Ling Feng

Jiajie Zhang, Yushi Bai, Xin Lv et al. · tsinghua

Though current long-context large language models (LLMs) have demonstrated impressive capacities in answering user questions based on extensive text, the lack of citations in their responses makes user verification difficult, leading to concerns about their trustworthiness due to their potential hallucinations. In this work, we aim to enable long-context LLMs to generate responses with fine-grained sentence-level citations, improving their faithfulness and verifiability. We first introduce LongBench-Cite, an automated benchmark for assessing current LLMs' performance in Long-Context Question Answering with Citations (LQAC), revealing considerable room for improvement. To this end, we propose CoF (Coarse to Fine), a novel pipeline that utilizes off-the-shelf LLMs to automatically generate long-context QA instances with precise sentence-level citations, and leverage this pipeline to construct LongCite-45k, a large-scale SFT dataset for LQAC. Finally, we train LongCite-8B and LongCite-9B using the LongCite-45k dataset, successfully enabling their generation of accurate responses and fine-grained sentence-level citations in a single output. The evaluation results on LongBench-Cite show that our trained models achieve state-of-the-art citation quality, surpassing advanced proprietary models including GPT-4o.

Jiajie Zhang, Xin Lv, Ling Feng et al.

Reinforcement learning (RL) has emerged as a critical technique for enhancing LLM-based deep search agents. However, existing approaches primarily rely on binary outcome rewards, which fail to capture the comprehensiveness and factuality of agents' reasoning process, and often lead to undesirable behaviors such as shortcut exploitation and hallucinations. To address these limitations, we propose \textbf{Citation-aware Rubric Rewards (CaRR)}, a fine-grained reward framework for deep search agents that emphasizes reasoning comprehensiveness, factual grounding, and evidence connectivity. CaRR decomposes complex questions into verifiable single-hop rubrics and requires agents to satisfy these rubrics by explicitly identifying hidden entities, supporting them with correct citations, and constructing complete evidence chains that link to the predicted answer. We further introduce \textbf{Citation-aware Group Relative Policy Optimization (C-GRPO)}, which combines CaRR and outcome rewards for training robust deep search agents. Experiments show that C-GRPO consistently outperforms standard outcome-based RL baselines across multiple deep search benchmarks. Our analysis also validates that C-GRPO effectively discourages shortcut exploitation, promotes comprehensive, evidence-grounded reasoning, and exhibits strong generalization to open-ended deep research tasks. Our code and data are available at https://github.com/THUDM/CaRR.

Jiajie Zhang, Nianyi Lin, Lei Hou et al.

Recently, large reasoning models have achieved impressive performance on various tasks by employing human-like deep thinking. However, the lengthy thinking process substantially increases inference overhead, making efficiency a critical bottleneck. In this work, we first demonstrate that NoThinking, which prompts the reasoning model to skip thinking and directly generate the final solution, is a better choice for relatively simple tasks in terms of both performance and efficiency. Motivated by this, we propose AdaptThink, a novel RL algorithm to teach reasoning models to choose the optimal thinking mode adaptively based on problem difficulty. Specifically, AdaptThink features two core components: (1) a constrained optimization objective that encourages the model to choose NoThinking while maintaining the overall performance; (2) an importance sampling strategy that balances Thinking and NoThinking samples during on-policy training, thereby enabling cold start and allowing the model to explore and exploit both thinking modes throughout the training process. Our experiments indicate that AdaptThink significantly reduces the inference costs while further enhancing performance. Notably, on three math datasets, AdaptThink reduces the average response length of DeepSeek-R1-Distill-Qwen-1.5B by 53% and improves its accuracy by 2.4%, highlighting the promise of adaptive thinking-mode selection for optimizing the balance between reasoning quality and efficiency. Our codes and models are available at https://github.com/THU-KEG/AdaptThink.

Weijie Wei, Zhipeng Luo, Ling Feng et al.

While Vision-Language Models (VLMs) show significant promise for end-to-end autonomous driving by leveraging the common sense embedded in language models, their reliance on 2D image cues for complex scene understanding and decision-making presents a critical bottleneck for safety and reliability. Current image-based methods struggle with accurate metric spatial reasoning and geometric inference, leading to unreliable driving policies. To bridge this gap, we propose LVLDrive (LiDAR-Vision-Language), a novel framework specifically designed to upgrade existing VLMs with robust 3D metric spatial understanding for autonomous driving by incoperating LiDAR point cloud as an extra input modality. A key challenge lies in mitigating the catastrophic disturbance introduced by disparate 3D data to the pre-trained VLMs. To this end, we introduce a Gradual Fusion Q-Former that incrementally injects LiDAR features, ensuring the stability and preservation of the VLM's existing knowledge base. Furthermore, we develop a spatial-aware question-answering (SA-QA) dataset to explicitly teach the model advanced 3D perception and reasoning capabilities. Extensive experiments on driving benchmarks demonstrate that LVLDrive achieves superior performance compared to vision-only counterparts across scene understanding, metric spatial perception, and reliable driving decision-making. Our work highlights the necessity of explicit 3D metric data for building trustworthy VLM-based autonomous systems.

Ling Feng, Tianhao Wu, Xiangrong Ren et al.

This paper introduces a new knowledge distillation method, called education distillation (ED), which is inspired by the structured and progressive nature of human learning. ED mimics the educational stages of primary school, middle school, and university and designs teaching reference blocks. The student model is split into a main body and multiple teaching reference blocks to learn from teachers step by step. This promotes efficient knowledge distillation while maintaining the architecture of the student model. Experimental results on the CIFAR100, Tiny Imagenet, Caltech and Food-101 datasets show that the teaching reference blocks can effectively avoid the problem of forgetting. Compared with conventional single-teacher and multi-teacher knowledge distillation methods, ED significantly improves the accuracy and generalization ability of the student model. These findings highlight the potential of ED to improve model performance across different architectures and datasets, indicating its value in various deep learning scenarios. Code examples can be obtained at: https://github.com/Revolutioner1/ED.git.

Jiajie Zhang, Shulin Cao, Linmei Hu et al.

Program induction (PI) has become a promising paradigm for using knowledge bases (KBs) to help large language models (LLMs) answer complex knowledge-intensive questions. Nonetheless, PI typically relies on a large number of parallel question-program pairs to make the LLM aware of the schema of the given KB, and is thus challenging for many low-resourced KBs that lack annotated data. To this end, we propose KB-Plugin, a plug-and-play framework that enables LLMs to induce programs over any low-resourced KB. Firstly, KB-Plugin adopts self-supervised learning to encode the detailed schema information of a given KB into a pluggable module, namely schema plugin. Secondly, KB-Plugin utilizes abundant annotated data from a rich-resourced KB to train another pluggable module, namely PI plugin, which can help the LLM extract question-relevant schema information from the schema plugin of any KB and utilize this information to induce programs over this KB. Experiments on five heterogeneous KBQA datasets show that KB-Plugin achieves better or comparable performance with 25$\times$ smaller backbone LLM compared to SoTA PI methods for low-resourced KBs, and even approaches the performance of supervised methods. Our code and data are available at https://github.com/THU-KEG/KB-Plugin.

Haibo Chen, Xin Wang, Jiaheng Chao et al.

Leveraging Graph Neural Networks (GNNs) as graph encoders and aligning the resulting representations with Large Language Models (LLMs) through alignment instruction tuning has become a mainstream paradigm for constructing Graph Language Models (GLMs), combining the generalization ability of LLMs with the structural modeling capacity of GNNs. However, existing GLMs that adopt GNNs as graph encoders largely overlook the problem of aligning GNN-encoded representations across domains and tasks with the LLM token space to obtain unified graph tokens, thereby limiting their ability to generalize across diverse graph data. To bridge this gap, we aim to incorporate a multi-domain, multi-task GNN encoder into GLMs and align its representations with LLMs to enable multi-domain, multi-task graph alignment instruction tuning. This alignment problem remains underexplored and poses two key challenges: 1) learning GNN-encoded representations that are simultaneously generalizable across domains and tasks and well aligned with textual semantics is difficult, due to substantial variations in graph structures, feature distributions, and supervision signals, together with the lack of textual-semantic alignment guidance in task-specific GNN training; 2) diverse graph data and task-specific instructions can exhibit different degrees of compatibility with the LLM token space during instruction tuning, leading to varying alignment difficulty and rendering a fixed alignment strategy suboptimal. To tackle these challenges, we propose UniGraphLM, a Unified Graph Language Model that incorporates a multi-domain, multi-task GNN encoder to learn generalizable graph representations aligned with textual semantics, and then adaptively aligns these representations with the LLM.

Xin Wang, Ling Feng, Huijun Zhang et al.

Stress haunts people in modern society, which may cause severe health issues if left unattended. With social media becoming an integral part of daily life, leveraging social media to detect stress has gained increasing attention. While the majority of the work focuses on classifying stress states and stress categories, this study introduce a new task aimed at estimating more specific stressors (like exam, writing paper, etc.) through users' posts on social media. Unfortunately, the diversity of stressors with many different classes but a few examples per class, combined with the consistent arising of new stressors over time, hinders the machine understanding of stressors. To this end, we cast the stressor estimation problem within a practical scenario few-shot learning setting, and propose a novel meta-learning based stressor estimation framework that is enhanced by a meta-knowledge inheritance mechanism. This model can not only learn generic stressor context through meta-learning, but also has a good generalization ability to estimate new stressors with little labeled data. A fundamental breakthrough in our approach lies in the inclusion of the meta-knowledge inheritance mechanism, which equips our model with the ability to prevent catastrophic forgetting when adapting to new stressors. The experimental results show that our model achieves state-of-the-art performance compared with the baselines. Additionally, we construct a social media-based stressor estimation dataset that can help train artificial intelligence models to facilitate human well-being. The dataset is now public at \href{https://www.kaggle.com/datasets/xinwangcs/stressor-cause-of-mental-health-problem-dataset}{\underline{Kaggle}} and \href{https://huggingface.co/datasets/XinWangcs/Stressor}{\underline{Hugging Face}}.

Sirui Hu, Ling Feng, Xiaohan Yang et al.

Federated learning is widely used to perform decentralized training of a global model on multiple devices while preserving the data privacy of each device. However, it suffers from heterogeneous local data on each training device which increases the difficulty to reach the same level of accuracy as the centralized training. Supervised Contrastive Learning which outperform cross-entropy tries to minimizes the difference between feature space of points belongs to the same class and pushes away points from different classes. We propose Supervised Contrastive Federated Learning in which devices can share the learned class-wise feature spaces with each other and add the supervised-contrastive learning loss as a regularization term to foster the feature space learning. The loss tries to minimize the cosine similarity distance between the feature map and the averaged feature map from another device in the same class and maximizes the distance between the feature map and that in a different class. This new regularization term when added on top of the moon regularization term is found to outperform the other state-of-the-art regularization terms in solving the heterogeneous data distribution problem.

Jeremy Oon, Rakhi Manohar Mepparambath, Ling Feng

Despite the significant progress of deep learning models in multitude of applications, their adaption in planning and policy related areas remains challenging due to the black-box nature of these models. In this work, we develop a set of DeepLogit models that follow a novel sequentially constrained approach in estimating deep learning models for transport policy analysis. In the first step of the proposed approach, we estimate a convolutional neural network (CNN) model with only linear terms, which is equivalent of a linear-in-parameter multinomial logit model. We then estimate other deep learning models by constraining the parameters that need interpretability at the values obtained in the linear-in-parameter CNN model and including higher order terms or by introducing advanced deep learning architectures like Transformers. Our approach can retain the interpretability of the selected parameters, yet provides significantly improved model accuracy than the discrete choice model. We demonstrate our approach on a transit route choice example using real-world transit smart card data from Singapore. This study shows the potential for a unifying approach, where theory-based discrete choice model (DCM) and data-driven AI models can leverage each other's strengths in interpretability and predictive power. With the availability of larger datasets and more complex constructions, such approach can lead to more accurate models using discrete choice models while maintaining its applicability in planning and policy-related areas. Our code is available on https://github.com/jeremyoon/route-choice/ .

Jiajie Zhang, Zhongni Hou, Xin Lv et al.

Though significant advancements have been achieved in developing long-context large language models (LLMs), the compromised quality of LLM-synthesized data for supervised fine-tuning (SFT) often affects the long-context performance of SFT models and leads to inherent limitations. In principle, reinforcement learning (RL) with appropriate reward signals can further enhance models' capacities. However, how to obtain reliable rewards in long-context scenarios remains unexplored. To this end, we propose LongReward, a novel method that utilizes an off-the-shelf LLM to provide rewards for long-context model responses from four human-valued dimensions: helpfulness, logicality, faithfulness, and completeness, each with a carefully designed assessment pipeline. By combining LongReward and offline RL algorithm DPO, we are able to effectively improve long-context SFT models. Our experiments indicate that LongReward not only significantly improves models' long-context performance but also enhances their ability to follow short instructions. We also find that long-context DPO with LongReward and conventional short-context DPO can be used together without hurting either one's performance.

Nicholas Chong Jia Le, Ling Feng

The presence of $1/f$ noise, also known as pink noise, is a well-established phenomenon in biological neural networks, and is thought to play an important role in information processing in the brain. In this study, we find that such $1/f$ noise is also found in deep neural networks trained on natural language, resembling that of their biological counterparts. Specifically, we trained Long Short-Term Memory (LSTM) networks on the `IMDb' AI benchmark dataset, then measured the neuron activations. The detrended fluctuation analysis (DFA) on the time series of the different neurons demonstrate clear $1/f$ patterns, which is absent in the time series of the inputs to the LSTM. Interestingly, when the neural network is at overcapacity, having more than enough neurons to achieve the learning task, the activation patterns deviate from $1/f$ noise and shifts towards white noise. This is because many of the neurons are not effectively used, showing little fluctuations when fed with input data. We further examine the exponent values in the $1/f$ noise in ``internal" and ``external" activations in the LSTM cell, finding some resemblance in the variations of the exponents in fMRI signals of the human brain. Our findings further supports the hypothesis that $1/f$ noise is a signature of optimal learning. With deep learning models approaching or surpassing humans in certain tasks, and being more ``experimentable'' than their biological counterparts, our study suggests that they are good candidates to understand the fundamental origins of $1/f$ noise.

Jun Li, Xiangmeng Wang, Haoyang Li et al.

Suicide is a critical global health issue that requires urgent attention. Even though prior work has revealed valuable insights into detecting current suicide risk on social media, little attention has been paid to developing models that can predict subsequent suicide risk over time, limiting their ability to capture rapid fluctuations in individuals' mental state transitions. In addition, existing work ignores protective factors that play a crucial role in suicide risk prediction, focusing predominantly on risk factors alone. Protective factors such as social support and coping strategies can mitigate suicide risk by moderating the impact of risk factors. Therefore, this study proposes a novel framework for predicting subsequent suicide risk by jointly learning the dynamic influence of both risk factors and protective factors on users' suicide risk transitions. We propose a novel Protective Factor-Aware Dataset, which is built from 12 years of Reddit posts along with comprehensive annotations of suicide risk and both risk and protective factors. We also introduce a Dynamic Factors Influence Learning approach that captures the varying impact of risk and protective factors on suicide risk transitions, recognizing that suicide risk fluctuates over time according to established psychological theories. Our thorough experiments demonstrate that the proposed model significantly outperforms state-of-the-art models and large language models across three datasets. In addition, the proposed Dynamic Factors Influence Learning provides interpretable weights, helping clinicians better understand suicidal patterns and enabling more targeted intervention strategies.

Ling Feng, Tianyu Xie, Wei Ma et al.

The modernization of smart farming is a way to improve agricultural production efficiency, and improve the agricultural production environment. Although many large models have achieved high accuracy in the task of object recognition and segmentation, they cannot really be put into use in the farming industry due to their own poor interpretability and limitations in computational volume. In this paper, we built AnYue Shelduck Dateset, which contains a total of 1951 Shelduck datasets, and performed target detection and segmentation annotation with the help of professional annotators. Based on AnYue ShelduckDateset, this paper describes DuckProcessing, an efficient and powerful module for duck identification based on real shelduckfarms. First of all, using the YOLOv8 module designed to divide the mahjong between them, Precision reached 98.10%, Recall reached 96.53% and F1 score reached 0.95 on the test set. Again using the DuckSegmentation segmentation model, DuckSegmentation reached 96.43% mIoU. Finally, the excellent DuckSegmentation was used as the teacher model, and through knowledge distillation, Deeplabv3 r50 was used as the student model, and the final student model achieved 94.49% mIoU on the test set. The method provides a new way of thinking in practical sisal duck smart farming.

Ling Feng, SK Yang

In Rectified Flow, by obtaining the rectified flow several times, the mapping relationship between distributions can be distilled into a neural network, and the target distribution can be directly predicted by the straight lines of the flow. However, during the pairing process of the mapping relationship, a large amount of error accumulation will occur, resulting in a decrease in performance after multiple rectifications. In the field of flow models, knowledge distillation of multi - teacher diffusion models is also a problem worthy of discussion in accelerating sampling. I intend to combine multi - teacher knowledge distillation with Bezier curves to solve the problem of error accumulation. Currently, the related paper is being written by myself.

Kaisheng Zeng, Zhenhao Dong, Lei Hou et al.

Self-supervised entity alignment (EA) aims to link equivalent entities across different knowledge graphs (KGs) without seed alignments. The current SOTA self-supervised EA method draws inspiration from contrastive learning, originally designed in computer vision based on instance discrimination and contrastive loss, and suffers from two shortcomings. Firstly, it puts unidirectional emphasis on pushing sampled negative entities far away rather than pulling positively aligned pairs close, as is done in the well-established supervised EA. Secondly, KGs contain rich side information (e.g., entity description), and how to effectively leverage those information has not been adequately investigated in self-supervised EA. In this paper, we propose an interactive contrastive learning model for self-supervised EA. The model encodes not only structures and semantics of entities (including entity name, entity description, and entity neighborhood), but also conducts cross-KG contrastive learning by building pseudo-aligned entity pairs. Experimental results show that our approach outperforms previous best self-supervised results by a large margin (over 9% average improvement) and performs on par with previous SOTA supervised counterparts, demonstrating the effectiveness of the interactive contrastive learning for self-supervised EA.

Lin Zhang, Ling Feng, Kan Chen et al.

The success of deep neural networks in real-world problems has prompted many attempts to explain their training dynamics and generalization performance, but more guiding principles for the training of neural networks are still needed. Motivated by the edge of chaos principle behind the optimal performance of neural networks, we study the role of various hyperparameters in modern neural network training algorithms in terms of the order-chaos phase diagram. In particular, we study a fully analytical feedforward neural network trained on the widely adopted Fashion-MNIST dataset, and study the dynamics associated with the hyperparameters in back-propagation during the training process. We find that for the basic algorithm of stochastic gradient descent with momentum, in the range around the commonly used hyperparameter values, clear scaling relations are present with respect to the training time during the ordered phase in the phase diagram, and the model's optimal generalization power at the edge of chaos is similar across different training parameter combinations. In the chaotic phase, the same scaling no longer exists. The scaling allows us to choose the training parameters to achieve faster training without sacrificing performance. In addition, we find that the commonly used model regularization method - weight decay - effectively pushes the model towards the ordered phase to achieve better performance. Leveraging on this fact and the scaling relations in the other hyperparameters, we derived a principled guideline for hyperparameter determination, such that the model can achieve optimal performance by saturating it at the edge of chaos. Demonstrated on this simple neural network model and training algorithm, our work improves the understanding of neural network training dynamics, and can potentially be extended to guiding principles of more complex model architectures and algorithms.

Lei Cao, Huijun Zhang, Ling Feng

A large number of individuals are suffering from suicidal ideation in the world. There are a number of causes behind why an individual might suffer from suicidal ideation. As the most popular platform for self-expression, emotion release, and personal interaction, individuals may exhibit a number of symptoms of suicidal ideation on social media. Nevertheless, challenges from both data and knowledge aspects remain as obstacles, constraining the social media-based detection performance. Data implicitness and sparsity make it difficult to discover the inner true intentions of individuals based on their posts. Inspired by psychological studies, we build and unify a high-level suicide-oriented knowledge graph with deep neural networks for suicidal ideation detection on social media. We further design a two-layered attention mechanism to explicitly reason and establish key risk factors to individual's suicidal ideation. The performance study on microblog and Reddit shows that: 1) with the constructed personal knowledge graph, the social media-based suicidal ideation detection can achieve over 93% accuracy; and 2) among the six categories of personal factors, post, personality, and experience are the top-3 key indicators. Under these categories, posted text, stress level, stress duration, posted image, and ruminant thinking contribute to one's suicidal ideation detection.

Chi Zhang, Yong Sheng Soh, Ling Feng et al.

While current machine learning models have impressive performance over a wide range of applications, their large size and complexity render them unsuitable for tasks such as remote monitoring on edge devices with limited storage and computational power. A naive approach to resolve this on the model level is to use simpler architectures, but this sacrifices prediction accuracy and is unsuitable for monitoring applications requiring accurate detection of the onset of adverse events. In this paper, we propose an alternative solution to this problem by decomposing the predictive model as the sum of a simple function which serves as a local monitoring tool, and a complex correction term to be evaluated on the server. A sign requirement is imposed on the latter to ensure that the local monitoring function is safe, in the sense that it can effectively serve as an early warning system. Our analysis quantifies the trade-offs between model complexity and performance, and serves as a guidance for architecture design. We validate our proposed framework on a series of monitoring experiments, where we succeed at learning monitoring models with significantly reduced complexity that minimally violate the safety requirement. More broadly, our framework is useful for learning classifiers in applications where false negatives are significantly more costly compared to false positives.

Lei Cao, Huijun Zhang, Ling Feng et al.

Despite detection of suicidal ideation on social media has made great progress in recent years, people's implicitly and anti-real contrarily expressed posts still remain as an obstacle, constraining the detectors to acquire higher satisfactory performance. Enlightened by the hidden "tree holes" phenomenon on microblog, where people at suicide risk tend to disclose their inner real feelings and thoughts to the microblog space whose authors have committed suicide, we explore the use of tree holes to enhance microblog-based suicide risk detection from the following two perspectives. (1) We build suicide-oriented word embeddings based on tree hole contents to strength the sensibility of suicide-related lexicons and context based on tree hole contents. (2) A two-layered attention mechanism is deployed to grasp intermittently changing points from individual's open blog streams, revealing one's inner emotional world more or less. Our experimental results show that with suicide-oriented word embeddings and attention, microblog-based suicide risk detection can achieve over 91\% accuracy. A large-scale well-labelled suicide data set is also reported in the paper.

Ling Feng, Lin Zhang, Choy Heng Lai

It has long been suggested that the biological brain operates at some critical point between two different phases, possibly order and chaos. Despite many indirect empirical evidence from the brain and analytical indication on simple neural networks, the foundation of this hypothesis on generic non-linear systems remains unclear. Here we develop a general theory that reveals the exact edge of chaos is the boundary between the chaotic phase and the (pseudo)periodic phase arising from Neimark-Sacker bifurcation. This edge is analytically determined by the asymptotic Jacobian norm values of the non-linear operator and influenced by the dimensionality of the system. The optimality at the edge of chaos is associated with the highest information transfer between input and output at this point similar to that of the logistic map. As empirical validations, our experiments on the various deep learning models in computer vision demonstrate the optimality of the models near the edge of chaos, and we observe that the state-of-art training algorithms push the models towards such edge as they become more accurate. We further establishes the theoretical understanding of deep learning model generalization through asymptotic stability.

Gangli Liu, Ling Feng

Re-finding electronic documents from a personal computer is a frequent demand to users. In a simple re-finding task, people can use many methods to retrieve a document, such as navigating directly to the document's folder, searching with a desktop search engine, or checking the Recent Files List. However, when encountering a difficult re-finding task, people usually cannot remember the attributes used by conventional re-finding methods, such as file path, file name, keywords etc., the re-finding would fail. We propose a new method to support difficult re-finding tasks. When a user is reading a document, we collect all kinds of possible memory pieces of the user about the document, such as number of pages, number of images, number of math formulas, cumulative reading time, reading frequency, printing experiences etc. If the user wants to re-find a document later, we use these collected attributes to filter out the target document. To alleviate the user's cognitive burden, we use a question and answer wizard interface and provide recommendations to the answers for the user, the recommendations are generated by analyzing the collected attributes of each document and the user's experiences about them.