Bowen Cao

Bowen Cao, Weibin Liao, Yushi Sun et al.

Text-to-SQL systems powered by Large Language Models have excelled on academic benchmarks but struggle in complex enterprise environments. The primary limitation lies in their reliance on static schema representations, which fails to resolve semantic ambiguity and scale effectively to large, complex databases. To address this, we propose APEX-SQL, an Agentic Text-to-SQL Framework that shifts the paradigm from passive translation to agentic exploration. Our framework employs a hypothesis-verification loop to ground model reasoning in real data. In the schema linking phase, we use logical planning to verbalize hypotheses, dual-pathway pruning to reduce the search space, and parallel data profiling to validate column roles against real data, followed by global synthesis to ensure topological connectivity. For SQL generation, we introduce a deterministic mechanism to retrieve exploration directives, allowing the agent to effectively explore data distributions, refine hypotheses, and generate semantically accurate SQLs. Experiments on BIRD (70.65% execution accuracy) and Spider 2.0-Snow (51.01% execution accuracy) demonstrate that APEX-SQL outperforms competitive baselines with reduced token consumption. Further analysis reveals that agentic exploration acts as a performance multiplier, unlocking the latent reasoning potential of foundation models in enterprise settings. Ablation studies confirm the critical contributions of each component in ensuring robust and accurate data analysis. Our code is released at https://github.com/Tencent/APEX-SQL-Project.

Bowen Cao, Qichen Ye, Weiyuan Xu et al.

Learning representations for graph-structured data is essential for graph analytical tasks. While remarkable progress has been made on static graphs, researches on temporal graphs are still in its beginning stage. The bottleneck of the temporal graph representation learning approach is the neighborhood aggregation strategy, based on which graph attributes share and gather information explicitly. Existing neighborhood aggregation strategies fail to capture either the short-term features or the long-term features of temporal graph attributes, leading to unsatisfactory model performance and even poor robustness and domain generality of the representation learning method. To address this problem, we propose a Frame-level Timeline Modeling (FTM) method that helps to capture both short-term and long-term features and thus learns more informative representations on temporal graphs. In particular, we present a novel link-based framing technique to preserve the short-term features and then incorporate a timeline aggregator module to capture the intrinsic dynamics of graph evolution as long-term features. Our method can be easily assembled with most temporal GNNs. Extensive experiments on common datasets show that our method brings great improvements to the capability, robustness, and domain generality of backbone methods in downstream tasks. Our code can be found at https://github.com/yeeeqichen/FTM.

Xuxin Cheng, Bowen Cao, Qichen Ye et al.

Spoken language understanding (SLU) is a fundamental task in the task-oriented dialogue systems. However, the inevitable errors from automatic speech recognition (ASR) usually impair the understanding performance and lead to error propagation. Although there are some attempts to address this problem through contrastive learning, they (1) treat clean manual transcripts and ASR transcripts equally without discrimination in fine-tuning; (2) neglect the fact that the semantically similar pairs are still pushed away when applying contrastive learning; (3) suffer from the problem of Kullback-Leibler (KL) vanishing. In this paper, we propose Mutual Learning and Large-Margin Contrastive Learning (ML-LMCL), a novel framework for improving ASR robustness in SLU. Specifically, in fine-tuning, we apply mutual learning and train two SLU models on the manual transcripts and the ASR transcripts, respectively, aiming to iteratively share knowledge between these two models. We also introduce a distance polarization regularizer to avoid pushing away the intra-cluster pairs as much as possible. Moreover, we use a cyclical annealing schedule to mitigate KL vanishing issue. Experiments on three datasets show that ML-LMCL outperforms existing models and achieves new state-of-the-art performance.

Qichen Ye, Bowen Cao, Nuo Chen et al.

Knowledge-aware question answering (KAQA) requires the model to answer questions over a knowledge base, which is essential for both open-domain QA and domain-specific QA, especially when language models alone cannot provide all the knowledge needed. Despite the promising result of recent KAQA systems which tend to integrate linguistic knowledge from pre-trained language models (PLM) and factual knowledge from knowledge graphs (KG) to answer complex questions, a bottleneck exists in effectively fusing the representations from PLMs and KGs because of (i) the semantic and distributional gaps between them, and (ii) the difficulties in joint reasoning over the provided knowledge from both modalities. To address the above two problems, we propose a Fine-grained Two-stage training framework (FiTs) to boost the KAQA system performance: The first stage aims at aligning representations from the PLM and the KG, thus bridging the modality gaps between them, named knowledge adaptive post-training. The second stage, called knowledge-aware fine-tuning, aims to improve the model's joint reasoning ability based on the aligned representations. In detail, we fine-tune the post-trained model via two auxiliary self-supervised tasks in addition to the QA supervision. Extensive experiments demonstrate that our approach achieves state-of-the-art performance on three benchmarks in the commonsense reasoning (i.e., CommonsenseQA, OpenbookQA) and medical question answering (i.e., MedQA-USMILE) domains.

Bowen Cao, Dongdong Zhang, Yixia Li et al.

Large language models now solve many benchmark math problems at near-expert levels, yet this progress has not fully translated into reliable performance in real-world applications. We study this gap through contextual mathematical reasoning, where the mathematical core must be formulated from descriptive scenarios. We introduce ContextMATH, a benchmark that repurposes AIME and MATH-500 problems into two contextual settings: Scenario Grounding (SG), which embeds abstract problems into realistic narratives without increasing reasoning complexity, and Complexity Scaling (CS), which transforms explicit conditions into sub-problems to capture how constraints often appear in practice. Evaluating 61 proprietary and open-source models, we observe sharp drops: on average, open-source models decline by 13 and 34 points on SG and CS, while proprietary models drop by 13 and 20. Error analysis shows that errors are dominated by incorrect problem formulation, with formulation accuracy declining as original problem difficulty increases. Correct formulation emerges as a prerequisite for success, and its sufficiency improves with model scale, indicating that larger models advance in both understanding and reasoning. Nevertheless, formulation and reasoning remain two complementary bottlenecks that limit contextual mathematical problem solving. Finally, we find that fine-tuning with scenario data improves performance, whereas formulation-only training is ineffective. However, performance gaps are only partially alleviated, highlighting contextual mathematical reasoning as a central unsolved challenge for LLMs.

Renchu Guan, Yajun Wang, Chunli Guo et al.

Graph few-shot learning, which aims to classify nodes from novel classes with only a few labeled examples, is a widely studied problem in graph learning. However, existing methods often face two key limitations. First, the predominant graph few-shot learning paradigm relies on supervised tasks, failing to leverage the vast number of unlabeled nodes in the graph. Second, many approaches require complex task adaptation or fine-tuning during inference, limiting their efficiency and applicability. Inspired by the powerful in-context learning capabilities of large language models, we propose a novel model named VISION for adVancIng graph few-Shot learning via In-cOntext LearNing to address these challenges. Our model reframes graph few-shot learning as a fine-tuning-free sequence reasoning problem. At its core is a context-aware network that initializes nodes with role embeddings and employs a dual-context fusion module to synergistically integrate local topological structures and global task-level dependencies. This allows our model to dynamically generate class-aware representations for the query set conditioned on the support set context in a single forward pass. To effectively train our model, we introduce an unsupervised task generator that creates structure-adaptive features and constructs diverse pseudo-tasks from abundant unlabeled data. Our method unifies unsupervised meta-learning with graph in-context learning, achieving efficient inference. Extensive experiments on multiple benchmark datasets demonstrate the superiority of our model. Our public code can be found

Bowen Cao, Deng Cai, Zhisong Zhang et al.

The performance of large language models (LLMs) is acutely sensitive to the phrasing of prompts, which raises significant concerns about their reliability in real-world scenarios. Existing studies often divide prompts into task-level instructions and case-level inputs and primarily focus on evaluating and improving robustness against variations in tasks-level instructions. However, this setup fails to fully address the diversity of real-world user queries and assumes the existence of task-specific datasets. To address these limitations, we introduce RobustAlpacaEval, a new benchmark that consists of semantically equivalent case-level queries and emphasizes the importance of using the worst prompt performance to gauge the lower bound of model performance. Extensive experiments on RobustAlpacaEval with ChatGPT and six open-source LLMs from the Llama, Mistral, and Gemma families uncover substantial variability in model performance; for instance, a difference of 45.48% between the worst and best performance for the Llama-2-70B-chat model, with its worst performance dipping as low as 9.38%. We further illustrate the difficulty in identifying the worst prompt from both model-agnostic and model-dependent perspectives, emphasizing the absence of a shortcut to characterize the worst prompt. We also attempt to enhance the worst prompt performance using existing prompt engineering and prompt consistency methods, but find that their impact is limited. These findings underscore the need to create more resilient LLMs that can maintain high performance across diverse prompts. Data and code are available at https://github.com/cbwbuaa/On-the-Worst-Prompt- Performance-of-LLMs.

Nuo Chen, Linjun Shou, Ming Gong et al.

Currently, learning better unsupervised sentence representations is the pursuit of many natural language processing communities. Lots of approaches based on pre-trained language models (PLMs) and contrastive learning have achieved promising results on this task. Experimentally, we observe that the over-smoothing problem reduces the capacity of these powerful PLMs, leading to sub-optimal sentence representations. In this paper, we present a Simple method named Self-Contrastive Learning (SSCL) to alleviate this issue, which samples negatives from PLMs intermediate layers, improving the quality of the sentence representation. Our proposed method is quite simple and can be easily extended to various state-of-the-art models for performance boosting, which can be seen as a plug-and-play contrastive framework for learning unsupervised sentence representation. Extensive results prove that SSCL brings the superior performance improvements of different strong baselines (e.g., BERT and SimCSE) on Semantic Textual Similarity and Transfer datasets. Our codes are available at https://github.com/nuochenpku/SSCL.

Yushi Sun, Bowen Cao, Dong Fang et al.

Long-horizon conversational agents rely on memory systems with increasingly sophisticated retrieval mechanisms. However, retrieved fragments are typically fed to the language model as unstructured text, lacking the relational, temporal, and thematic structures essential for complex reasoning. To bridge this reasoning gap, we introduce GRAVITY (\textbf{G}eneration-time \textbf{R}elational \textbf{A}nchoring \textbf{V}ia \textbf{I}njected \textbf{T}opological Memor\textbf{Y}), a plug-and-play structured memory module. GRAVITY extracts three complementary knowledge representations from raw conversational utterances: entity profiles grounded in relational graphs, temporal event tuples linked into causal traces, and cross-session topic summaries. At generation time, it injects these representations into the host system's prompt as structured anchoring contexts. This approach effectively synthesizes scattered evidence into a coherent, query-relevant context without requiring any architectural modifications to the host model. Extensive evaluations across five diverse memory systems on the LongMemEval and LoCoMo benchmarks demonstrate the efficacy of our approach. On average, GRAVITY improves LLM-judge accuracy by 7.5--10.1%. Gains are inversely correlated with baseline strength: the weakest host improves by 12.2% while the strongest still gains 3.8--5.7%. These findings establish structured context anchoring as a broadly effective, architecture-agnostic augmentation paradigm for long-horizon conversational memory.

Shijue Huang, Hongru Wang, Wanjun Zhong et al. · pku

Modern large reasoning models demonstrate impressive problem-solving capabilities by employing sophisticated reasoning strategies. However, they often struggle to balance efficiency and effectiveness, frequently generating unnecessarily lengthy reasoning chains for simple problems. In this work, we propose AdaCtrl, a novel framework to support both difficulty-aware adaptive reasoning budget allocation and explicit user control over reasoning depth. AdaCtrl dynamically adjusts its reasoning length based on self-assessed problem difficulty, while also allowing users to manually control the budget to prioritize either efficiency or effectiveness. This is achieved through a two-stage training pipeline: an initial cold-start fine-tuning phase to instill the ability to self-aware difficulty and adjust reasoning budget, followed by a difficulty-aware reinforcement learning (RL) stage that refines the model's adaptive reasoning strategies and calibrates its difficulty assessments based on its evolving capabilities during online training. To enable intuitive user interaction, we design explicit length-triggered tags that function as a natural interface for budget control. Empirical results show that AdaCtrl adapts reasoning length based on estimated difficulty, compared to the standard training baseline that also incorporates fine-tuning and RL, it yields performance improvements and simultaneously reduces response length by 10.06% and 12.14% on the more challenging AIME2024 and AIME2025 datasets, which require elaborate reasoning, and by 62.05% and 91.04% on the MATH500 and GSM8K datasets, where more concise responses are sufficient. Furthermore, AdaCtrl enables precise user control over the reasoning budget, allowing for tailored responses to meet specific needs.

Dingchen Yang, Bowen Cao, Guang Chen et al.

Multi-modal Large Language Models (MLLMs) demonstrate remarkable success across various vision-language tasks. However, they suffer from visual hallucination, where the generated responses diverge from the provided image. Are MLLMs oblivious to the accurate visual cues when they hallucinate? Our investigation reveals that the visual branch may equally advocate both accurate and erroneous content. To address this issue, we propose Pensieve, a training-free method that leverages the analogous visual hallucinations, which are induced by images sharing common semantic and appearance characteristics, to mitigate hallucination. Specifically, Pensieve enables MLLMs to retrospect relevant images as references and compare their visual content with the test image via confidence score subtraction. Moreover, our paradigm balances the effects of addressing errors from both the visual and textual branches by adaptively scaling the subtracted scores. Experiments on Whoops, LLaVA Bench, POPE, and MME demonstrate the efficacy of Pensieve in mitigating visual hallucination, surpassing other advanced decoding strategies. Pensieve also aids MLLMs in identifying visual details and enhance the specificity of generated image descriptions.

Bowen Cao, Deng Cai, Leyang Cui et al.

Standard language models generate text by selecting tokens from a fixed, finite, and standalone vocabulary. We introduce a novel method that selects context-aware phrases from a collection of supporting documents. One of the most significant challenges for this paradigm shift is determining the training oracles, because a string of text can be segmented in various ways and each segment can be retrieved from numerous possible documents. To address this, we propose to initialize the training oracles using linguistic heuristics and, more importantly, bootstrap the oracles through iterative self-reinforcement. Extensive experiments show that our model not only outperforms standard language models on a variety of knowledge-intensive tasks but also demonstrates improved generation quality in open-ended text generation. For instance, compared to the standard language model counterpart, our model raises the accuracy from 23.47% to 36.27% on OpenbookQA, and improves the MAUVE score from 42.61% to 81.58% in open-ended text generation. Remarkably, our model also achieves the best performance and the lowest latency among several retrieval-augmented baselines. In conclusion, we assert that retrieval is more accurate generation and hope that our work will encourage further research on this new paradigm shift.

Hongyuan Adam Lu, Z. L., Victor Wei et al.

While textual frequency has been validated as relevant to human cognition in reading speed, its relatedness to Large Language Models (LLMs) is seldom studied. We propose a novel research direction in terms of textual data frequency, which is an understudied topic, to the best of our knowledge. Our framework is composed of three units. First, this paper proposes Textual Frequency Law (TFL), which indicates that frequent textual data should be preferred for LLMs for both prompting and fine-tuning. Since many LLMs are closed-source in their training data, we propose using online resources to estimate the sentence-level frequency. We then utilize an input paraphraser to paraphrase the input into a more frequent textual expression. Next, we propose Textual Frequency Distillation (TFD) by querying LLMs to conduct story completion by further extending the sentences in the datasets, and the resulting corpora are used to adjust the initial estimation. Finally, we propose Curriculum Textual Frequency Training (CTFT) that fine-tunes LLMs in an increasing order of sentence-level frequency. Experiments are conducted on our curated dataset Textual Frequency Paired Dataset (TFPD) on math reasoning, machine translation, commonsense reasoning and agentic tool calling. Results show the effectiveness of our framework.

Bowen Cao, Deng Cai, Wai Lam

In-context learning (ICL) is critical for large language models (LLMs), but its effectiveness is constrained by finite context windows, particularly in ultra-long contexts. To overcome this, we introduce InfiniteICL, a framework that parallels context and parameters in LLMs with short- and long-term memory in human cognitive systems, focusing on transforming temporary context knowledge into permanent parameter updates. This approach significantly reduces memory usage, maintains robust performance across varying input lengths, and theoretically enables infinite context integration through the principles of context knowledge elicitation, selection, and consolidation. Evaluations demonstrate that our method reduces context length by 90% while achieving 103% average performance of full-context prompting across fact recall, grounded reasoning, and skill acquisition tasks. When conducting sequential multi-turn transformations on complex, real-world contexts (with length up to 2M tokens), our approach surpasses full-context prompting while using only 0.4% of the original contexts. These findings highlight InfiniteICL's potential to enhance the scalability and efficiency of LLMs by breaking the limitations of conventional context window sizes.

Dingchen Yang, Bowen Cao, Anran Zhang et al.

Multi-modal Large Langue Models (MLLMs) often process thousands of visual tokens, which consume a significant portion of the context window and impose a substantial computational burden. Prior work has empirically explored visual token pruning methods based on MLLMs' intermediate states (e.g., attention scores). However, they have limitations in precisely defining visual redundancy due to their inability to capture the influence of visual tokens on MLLMs' visual understanding (i.e., the predicted probabilities for textual token candidates). To address this issue, we manipulate the visual input and investigate variations in the textual output from both token-centric and context-centric perspectives, achieving intuitive and comprehensive analysis. Experimental results reveal that visual tokens with low ViT-[cls] association and low text-to-image attention scores can contain recognizable information and significantly contribute to images' overall information. To develop a more reliable method for identifying and pruning redundant visual tokens, we integrate these two perspectives and introduce a context-independent condition to identify redundant prototypes from training images, which probes the redundancy of each visual token during inference. Extensive experiments on single-image, multi-image and video comprehension tasks demonstrate the effectiveness of our method, notably achieving 90% to 110% of the performance while pruning 80% to 90% of visual tokens.

Junpeng Liu, Yuzhong Zhao, Bowen Cao et al.

Recent advances in agentic workflows have enabled the automation of tasks such as professional document generation. However, they primarily focus on textual quality, neglecting visual structure and style, which are crucial for readability and engagement. This gap arises mainly from the absence of suitable reward models to guide agentic workflows toward producing documents with stronger structural and stylistic quality. To address this, we propose DocReward, a document reward model that evaluates documents based on their structure and style. We construct a multi-domain dataset DocPair of 117K paired documents, covering 32 domains and 267 document types, each including a high- and low-professionalism document with identical content but different structure and style. This enables the model to evaluate professionalism comprehensively, and in a textual-quality-agnostic way. DocReward is trained using the Bradley-Terry loss to score documents, penalizing predictions that contradict the annotated ranking. To assess the performance of reward models, we create a test dataset containing document bundles ranked by well-educated human evaluators. Notably, DocReward outperforms GPT-4o and GPT-5 in accuracy by 30.6 and 19.4 percentage points, respectively, demonstrating its superiority over baselines. In an extrinsic evaluation of document generation, DocReward achieves a significantly higher win rate of 60.8%, compared to GPT-5's 37.7% win rate, demonstrating its utility in guiding generation agents toward producing human-preferred documents.

Deng Cai, Huayang Li, Tingchen Fu et al.

Despite the general capabilities of pre-trained large language models (LLMs), they still need further adaptation to better serve practical applications. In this paper, we demonstrate the interchangeability of three popular and distinct adaptation tools: parameter updating, reward modeling, and in-context prompting. This interchangeability establishes a triangular framework with six transformation directions, each of which facilitates a variety of applications. Our work offers a holistic view that unifies numerous existing studies and suggests potential research directions. We envision our work as a useful roadmap for future research on LLMs.

Aishan Liu, Jiakai Wang, Xianglong Liu et al.

Adversarial examples are inputs with imperceptible perturbations that easily misleading deep neural networks(DNNs). Recently, adversarial patch, with noise confined to a small and localized patch, has emerged for its easy feasibility in real-world scenarios. However, existing strategies failed to generate adversarial patches with strong generalization ability. In other words, the adversarial patches were input-specific and failed to attack images from all classes, especially unseen ones during training. To address the problem, this paper proposes a bias-based framework to generate class-agnostic universal adversarial patches with strong generalization ability, which exploits both the perceptual and semantic bias of models. Regarding the perceptual bias, since DNNs are strongly biased towards textures, we exploit the hard examples which convey strong model uncertainties and extract a textural patch prior from them by adopting the style similarities. The patch prior is more close to decision boundaries and would promote attacks. To further alleviate the heavy dependency on large amounts of data in training universal attacks, we further exploit the semantic bias. As the class-wise preference, prototypes are introduced and pursued by maximizing the multi-class margin to help universal training. Taking AutomaticCheck-out (ACO) as the typical scenario, extensive experiments including white-box and black-box settings in both digital-world(RPC, the largest ACO related dataset) and physical-world scenario(Taobao and JD, the world' s largest online shopping platforms) are conducted. Experimental results demonstrate that our proposed framework outperforms state-of-the-art adversarial patch attack methods.