Mozhi Zhang

Qinyuan Cheng, Tianxiang Sun, Wenwei Zhang et al.

In this paper, we establish a benchmark named HalluQA (Chinese Hallucination Question-Answering) to measure the hallucination phenomenon in Chinese large language models. HalluQA contains 450 meticulously designed adversarial questions, spanning multiple domains, and takes into account Chinese historical culture, customs, and social phenomena. During the construction of HalluQA, we consider two types of hallucinations: imitative falsehoods and factual errors, and we construct adversarial samples based on GLM-130B and ChatGPT. For evaluation, we design an automated evaluation method using GPT-4 to judge whether a model output is hallucinated. We conduct extensive experiments on 24 large language models, including ERNIE-Bot, Baichuan2, ChatGLM, Qwen, SparkDesk and etc. Out of the 24 models, 18 achieved non-hallucination rates lower than 50%. This indicates that HalluQA is highly challenging. We analyze the primary types of hallucinations in different types of models and their causes. Additionally, we discuss which types of hallucinations should be prioritized for different types of models.

MiniMax, Aili Chen, Aonian Li et al.

We introduce MiniMax-M1, the world's first open-weight, large-scale hybrid-attention reasoning model. MiniMax-M1 is powered by a hybrid Mixture-of-Experts (MoE) architecture combined with a lightning attention mechanism. The model is developed based on our previous MiniMax-Text-01 model, which contains a total of 456 billion parameters with 45.9 billion parameters activated per token. The M1 model natively supports a context length of 1 million tokens, 8x the context size of DeepSeek R1. Furthermore, the lightning attention mechanism in MiniMax-M1 enables efficient scaling of test-time compute. These properties make M1 particularly suitable for complex tasks that require processing long inputs and thinking extensively. MiniMax-M1 is trained using large-scale reinforcement learning (RL) on diverse problems including sandbox-based, real-world software engineering environments. In addition to M1's inherent efficiency advantage for RL training, we propose CISPO, a novel RL algorithm to further enhance RL efficiency. CISPO clips importance sampling weights rather than token updates, outperforming other competitive RL variants. Combining hybrid-attention and CISPO enables MiniMax-M1's full RL training on 512 H800 GPUs to complete in only three weeks, with a rental cost of just $534,700. We release two versions of MiniMax-M1 models with 40K and 80K thinking budgets respectively, where the 40K model represents an intermediate phase of the 80K training. Experiments on standard benchmarks show that our models are comparable or superior to strong open-weight models such as the original DeepSeek-R1 and Qwen3-235B, with particular strengths in complex software engineering, tool utilization, and long-context tasks. We publicly release MiniMax-M1 at https://github.com/MiniMax-AI/MiniMax-M1.

Kyle Seelman, Mozhi Zhang, Jordan Boyd-Graber

Topic models are valuable for understanding extensive document collections, but they don't always identify the most relevant topics. Classical probabilistic and anchor-based topic models offer interactive versions that allow users to guide the models towards more pertinent topics. However, such interactive features have been lacking in neural topic models. To correct this lacuna, we introduce a user-friendly interaction for neural topic models. This interaction permits users to assign a word label to a topic, leading to an update in the topic model where the words in the topic become closely aligned with the given label. Our approach encompasses two distinct kinds of neural topic models. The first includes models where topic embeddings are trainable and evolve during the training process. The second kind involves models where topic embeddings are integrated post-training, offering a different approach to topic refinement. To facilitate user interaction with these neural topic models, we have developed an interactive interface. This interface enables users to engage with and re-label topics as desired. We evaluate our method through a human study, where users can relabel topics to find relevant documents. Using our method, user labeling improves document rank scores, helping to find more relevant documents to a given query when compared to no user labeling.

Junteng Liu, Yuanxiang Fan, Zhuo Jiang et al.

Recent advances such as OpenAI-o1 and DeepSeek R1 have demonstrated the potential of Reinforcement Learning (RL) to enhance reasoning abilities in Large Language Models (LLMs). While open-source replication efforts have primarily focused on mathematical and coding domains, methods and resources for developing general reasoning capabilities remain underexplored. This gap is partly due to the challenge of collecting diverse and verifiable reasoning data suitable for RL. We hypothesize that logical reasoning is critical for developing general reasoning capabilities, as logic forms a fundamental building block of reasoning. In this work, we present SynLogic, a data synthesis framework and dataset that generates diverse logical reasoning data at scale, encompassing 35 diverse logical reasoning tasks. The SynLogic approach enables controlled synthesis of data with adjustable difficulty and quantity. Importantly, all examples can be verified by simple rules, making them ideally suited for RL with verifiable rewards. In our experiments, we validate the effectiveness of RL training on the SynLogic dataset based on 7B and 32B models. SynLogic leads to state-of-the-art logical reasoning performance among open-source datasets, surpassing DeepSeek-R1-Distill-Qwen-32B by 6 points on BBEH. Furthermore, mixing SynLogic data with mathematical and coding tasks improves the training efficiency of these domains and significantly enhances reasoning generalization. Notably, our mixed training model outperforms DeepSeek-R1-Zero-Qwen-32B across multiple benchmarks. These findings position SynLogic as a valuable resource for advancing the broader reasoning capabilities of LLMs. We open-source both the data synthesis pipeline and the SynLogic dataset at https://github.com/MiniMax-AI/SynLogic.

Mozhi Zhang, Wei Wang, Budhaditya Deb et al.

Reply suggestion models help users process emails and chats faster. Previous work only studies English reply suggestion. Instead, we present MRS, a multilingual reply suggestion dataset with ten languages. MRS can be used to compare two families of models: 1) retrieval models that select the reply from a fixed set and 2) generation models that produce the reply from scratch. Therefore, MRS complements existing cross-lingual generalization benchmarks that focus on classification and sequence labeling tasks. We build a generation model and a retrieval model as baselines for MRS. The two models have different strengths in the monolingual setting, and they require different strategies to generalize across languages. MRS is publicly available at https://github.com/zhangmozhi/mrs.

MiniMax, Aonian Li, Bangwei Gong et al.

We introduce MiniMax-01 series, including MiniMax-Text-01 and MiniMax-VL-01, which are comparable to top-tier models while offering superior capabilities in processing longer contexts. The core lies in lightning attention and its efficient scaling. To maximize computational capacity, we integrate it with Mixture of Experts (MoE), creating a model with 32 experts and 456 billion total parameters, of which 45.9 billion are activated for each token. We develop an optimized parallel strategy and highly efficient computation-communication overlap techniques for MoE and lightning attention. This approach enables us to conduct efficient training and inference on models with hundreds of billions of parameters across contexts spanning millions of tokens. The context window of MiniMax-Text-01 can reach up to 1 million tokens during training and extrapolate to 4 million tokens during inference at an affordable cost. Our vision-language model, MiniMax-VL-01 is built through continued training with 512 billion vision-language tokens. Experiments on both standard and in-house benchmarks show that our models match the performance of state-of-the-art models like GPT-4o and Claude-3.5-Sonnet while offering 20-32 times longer context window. We publicly release MiniMax-01 at https://github.com/MiniMax-AI.

Mozhi Zhang, Mianqiu Huang, Rundong Shi et al.

Large language models optimized with techniques like RLHF have achieved good alignment in being helpful and harmless. However, post-alignment, these language models often exhibit overconfidence, where the expressed confidence does not accurately calibrate with their correctness rate. In this paper, we decompose the language model confidence into the \textit{Uncertainty} about the question and the \textit{Fidelity} to the answer generated by language models. Then, we propose a plug-and-play method to estimate the confidence of language models. Our method has shown good calibration performance by conducting experiments with 6 RLHF-LMs on four MCQA datasets. Moreover, we propose two novel metrics, IPR and CE, to evaluate the calibration of the model, and we have conducted a detailed discussion on \textit{Truly Well-Calibrated Confidence}. Our method could serve as a strong baseline, and we hope that this work will provide some insights into the model confidence calibration.

Mozhi Zhang, Pengyu Wang, Chenkun Tan et al.

Large Language Models (LLMs) acquire extensive knowledge and remarkable abilities from extensive text corpora, making them powerful tools for various applications. To make LLMs more usable, aligning them with human preferences is essential. Existing alignment techniques, such as Reinforcement Learning from Human Feedback (RLHF) and Direct Preference Optimization (DPO), typically embed predefined preferences directly within the model's parameters. These methods, however, often result in a static alignment that can not account for the diversity of human preferences in practical applications. In response to this challenge, we propose an effective method, \textbf{MetaAlign}, which aims to help LLMs dynamically align with various explicit or implicit preferences specified at inference time. Experimental results show that LLMs optimized on our meticulously constructed MetaAlign Dataset can effectively align with any preferences specified at the inference stage, validating the feasibility of MetaAlign. We hope that our work can provide some insights into the alignment of language models.

Mianqiu Huang, Xiaoran Liu, Shaojun Zhou et al.

Recent advancements in model architectures and length extrapolation techniques have significantly extended the context length of large language models (LLMs), paving the way for their application in increasingly complex tasks. However, despite the growing capabilities of long-context LLMs, the safety issues in long-context scenarios remain underexplored. While safety alignment in short context has been widely studied, the safety concerns of long-context LLMs have not been adequately addressed. In this work, we introduce \textbf{LongSafety}, a comprehensive safety alignment dataset for long-context LLMs, containing 10 tasks and 17k samples, with an average length of 40.9k tokens. Our experiments demonstrate that training with LongSafety can enhance long-context safety performance while enhancing short-context safety and preserving general capabilities. Furthermore, we demonstrate that long-context safety does not equal long-context alignment with short-context safety data and LongSafety has generalizing capabilities in context length and long-context safety scenarios.

Mozhi Zhang, Howe Tissue, Lu Wang et al.

We introduce~\textsc{Domain2Vec}, a novel approach that decomposes any dataset into a linear combination of several \emph{meta-domains}, a new concept designed to capture the key underlying features of datasets. \textsc{Domain2Vec} maintains a vocabulary of meta-domains and uses a classifier to decompose any given dataset into a domain vector that corresponds to a distribution over this vocabulary. These domain vectors enable the identification of the optimal data mixture for language model (LM) pretraining in a training-free manner under the \emph{\textbf{D}istribution \textbf{A}lignment \textbf{A}ssumption} (DA$^{2}$), which suggests that when the data distributions of the training set and the validation set are better aligned, a lower validation loss is achieved. Moreover, \textsc{Domain2vec} can be seamlessly integrated into previous works to model the relationship between domain vectors and LM performance, greatly enhancing the efficiency and scalability of previous methods. Extensive experiments demonstrate that \textsc{Domain2Vec} helps find the data mixture that enhances downstream task performance with minimal computational overhead. Specifically, \textsc{Domain2Vec} achieves the same validation loss on Pile-CC using only $51.5\%$ of the computation required when training on the original mixture of The Pile dataset. Under equivalent compute budget, \textsc{Domain2Vec} improves downstream performance by an average of $2.83\%$.

Mozhi Zhang, Hang Yan, Yaqian Zhou et al.

Few-shot Named Entity Recognition (NER) is a task aiming to identify named entities via limited annotated samples. Recently, prototypical networks have shown promising performance in few-shot NER. Most of prototypical networks will utilize the entities from the support set to construct label prototypes and use the query set to compute span-level similarities and optimize these label prototype representations. However, these methods are usually unsuitable for fine-tuning in the target domain, where only the support set is available. In this paper, we propose PromptNER: a novel prompting method for few-shot NER via k nearest neighbor search. We use prompts that contains entity category information to construct label prototypes, which enables our model to fine-tune with only the support set. Our approach achieves excellent transfer learning ability, and extensive experiments on the Few-NERD and CrossNER datasets demonstrate that our model achieves superior performance over state-of-the-art methods.

Keyulu Xu, Mozhi Zhang, Stefanie Jegelka et al.

Graph Neural Networks (GNNs) have been studied through the lens of expressive power and generalization. However, their optimization properties are less well understood. We take the first step towards analyzing GNN training by studying the gradient dynamics of GNNs. First, we analyze linearized GNNs and prove that despite the non-convexity of training, convergence to a global minimum at a linear rate is guaranteed under mild assumptions that we validate on real-world graphs. Second, we study what may affect the GNNs' training speed. Our results show that the training of GNNs is implicitly accelerated by skip connections, more depth, and/or a good label distribution. Empirical results confirm that our theoretical results for linearized GNNs align with the training behavior of nonlinear GNNs. Our results provide the first theoretical support for the success of GNNs with skip connections in terms of optimization, and suggest that deep GNNs with skip connections would be promising in practice.

Jingling Li, Mozhi Zhang, Keyulu Xu et al.

Noisy labels are inevitable in large real-world datasets. In this work, we explore an area understudied by previous works -- how the network's architecture impacts its robustness to noisy labels. We provide a formal framework connecting the robustness of a network to the alignments between its architecture and target/noise functions. Our framework measures a network's robustness via the predictive power in its representations -- the test performance of a linear model trained on the learned representations using a small set of clean labels. We hypothesize that a network is more robust to noisy labels if its architecture is more aligned with the target function than the noise. To support our hypothesis, we provide both theoretical and empirical evidence across various neural network architectures and different domains. We also find that when the network is well-aligned with the target function, its predictive power in representations could improve upon state-of-the-art (SOTA) noisy-label-training methods in terms of test accuracy and even outperform sophisticated methods that use clean labels.

Keyulu Xu, Mozhi Zhang, Jingling Li et al.

We study how neural networks trained by gradient descent extrapolate, i.e., what they learn outside the support of the training distribution. Previous works report mixed empirical results when extrapolating with neural networks: while feedforward neural networks, a.k.a. multilayer perceptrons (MLPs), do not extrapolate well in certain simple tasks, Graph Neural Networks (GNNs) -- structured networks with MLP modules -- have shown some success in more complex tasks. Working towards a theoretical explanation, we identify conditions under which MLPs and GNNs extrapolate well. First, we quantify the observation that ReLU MLPs quickly converge to linear functions along any direction from the origin, which implies that ReLU MLPs do not extrapolate most nonlinear functions. But, they can provably learn a linear target function when the training distribution is sufficiently "diverse". Second, in connection to analyzing the successes and limitations of GNNs, these results suggest a hypothesis for which we provide theoretical and empirical evidence: the success of GNNs in extrapolating algorithmic tasks to new data (e.g., larger graphs or edge weights) relies on encoding task-specific non-linearities in the architecture or features. Our theoretical analysis builds on a connection of over-parameterized networks to the neural tangent kernel. Empirically, our theory holds across different training settings.

Mozhi Zhang, Yoshinari Fujinuma, Michael J. Paul et al.

Cross-lingual word embeddings (CLWE) are often evaluated on bilingual lexicon induction (BLI). Recent CLWE methods use linear projections, which underfit the training dictionary, to generalize on BLI. However, underfitting can hinder generalization to other downstream tasks that rely on words from the training dictionary. We address this limitation by retrofitting CLWE to the training dictionary, which pulls training translation pairs closer in the embedding space and overfits the training dictionary. This simple post-processing step often improves accuracy on two downstream tasks, despite lowering BLI test accuracy. We also retrofit to both the training dictionary and a synthetic dictionary induced from CLWE, which sometimes generalizes even better on downstream tasks. Our results confirm the importance of fully exploiting training dictionary in downstream tasks and explains why BLI is a flawed CLWE evaluation.

Michelle Yuan, Mozhi Zhang, Benjamin Van Durme et al.

Cross-lingual word embeddings transfer knowledge between languages: models trained on high-resource languages can predict in low-resource languages. We introduce CLIME, an interactive system to quickly refine cross-lingual word embeddings for a given classification problem. First, CLIME ranks words by their salience to the downstream task. Then, users mark similarity between keywords and their nearest neighbors in the embedding space. Finally, CLIME updates the embeddings using the annotations. We evaluate CLIME on identifying health-related text in four low-resource languages: Ilocano, Sinhalese, Tigrinya, and Uyghur. Embeddings refined by CLIME capture more nuanced word semantics and have higher test accuracy than the original embeddings. CLIME often improves accuracy faster than an active learning baseline and can be easily combined with active learning to improve results.

Mozhi Zhang, Keyulu Xu, Ken-ichi Kawarabayashi et al.

Cross-lingual word embeddings (CLWE) underlie many multilingual natural language processing systems, often through orthogonal transformations of pre-trained monolingual embeddings. However, orthogonal mapping only works on language pairs whose embeddings are naturally isomorphic. For non-isomorphic pairs, our method (Iterative Normalization) transforms monolingual embeddings to make orthogonal alignment easier by simultaneously enforcing that (1) individual word vectors are unit length, and (2) each language's average vector is zero. Iterative Normalization consistently improves word translation accuracy of three CLWE methods, with the largest improvement observed on English-Japanese (from 2% to 44% test accuracy).

Keyulu Xu, Jingling Li, Mozhi Zhang et al.

Neural networks have succeeded in many reasoning tasks. Empirically, these tasks require specialized network structures, e.g., Graph Neural Networks (GNNs) perform well on many such tasks, but less structured networks fail. Theoretically, there is limited understanding of why and when a network structure generalizes better than others, although they have equal expressive power. In this paper, we develop a framework to characterize which reasoning tasks a network can learn well, by studying how well its computation structure aligns with the algorithmic structure of the relevant reasoning process. We formally define this algorithmic alignment and derive a sample complexity bound that decreases with better alignment. This framework offers an explanation for the empirical success of popular reasoning models, and suggests their limitations. As an example, we unify seemingly different reasoning tasks, such as intuitive physics, visual question answering, and shortest paths, via the lens of a powerful algorithmic paradigm, dynamic programming (DP). We show that GNNs align with DP and thus are expected to solve these tasks. On several reasoning tasks, our theory is supported by empirical results.

Mozhi Zhang, Yoshinari Fujinuma, Jordan Boyd-Graber

Text classification must sometimes be applied in a low-resource language with no labeled training data. However, training data may be available in a related language. We investigate whether character-level knowledge transfer from a related language helps text classification. We present a cross-lingual document classification framework (CACO) that exploits cross-lingual subword similarity by jointly training a character-based embedder and a word-based classifier. The embedder derives vector representations for input words from their written forms, and the classifier makes predictions based on the word vectors. We use a joint character representation for both the source language and the target language, which allows the embedder to generalize knowledge about source language words to target language words with similar forms. We propose a multi-task objective that can further improve the model if additional cross-lingual or monolingual resources are available. Experiments confirm that character-level knowledge transfer is more data-efficient than word-level transfer between related languages.