Large language models have become increasingly capable agents for low-level code and kernel optimization, but isolated kernel benchmarks provide only a proxy for the deployment behavior that matters in language-model inference. We identify a benchmark-to-deployment gap: candidate kernels that appear correct and fast in standalone harnesses can exhibit different performance, safety, or phase behavior after integration into a real inference workload. We introduce LLM4LLM, a deployment-aware closed-loop optimization framework that starts from a target inference script, extracts phase-aware optimization tasks, searches with an experience-guided episodic agent, and accepts patches through in-model validation. Across ten language-model inference workloads on A100 and H100 GPUs, LLM4LLM improves end-to-end latency for every evaluated model, achieving 3.91x/6.98x geometric-mean speedups on A100/H100; as supporting kernel-level evidence, it also attains up to 2.745× GeoMean speedup on KernelBench Level 2.

This is LanguageX (ZengHuiMT)’s submission to WMT 2023 General Machine Translation task for 13 language directions. We initially employ an encoder-decoder model to train on all 13 competition translation directions as our baseline system. Subsequently, we adopt a decoder-only architecture and fine-tune a multilingual language model by partially sampling data from diverse multilingual datasets such as CC100 and WuDaoCorpora. This is further refined using carefully curated high-quality parallel corpora across multiple translation directions to enable the model to perform translation tasks. As per automated evaluation metrics, our model ranks first in the translation directions from English to Russian, English to German, and English to Ukrainian. It secures the second position in the directions from English to Czech, English to Hebrew, Hebrew to English, and Ukrainian to English, and ranks third in German to English, Japanese to English, and Russian to English among all participating teams. Our best-performing model, covering 13 translation directions, stands on par with GPT-4. Among all 13 translation directions, our multilingual model surpasses GPT-4 in bleu scores for 7 translation directions.

We participated in the WMT General MT task and focus on four high resource language pairs: English to Chinese, Chinese to English, English to Japanese and Japanese to English). The submitted systems (LanguageX) focus on data cleaning, data selection, data mixing and TM-augmented NMT. Rules and multilingual language model are used for data filtering and data selection. In the automatic evaluation, our best submitted English to Chinese system achieved 54.3 BLEU score and 63.8 COMET score, which is the highest among all the submissions.

I participated in the WMT shared news translation task and focus on one high resource language pair: English and Chinese (two directions, Chinese to English and English to Chinese). The submitted systems (ZengHuiMT) focus on data cleaning, data selection, back translation and model ensemble. The techniques I used for data filtering and selection include filtering by rules, language model and word alignment. I used a base translation model trained on initial corpus to obtain the target versions of the WMT21 test sets, then I used language models to find out the monolingual data that is most similar to the target version of test set, such monolingual data was then used to do back translation. On the test set, my best submitted systems achieve 35.9 and 32.2 BLEU for English to Chinese and Chinese to English directions respectively, which are quite high for a small model.