Published in ACL ARR 2026 May Submission (under review), 2026
Under review at ACL ARR 2026 May Submission. IntentEval is a benchmark for evaluating whether LLMs answer the user’s intended question, using human-derived probability distributions over interpretations and an intent-aware judge. It achieves 0.90 correlation with LMArena rankings.
Published in NeurIPS 2026 Conference Submission (under review), 2026
Under review at NeurIPS 2026. Fusion-R1 is an LLM-based fusion proposal layer for TorchInductor trained with GRPO, improving speedup over PyTorch eager mode from 1.74× to 1.83× and reducing average kernel count from 99.10 to 77.18.
Published in ACL ARR 2026 March Submission (under review), 2026
Under review at ACL ARR 2026 March Submission. This is my first paper and it studies reinforcement learning for hierarchical syntactic prediction in large language models.