Xuefeng Li

Dongcheng Zou, Senzhang Wang, Xuefeng Li et al.

Traffic forecasting is a complex multivariate time-series regression task of paramount importance for traffic management and planning. However, existing approaches often struggle to model complex multi-range dependencies using local spatiotemporal features and road network hierarchical knowledge. To address this, we propose MultiSPANS. First, considering that an individual recording point cannot reflect critical spatiotemporal local patterns, we design multi-filter convolution modules for generating informative ST-token embeddings to facilitate attention computation. Then, based on ST-token and spatial-temporal position encoding, we employ the Transformers to capture long-range temporal and spatial dependencies. Furthermore, we introduce structural entropy theory to optimize the spatial attention mechanism. Specifically, The structural entropy minimization algorithm is used to generate optimal road network hierarchies, i.e., encoding trees. Based on this, we propose a relative structural entropy-based position encoding and a multi-head attention masking scheme based on multi-layer encoding trees. Extensive experiments demonstrate the superiority of the presented framework over several state-of-the-art methods in real-world traffic datasets, and the longer historical windows are effectively utilized. The code is available at https://github.com/SELGroup/MultiSPANS.

Guanting Dong, Zechen Wang, Jinxu Zhao et al.

The objective of few-shot named entity recognition is to identify named entities with limited labeled instances. Previous works have primarily focused on optimizing the traditional token-wise classification framework, while neglecting the exploration of information based on NER data characteristics. To address this issue, we propose a Multi-Task Semantic Decomposition Framework via Joint Task-specific Pre-training (MSDP) for few-shot NER. Drawing inspiration from demonstration-based and contrastive learning, we introduce two novel pre-training tasks: Demonstration-based Masked Language Modeling (MLM) and Class Contrastive Discrimination. These tasks effectively incorporate entity boundary information and enhance entity representation in Pre-trained Language Models (PLMs). In the downstream main task, we introduce a multi-task joint optimization framework with the semantic decomposing method, which facilitates the model to integrate two different semantic information for entity classification. Experimental results of two few-shot NER benchmarks demonstrate that MSDP consistently outperforms strong baselines by a large margin. Extensive analyses validate the effectiveness and generalization of MSDP.

Xuefeng Li, Liwen Wang, Guanting Dong et al.

Zero-shot cross-domain slot filling aims to transfer knowledge from the labeled source domain to the unlabeled target domain. Existing models either encode slot descriptions and examples or design handcrafted question templates using heuristic rules, suffering from poor generalization capability or robustness. In this paper, we propose a generative zero-shot prompt learning framework for cross-domain slot filling, both improving generalization and robustness than previous work. Besides, we introduce a novel inverse prompting strategy to distinguish different slot types to avoid the multiple prediction problem, and an efficient prompt-tuning strategy to boost higher performance by only training fewer prompt parameters. Experiments and analysis demonstrate the effectiveness of our proposed framework, especially huge improvements (+13.44% F1) on the unseen slots.

Guanting Dong, Daichi Guo, Liwen Wang et al.

Most existing slot filling models tend to memorize inherent patterns of entities and corresponding contexts from training data. However, these models can lead to system failure or undesirable outputs when being exposed to spoken language perturbation or variation in practice. We propose a perturbed semantic structure awareness transferring method for training perturbation-robust slot filling models. Specifically, we introduce two MLM-based training strategies to respectively learn contextual semantic structure and word distribution from unsupervised language perturbation corpus. Then, we transfer semantic knowledge learned from upstream training procedure into the original samples and filter generated data by consistency processing. These procedures aim to enhance the robustness of slot filling models. Experimental results show that our method consistently outperforms the previous basic methods and gains strong generalization while preventing the model from memorizing inherent patterns of entities and contexts.

Daichi Guo, Guanting Dong, Dayuan Fu et al.

In real dialogue scenarios, the existing slot filling model, which tends to memorize entity patterns, has a significantly reduced generalization facing Out-of-Vocabulary (OOV) problems. To address this issue, we propose an OOV robust slot filling model based on multi-level data augmentations to solve the OOV problem from both word and slot perspectives. We present a unified contrastive learning framework, which pull representations of the origin sample and augmentation samples together, to make the model resistant to OOV problems. We evaluate the performance of the model from some specific slots and carefully design test data with OOV word perturbation to further demonstrate the effectiveness of OOV words. Experiments on two datasets show that our approach outperforms the previous sota methods in terms of both OOV slots and words.

Guanting Dong, Zechen Wang, Liwen Wang et al.

Few-shot named entity recognition (NER) aims at identifying named entities based on only few labeled instances. Most existing prototype-based sequence labeling models tend to memorize entity mentions which would be easily confused by close prototypes. In this paper, we proposed a Prototypical Semantic Decoupling method via joint Contrastive learning (PSDC) for few-shot NER. Specifically, we decouple class-specific prototypes and contextual semantic prototypes by two masking strategies to lead the model to focus on two different semantic information for inference. Besides, we further introduce joint contrastive learning objectives to better integrate two kinds of decoupling information and prevent semantic collapse. Experimental results on two few-shot NER benchmarks demonstrate that PSDC consistently outperforms the previous SOTA methods in terms of overall performance. Extensive analysis further validates the effectiveness and generalization of PSDC.

Xuefeng Li, Hao Lei, Liwen Wang et al.

Multi-domain text classification can automatically classify texts in various scenarios. Due to the diversity of human languages, texts with the same label in different domains may differ greatly, which brings challenges to the multi-domain text classification. Current advanced methods use the private-shared paradigm, capturing domain-shared features by a shared encoder, and training a private encoder for each domain to extract domain-specific features. However, in realistic scenarios, these methods suffer from inefficiency as new domains are constantly emerging. In this paper, we propose a robust contrastive alignment method to align text classification features of various domains in the same feature space by supervised contrastive learning. By this means, we only need two universal feature extractors to achieve multi-domain text classification. Extensive experimental results show that our method performs on par with or sometimes better than the state-of-the-art method, which uses the complex multi-classifier in a private-shared framework.

Tengfei Xue, Xuefeng Li, Tahir Azim et al.

Large language models (LLMs) have significantly improved code generation, particularly in one-pass code generation. However, most existing approaches focus solely on generating code in a single programming language, overlooking the potential of leveraging the multi-language capabilities of LLMs. LLMs have varying patterns of errors across different languages, suggesting that a more robust approach could be developed by leveraging these multi-language outputs. In this study, we propose Multi-Programming Language Ensemble (MPLE), a novel ensemble-based method that utilizes code generation across multiple programming languages to enhance overall performance. By treating each language-specific code generation process as an individual "weak expert" and effectively integrating their outputs, our method mitigates language-specific errors and biases. This multi-language ensemble strategy leverages the complementary strengths of different programming languages, enabling the model to produce more accurate and robust code. Our approach can be seamlessly integrated with commonly used techniques such as the reflection algorithm and Monte Carlo tree search to improve code generation quality further. Experimental results show that our framework consistently enhances baseline performance by up to 17.92% on existing benchmarks (HumanEval and HumanEval-plus), with a standout result of 96.25% accuracy on the HumanEval benchmark, achieving new state-of-the-art results across various LLM models. The code will be released at https://github.com/NinjaTech-AI/MPLE

Ting Wu, Xuefeng Li, Pengfei Liu

Self-improvement through post-training methods such as iterative preference learning has been acclaimed for enhancing the problem-solving capabilities (e.g., mathematical reasoning) of Large Language Models (LLMs) without human intervention. However, as exploration deepens, it becomes crucial to assess whether these improvements genuinely signify progress in solving more challenging problems or if they could lead to unintended regressions. To address this, we propose a comprehensive evaluative framework that goes beyond the superficial pass@1 metric to scrutinize the underlying enhancements of post-training paradigms for self-improvement. Through rigorous experimentation and analysis across diverse problem-solving tasks, the empirical results point out the phenomenon of \emph{self-improvement reversal}, where models showing improved performance across benchmarks will paradoxically exhibit declines in broader, essential capabilities, like output diversity and out-of-distribution (OOD) generalization. These findings indicate that current self-improvement practices through post-training are inadequate for equipping models to tackle more complex problems. Furthermore, they underscore the necessity of our critical evaluation metrics in discerning the \emph{progress or regress} dichotomy for self-improving LLMs.

Xuefeng Li, Haoyang Zou, Pengfei Liu

In this paper, we ask: what truly determines the effectiveness of RL training data for enhancing language models' reasoning capabilities? While recent advances like o1, Deepseek R1, and Kimi1.5 demonstrate RL's potential, the lack of transparency about training data requirements has hindered systematic progress. Starting directly from base models without distillation, we challenge the assumption that scaling up RL training data inherently improves performance. we demonstrate that a strategically selected subset of just 1,389 samples can outperform the full 8,523-sample dataset. We introduce Learning Impact Measurement (LIM), an automated method to evaluate and prioritize training samples based on their alignment with model learning trajectories, enabling efficient resource utilization and scalable implementation. Our method achieves comparable or even superior performance using only 1,389 samples versus the full 8,523 samples dataset. Notably, while recent data-efficient approaches (e.g., LIMO and s1) show promise with 32B-scale models, we find it significantly underperforms at 7B-scale through supervised fine-tuning (SFT). In contrast, our RL-based LIMR achieves 16.7% higher accuracy on AIME24 and outperforms LIMO and s1 by 13.0% and 22.2% on MATH500. These results fundamentally reshape our understanding of RL scaling in LLMs, demonstrating that precise sample selection, rather than data scale, may be the key to unlocking enhanced reasoning capabilities. For reproducible research and future innovation, we are open-sourcing LIMR, including implementation of LIM, training and evaluation code, curated datasets, and trained models at https://github.com/GAIR-NLP/LIMR.

Zengzhi Wang, Xuefeng Li, Rui Xia et al.

High-quality, large-scale corpora are the cornerstone of building foundation models. In this work, we introduce MathPile, a diverse and high-quality math-centric corpus comprising about 9.5 billion tokens. Throughout its creation, we adhered to the principle of "less is more", firmly believing in the supremacy of data quality over quantity, even in the pre-training phase. Our meticulous data collection and processing efforts included a complex suite of preprocessing, prefiltering, language identification, cleaning, filtering, and deduplication, ensuring the high quality of our corpus. Furthermore, we performed data contamination detection on downstream benchmark test sets to eliminate duplicates and conducted continual pre-training experiments, booting the performance on common mathematical reasoning benchmarks. We aim for our MathPile to boost language models' mathematical reasoning abilities and open-source its different versions and processing scripts to advance the field.

Zengzhi Wang, Fan Zhou, Xuefeng Li et al.

Different base language model families, such as Llama and Qwen, exhibit divergent behaviors during post-training with reinforcement learning (RL), especially on reasoning-intensive tasks. What makes a base language model suitable for reinforcement learning? Gaining deeper insight into this question is essential for developing RL-scalable foundation models of the next generation. In this work, we investigate how mid-training strategies shape RL dynamics, focusing on two representative model families: Qwen and Llama. Our study reveals that (1) high-quality mathematical corpora, such as MegaMath-Web-Pro, significantly improve both base model and RL performance, while existing alternatives (e.g., FineMath-4plus) fail to do so; (2) further adding QA-style data, particularly long chain-of-thought (CoT) reasoning examples, enhances RL outcomes, and instruction data further unlocks this effect; (3) while long-CoT improves reasoning depth, it can also induce verbosity of model responses and unstability of RL training, underscoring the importance of data formatting; (4) scaling mid-training consistently leads to stronger downstream RL performance. Building on these insights, we introduce a two-stage mid-training strategy, Stable-then-Decay, in which base models are first trained on 200B tokens with a constant learning rate, followed by 20B tokens across three CoT-focused branches with learning rate decay. This yields OctoThinker, a family of models demonstrating strong RL compatibility and closing the performance gap with more RL-friendly model families, i.e., Qwen. We hope our work will help shape pre-training strategies for foundation models in the RL era. To support further research, we release our open-source models along with a curated math reasoning-intensive corpus of over 70 billion tokens (i.e., MegaMath-Web-Pro-Max).

Shijie Xia, Xuefeng Li, Yixin Liu et al.

The leaderboard of Large Language Models (LLMs) in mathematical tasks has been continuously updated. However, the majority of evaluations focus solely on the final results, neglecting the quality of the intermediate steps. This oversight can mask underlying problems, such as logical errors or unnecessary steps in the reasoning process. To measure reasoning beyond final-answer accuracy, we introduce ReasonEval, a new methodology for evaluating the quality of reasoning steps. ReasonEval employs validity and redundancy to characterize the reasoning quality, as well as accompanying LLMs to assess them automatically. We explore different design options for the LLM-based evaluators and empirically demonstrate that ReasonEval, when instantiated with base models possessing strong mathematical knowledge and trained with high-quality labeled data, consistently outperforms baseline methods in the meta-evaluation datasets. We also highlight the strong generalization capabilities of ReasonEval. By utilizing ReasonEval to evaluate LLMs specialized in math, we find that an increase in final-answer accuracy does not necessarily guarantee an improvement in the overall quality of the reasoning steps for challenging mathematical problems. Additionally, we observe that ReasonEval can play a significant role in data selection. We open-source the best-performing model, meta-evaluation script, and all evaluation results to facilitate future research.

Yiyuan Li, Zhen Huang, Yanan Wu et al.

The reasoning ability of large language models (LLMs) can be unleashed with reinforcement learning (RL) (OpenAI, 2024; DeepSeek-AI et al., 2025a; Zeng et al., 2025). The success of existing RL attempts in LLMs usually rely on high-quality samples of large volumes. In this paper, we challenge conventional assumptions about data requirements in RL for LLMs by demonstrating the effectiveness of one-shot reinforcement learning. Specifically, we introduce polymath learning, a framework for designing one training sample that elicits multidisciplinary reasoning improvement. We present three key findings: (1) A single, strategically selected math reasoning sample can produce significant performance improvements across multiple domains, including physics, chemistry, and biology; (2) Analysis of salient mathematical skills provides insight into the characteristics associated with effective polymath samples; and (3) An engineered synthetic sample that integrates multidisciplinary elements and broader skill coverage achieves stronger performance than naturally occurring individual samples. Across various reasoning benchmarks, polymath learning achieves stronger performance than larger datasets, demonstrating that reasoning structure and skills in samples, rather than quantity, may be the key to unlock enhanced reasoning capabilities in language models. Our results suggest a shift, dubbed as sample engineering, toward precision engineering of samples that complements simply increasing data volume.

Ji Zeng, Dayuan Fu, Tiantian Mi et al.

Recently, the frontier of Large Language Model (LLM) capabilities has shifted from single-turn code generation to agentic software engineering-a paradigm where models autonomously navigate, edit, and test complex repositories. While post-training methods have become the de facto approach for code agents, **agentic mid-training**-mid-training (MT) on large-scale data that mirrors authentic agentic workflows-remains critically underexplored due to substantial resource requirements, despite offering a more scalable path to instilling foundational agentic behaviors than relying solely on expensive reinforcement learning. A central challenge in realizing effective agentic mid-training is the distribution mismatch between static training data and the dynamic, feedback-rich environment of real development. To address this, we present a systematic study of agentic mid-training, establishing both the data synthesis principles and training methodology for effective agent development at scale. Central to our approach is **agent-native data**-supervision comprising two complementary types of trajectories: **contextually-native trajectories** that preserve the complete information flow an agent experiences, offering broad coverage and diversity; and **environmentally-native trajectories** collected from executable repositories where observations stem from actual tool invocations and test executions, providing depth and interaction authenticity. We verify the model's agentic capabilities on `SWE-Bench Verified`. We demonstrate our superiority over the previous open software engineering mid-training recipe `Kimi-Dev` under two post-training settings with an aligned base model and agentic scaffold, while using less than half mid-training tokens (73.1B). Besides relative advantage, our best performing 32B and 72B models achieve **56.1%** and **58.5%** resolution rates, respectively, which are ...

Run-Ze Fan, Xuefeng Li, Haoyang Zou et al.

The quality of finetuning data is crucial for aligning large language models (LLMs) with human values. Current methods to improve data quality are either labor-intensive or prone to factual errors caused by LLM hallucinations. This paper explores elevating the quality of existing instruction data to better align with human values, introducing a simple and effective approach named ReAlign, which reformats the responses of instruction data into a format that better aligns with pre-established criteria and the collated evidence. This approach minimizes human annotation, hallucination, and the difficulty in scaling, remaining orthogonal to existing alignment techniques. Experimentally, ReAlign significantly boosts the general alignment ability, math reasoning, factuality, and readability of the LLMs. Encouragingly, without introducing any additional data or advanced training techniques, and merely by reformatting the response, LLaMA-2-13B's mathematical reasoning ability on GSM8K can be improved from 46.77% to 56.63% in accuracy. Additionally, a mere 5% of ReAlign data yields a 67% boost in general alignment ability measured by the Alpaca dataset. This work highlights the need for further research into the science and mechanistic interpretability of LLMs. We have made the associated code and data publicly accessible to support future studies at https://github.com/GAIR-NLP/ReAlign.

Shijie Xia, Yiwei Qin, Xuefeng Li et al.

The first generation of Large Language Models - what might be called "Act I" of generative AI (2020-2023) - achieved remarkable success through massive parameter and data scaling, yet exhibited fundamental limitations such as knowledge latency, shallow reasoning, and constrained cognitive processes. During this era, prompt engineering emerged as our primary interface with AI, enabling dialogue-level communication through natural language. We now witness the emergence of "Act II" (2024-present), where models are transitioning from knowledge-retrieval systems (in latent space) to thought-construction engines through test-time scaling techniques. This new paradigm establishes a mind-level connection with AI through language-based thoughts. In this paper, we clarify the conceptual foundations of cognition engineering and explain why this moment is critical for its development. We systematically break down these advanced approaches through comprehensive tutorials and optimized implementations, democratizing access to cognition engineering and enabling every practitioner to participate in AI's second act. We provide a regularly updated collection of papers on test-time scaling in the GitHub Repository: https://github.com/GAIR-NLP/cognition-engineering

Qianyu He, Siyu Yuan, Xuefeng Li et al.

Large language models (LLMs) with chain-of-thought reasoning have demonstrated remarkable problem-solving capabilities, but controlling their computational effort remains a significant challenge for practical deployment. Recent proprietary systems like OpenAI's gpt-oss series have introduced discrete operational modes for intuitive reasoning control, but the open-source community has largely failed to achieve such capabilities. In this paper, we introduce ThinkDial, the first open-recipe end-to-end framework that successfully implements gpt-oss-style controllable reasoning through discrete operational modes. Our system enables seamless switching between three distinct reasoning regimes: High mode (full reasoning capability), Medium mode (50 percent token reduction with <10 percent performance degradation), and Low mode (75 percent token reduction with <15 percent performance degradation). We achieve this through an end-to-end training paradigm that integrates budget-mode control throughout the entire pipeline: budget-mode supervised fine-tuning that embeds controllable reasoning capabilities directly into the learning process, and two-phase budget-aware reinforcement learning with adaptive reward shaping. Extensive experiments demonstrate that ThinkDial achieves target compression-performance trade-offs with clear response length reductions while maintaining performance thresholds. The framework also exhibits strong generalization capabilities on out-of-distribution tasks.

Zixian Wu, Xuebao Li, Yanfang Zheng et al.

In this work, we develop, for the first time, a supervised classification framework with class-dependent rewards (CDR) to predict $\geq$MM flares within 24 hr. We construct multiple datasets, covering knowledge-informed features and line-of sight (LOS) magnetograms. We also apply three deep learning models (CNN, CNN-BiLSTM, and Transformer) and three CDR counterparts (CDR-CNN, CDR-CNN-BiLSTM, and CDR-Transformer). First, we analyze the importance of LOS magnetic field parameters with the Transformer, then compare its performance using LOS-only, vector-only, and combined magnetic field parameters. Second, we compare flare prediction performance based on CDR models versus deep learning counterparts. Third, we perform sensitivity analysis on reward engineering for CDR models. Fourth, we use the SHAP method for model interpretability. Finally, we conduct performance comparison between our models and NASA/CCMC. The main findings are: (1)Among LOS feature combinations, R_VALUE and AREA_ACR consistently yield the best results. (2)Transformer achieves better performance with combined LOS and vector magnetic field data than with either alone. (3)Models using knowledge-informed features outperform those using magnetograms. (4)While CNN and CNN-BiLSTM outperform their CDR counterparts on magnetograms, CDR-Transformer is slightly superior to its deep learning counterpart when using knowledge-informed features. Among all models, CDR-Transformer achieves the best performance. (5)The predictive performance of the CDR models is not overly sensitive to the reward choices.(6)Through SHAP analysis, the CDR model tends to regard TOTUSJH as more important, while the Transformer tends to prioritize R_VALUE more.(7)Under identical prediction time and active region (AR) number, the CDR-Transformer shows superior predictive capabilities compared to NASA/CCMC.

Tengfei Xue, Xuefeng Li, Roman Smirnov et al.

Retrieval-augmented generation (RAG) techniques are widely used today to retrieve and present information in a conversational format. This paper presents a set of enhancements to traditional RAG techniques, focusing on large language models (LLMs) fine-tuned and hosted on AWS Trainium and Inferentia2 AI chips via SageMaker. These chips are characterized by their elasticity, affordability, and efficient performance for AI compute tasks. Besides enabling deployment on these chips, this work aims to improve tool usage, add citation capabilities, and mitigate the risks of hallucinations and unsafe responses due to context bias. We benchmark our RAG system's performance on the Natural Questions and HotPotQA datasets, achieving an accuracy of 62% and 59% respectively, exceeding other models such as DBRX and Mixtral Instruct.

Zhen Huang, Zengzhi Wang, Shijie Xia et al.

The evolution of Artificial Intelligence (AI) has been significantly accelerated by advancements in Large Language Models (LLMs) and Large Multimodal Models (LMMs), gradually showcasing potential cognitive reasoning abilities in problem-solving and scientific discovery (i.e., AI4Science) once exclusive to human intellect. To comprehensively evaluate current models' performance in cognitive reasoning abilities, we introduce OlympicArena, which includes 11,163 bilingual problems across both text-only and interleaved text-image modalities. These challenges encompass a wide range of disciplines spanning seven fields and 62 international Olympic competitions, rigorously examined for data leakage. We argue that the challenges in Olympic competition problems are ideal for evaluating AI's cognitive reasoning due to their complexity and interdisciplinary nature, which are essential for tackling complex scientific challenges and facilitating discoveries. Beyond evaluating performance across various disciplines using answer-only criteria, we conduct detailed experiments and analyses from multiple perspectives. We delve into the models' cognitive reasoning abilities, their performance across different modalities, and their outcomes in process-level evaluations, which are vital for tasks requiring complex reasoning with lengthy solutions. Our extensive evaluations reveal that even advanced models like GPT-4o only achieve a 39.97% overall accuracy, illustrating current AI limitations in complex reasoning and multimodal integration. Through the OlympicArena, we aim to advance AI towards superintelligence, equipping it to address more complex challenges in science and beyond. We also provide a comprehensive set of resources to support AI research, including a benchmark dataset, an open-source annotation platform, a detailed evaluation tool, and a leaderboard with automatic submission features.

Hanwei Liu, Huiling Cai, Qingcheng Lin et al.

Annotation ambiguity caused by the inherent subjectivity of visual judgment has always been a major challenge for Facial Expression Recognition (FER) tasks, particularly for largescale datasets from in-the-wild scenarios. A potential solution is the evaluation of relatively objective emotional distributions to help mitigate the ambiguity of subjective annotations. To this end, this paper proposes a novel Prior-based Objective Inference (POI) network. This network employs prior knowledge to derive a more objective and varied emotional distribution and tackles the issue of subjective annotation ambiguity through dynamic knowledge transfer. POI comprises two key networks: Firstly, the Prior Inference Network (PIN) utilizes the prior knowledge of AUs and emotions to capture intricate motion details. To reduce over-reliance on priors and facilitate objective emotional inference, PIN aggregates inferential knowledge from various key facial subregions, encouraging mutual learning. Secondly, the Target Recognition Network (TRN) integrates subjective emotion annotations and objective inference soft labels provided by the PIN, fostering an understanding of inherent facial expression diversity, thus resolving annotation ambiguity. Moreover, we introduce an uncertainty estimation module to quantify and balance facial expression confidence. This module enables a flexible approach to dealing with the uncertainties of subjective annotations. Extensive experiments show that POI exhibits competitive performance on both synthetic noisy datasets and multiple real-world datasets. All codes and training logs will be publicly available at https://github.com/liuhw01/POI.

Zhen Huang, Haoyang Zou, Xuefeng Li et al.

This paper presents a critical examination of current approaches to replicating OpenAI's O1 model capabilities, with particular focus on the widespread but often undisclosed use of knowledge distillation techniques. While our previous work explored the fundamental technical path to O1 replication, this study reveals how simple distillation from O1's API, combined with supervised fine-tuning, can achieve superior performance on complex mathematical reasoning tasks. Through extensive experiments, we show that a base model fine-tuned on simply tens of thousands of samples O1-distilled long-thought chains outperforms O1-preview on the American Invitational Mathematics Examination (AIME) with minimal technical complexity. Moreover, our investigation extends beyond mathematical reasoning to explore the generalization capabilities of O1-distilled models across diverse tasks: hallucination, safety and open-domain QA. Notably, despite training only on mathematical problem-solving data, our models demonstrated strong generalization to open-ended QA tasks and became significantly less susceptible to sycophancy after fine-tuning. We deliberately make this finding public to promote transparency in AI research and to challenge the current trend of obscured technical claims in the field. Our work includes: (1) A detailed technical exposition of the distillation process and its effectiveness, (2) A comprehensive benchmark framework for evaluating and categorizing O1 replication attempts based on their technical transparency and reproducibility, (3) A critical discussion of the limitations and potential risks of over-relying on distillation approaches, our analysis culminates in a crucial bitter lesson: while the pursuit of more capable AI systems is important, the development of researchers grounded in first-principles thinking is paramount.

Xuefeng Li, Haoyang Zou, Pengfei Liu

We introduce ToRL (Tool-Integrated Reinforcement Learning), a framework for training large language models (LLMs) to autonomously use computational tools via reinforcement learning. Unlike supervised fine-tuning, ToRL allows models to explore and discover optimal strategies for tool use. Experiments with Qwen2.5-Math models show significant improvements: ToRL-7B reaches 43.3\% accuracy on AIME~24, surpassing reinforcement learning without tool integration by 14\% and the best existing Tool-Integrated Reasoning (TIR) model by 17\%. Further analysis reveals emergent behaviors such as strategic tool invocation, self-regulation of ineffective code, and dynamic adaptation between computational and analytical reasoning, all arising purely through reward-driven learning.

Yang Xiao, Mohan Jiang, Jie Sun et al.

We define Agency as the emergent capacity of AI systems to function as autonomous agents actively discovering problems, formulating hypotheses, and executing solutions through self-directed engagement with environments and tools. This fundamental capability marks the dawn of the Age of AI Agency, driven by a critical industry shift: the urgent need for AI systems that don't just think, but work. While current AI excels at reasoning and generating responses, industries demand autonomous agents that can execute tasks, operate tools, and drive real-world outcomes. As agentic intelligence becomes the defining characteristic separating cognitive systems from productive workers, efficiently cultivating machine autonomy becomes paramount. Current approaches assume that more data yields better agency, following traditional scaling laws from language modeling. We fundamentally challenge this paradigm. LIMI (Less Is More for Intelligent Agency) demonstrates that agency follows radically different development principles. Through strategic focus on collaborative software development and scientific research workflows, we show that sophisticated agentic intelligence can emerge from minimal but strategically curated demonstrations of autonomous behavior. Using only 78 carefully designed training samples, LIMI achieves 73.5% on comprehensive agency benchmarks, dramatically outperforming state-of-the-art models: Kimi-K2-Instruct (24.1%), DeepSeek-V3.1 (11.9%), Qwen3-235B-A22B-Instruct (27.5%), and GLM-4.5 (45.1%). Most strikingly, LIMI demonstrates 53.7% improvement over models trained on 10,000 samples-achieving superior agentic intelligence with 128 times fewer samples. Our findings establish the Agency Efficiency Principle: machine autonomy emerges not from data abundance but from strategic curation of high-quality agentic demonstrations.

Peng Zhang, Xuefeng Li, Xiaoqi Wang et al.

Network visualization has traditionally relied on heuristic metrics, such as stress, under the assumption that optimizing them leads to aesthetic and informative layouts. However, no single metric consistently produces the most effective results. A data-driven alternative is to learn from human preferences, where annotators select their favored visualization among multiple layouts of the same graphs. These human-preference labels can then be used to train a generative model that approximates human aesthetic preferences. However, obtaining human labels at scale is costly and time-consuming. As a result, this generative approach has so far been tested only with machine-labeled data. In this paper, we explore the use of large language models (LLMs) and vision models (VMs) as proxies for human judgment. Through a carefully designed user study involving 27 participants, we curated a large set of human preference labels. We used this data both to better understand human preferences and to bootstrap LLM/VM labelers. We show that prompt engineering that combines few-shot examples and diverse input formats, such as image embeddings, significantly improves LLM-human alignment, and additional filtering by the confidence score of the LLM pushes the alignment to human-human levels. Furthermore, we demonstrate that carefully trained VMs can achieve VM-human alignment at a level comparable to that between human annotators. Our results suggest that AI can feasibly serve as a scalable proxy for human labelers.

Mohan Jiang, Dayuan Fu, Junhao Shi et al.

While Large Language Models (LLMs) excel at short-term tasks, scaling them to long-horizon agentic workflows remains challenging. The core bottleneck lies in the scarcity of training data that captures authentic long-dependency structures and cross-stage evolutionary dynamics--existing synthesis methods either confine to single-feature scenarios constrained by model distribution, or incur prohibitive human annotation costs, failing to provide scalable, high-quality supervision. We address this by reconceptualizing data synthesis through the lens of real-world software evolution. Our key insight: Pull Request (PR) sequences naturally embody the supervision signals for long-horizon learning. They decompose complex objectives into verifiable submission units, maintain functional coherence across iterations, and encode authentic refinement patterns through bug-fix histories. Building on this, we propose daVinci-Agency, which systematically mines structured supervision from chain-of-PRs through three interlocking mechanisms: (1) progressive task decomposition via continuous commits, (2) long-term consistency enforcement through unified functional objectives, and (3) verifiable refinement from authentic bug-fix trajectories. Unlike synthetic trajectories that treat each step independently, daVinci-Agency's PR-grounded structure inherently preserves the causal dependencies and iterative refinements essential for teaching persistent goal-directed behavior and enables natural alignment with project-level, full-cycle task modeling. The resulting trajectories are substantial--averaging 85k tokens and 116 tool calls--yet remarkably data-efficient: fine-tuning GLM-4.6 on 239 daVinci-Agency samples yields broad improvements across benchmarks, notably achieving a 47% relative gain on Toolathlon. Beyond benchmark performance, our analysis confirms...

Weixuan Kong, Jinpeng Yu, Zijun Li et al.

Automatic personality recognition is a research hotspot in the intersection of computer science and psychology, and in human-computer interaction, personalised has a wide range of applications services and other scenarios. In this paper, an end-to-end multimodal performance personality is established for both visual and auditory modal datarecognition network , and the through feature-level fusion , which effectively of the two modalities is carried out the cross-attention mechanismfuses the features of the two modal data; and a is proposed multiscale feature enhancement modalitiesmodule , which enhances for visual and auditory boththe expression of the information of effective the features and suppresses the interference of the redundant information. In addition, during the training process, this paper proposes a modal enhancement training strategy to simulate non-ideal such as modal loss and noise interferencedata situations , which enhances the adaptability ofand the model to non-ideal data scenarios improves the robustness of the model. Experimental results show that the method proposed in this paper is able to achieve an average Big Five personality accuracy of , which outperforms existing 0.916 on the personality analysis dataset ChaLearn First Impressionother methods based on audiovisual and audio-visual both modalities. The ablation experiments also validate our proposed , respectivelythe contribution of module and modality enhancement strategy to the model performance. Finally, we simulate in the inference phase multi-scale feature enhancement six non-ideal data scenarios to verify the modal enhancement strategy's improvement in model robustness.

Pengchao Wu, Xuefeng Li, Jinghang Gu et al.

Biomedical event extraction is an information extraction task to obtain events from biomedical text, whose targets include the type, the trigger, and the respective arguments involved in an event. Traditional biomedical event extraction usually adopts a pipelined approach, which contains trigger identification, argument role recognition, and finally event construction either using specific rules or by machine learning. In this paper, we propose an n-ary relation extraction method based on the BERT pre-training model to construct Binding events, in order to capture the semantic information about an event's context and its participants. The experimental results show that our method achieves promising results on the GE11 and GE13 corpora of the BioNLP shared task with F1 scores of 63.14% and 59.40%, respectively. It demonstrates that by significantly improving theperformance of Binding events, the overall performance of the pipelined event extraction approach or even exceeds those of current joint learning methods.

Liwen Wang, Xuefeng Li, Jiachi Liu et al.

Zero-shot cross-domain slot filling alleviates the data dependence in the case of data scarcity in the target domain, which has aroused extensive research. However, as most of the existing methods do not achieve effective knowledge transfer to the target domain, they just fit the distribution of the seen slot and show poor performance on unseen slot in the target domain. To solve this, we propose a novel approach based on prototypical contrastive learning with a dynamic label confusion strategy for zero-shot slot filling. The prototypical contrastive learning aims to reconstruct the semantic constraints of labels, and we introduce the label confusion strategy to establish the label dependence between the source domains and the target domain on-the-fly. Experimental results show that our model achieves significant improvement on the unseen slots, while also set new state-of-the-arts on slot filling task.

Xuefeng Li, Alan Blair

Several regularization methods have recently been introduced which force the latent activations of an autoencoder or deep neural network to conform to either a Gaussian or hyperspherical distribution, or to minimize the implicit rank of the distribution in latent space. In the present work, we introduce a novel regularizing loss function which simulates a pairwise repulsive force between items and an attractive force of each item toward the origin. We show that minimizing this loss function in isolation achieves a hyperspherical distribution. Moreover, when used as a regularizing term, the scaling factor can be adjusted to allow greater flexibility and tolerance of eccentricity, thus allowing the latent variables to be stratified according to their relative importance, while still promoting diversity. We apply this method of Eccentric Regularization to an autoencoder, and demonstrate its effectiveness in image generation, representation learning and downstream classification tasks.

Xuefeng Li, Tongliang Liu, Bo Han et al.

In label-noise learning, the transition matrix plays a key role in building statistically consistent classifiers. Existing consistent estimators for the transition matrix have been developed by exploiting anchor points. However, the anchor-point assumption is not always satisfied in real scenarios. In this paper, we propose an end-to-end framework for solving label-noise learning without anchor points, in which we simultaneously optimize two objectives: the cross entropy loss between the noisy label and the predicted probability by the neural network, and the volume of the simplex formed by the columns of the transition matrix. Our proposed framework can identify the transition matrix if the clean class-posterior probabilities are sufficiently scattered. This is by far the mildest assumption under which the transition matrix is provably identifiable and the learned classifier is statistically consistent. Experimental results on benchmark datasets demonstrate the effectiveness and robustness of the proposed method.