles to generate correct high-performance GPU kernels, reporting 2.13x average speedup over PyTorch on KernelBench and real workloads.
-
SpecGen: Accelerating Agentic Kernel Optimization with Speculative Generation
SpecGen introduces speculative generation to fork non-reasoning kernel candidates during LLM reasoning traces, enabling early termination and parallel profiling to reduce end-to-end optimization time on H200 GPUs.
-
daVinci-kernel: Co-Evolving Skill Selection, Summarization, and Utilization via RL for GPU Kernel Optimization
daVinci-kernel trains one LLM to select, use, and summarize reusable GPU-kernel optimization skills in a single RL loop, reaching 37.2%, 70.6%, and 32.2% Fast-1 pass rates on KernelBench Levels 1-3 at 14B.
-
AscendOptimizer: Episodic Agent for Ascend NPU Operator Optimization
AscendOptimizer combines kernel rewinding for reusable experience with evolutionary search on hardware feedback to optimize Ascend NPU operators, delivering 1.21x geometric-mean speedup and faster performance on 53.47% of 101 tested operators versus baseline.