Published in Under review, 2026
Under review. GraphFuse-R1 trains an 8B LLM with SFT and GRPO to propose whole-graph operator-fusion groups inside PyTorch’s TorchInductor, compiling 28.5% fewer kernels and fusing 39% more nodes than the hand-written heuristic, and derives a compiler-side bound on how much fusion headroom remains.
Published in Under review, 2026
Under review. A Muon-family optimizer that shifts update energy away from weight rows whose recent gradients conflict with the longer-term trend, reaching NorMuon’s loss with 1.07–1.10x higher token efficiency from 130M to 1.1B parameters at under 5% extra wall-clock time.
Published in NeurIPS 2026 Workshop on Neural Artifacts (NeuralArtifacts), Poster, 2026
NeurIPS 2026 NeuralArtifacts Workshop (Poster). A training-free, merge-operator-agnostic wrapper that provably never underperforms the base merge on validation. It gives the best result in every CLIP ViT and LoRA setting tested, reaching 94.29 on the LLaMA-3 8B six-NLI benchmark.
Published in Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing (EMNLP 2026), Main Conference, 2026
Accepted to EMNLP 2026 (Main Conference). Syntax-R1 is a two-stage post-training framework that uses reinforcement learning to improve hierarchical syntactic prediction in large language models, achieving state-of-the-art results on constituency parsing and CCG supertagging.