Chao Ni

Aakshita Chandiramani, Aaron Blakeman, Abdullahi Olaoye et al. · amazon-science, cmu

We describe the pre-training, post-training, and quantization of Nemotron 3 Super, a 120 billion (active 12 billion) parameter hybrid Mamba-Attention Mixture-of-Experts model. Nemotron 3 Super is the first model in the Nemotron 3 family to 1) be pre-trained in NVFP4, 2) leverage LatentMoE, a new Mixture-of-Experts architecture that optimizes for both accuracy per FLOP and accuracy per parameter, and 3) include MTP layers for inference acceleration through native speculative decoding. We pre-trained Nemotron 3 Super on 25 trillion tokens followed by post-training using supervised fine tuning (SFT) and reinforcement learning (RL). The final model supports up to 1M context length and achieves comparable accuracy on common benchmarks, while also achieving up to 2.2x and 7.5x higher inference throughput compared to GPT-OSS-120B and Qwen3.5-122B, respectively. Nemotron 3 Super datasets, along with the base, post-trained, and quantized checkpoints, are open-sourced on HuggingFace.

Felix Abecassis, Anjulie Agrusa, Dong Ahn et al. · nvidia

Large Language Models (LLMs) today are powerful problem solvers across many domains, and they continue to get stronger as they scale in model size, training set size, and training set quality, as shown by extensive research and experimentation across the industry. Training a frontier model today requires on the order of tens to hundreds of yottaflops, which is a massive investment of time, compute, and energy. Improving pretraining efficiency is therefore essential to enable the next generation of even more capable LLMs. While 8-bit floating point (FP8) training is now widely adopted, transitioning to even narrower precision, such as 4-bit floating point (FP4), could unlock additional improvements in computational speed and resource utilization. However, quantization at this level poses challenges to training stability, convergence, and implementation, notably for large-scale models trained on long token horizons. In this study, we introduce a novel approach for stable and accurate training of large language models (LLMs) using the NVFP4 format. Our method integrates Random Hadamard transforms (RHT) to bound block-level outliers, employs a two-dimensional quantization scheme for consistent representations across both the forward and backward passes, utilizes stochastic rounding for unbiased gradient estimation, and incorporates selective high-precision layers. We validate our approach by training a 12-billion-parameter model on 10 trillion tokens -- the longest publicly documented training run in 4-bit precision to date. Our results show that the model trained with our NVFP4-based pretraining technique achieves training loss and downstream task accuracies comparable to an FP8 baseline. These findings highlight that NVFP4, when combined with our training approach, represents a major step forward in narrow-precision LLM training algorithms.

Yunjie Ji, Yan Gong, Yiping Peng et al.

As a natural language assistant, ChatGPT is capable of performing various tasks, including but not limited to article generation, code completion, and data analysis. Furthermore, ChatGPT has consistently demonstrated a remarkable level of accuracy and reliability in terms of content evaluation, exhibiting the capability of mimicking human preferences. To further explore ChatGPT's potential in this regard, a study is conducted to assess its ability to rank content. In order to do so, a test set consisting of prompts is created, covering a wide range of use cases, and five models are utilized to generate corresponding responses. ChatGPT is then instructed to rank the responses generated by these models. The results on the test set show that ChatGPT's ranking preferences are consistent with human to a certain extent. This preliminary experimental finding implies that ChatGPT's zero-shot ranking capability could be used to reduce annotation pressure in a number of ranking tasks.

Xin Yin, Chao Ni, Tien N. Nguyen et al.

Software migration is garnering increasing attention with the evolution of software and society. Early studies mainly relied on handcrafted translation rules to translate between two languages, the translation process is error-prone and time-consuming. In recent years, researchers have begun to explore the use of pre-trained large language models (LLMs) in code translation. However, code translation is a complex task that LLMs would generate mistakes during code translation, they all produce certain types of errors when performing code translation tasks, which include (1) compilation error, (2) runtime error, (3) functional error, and (4) non-terminating execution. We found that the root causes of these errors are very similar (e.g. failure to import packages, errors in loop boundaries, operator errors, and more). In this paper, we propose a general corrector, namely Rectifier, which is a micro and universal model for repairing translation errors. It learns from errors generated by existing LLMs and can be widely applied to correct errors generated by any LLM. The experimental results on translation tasks between C++, Java, and Python show that our model has effective repair ability, and cross experiments also demonstrate the robustness of our method.

Chao Ni, Liyu Shen, Xiaodan Xu et al.

Though many deep learning-based models have made great progress in vulnerability detection, we have no good understanding of these models, which limits the further advancement of model capability, understanding of the mechanism of model detection, and efficiency and safety of practical application of models. In this paper, we extensively and comprehensively investigate two types of state-of-the-art learning-based approaches (sequence-based and graph-based) by conducting experiments on a recently built large-scale dataset. We investigate seven research questions from five dimensions, namely model capabilities, model interpretation, model stability, ease of use of model, and model economy. We experimentally demonstrate the priority of sequence-based models and the limited abilities of both LLM (ChatGPT) and graph-based models. We explore the types of vulnerability that learning-based models skilled in and reveal the instability of the models though the input is subtlely semantical-equivalently changed. We empirically explain what the models have learned. We summarize the pre-processing as well as requirements for easily using the models. Finally, we initially induce the vital information for economically and safely practical usage of these models.

Le Deng, Xiaoxue Ren, Chao Ni et al.

Project-specific code completion is a critical task that leverages context from a project to generate accurate code. State-of-the-art methods use retrieval-augmented generation (RAG) with large language models (LLMs) and project information for code completion. However, they often struggle to incorporate internal API information, which is crucial for accuracy, especially when APIs are not explicitly imported in the file. To address this, we propose a method to infer internal API information without relying on imports. Our method extends the representation of APIs by constructing usage examples and semantic descriptions, building a knowledge base for LLMs to generate relevant completions. We also introduce ProjBench, a benchmark that avoids leaked imports and consists of large-scale real-world projects. Experiments on ProjBench and CrossCodeEval show that our approach significantly outperforms existing methods, improving code exact match by 22.72% and identifier exact match by 18.31%. Additionally, integrating our method with existing baselines boosts code match by 47.80% and identifier match by 35.55%.

Xin Yin, Chao Ni, Liushan Chen et al.

Large Language Models (LLMs) have demonstrated remarkable potential in automating software development tasks. While recent advances leverage Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO) to align models with human preferences, the optimal training strategy remains unclear across diverse code preference scenarios. This paper systematically investigates the roles of SFT and DPO in aligning LLMs with different code preferences. Through both theoretical analysis and empirical observation, we hypothesize that SFT excels in scenarios with objectively verifiable optimal solutions, while applying SFT followed by DPO (S&D) enables models to explore superior solutions in scenarios without objectively verifiable optimal solutions. Based on the analysis and experimental evidence, we propose Adaptive Preference Optimization (APO), a dynamic integration approach that adaptively amplifies preferred responses, suppresses dispreferred ones, and encourages exploration of potentially superior solutions during training. Extensive experiments across six representative code preference tasks validate our theoretical hypotheses and demonstrate that APO consistently matches or surpasses the performance of existing SFT and S&D strategies. Our work provides both theoretical foundations and practical guidance for selecting appropriate training strategies in different code preference alignment scenarios.

Dianshu Liao, Xin Yin, Shidong Pan et al.

Unit testing is essential for software quality assurance, yet writing and maintaining tests remains time-consuming and error-prone. To address this challenge, researchers have proposed various techniques for automating unit test generation, including traditional heuristic-based methods and more recent approaches that leverage large language models (LLMs). However, these existing approaches are inherently path-insensitive because they rely on fixed heuristics or limited contextual information and fail to reason about deep control-flow structures. As a result, they often struggle to achieve adequate coverage, particularly for deep or complex execution paths. In this work, we present a path-sensitive framework, JUnitGenie, to fill this gap by combining code knowledge with the semantic capabilities of LLMs in guiding context-aware unit test generation. After extracting code knowledge from Java projects, JUnitGenie distills this knowledge into structured prompts to guide the generation of high-coverage unit tests. We evaluate JUnitGenie on 2,258 complex focal methods from ten real-world Java projects. The results show that JUnitGenie generates valid tests and improves branch and line coverage by 29.60% and 31.00% on average over both heuristic and LLM-based baselines. We further demonstrate that the generated test cases can uncover real-world bugs, which were later confirmed and fixed by developers.

Lukas Schmid, Chao Ni, Yuliang Zhong et al.

Exploration is a fundamental problem in robotics. While sampling-based planners have shown high performance, they are oftentimes compute intensive and can exhibit high variance. To this end, we propose to directly learn the underlying distribution of informative views based on the spatial context in the robot's map. We further explore a variety of methods to also learn the information gain. We show in thorough experimental evaluation that our proposed system improves exploration performance by up to 28% over classical methods, and find that learning the gains in addition to the sampling distribution can provide favorable performance vs. compute trade-offs for compute-constrained systems. We demonstrate in simulation and on a low-cost mobile robot that our system generalizes well to varying environments.

Xiang Chen, Yanzhou Mu, Chao Ni et al.

Until now, researchers have proposed several novel heterogeneous defect prediction HDP methods with promising performance. To the best of our knowledge, whether HDP methods can perform significantly better than unsupervised methods has not yet been thoroughly investigated. In this article, we perform a replication study to have a holistic look in this issue. In particular, we compare state-of-the-art five HDP methods with five unsupervised methods. Final results surprisingly show that these HDP methods do not perform significantly better than some of unsupervised methods (especially the simple unsupervised methods proposed by Zhou et al.) in terms of two non-effort-aware performance measures and four effort-aware performance measures. Then, we perform diversity analysis on defective modules via McNemar's test and find the prediction diversity is more obvious when the comparison is performed between the HDP methods and the unsupervised methods than the comparisons only between the HDP methods or between the unsupervised methods. This shows the HDP methods and the unsupervised methods are complementary to each other in identifying defective models to some extent. Finally, we investigate the feasibility of five HDP methods by considering two satisfactory criteria recommended by previous CPDP studies and find the satisfactory ratio of these HDP methods is still pessimistic. The above empirical results implicate there is still a long way for heterogeneous defect prediction to go. More effective HDP methods need to be designed and the unsupervised methods should be considered as baselines.