Shuangyong Song

Xiang Li, Yiqun Yao, Xin Jiang et al.

Large Language Models (LLMs) represent a significant stride toward Artificial General Intelligence. As scaling laws underscore the potential of increasing model sizes, the academic community has intensified its investigations into LLMs with capacities exceeding 50 billion parameters. This technical report builds on our prior work with Tele-FLM (also known as FLM-2), a publicly available 52-billion-parameter model. We delve into two primary areas: we first discuss our observation of Supervised Fine-tuning (SFT) on Tele-FLM-52B, which supports the "less is more" approach for SFT data construction; second, we demonstrate our experiments and analyses on the best practices for progressively growing a model from 52 billion to 102 billion, and subsequently to 1 trillion parameters. We will open-source a 1T model checkpoint, namely Tele-FLM-1T, to advance further training and research.

Xiang Li, Yiqun Yao, Xin Jiang et al.

Large language models (LLMs) have showcased profound capabilities in language understanding and generation, facilitating a wide array of applications. However, there is a notable paucity of detailed, open-sourced methodologies on efficiently scaling LLMs beyond 50 billion parameters with minimum trial-and-error cost and computational resources. In this report, we introduce Tele-FLM (aka FLM-2), a 52B open-sourced multilingual large language model that features a stable, efficient pre-training paradigm and enhanced factual judgment capabilities. Tele-FLM demonstrates superior multilingual language modeling abilities, measured by BPB on textual corpus. Besides, in both English and Chinese foundation model evaluation, it is comparable to strong open-sourced models that involve larger pre-training FLOPs, such as Llama2-70B and DeepSeek-67B. In addition to the model weights, we share the core designs, engineering practices, and training details, which we expect to benefit both the academic and industrial communities.

Zhongjiang He, Zihan Wang, Xinzhang Liu et al.

In this technical report, we present TeleChat, a collection of large language models (LLMs) with parameters of 3 billion, 7 billion and 12 billion. It includes pretrained language models as well as fine-tuned chat models that is aligned with human preferences. TeleChat is initially pretrained on an extensive corpus containing a diverse collection of texts from both English and Chinese languages, including trillions of tokens. Subsequently, the model undergoes fine-tuning to align with human preferences, following a detailed methodology that we describe. We evaluate the performance of TeleChat on various tasks, including language understanding, mathematics, reasoning, code generation, and knowledge-based question answering. Our findings indicate that TeleChat achieves comparable performance to other open-source models of similar size across a wide range of public benchmarks. To support future research and applications utilizing LLMs, we release the fine-tuned model checkpoints of TeleChat's 7B and 12B variant, along with code and a portion of our pretraining data, to the public community.

Zhenhe Wu, Zhongqiu Li, Jie Zhang et al.

Large language models (LLMs) with in-context learning have significantly improved the performance of text-to-SQL task. Previous works generally focus on using exclusive SQL generation prompt to improve the LLMs' reasoning ability. However, they are mostly hard to handle large databases with numerous tables and columns, and usually ignore the significance of pre-processing database and extracting valuable information for more efficient prompt engineering. Based on above analysis, we propose RB-SQL, a novel retrieval-based LLM framework for in-context prompt engineering, which consists of three modules that retrieve concise tables and columns as schema, and targeted examples for in-context learning. Experiment results demonstrate that our model achieves better performance than several competitive baselines on public datasets BIRD and Spider.

Zhenhe Wu, Jian Yang, Jiaheng Liu et al.

Tables present unique challenges for language models due to their structured row-column interactions, necessitating specialized approaches for effective comprehension. While large language models (LLMs) have demonstrated potential in table reasoning through prompting and techniques like chain-of-thought (CoT) and program-of-thought (PoT), optimizing their performance for table question answering remains underexplored. In this paper, we introduce region-based Table-R1, a novel reinforcement learning approach that enhances LLM table understanding by integrating region evidence into reasoning steps. Our method employs Region-Enhanced Supervised Fine-Tuning (RE-SFT) to guide models in identifying relevant table regions before generating answers, incorporating textual, symbolic, and program-based reasoning. Additionally, Table-Aware Group Relative Policy Optimization (TARPO) introduces a mixed reward system to dynamically balance region accuracy and answer correctness, with decaying region rewards and consistency penalties to align reasoning steps. Experiments show that Table-R1 achieves an average performance improvement of 14.36 points across multiple base models on three benchmark datasets, even outperforming baseline models with ten times the parameters, while TARPO reduces response token consumption by 67.5% compared to GRPO, significantly advancing LLM capabilities in efficient tabular reasoning.

Zihan Wang, Jiayu Xiao, Mengxiang Li et al.

In our dynamic world where data arrives in a continuous stream, continual learning enables us to incrementally add new tasks/domains without the need to retrain from scratch. A major challenge in continual learning of language model is catastrophic forgetting, the tendency of models to forget knowledge from previously trained tasks/domains when training on new ones. This paper studies dialog generation under the continual learning setting. We propose a novel method that 1) uses \textit{Text-Mixup} as data augmentation to avoid model overfitting on replay memory and 2) leverages Batch-Nuclear Norm Maximization (BNNM) to alleviate the problem of mode collapse. Experiments on a $37$-domain task-oriented dialog dataset and DailyDialog (a $10$-domain chitchat dataset) demonstrate that our proposed approach outperforms the state-of-the-art in continual learning.

Hongyan Xie, Yitong Yao, Yikun Ban et al.

Large language models (LLMs) excel at reasoning tasks but are expensive to deploy. Thus small language models (SLMs) are fine-tuned on CoT data generated by LLMs to copy LLMs' abilities. However, these CoT data may include noisy rationales that either fail to substantiate the answers or contribute no additional information to support answer prediction, which leads SLMs to capture spurious correlations between questions and answers and compromise the quality of reasoning. In this work, we propose Chain-of-Thought Correctness Perception Distillation (CoPeD), which aims to improve the reasoning quality of the student model from the perspectives of task setting and data utilization. Firstly, we introduce a correctness-aware task setting that encourages the student model to predict answers based on correct rationales and revise them when they are incorrect. This setting improves the faithfulness of reasoning and allows the model to learn from its mistakes. Then, we propose a Correctness-Aware Weighted loss, which dynamically adjusts the contribution of each training instance based on the combined loss of the rationale and the answer. This strategy encourages the model to focus more on samples where the rationale offers stronger support for the correct answer. Experiments have shown that CoPeD is effective on both in-distribution (IND) and out-of-distribution (OOD) benchmark reasoning datasets.

Jian Yang, Xianglong Liu, Weifeng Lv et al.

Large language models (LLMs) have fundamentally transformed automated software development by enabling direct translation of natural language descriptions into functional code, driving commercial adoption through tools like Github Copilot (Microsoft), Cursor (Anysphere), Trae (ByteDance), and Claude Code (Anthropic). While the field has evolved dramatically from rule-based systems to Transformer-based architectures, achieving performance improvements from single-digit to over 95\% success rates on benchmarks like HumanEval. In this work, we provide a comprehensive synthesis and practical guide (a series of analytic and probing experiments) about code LLMs, systematically examining the complete model life cycle from data curation to post-training through advanced prompting paradigms, code pre-training, supervised fine-tuning, reinforcement learning, and autonomous coding agents. We analyze the code capability of the general LLMs (GPT-4, Claude, LLaMA) and code-specialized LLMs (StarCoder, Code LLaMA, DeepSeek-Coder, and QwenCoder), critically examining the techniques, design decisions, and trade-offs. Further, we articulate the research-practice gap between academic research (e.g., benchmarks and tasks) and real-world deployment (e.g., software-related code tasks), including code correctness, security, contextual awareness of large codebases, and integration with development workflows, and map promising research directions to practical needs. Last, we conduct a series of experiments to provide a comprehensive analysis of code pre-training, supervised fine-tuning, and reinforcement learning, covering scaling law, framework selection, hyperparameter sensitivity, model architectures, and dataset comparisons.

Sishi Xiong, Ziyang He, Zhongjiang He et al.

While large language models (LLMs) have shown promise in the table question answering (TQA) task through prompt engineering, they face challenges in industrial applications, including structural heterogeneity, difficulties in target data localization, and bottlenecks in complex reasoning. To address these limitations, this paper presents TableZoomer, a novel LLM-powered, programming-based agent framework. It introduces three key innovations: (1) replacing the original fully verbalized table with structured table schema to bridge the semantic gap and reduce computational complexity; (2) a query-aware table zooming mechanism that dynamically generates sub-table schema through column selection and entity linking, significantly improving target localization efficiency; and (3) a Program-of-Thoughts (PoT) strategy that transforms queries into executable code to mitigate numerical hallucination. Additionally, we integrate the reasoning workflow with the ReAct paradigm to enable iterative reasoning. Extensive experiments demonstrate that our framework maintains the usability advantages while substantially enhancing performance and scalability across tables of varying scales. When implemented with the Qwen3-8B-Instruct LLM, TableZoomer achieves accuracy improvements of 19.34% and 25% over conventional PoT methods on the large-scale DataBench dataset and the small-scale Fact Checking task of TableBench dataset, respectively.

Wenhan Chang, Tianqing Zhu, Yu Zhao et al.

In the era of rapid generative AI development, interactions with large language models (LLMs) pose increasing risks of misuse. Prior research has primarily focused on attacks using template-based prompts and optimization-oriented methods, while overlooking the fact that LLMs possess strong unconstrained deceptive capabilities to attack other LLMs. This paper introduces a novel jailbreaking method inspired by the Chain-of-Thought mechanism. The attacker employs mission transfer to conceal harmful user intent within dialogue and generates a progressive chain of lure questions without relying on predefined templates, enabling successful jailbreaks. To further improve the attack's strength, we incorporate a helper LLM model that performs randomized narrative optimization over multi-turn interactions, enhancing the attack performance while preserving alignment with the original intent. We also propose a toxicity-based framework using third-party LLMs to evaluate harmful content and its alignment with malicious intent. Extensive experiments demonstrate that our method consistently achieves high attack success rates and elevated toxicity scores across diverse types of LLMs under black-box API settings. These findings reveal the intrinsic potential of LLMs to perform unrestricted attacks in the absence of robust alignment constraints. Our approach offers data-driven insights to inform the design of future alignment mechanisms. Finally, we propose two concrete defense strategies to support the development of safer generative models.

Yuxin Dong, Tieliang Gong, Hong Chen et al.

Domain generalization aims to learn invariance across multiple training domains, thereby enhancing generalization against out-of-distribution data. While gradient or representation matching algorithms have achieved remarkable success, these methods generally lack generalization guarantees or depend on strong assumptions, leaving a gap in understanding the underlying mechanism of distribution matching. In this work, we formulate domain generalization from a novel probabilistic perspective, ensuring robustness while avoiding overly conservative solutions. Through comprehensive information-theoretic analysis, we provide key insights into the roles of gradient and representation matching in promoting generalization. Our results reveal the complementary relationship between these two components, indicating that existing works focusing solely on either gradient or representation alignment are insufficient to solve the domain generalization problem. In light of these theoretical findings, we introduce IDM to simultaneously align the inter-domain gradients and representations. Integrated with the proposed PDM method for complex distribution matching, IDM achieves superior performance over various baseline methods.

Shuangyong Song

Knowledge graphs (KGs) have been widely used for question answering (QA) applications, especially the entity based QA. However, searching an-swers from an entire large-scale knowledge graph is very time-consuming and it is hard to meet the speed need of real online QA systems. In this pa-per, we design a sub-graph searching mechanism to solve this problem by creating sub-graph index, and each answer generation step is restricted in the sub-graph level. We use this mechanism into a real online QA chat system, and it can bring obvious improvement on question coverage by well answer-ing entity based questions, and it can be with a very high speed, which en-sures the user experience of online QA.

Shuangyong Song, Kexin Wang, Chao Wang et al.

In response generation task, proper sentimental expressions can obviously improve the human-like level of the responses. However, for real application in online systems, high QPS (queries per second, an indicator of the flow capacity of on-line systems) is required, and a dynamic vocabulary mechanism has been proved available in improving speed of generative models. In this paper, we proposed an emotion-controlled dialog response generation model based on the dynamic vocabulary mechanism, and the experimental results show the benefit of this model.

Shuangyong Song, Chao Wang

Automatic question-answering (QA) systems have boomed during last few years, and commonly used techniques can be roughly categorized into Information Retrieval (IR)-based and generation-based. A key solution to the IR based models is to retrieve the most similar knowledge entries of a given query from a QA knowledge base, and then rerank those knowledge entries with semantic matching models. In this paper, we aim to improve an IR based e-commerce QA system-AliMe with proposed text matching models, including a basic Triple Convolutional Neural Network (TCNN) model and two Attention-based TCNN (ATCNN) models. Experimental results show their effect.

Shuangyong Song, Chao Wang, Qianqian Xie et al.

Conversational context understanding aims to recognize the real intention of user from the conversation history, which is critical for building the dialogue system. However, the multi-turn conversation understanding in open domain is still quite challenging, which requires the system extracting the important information and resolving the dependencies in contexts among a variety of open topics. In this paper, we propose the conversational query rewriting model - MLR, which is a Multi-task model on sequence Labeling and query Rewriting. MLR reformulates the multi-turn conversational queries into a single turn query, which conveys the true intention of users concisely and alleviates the difficulty of the multi-turn dialogue modeling. In the model, we formulate the query rewriting as a sequence generation problem and introduce word category information via the auxiliary word category label predicting task. To train our model, we construct a new Chinese query rewriting dataset and conduct experiments on it. The experimental results show that our model outperforms compared models, and prove the effectiveness of the word category information in improving the rewriting performance.

Jianfei Yu, Minghui Qiu, Jing Jiang et al.

In this paper, we study transfer learning for the PI and NLI problems, aiming to propose a general framework, which can effectively and efficiently adapt the shared knowledge learned from a resource-rich source domain to a resource- poor target domain. Specifically, since most existing transfer learning methods only focus on learning a shared feature space across domains while ignoring the relationship between the source and target domains, we propose to simultaneously learn shared representations and domain relationships in a unified framework. Furthermore, we propose an efficient and effective hybrid model by combining a sentence encoding- based method and a sentence interaction-based method as our base model. Extensive experiments on both paraphrase identification and natural language inference demonstrate that our base model is efficient and has promising performance compared to the competing models, and our transfer learning method can help to significantly boost the performance. Further analysis shows that the inter-domain and intra-domain relationship captured by our model are insightful. Last but not least, we deploy our transfer learning model for PI into our online chatbot system, which can bring in significant improvements over our existing system. Finally, we launch our new system on the chatbot platform Eva in our E-commerce site AliExpress.

Shuangyong Song, Yao Meng, Zhongguang Zheng et al.

Our FRDC_QA team participated in the QA-Lab English subtask of the NTCIR-11. In this paper, we describe our system for solving real-world university entrance exam questions, which are related to world history. Wikipedia is used as the main external resource for our system. Since problems with choosing right/wrong sentence from multiple sentence choices account for about two-thirds of the total, we individually design a classification based model for solving this type of questions. For other types of questions, we also design some simple methods.

Shuangyong Song, Yao Meng

In microblogging, hashtags are used to be topical markers, and they are adopted by users that contribute similar content or express a related idea. However, hashtags are created in a free style and there is no domain category information about them, which make users hard to get access to organized hashtag presentation. In this paper, we propose an approach that classifies hashtags with news categories, and then carry out a domain-sensitive popularity ranking to get hot hashtags in each domain. The proposed approach first trains a domain classification model with news content and news category information, then detects microblogs related to a hashtag to be its representative text, based on which we can classify this hashtag with a domain. Finally, we calculate the domain-sensitive popularity of each hashtag with multiple factors, to get most hotly discussed hashtags in each domain. Preliminary experimental results on a dataset from Sina Weibo, one of the largest Chinese microblogging websites, show usefulness of the proposed approach on describing hashtags.

Shuangyong Song, Yao Meng

In this paper, we propose a Concept-level Emotion Cause Model (CECM), instead of the mere word-level models, to discover causes of microblogging users' diversified emotions on specific hot event. A modified topic-supervised biterm topic model is utilized in CECM to detect emotion topics' in event-related tweets, and then context-sensitive topical PageRank is utilized to detect meaningful multiword expressions as emotion causes. Experimental results on a dataset from Sina Weibo, one of the largest microblogging websites in China, show CECM can better detect emotion causes than baseline methods.