Keqin Bao

Keqin Bao, Yu Wan, Dayiheng Liu et al.

In this paper, we present our submission to the sentence-level MQM benchmark at Quality Estimation Shared Task, named UniTE (Unified Translation Evaluation). Specifically, our systems employ the framework of UniTE, which combined three types of input formats during training with a pre-trained language model. First, we apply the pseudo-labeled data examples for the continuously pre-training phase. Notably, to reduce the gap between pre-training and fine-tuning, we use data pruning and a ranking-based score normalization strategy. For the fine-tuning phase, we use both Direct Assessment (DA) and Multidimensional Quality Metrics (MQM) data from past years' WMT competitions. Finally, we collect the source-only evaluation results, and ensemble the predictions generated by two UniTE models, whose backbones are XLM-R and InfoXLM, respectively. Results show that our models reach 1st overall ranking in the Multilingual and English-Russian settings, and 2nd overall ranking in English-German and Chinese-English settings, showing relatively strong performances in this year's quality estimation competition.

Keqin Bao, Yu Wan, Dayiheng Liu et al.

Fine-grained information on translation errors is helpful for the translation evaluation community. Existing approaches can not synchronously consider error position and type, failing to integrate the error information of both. In this paper, we propose Fine-Grained Translation Error Detection (FG-TED) task, aiming at identifying both the position and the type of translation errors on given source-hypothesis sentence pairs. Besides, we build an FG-TED model to predict the \textbf{addition} and \textbf{omission} errors -- two typical translation accuracy errors. First, we use a word-level classification paradigm to form our model and use the shortcut learning reduction to relieve the influence of monolingual features. Besides, we construct synthetic datasets for model training, and relieve the disagreement of data labeling in authoritative datasets, making the experimental benchmark concordant. Experiments show that our model can identify both error type and position concurrently, and gives state-of-the-art results on the restored dataset. Our model also delivers more reliable predictions on low-resource and transfer scenarios than existing baselines. The related datasets and the source code will be released in the future.

Yu Wan, Keqin Bao, Dayiheng Liu et al.

In this report, we present our submission to the WMT 2022 Metrics Shared Task. We build our system based on the core idea of UNITE (Unified Translation Evaluation), which unifies source-only, reference-only, and source-reference-combined evaluation scenarios into one single model. Specifically, during the model pre-training phase, we first apply the pseudo-labeled data examples to continuously pre-train UNITE. Notably, to reduce the gap between pre-training and fine-tuning, we use data cropping and a ranking-based score normalization strategy. During the fine-tuning phase, we use both Direct Assessment (DA) and Multidimensional Quality Metrics (MQM) data from past years' WMT competitions. Specially, we collect the results from models with different pre-trained language model backbones, and use different ensembling strategies for involved translation directions.

Xiaoyuan Li, Yubo Ma, Chengpeng Li et al.

Effectively training Large Language Models (LLMs) for complex, long-CoT reasoning is often bottlenecked by the need for massive high-quality reasoning data. Existing methods are either computationally expensive or fail to reliably distinguish high- from low-quality reasoning samples. To address this, we propose High-Entropy Sum (HES), a training-free metric that quantifies reasoning quality by summing only the entropy of the top (e.g., 0.5\%) highest-entropy tokens in each reasoning sample. We validate HES across three mainstream training paradigms: Supervised Fine-tuning (SFT), Rejection Fine-tuning (RFT), and Reinforcement Learning (RL), with extensive results demonstrating its consistent effectiveness and significantly reduced computational overhead. In SFT, training on the top 20\% HES-ranked data matches full-dataset performance, while using the lowest-HES data degrades it. In RFT, our HES-based training approach significantly outperforms baseline methods. In RL, HES-selected successful trajectories enable the model to learn strong reasoning patterns, significantly surpassing other compared methods. Our findings establish HES as a robust, training-free metric that enables a unified, effective, and efficient method for developing advanced reasoning in LLMs.

Xiaoyuan Li, Keqin Bao, Moxin Li et al.

Rubric-based rewards offer a promising way to extend reinforcement learning (RL) for large language models beyond tasks with automatically verifiable answers. However, scaling rubric-based RL remains challenging: existing approaches often rely on expert-written rubrics and manually constructed question sets, while fixed task-level rubrics may fail to capture the evaluation requirements of individual questions. We propose ARES (Automated Rubric synthEsis for Scalable RL), a framework for automatically constructing rubric-based RL data at scale. Starting from raw pretraining documents, ARES converts source knowledge into self-contained question-answer pairs and co-generates question-specific weighted rubrics, enabling instance-level reward supervision for open-ended responses. To improve diversity and quality, ARES conditions generation on domain labels and persona information, and applies validation filters for question self-containment, answer faithfulness, and rubric validity. Using ARES, we construct 100K rubric-annotated instances across ten domains. Experiments on seven benchmarks show that rubric-based RL trained with ARES, outperforms continual pretraining, supervised fine-tuning, and binary-reward RL, with the largest gains on multi-dimensional open-ended tasks such as healthcare and instruction following.

Yang Zhang, Juntao You, Yimeng Bai et al.

Recent advancements in recommender systems have focused on leveraging Large Language Models (LLMs) to improve user preference modeling, yielding promising outcomes. However, current LLM-based approaches struggle to fully leverage user behavior sequences, resulting in suboptimal preference modeling for personalized recommendations. In this study, we propose a novel Counterfactual Fine-Tuning (CFT) method to address this issue by explicitly emphasizing the role of behavior sequences when generating recommendations. Specifically, we employ counterfactual reasoning to identify the causal effects of behavior sequences on model output and introduce a task that directly fits the ground-truth labels based on these effects, achieving the goal of explicit emphasis. Additionally, we develop a token-level weighting mechanism to adjust the emphasis strength for different item tokens, reflecting the diminishing influence of behavior sequences from earlier to later tokens during predicting an item. Extensive experiments on real-world datasets demonstrate that CFT effectively improves behavior sequence modeling. Our codes are available at https://github.com/itsmeyjt/CFT.

An Yang, Anfeng Li, Baosong Yang et al. · tsinghua

In this work, we present Qwen3, the latest version of the Qwen model family. Qwen3 comprises a series of large language models (LLMs) designed to advance performance, efficiency, and multilingual capabilities. The Qwen3 series includes models of both dense and Mixture-of-Expert (MoE) architectures, with parameter scales ranging from 0.6 to 235 billion. A key innovation in Qwen3 is the integration of thinking mode (for complex, multi-step reasoning) and non-thinking mode (for rapid, context-driven responses) into a unified framework. This eliminates the need to switch between different models--such as chat-optimized models (e.g., GPT-4o) and dedicated reasoning models (e.g., QwQ-32B)--and enables dynamic mode switching based on user queries or chat templates. Meanwhile, Qwen3 introduces a thinking budget mechanism, allowing users to allocate computational resources adaptively during inference, thereby balancing latency and performance based on task complexity. Moreover, by leveraging the knowledge from the flagship models, we significantly reduce the computational resources required to build smaller-scale models, while ensuring their highly competitive performance. Empirical evaluations demonstrate that Qwen3 achieves state-of-the-art results across diverse benchmarks, including tasks in code generation, mathematical reasoning, agent tasks, etc., competitive against larger MoE models and proprietary models. Compared to its predecessor Qwen2.5, Qwen3 expands multilingual support from 29 to 119 languages and dialects, enhancing global accessibility through improved cross-lingual understanding and generation capabilities. To facilitate reproducibility and community-driven research and development, all Qwen3 models are publicly accessible under Apache 2.0.

Qwen, An Yang, Baosong Yang et al.

In this report, we introduce Qwen2.5, a comprehensive series of large language models (LLMs) designed to meet diverse needs. Compared to previous iterations, Qwen 2.5 has been significantly improved during both the pre-training and post-training stages. In terms of pre-training, we have scaled the high-quality pre-training datasets from the previous 7 trillion tokens to 18 trillion tokens. This provides a strong foundation for common sense, expert knowledge, and reasoning capabilities. In terms of post-training, we implement intricate supervised finetuning with over 1 million samples, as well as multistage reinforcement learning. Post-training techniques enhance human preference, and notably improve long text generation, structural data analysis, and instruction following. To handle diverse and varied use cases effectively, we present Qwen2.5 LLM series in rich sizes. Open-weight offerings include base and instruction-tuned models, with quantized versions available. In addition, for hosted solutions, the proprietary models currently include two mixture-of-experts (MoE) variants: Qwen2.5-Turbo and Qwen2.5-Plus, both available from Alibaba Cloud Model Studio. Qwen2.5 has demonstrated top-tier performance on a wide range of benchmarks evaluating language understanding, reasoning, mathematics, coding, human preference alignment, etc. Specifically, the open-weight flagship Qwen2.5-72B-Instruct outperforms a number of open and proprietary models and demonstrates competitive performance to the state-of-the-art open-weight model, Llama-3-405B-Instruct, which is around 5 times larger. Qwen2.5-Turbo and Qwen2.5-Plus offer superior cost-effectiveness while performing competitively against GPT-4o-mini and GPT-4o respectively. Additionally, as the foundation, Qwen2.5 models have been instrumental in training specialized models such as Qwen2.5-Math, Qwen2.5-Coder, QwQ, and multimodal models.

Shihao Cai, Jizhi Zhang, Keqin Bao et al.

Large language model-based agents are increasingly applied in the recommendation field due to their extensive knowledge and strong planning capabilities. While prior research has primarily focused on enhancing either the recommendation agent or the user agent individually, the collaborative interaction between the two has often been overlooked. Towards this research gap, we propose a novel framework that emphasizes the feedback loop process to facilitate the collaboration between the recommendation agent and the user agent. Specifically, the recommendation agent refines its understanding of user preferences by analyzing the feedback from the user agent on the item recommendation. Conversely, the user agent further identifies potential user interests based on the items and recommendation reasons provided by the recommendation agent. This iterative process enhances the ability of both agents to infer user behaviors, enabling more effective item recommendations and more accurate user simulations. Extensive experiments on three datasets demonstrate the effectiveness of the agentic feedback loop: the agentic feedback loop yields an average improvement of 11.52% over the single recommendation agent and 21.12% over the single user agent. Furthermore, the results show that the agentic feedback loop does not exacerbate popularity or position bias, which are typically amplified by the real-world feedback loop, highlighting its robustness. The source code is available at https://github.com/Lanyu0303/AFL.

Chengpeng Li, Zhengyang Tang, Ziniu Li et al.

Large Reasoning Models (LRMs) like o1 and DeepSeek-R1 have shown remarkable progress in natural language reasoning with long chain-of-thought (CoT), yet they remain inefficient or inaccurate when handling complex mathematical operations. Addressing these limitations through computational tools (e.g., computation libraries and symbolic solvers) is promising, but it introduces a technical challenge: Code Interpreter (CI) brings external knowledge beyond the model's internal text representations, thus the direct combination is not efficient. This paper introduces CoRT, a post-training framework for teaching LRMs to leverage CI effectively and efficiently. As a first step, we address the data scarcity issue by synthesizing code-integrated reasoning data through Hint-Engineering, which strategically inserts different hints at appropriate positions to optimize LRM-CI interaction. We manually create 30 high-quality samples, upon which we post-train models ranging from 1.5B to 32B parameters, with supervised fine-tuning, rejection fine-tuning and reinforcement learning. Our experimental results demonstrate that Hint-Engineering models achieve 4\% and 8\% absolute improvements on DeepSeek-R1-Distill-Qwen-32B and DeepSeek-R1-Distill-Qwen-1.5B respectively, across five challenging mathematical reasoning datasets. Furthermore, Hint-Engineering models use about 30\% fewer tokens for the 32B model and 50\% fewer tokens for the 1.5B model compared with the natural language models. The models and code are available at https://github.com/ChengpengLi1003/CoRT.

Xiaoyuan Li, Moxin Li, Keqin Bao et al.

Skill libraries enable large language model agents to reuse experience from past interactions, but most existing libraries store skills as isolated entries and retrieve them only by semantic similarity. This leads to two key challenges for compositional tasks. Firstly, an agent must identify not only relevant skills but also how they depend on and build upon each other. Secondly, it also makes library maintenance difficult, since the system lacks structural cues for deciding when skills should be merged, split, or removed. We propose SKILLGRAPH, a framework that represents reusable skills as nodes in a directed graph, with typed edges encoding prerequisite, enhancement, and co-occurrence relations. Given a new task, SKILLGRAPH retrieves not just individual skills, but an ordered skill subgraph that can guide multi-step decision making. The graph is continuously updated from agent trajectories and reinforcement learning feedback, allowing both the skill library and the agent policy to improve together. Experiments on ALFWorld, WebShop, and seven search-augmented QA tasks show that SKILLGRAPH achieves state-of-the-art performance against memory-augmented RL methods, with especially large gains on complex tasks that require composing multiple skills.

Xiaoyuan Li, Yuzhe Wang, Moxin Li et al.

Large Language Models (LLMs) achieve strong performance on standard knowledge evaluation benchmarks, yet recent work shows that their knowledge capabilities remain brittle under question variants that test the same knowledge in different forms. Robustness augmentation of existing knowledge evaluation benchmarks is therefore necessary, but current LLM-assisted generate-then-verify pipelines are costly and difficult to scale due to low-yield variant generation and unreliable variant verification. We propose SAGE (Scalable Automated Generation of Robustness BEnchmarks), a framework for scalable robustness augmentation of knowledge evaluation benchmarks using fine-tuned smaller models. SAGE consists of VariantQual, a rubric-based verifier trained on human-labeled seed data, and VariantGen, a variant generator initialized with supervised fine-tuning and further optimized with reinforcement learning using VariantQual as the reward model. Experiments on HellaSwag show that SAGE constructs a large-scale robustness-augmented benchmark with quality comparable to the human-annotated HellaSwag-Pro at substantially lower cost, while the fine-tuned models further generalize to MMLU without benchmark-specific fine-tuning.

Xiaoyuan Li, Yubo Ma, Kexin Yang et al.

The performance of Large Language Models (LLMs) on downstream tasks is fundamentally constrained by the capabilities acquired during pre-training. However, traditional benchmarks like MMLU often fail to reflect a base model's plasticity in complex open-ended scenarios, leading to inefficient model selection. We address this by introducing a new task of predicting post-training potential - forecasting a base model's performance before post-training. We propose RuDE (Rubric-based Discriminative Evaluation), a unified framework that bypasses the generation gap of base models by leveraging response discrimination. Guided by our systematic 4C Taxonomy, RuDE constructs controlled contrastive pairs across diverse domains by fine-grained rubric violations. Extensive experiments demonstrate a correlation greater than 90% with post-training performance. Crucially, validation via Reinforcement Learning (RL) confirms that RuDE effectively identifies high-potential smaller models that outperform larger counterparts, offering a compute-efficient mechanism for foundation model development.

Chengpeng Li, Zhengyang Tang, Ziniu Li et al.

Large reasoning models (LRMs) like OpenAI-o1 have shown impressive capabilities in natural language reasoning. However, these models frequently demonstrate inefficiencies or inaccuracies when tackling complex mathematical operations. While integrating computational tools such as Code Interpreters (CIs) offers a promising solution, it introduces a critical challenge: a conflict between the model's internal, probabilistic reasoning and the external, deterministic knowledge provided by the CI, which often leads models to unproductive deliberation. To overcome this, we introduce CoRT (Code-Optimized Reasoning Training), a post-training framework designed to teach LRMs to effectively utilize CIs. We propose \emph{Hint-Engineering}, a new data synthesis strategy that strategically injects diverse hints at optimal points within reasoning paths. This approach generates high-quality, code-integrated reasoning data specifically tailored to optimize LRM-CI interaction. Using this method, we have synthesized 30 high-quality samples to post-train models ranging from 1.5B to 32B parameters through supervised fine-tuning. CoRT further refines the multi-round interleaving of external CI usage and internal thinking by employing rejection sampling and reinforcement learning. Our experimental evaluations demonstrate CoRT's effectiveness, yielding absolute improvements of 4\% and 8\% on DeepSeek-R1-Distill-Qwen-32B and DeepSeek-R1-Distill-Qwen-1.5B, respectively, across five challenging mathematical reasoning datasets. Moreover, CoRT significantly enhances efficiency, reducing token usage by approximately 30\% for the 32B model and 50\% for the 1.5B model compared to pure natural language reasoning baselines. The models and code are available at: https://github.com/ChengpengLi1003/CoRT.

Shihao Cai, Keqin Bao, Hangyu Guo et al.

Large language models have seen widespread adoption in math problem-solving. However, in geometry problems that usually require visual aids for better understanding, even the most advanced multi-modal models currently still face challenges in effectively using image information. High-quality data is crucial for enhancing the geometric capabilities of multi-modal models, yet existing open-source datasets and related efforts are either too challenging for direct model learning or suffer from misalignment between text and images. To overcome this issue, we introduce a novel pipeline that leverages GPT-4 and GPT-4V to generate relatively basic geometry problems with aligned text and images, facilitating model learning. We have produced a dataset of 4.9K geometry problems and combined it with 19K open-source data to form our GeoGPT4V dataset. Experimental results demonstrate that the GeoGPT4V dataset significantly improves the geometry performance of various models on the MathVista and MathVision benchmarks. The code is available at https://github.com/Lanyu0303/GeoGPT4V_Project

Jizhi Zhang, Keqin Bao, Yang Zhang et al.

The remarkable achievements of Large Language Models (LLMs) have led to the emergence of a novel recommendation paradigm -- Recommendation via LLM (RecLLM). Nevertheless, it is important to note that LLMs may contain social prejudices, and therefore, the fairness of recommendations made by RecLLM requires further investigation. To avoid the potential risks of RecLLM, it is imperative to evaluate the fairness of RecLLM with respect to various sensitive attributes on the user side. Due to the differences between the RecLLM paradigm and the traditional recommendation paradigm, it is problematic to directly use the fairness benchmark of traditional recommendation. To address the dilemma, we propose a novel benchmark called Fairness of Recommendation via LLM (FaiRLLM). This benchmark comprises carefully crafted metrics and a dataset that accounts for eight sensitive attributes1 in two recommendation scenarios: music and movies. By utilizing our FaiRLLM benchmark, we conducted an evaluation of ChatGPT and discovered that it still exhibits unfairness to some sensitive attributes when generating recommendations. Our code and dataset can be found at https://github.com/jizhi-zhang/FaiRLLM.

Jizhi Zhang, Keqin Bao, Wenjie Wang et al.

The new kind of Agent-oriented information system, exemplified by GPTs, urges us to inspect the information system infrastructure to support Agent-level information processing and to adapt to the characteristics of Large Language Model (LLM)-based Agents, such as interactivity. In this work, we envisage the prospect of the recommender system on LLM-based Agent platforms and introduce a novel recommendation paradigm called Rec4Agentverse, comprised of Agent Items and Agent Recommender. Rec4Agentverse emphasizes the collaboration between Agent Items and Agent Recommender, thereby promoting personalized information services and enhancing the exchange of information beyond the traditional user-recommender feedback loop. Additionally, we prospect the evolution of Rec4Agentverse and conceptualize it into three stages based on the enhancement of the interaction and information exchange among Agent Items, Agent Recommender, and the user. A preliminary study involving several cases of Rec4Agentverse validates its significant potential for application. Lastly, we discuss potential issues and promising directions for future research.

Xiaoyuan Li, Keqin Bao, Yubo Ma et al.

Recent advances in Large Language Models (LLMs) have shown promising results in complex reasoning tasks. However, current evaluations predominantly focus on single-turn reasoning scenarios, leaving interactive tasks largely unexplored. We attribute it to the absence of comprehensive datasets and scalable automatic evaluation protocols. To fill these gaps, we present MTR-Bench for LLMs' Multi-Turn Reasoning evaluation. Comprising 4 classes, 40 tasks, and 3600 instances, MTR-Bench covers diverse reasoning capabilities, fine-grained difficulty granularity, and necessitates multi-turn interactions with the environments. Moreover, MTR-Bench features fully-automated framework spanning both dataset constructions and model evaluations, which enables scalable assessment without human interventions. Extensive experiments reveal that even the cutting-edge reasoning models fall short of multi-turn, interactive reasoning tasks. And the further analysis upon these results brings valuable insights for future research in interactive AI systems.

Xiaoyuan Li, Moxin Li, Rui Men et al.

Large language models (LLMs) have shown remarkable capabilities in commonsense reasoning; however, some variations in questions can trigger incorrect responses. Do these models truly understand commonsense knowledge, or just memorize expression patterns? To investigate this question, we present the first extensive robustness evaluation of LLMs in commonsense reasoning. We introduce HellaSwag-Pro, a large-scale bilingual benchmark consisting of 11,200 cases, by designing and compiling seven types of question variants. To construct this benchmark, we propose a two-stage method to develop Chinese HellaSwag, a finely annotated dataset comprising 12,000 instances across 56 categories. We conduct extensive experiments on 41 representative LLMs, revealing that these LLMs are far from robust in commonsense reasoning. Furthermore, this robustness varies depending on the language in which the LLM is tested. This work establishes a high-quality evaluation benchmark, with extensive experiments offering valuable insights to the community in commonsense reasoning for LLMs.

Nuo Chen, Zehua Li, Keqin Bao et al.

Building robust and general reasoning ability is a central goal in the development of large language models (LLMs). Recent efforts increasingly turn to code as a rich training source, given its inherent logical structure and diverse reasoning paradigms such as divide-and-conquer, topological ordering, and enumeration. However, reasoning in code is often expressed implicitly and entangled with syntactic or implementation noise, making direct training on raw code suboptimal.To address this, we introduce TracePile, a large-scale corpus of 2.6 million samples that transforms code execution into explicit, step-by-step chain-of-thought-style rationales, which we call Chain of Execution (CoE). The corpus spans domains including mathematics, classical algorithms and algorithmic competition, and is enriched with variable-tracing questions and code rewritings to enhance logical granularity and code diversity. We evaluate TracePile using three training setups: continue-pretraining, instruction tuning after pretraining, and two-stage finetuning. Experiments across four base models (LLaMA 3, LLaMA 3.1, Qwen-2.5, and Qwen-2.5 Coder) and 20 benchmarks covering math, code, logic, and algorithms demonstrate consistent improvements. Notably, TracePile boosts LLaMA3.1-8B by 7.1\% on average across nine math datasets and delivers clear gains on LiveCodeBench, CRUX, and MMLU under two-stage fine-tuning.

Keqin Bao, Nuo Chen, Xiaoyuan Li et al.

Enhancing reasoning capabilities remains a central focus in the LLM reasearch community. A promising direction involves requiring models to simulate code execution step-by-step to derive outputs for given inputs. However, as code is often designed for large-scale systems, direct application leads to over-reliance on complex data structures and algorithms, even for simple cases, resulting in overfitting to algorithmic patterns rather than core reasoning structures. To address this, we propose TeaR, which aims at teaching LLMs to reason better. TeaR leverages careful data curation and reinforcement learning to guide models in discovering optimal reasoning paths through code-related tasks, thereby improving general reasoning abilities. We conduct extensive experiments using two base models and three long-CoT distillation models, with model sizes ranging from 1.5 billion to 32 billion parameters, and across 17 benchmarks spanning Math, Knowledge, Code, and Logical Reasoning. The results consistently show significant performance improvements. Notably, TeaR achieves a 35.9% improvement on Qwen2.5-7B and 5.9% on R1-Distilled-7B.