Xuefeng Bai

Xuefeng Bai, Yulong Chen, Yue Zhang · cambridge

Abstract meaning representation (AMR) highlights the core semantic information of text in a graph structure. Recently, pre-trained language models (PLMs) have advanced tasks of AMR parsing and AMR-to-text generation, respectively. However, PLMs are typically pre-trained on textual data, thus are sub-optimal for modeling structural knowledge. To this end, we investigate graph self-supervised training to improve the structure awareness of PLMs over AMR graphs. In particular, we introduce two graph auto-encoding strategies for graph-to-graph pre-training and four tasks to integrate text and graph information during pre-training. We further design a unified framework to bridge the gap between pre-training and fine-tuning tasks. Experiments on both AMR parsing and AMR-to-text generation show the superiority of our model. To our knowledge, we are the first to consider pre-training on semantic graphs.

Yulong Chen, Ming Zhong, Xuefeng Bai et al. · cambridge

We propose the shared task of cross-lingual conversation summarization, \emph{ConvSumX Challenge}, opening new avenues for researchers to investigate solutions that integrate conversation summarization and machine translation. This task can be particularly useful due to the emergence of online meetings and conferences. We construct a new benchmark, covering 2 real-world scenarios and 3 language directions, including a low-resource language. We hope that \emph{ConvSumX} can motivate researches to go beyond English and break the barrier for non-English speakers to benefit from recent advances of conversation summarization.

Yulong Chen, Huajian Zhang, Yijie Zhou et al. · cambridge, tencent-ai

Most existing cross-lingual summarization (CLS) work constructs CLS corpora by simply and directly translating pre-annotated summaries from one language to another, which can contain errors from both summarization and translation processes. To address this issue, we propose ConvSumX, a cross-lingual conversation summarization benchmark, through a new annotation schema that explicitly considers source input context. ConvSumX consists of 2 sub-tasks under different real-world scenarios, with each covering 3 language directions. We conduct thorough analysis on ConvSumX and 3 widely-used manually annotated CLS corpora and empirically find that ConvSumX is more faithful towards input text. Additionally, based on the same intuition, we propose a 2-Step method, which takes both conversation and summary as input to simulate human annotation process. Experimental results show that 2-Step method surpasses strong baselines on ConvSumX under both automatic and human evaluation. Analysis shows that both source input text and summary are crucial for modeling cross-lingual summaries.

Yulong Chen, Yang Liu, Jianhao Yan et al. · cambridge, tencent-ai

The impressive performance of Large Language Models (LLMs) has consistently surpassed numerous human-designed benchmarks, presenting new challenges in assessing the shortcomings of LLMs. Designing tasks and finding LLMs' limitations are becoming increasingly important. In this paper, we investigate the question of whether an LLM can discover its own limitations from the errors it makes. To this end, we propose a Self-Challenge evaluation framework with human-in-the-loop. Starting from seed instances that GPT-4 fails to answer, we prompt GPT-4 to summarize error patterns that can be used to generate new instances and incorporate human feedback on them to refine these patterns for generating more challenging data, iteratively. We end up with 8 diverse patterns, such as text manipulation and questions with assumptions. We then build a benchmark, SC-G4, consisting of 1,835 instances generated by GPT-4 using these patterns, with human-annotated gold responses. The SC-G4 serves as a challenging benchmark that allows for a detailed assessment of LLMs' abilities. Our results show that only 44.96\% of instances in SC-G4 can be answered correctly by GPT-4. Interestingly, our pilot study indicates that these error patterns also challenge other LLMs, such as Claude-3 and Llama-3, and cannot be fully resolved through fine-tuning. Our work takes the first step to demonstrate that LLMs can autonomously identify their inherent flaws and provide insights for future dynamic and automatic evaluation.

Xuefeng Bai, Linfeng Song, Yue Zhang

Pre-trained language models have made great progress on dialogue tasks. However, these models are typically trained on surface dialogue text, thus are proven to be weak in understanding the main semantic meaning of a dialogue context. We investigate Abstract Meaning Representation (AMR) as explicit semantic knowledge for pre-training models to capture the core semantic information in dialogues during pre-training. In particular, we propose a semantic-based pre-training framework that extends the standard pre-training framework (Devlin et al., 2019) by three tasks for learning 1) core semantic units, 2) semantic relations and 3) the overall semantic representation according to AMR graphs. Experiments on the understanding of both chit-chats and task-oriented dialogues show the superiority of our model. To our knowledge, we are the first to leverage a deep semantic representation for dialogue pre-training.

Xuefeng Bai, Seng Yang, Leyang Cui et al.

Abstract Meaning Representation (AMR) parsing aims to predict an AMR graph from textual input. Recently, there has been notable growth in AMR parsing performance. However, most existing work focuses on improving the performance in the specific domain, ignoring the potential domain dependence of AMR parsing systems. To address this, we extensively evaluate five representative AMR parsers on five domains and analyze challenges to cross-domain AMR parsing. We observe that challenges to cross-domain AMR parsing mainly arise from the distribution shift of words and AMR concepts. Based on our observation, we investigate two approaches to reduce the domain distribution divergence of text and AMR features, respectively. Experimental results on two out-of-domain test sets show the superiority of our method.

Xuefeng Bai, Jialong Wu, Yulong Chen et al. · cambridge

Constituency parsing is a fundamental yet unsolved challenge in natural language processing. In this paper, we examine the potential of recent large language models (LLMs) to address this challenge. We reformat constituency parsing as a sequence-to-sequence generation problem and evaluate the performance of a diverse range of LLMs under zero-shot, few-shot, and supervised fine-tuning learning paradigms. We observe that while LLMs achieve acceptable improvements, they still encounter substantial limitations, due to the absence of mechanisms to guarantee the validity and faithfulness of the generated constituent trees. Motivated by this observation, we propose two strategies to guide LLMs to generate more accurate constituent trees by learning from erroneous samples and refining outputs in a multi-agent collaboration way, respectively. The experimental results demonstrate that our methods effectively reduce the occurrence of invalid and unfaithful trees, thereby enhancing overall parsing performance and achieving promising results across different learning paradigms.

Andong Chen, Lianzhang Lou, Kehai Chen et al.

Different from the traditional translation tasks, classical Chinese poetry translation requires both adequacy and fluency in translating culturally and historically significant content and linguistic poetic elegance. Large language models (LLMs) with impressive multilingual capabilities may bring a ray of hope to achieve this extreme translation demand. This paper first introduces a suitable benchmark (PoetMT) where each Chinese poetry has a recognized elegant translation. Meanwhile, we propose a new metric based on GPT-4 to evaluate the extent to which current LLMs can meet these demands. Our empirical evaluation reveals that the existing LLMs fall short in the challenging task. Hence, we propose a Retrieval-Augmented Machine Translation (RAT) method which incorporates knowledge related to classical poetry for advancing the translation of Chinese Poetry in LLMs. Experimental results show that RAT consistently outperforms all comparison methods regarding wildly used BLEU, COMET, BLEURT, our proposed metric, and human evaluation.

Zelin Li, Kehai Chen, Lemao Liu et al.

With the great advancements in large language models (LLMs), adversarial attacks against LLMs have recently attracted increasing attention. We found that pre-existing adversarial attack methodologies exhibit limited transferability and are notably inefficient, particularly when applied to LLMs. In this paper, we analyze the core mechanisms of previous predominant adversarial attack methods, revealing that 1) the distributions of importance score differ markedly among victim models, restricting the transferability; 2) the sequential attack processes induces substantial time overheads. Based on the above two insights, we introduce a new scheme, named TF-Attack, for Transferable and Fast adversarial attacks on LLMs. TF-Attack employs an external LLM as a third-party overseer rather than the victim model to identify critical units within sentences. Moreover, TF-Attack introduces the concept of Importance Level, which allows for parallel substitutions of attacks. We conduct extensive experiments on 6 widely adopted benchmarks, evaluating the proposed method through both automatic and human metrics. Results show that our method consistently surpasses previous methods in transferability and delivers significant speed improvements, up to 20 times faster than earlier attack strategies.

Yu Zhang, Chuyang Sun, Kehai Chen et al.

Large Vision-Language Models (LVLMs) still struggle with vision hallucination, where generated responses are inconsistent with the visual input. Existing methods either rely on large-scale annotated data for fine-tuning, which incurs massive computational overhead, or employ static post-hoc strategies that overlook the dynamic nature of hallucination emergence. To address these, we introduce a new self-rewarding framework, enabling dynamic hallucination mitigation at inference time without external supervision. On the empirical side, we reveal that visual hallucination exhibits phase-wise dynamic patterns, peaking at the onset of each semantic phase. Drawing on these insights, we propose \textbf{PSRD} (\textbf{Phase-wise \textbf{S}elf-\textbf{R}eward \textbf{D}ecoding) for online hallucination correction guided by phase-wise self-reward signals. To reduce the cost of repeated self-evaluation during decoding, we distill the hallucination guidance signal from LVLMs into a lightweight reward model. The reward model subsequently provides on-the-fly guidance for targeted intervention during the decoding process, enabling precise hallucination suppression. The proposed PSRD significantly reduces the hallucination rate of LLaVA-1.5-7B by 50.0% and consistently outperforms existing post-hoc methods across five hallucination evaluation benchmarks for four LVLMs. Further analysis confirms that PSRD effectively mitigates hallucination propagation and achieves a highly controllable trade-off between strong performance and inference efficiency.

Zirui Li, Xuefeng Bai, Kehai Chen et al.

Latent or continuous chain-of-thought methods replace explicit textual rationales with a number of internal latent steps, but these intermediate computations are difficult to evaluate beyond correlation-based probes. In this paper, we view latent chain-of-thought as a manipulable causal process in representation space by modeling latent steps as variables in a structural causal model (SCM) and analyzing their effects through step-wise $\mathrm{do}$-interventions. We study two representative paradigms (i.e., Coconut and CODI) on both mathematical and general reasoning tasks to investigate three key questions: (1) which steps are causally necessary for correctness and when answers become decidable early; (2) how does influence propagate across steps, and how does this structure compare to explicit CoT; and (3) do intermediate trajectories retain competing answer modes, and how does output-level commitment differ from representational commitment across steps. We find that latent-step budgets behave less like homogeneous extra depth and more like staged functionality with non-local routing, and we identify a persistent gap between early output bias and late representational commitment. These results motivate mode-conditional and stability-aware analyses -- and corresponding training/decoding objectives -- as more reliable tools for interpreting and improving latent reasoning systems.

Bingzheng QU, Kehai Chen, Xuefeng Bai et al.

Recent developments in video translation have further enhanced cross-lingual access to video content, with multimodal large language models (MLLMs) playing an increasingly important supporting role. With strong multimodal understanding, reasoning, and generation capabilities, MLLMs-based video translation systems are overcoming the limitations of traditional cascaded pipelines that separately handle automatic speech recognition, machine translation, text-to-speech and lip synchronization. These MLLM-powered approaches not only achieve competitive or superior translation quality, but also demonstrate stronger robustness in zero-shot settings and multi-speaker scenarios, while jointly modeling semantic fidelity, timing, speaker identity, and emotional consistency. However, despite the rapid progress of MLLMs and extensive surveys on general video-language understanding, a focused and systematic review of how MLLMs empower video translation tasks is still lacking. To fill this gap, we provide the first comprehensive overview of MLLMs-based video translation, organized around a three-role taxonomy: 1) Semantic Reasoner, which characterizes how MLLMs perform video understanding, temporal reasoning, and multimodal fusion; 2) Expressive Performer, which analyzes LLM-driven and LLM-augmented techniques for expressive, controllable speech generation; and 3) Visual Synthesizer, which examines different types of video generators for high-fidelity lip-sync and visual alignment. Finally, we discuss open challenges in video understanding, temporal modeling, and multimodal alignment, and outline promising future research directions for MLLMs-powered video translation.

Yu Zhang, Mufan Xu, Xuefeng Bai et al.

Modality following serves as the capacity of multimodal large language models (MLLMs) to selectively utilize multimodal contexts based on user instructions. It is fundamental to ensuring safety and reliability in real-world deployments. However, the underlying mechanisms governing this decision-making process remain poorly understood. In this paper, we investigate its working mechanism through an information flow lens. Our findings reveal that instruction tokens function as structural anchors for modality arbitration: Shallow attention layers perform non-selective information transfer, routing multimodal cues to these anchors as a latent buffer; Modality competition is resolved within deep attention layers guided by the instruction intent, while MLP layers exhibit semantic inertia, acting as an adversarial force. Furthermore, we identify a sparse set of specialized attention heads that drive this arbitration. Causal interventions demonstrate that manipulating a mere $5\%$ of these critical heads can decrease the modality-following ratio by $60\%$ through blocking, or increase it by $60\%$ through targeted amplification of failed samples. Our work provides a substantial step toward model transparency and offers a principled framework for the orchestration of multimodal information in MLLMs.

Yihong Tang, Kehai Chen, Xuefeng Bai et al.

Current role-playing agents (RPAs) are typically constructed by imitating surface-level behaviors, but this approach lacks internal cognitive consistency, often causing out-of-character errors in complex situations. To address this, we propose Character-R1, a framework designed to provide comprehensive verifiable reward signals for effective role-aware reasoning, which are missing in recent studies. Specifically, our framework comprises three core designs: (1) Cognitive Focus Reward, which enforces explicit label-based analysis of 10 character elements (e.g., worldview) to structure internal cognition; (2) Reference-Guided Reward, which utilizes overlap-based metrics with reference responses as optimization anchors to enhance exploration and performance; and (3) Character-Conditioned Reward Normalization, which adjusts reward distributions based on character categories to ensure robust optimization across heterogeneous roles. Extensive experiments demonstrate that Character-R1 significantly outperforms existing methods in knowledge, memory and others.

Yingjie Zhu, Xuefeng Bai, Kehai Chen et al.

Reasoning over table images remains challenging for Large Vision-Language Models (LVLMs) due to complex layouts and tightly coupled structure-content information. Existing solutions often depend on expensive supervised training, reinforcement learning, or external tools, limiting efficiency and scalability. This work addresses a key question: how to adapt LVLMs to table reasoning with minimal annotation and no external tools? Specifically, we first introduce DiSCo, a Disentangled Structure-Content alignment framework that explicitly separates structural abstraction from semantic grounding during multimodal alignment, efficiently adapting LVLMs to tables structures. Building on DiSCo, we further present Table-GLS, a Global-to-Local Structure-guided reasoning framework that performs table reasoning via structured exploration and evidence-grounded inference. Extensive experiments across diverse benchmarks demonstrate that our framework efficiently enhances LVLM's table understanding and reasoning capabilities, particularly generalizing to unseen table structures.

Jinlong Ma, Yu Zhang, Xuefeng Bai et al.

Grounded Multimodal Named Entity Recognition (GMNER) aims to extract text-based entities, assign them semantic categories, and ground them to corresponding visual regions. In this work, we explore the potential of Multimodal Large Language Models (MLLMs) to perform GMNER in an end-to-end manner, moving beyond their typical role as auxiliary tools within cascaded pipelines. Crucially, our investigation reveals a fundamental challenge: MLLMs exhibit $\textbf{modality bias}$, including visual bias and textual bias, which stems from their tendency to take unimodal shortcuts rather than rigorous cross-modal verification. To address this, we propose Modality-aware Consistency Reasoning ($\textbf{MCR}$), which enforces structured cross-modal reasoning through Multi-style Reasoning Schema Injection (MRSI) and Constraint-guided Verifiable Optimization (CVO). MRSI transforms abstract constraints into executable reasoning chains, while CVO empowers the model to dynamically align its reasoning trajectories with Group Relative Policy Optimization (GRPO). Experiments on GMNER and visual grounding tasks demonstrate that MCR effectively mitigates modality bias and achieves superior performance compared to existing baselines.

Bingzheng Qu, Kehai Chen, Xuefeng Bai et al.

Despite the remarkable progress in text-driven video editing, generating coherent non-rigid deformations remains a critical challenge, often plagued by physical distortion and temporal flicker. To bridge this gap, we propose NRVBench, the first dedicated and comprehensive benchmark designed to evaluate non-rigid video editing. First, we curate a high-quality dataset consisting of 180 non-rigid motion videos from six physics-based categories, equipped with 2,340 fine-grained task instructions and 360 multiple-choice questions. Second, we propose NRVE-Acc, a novel evaluation metric based on Vision-Language Models that can rigorously assess physical compliance, temporal consistency, and instruction alignment, overcoming the limitations of general metrics in capturing complex dynamics. Third, we introduce a training-free baseline, VM-Edit, which utilizes a dual-region denoising mechanism to achieve structure-aware control, balancing structural preservation and dynamic deformation. Extensive experiments demonstrate that while current methods have shortcomings in maintaining physical plausibility, our method achieves excellent performance across both standard and proposed metrics. We believe the benchmark could serve as a standard testing platform for advancing physics-aware video editing.

Zhenyu Li, Kehai Chen, Yunfei Long et al.

Large Language Models (LLMs) have demonstrated remarkable instruction-following capabilities across various applications. However, their performance in multilingual settings lacks systematic investigation, with existing evaluations lacking fine-grained constraint analysis across diverse linguistic contexts. We introduce XIFBench, a comprehensive constraint-based benchmark for evaluating multilingual instruction-following abilities of LLMs, comprising 558 instructions with 0-5 additional constraints across five categories (Content, Style, Situation, Format, and Numerical) in six languages spanning different resource levels. To support reliable and consistent cross-lingual evaluation, we implement three methodological innovations: cultural accessibility annotation, constraint-level translation validation, and requirement-based evaluation using English requirements as semantic anchors across languages. Extensive experiments with various LLMs not only quantify performance disparities across resource levels but also provide detailed insights into how language resources, constraint categories, instruction complexity, and cultural specificity influence multilingual instruction-following. Our code and data are available at https://github.com/zhenyuli801/XIFBench.

Mufan Xu, Kehai Chen, Xuefeng Bai et al.

Existing policy-gradient methods for auto-regressive language models typically select subsequent tokens one at a time as actions in the policy. While effective for many generation tasks, such an approach may not fully capture the structure of complex reasoning tasks, where a single semantic decision is often realized across multiple tokens--for example, when defining variables or composing equations. This introduces a potential mismatch between token-level optimization and the inherently block-level nature of reasoning in these settings. To bridge this gap, we propose Multi-token Policy Gradient Optimization (MPO), a framework that treats sequences of K consecutive tokens as unified semantic actions. This block-level perspective enables our method to capture the compositional structure of reasoning trajectories and supports optimization over coherent, higher-level objectives. Experiments on mathematical reasoning and coding benchmarks show that MPO outperforms standard token-level policy gradient baselines, highlight the limitations of token-level policy gradients for complex reasoning, motivating future research to look beyond token-level granularity for reasoning-intensive language tasks.

Yihong Tang, Kehai Chen, Xuefeng Bai et al.

The advancement of Multimodal Large Language Models (MLLMs) has expanded Role-Playing Agents (RPAs) into visually grounded environments. However, human vision is inherently subjective and identity-driven, whereas existing MLLMs extract objective, character-agnostic features for general tasks. In RPAs, this generic visual noise overpowers fragile character traits, causing Modality-Role Interference (MRI), where agents struggle to integrate visual grounding and character consistency. To address this, we introduce the training-free Character-Aware Visual Intervention (CAVI) framework, enabling agents to perceive the world through the lens of character. CAVI systematically targets MRI: macroscopically, Character-Guided Token Pruning (CTP) restricts the visual receptive field to role-relevant entities; microscopically, Orthogonal Feature Modulation (OFM) projects tokens onto a character-context subspace to extract aligned facts; and during decoding, Modality-Adaptive Role Steering (MARS) dynamically optimizes steering intensity based on visual reliance. Extensive experiments show CAVI effectively alleviates MRI, significantly enhancing character-consistent multimodal interactions.

Yuchen Song, Andong Chen, Wenxin Zhu et al.

Cultural awareness capabilities has emerged as a critical capability for Multimodal Large Language Models (MLLMs). However, current benchmarks lack progressed difficulty in their task design and are deficient in cross-lingual tasks. Moreover, current benchmarks often use real-world images. Each real-world image typically contains one culture, making these benchmarks relatively easy for MLLMs. Based on this, we propose C$^3$B ($\textbf{C}$omics $\textbf{C}$ross-$\textbf{C}$ultural $\textbf{B}$enchmark), a novel multicultural, multitask and multilingual cultural awareness capabilities benchmark. C$^3$B comprises over 2000 images and over 18000 QA pairs, constructed on three tasks with progressed difficulties, from basic visual recognition to higher-level cultural conflict understanding, and finally to cultural content generation. We conducted evaluations on 11 open-source MLLMs, revealing a significant performance gap between MLLMs and human performance. The gap demonstrates that C$^3$B poses substantial challenges for current MLLMs, encouraging future research to advance the cultural awareness capabilities of MLLMs.

Yuchen Wang, Xuefeng Bai, Xiucheng Li et al.

Adapting vision-language models (VLMs) to downstream tasks with pseudolabels has gained increasing attention. A major obstacle is that the pseudolabels generated by VLMs tend to be imbalanced, leading to inferior performance. While existing methods have explored various strategies to address this, the underlying causes of imbalance remain insufficiently investigated. To fill this gap, we delve into imbalanced pseudolabels and identify two primary contributing factors: concept mismatch and concept confusion. To mitigate these two issues, we propose a novel framework incorporating concept alignment and confusion-aware calibrated margin mechanisms. The core of our approach lies in enhancing underperforming classes and promoting balanced predictions across categories, thus mitigating imbalance. Extensive experiments on six benchmark datasets with three learning paradigms demonstrate that the proposed method effectively enhances the accuracy and balance of pseudolabels, achieving a relative improvement of 6.29% over the SoTA method. Our code is avaliable at https://anonymous.4open.science/r/CAP-C642/

Weiyang Huang, Xuefeng Bai, Kehai Chen et al.

Large Reasoning Models (LRMs) have revolutionized complex problem-solving, yet they exhibit a pervasive "overthinking", generating unnecessarily long reasoning chains. While current solutions improve token efficiency, they often sacrifice fine-grained control or risk disrupting the logical integrity of the reasoning process. To address this, we introduce Stepwise Adaptive Thinking (SAT), a framework that performs step-level, difficulty-aware pruning while preserving the core reasoning structure. SAT formulates reasoning as a Finite-State Machine (FSM) with distinct thinking modes (Slow, Normal, Fast, Skip). It navigates these states dynamically using a lightweight Process Reward Model (PRM), compressing easy steps while preserving depth for hard ones. Experiments across 9 LRMs and 7 benchmarks show that SAT achieves up to 40% reduction in reasoning tokens while generally maintaining or improving accuracy.

Yu Zhang, Jinlong Ma, Yongshuai Hou et al.

Multimodal large language models (MLLMs) have achieved remarkable performance on complex tasks with multimodal context. However, it is still understudied whether they exhibit modality preference when processing multimodal contexts. To study this question, we first build a \textbf{MC\textsuperscript{2}} benchmark under controlled evidence conflict scenarios to systematically evaluate modality preference, which is the tendency to favor one modality over another when making decisions based on multimodal conflicting evidence. Our extensive evaluation reveals that all 18 tested MLLMs generally demonstrate clear modality bias, and modality preference can be influenced by external interventions. An in-depth analysis reveals that the preference direction can be captured within the latent representations of MLLMs. Built on this, we propose a probing and steering method based on representation engineering to explicitly control modality preference without additional fine-tuning or carefully crafted prompts. Our method effectively amplifies modality preference toward a desired direction and applies to downstream tasks such as hallucination mitigation and multimodal machine translation, yielding promising improvements.

Yifan Duan, Yihong Tang, Xuefeng Bai et al.

Large language models (LLMs) excel in both closed tasks (including problem-solving, and code generation) and open tasks (including creative writing), yet existing explanations for their capabilities lack connections to real-world human intelligence. To fill this gap, this paper systematically investigates LLM intelligence through the lens of ``human simulation'', addressing three core questions: (1) \textit{How do personality traits affect problem-solving in closed tasks?} (2) \textit{How do traits shape creativity in open tasks?} (3) \textit{How does single-agent performance influence multi-agent collaboration?} By assigning Big Five personality traits to LLM agents and evaluating their performance in single- and multi-agent settings, we reveal that specific traits significantly influence reasoning accuracy (closed tasks) and creative output (open tasks). Furthermore, multi-agent systems exhibit collective intelligence distinct from individual capabilities, driven by distinguishing combinations of personalities.

Andong Chen, Kehai Chen, Yang Xiang et al.

The remarkable understanding and generation capabilities of large language models (LLMs) have greatly improved translation performance. However, incorrect understanding of the sentence to be translated can degrade translation quality. To address this issue, we proposed a novel Iterative Bilingual Understanding Translation (IBUT) method based on the cross-lingual capabilities of LLMs and the dual characteristics of translation tasks. The cross-lingual capability of LLMs enables the generation of contextual understanding for both the source and target languages separately. Furthermore, the dual characteristics allow IBUT to generate effective cross-lingual feedback, iteratively refining contextual understanding, thereby reducing errors and improving translation performance. Experimental results showed that the proposed IBUT outperforms several strong comparison methods, especially being generalized to multiple domains (e.g., news, commonsense, and cultural translation benchmarks).

Mufan Xu, Kehai Chen, Xuefeng Bai et al.

Large language models (LLMs) based on generative pre-trained Transformer have achieved remarkable performance on knowledge graph question-answering (KGQA) tasks. However, LLMs often produce ungrounded subgraph planning or reasoning results in KGQA due to the hallucinatory behavior brought by the generative paradigm. To tackle this issue, we propose READS to reformulate the KGQA process into discriminative subtasks, which simplifies the search space for each subtasks. Based on the subtasks, we design a new corresponding discriminative inference strategy to conduct the reasoning for KGQA, thereby alleviating hallucination and ungrounded reasoning issues in LLMs. Experimental results show that the proposed approach outperforms multiple strong comparison methods, along with achieving state-of-the-art performance on widely used benchmarks WebQSP and CWQ.

Yingjie Zhu, Xuefeng Bai, Kehai Chen et al.

Large Vision-Language Models (LVLMs) have demonstrated remarkable performance across diverse tasks. Despite great success, recent studies show that LVLMs encounter substantial limitations when engaging with visual graphs. To study the reason behind these limitations, we propose VGCure, a comprehensive benchmark covering 22 tasks for examining the fundamental graph understanding and reasoning capacities of LVLMs. Extensive evaluations conducted on 14 LVLMs reveal that LVLMs are weak in basic graph understanding and reasoning tasks, particularly those concerning relational or structurally complex information. Based on this observation, we propose a structure-aware fine-tuning framework to enhance LVLMs with structure learning abilities through three self-supervised learning tasks. Experiments validate the effectiveness of our method in improving LVLMs' performance on fundamental and downstream graph learning tasks, as well as enhancing their robustness against complex visual graphs.

Yihong Tang, Kehai Chen, Xuefeng Bai et al.

Large Language Models (LLMs) have made remarkable advances in role-playing dialogue agents, demonstrating their utility in character simulations. However, it remains challenging for these agents to balance character portrayal utility with content safety because this essential character simulation often comes with the risk of generating unsafe content. To address this issue, we first conduct a systematic exploration of the safety-utility trade-off across multiple LLMs. Our analysis reveals that risk scenarios created by villain characters and user queries (referred to as risk coupling) contribute to this trade-off. Building on this, we propose a novel Adaptive Dynamic Multi-Preference (ADMP) method, which dynamically adjusts safety-utility preferences based on the degree of risk coupling and guides the model to generate responses biased toward utility or safety. We further introduce Coupling Margin Sampling (CMS) into coupling detection to enhance the model's ability to handle high-risk scenarios. Experimental results demonstrate that our approach improves safety metrics while maintaining utility.

Xufei Lv, Kehai Chen, Haoyuan Sun et al.

Alignment of large language models (LLMs) with human values has recently garnered significant attention, with prominent examples including the canonical yet costly Reinforcement Learning from Human Feedback (RLHF) and the simple Direct Preference Optimization (DPO). In this work, we demonstrate that both RLHF and DPO can be interpreted from the perspective of mutual information (MI) maximization, uncovering a profound connection to contrastive learning. Within this framework, both RLHF and DPO can be interpreted as methods that performing contrastive learning based on the positive and negative samples derived from base model, leveraging the Donsker-Varadhan (DV) lower bound on MI (equivalently, the MINE estimator). Such paradigm further illuminates why RLHF may not intrinsically incentivize reasoning capacities in LLMs beyond what is already present in the base model. Building on the perspective, we replace the DV/MINE bound with the Jensen-Shannon (JS) MI estimator and propose the Mutual Information Optimization (MIO). Comprehensive theoretical analysis and extensive empirical evaluations demonstrate that MIO mitigates the late-stage decline in chosen-likelihood observed in DPO, achieving competitive or superior performance across various challenging reasoning and mathematical benchmarks.

Hongbin Zhang, Kehai Chen, Xuefeng Bai et al.

Large language models (LLMs) have succeeded remarkably in multilingual translation tasks. However, the inherent translation mechanisms of LLMs remain poorly understood, largely due to sophisticated architectures and vast parameter scales. In response to this issue, this study explores the translation mechanism of LLM from the perspective of computational components (e.g., attention heads and MLPs). Path patching is utilized to explore causal relationships between components, detecting those crucial for translation tasks and subsequently analyzing their behavioral patterns in human-interpretable terms. Comprehensive analysis reveals that translation is predominantly facilitated by a sparse subset of specialized attention heads (less than 5\%), which extract source language, indicator, and positional features. MLPs subsequently integrate and process these features by transiting towards English-centric latent representations. Notably, building on the above findings, targeted fine-tuning of only 64 heads achieves translation improvement comparable to full-parameter tuning while preserving general capabilities.

Henglyu Liu, Andong Chen, Kehai Chen et al.

Recent advancement of large language models (LLMs) has led to significant breakthroughs across various tasks, laying the foundation for the development of LLM-based speech translation systems. Existing methods primarily focus on aligning inputs and outputs across modalities while overlooking deeper semantic alignment within model representations. To address this limitation, we propose an Adaptive Inner Speech-Text Alignment (AI-STA) method to bridge the modality gap by explicitly aligning speech and text representations at selected layers within LLMs. To achieve this, we leverage the optimal transport (OT) theory to quantify fine-grained representation discrepancies between speech and text. Furthermore, we utilize the cross-modal retrieval technique to identify the layers that are best suited for alignment and perform joint training on these layers. Experimental results on speech translation (ST) tasks demonstrate that AI-STA significantly improves the translation performance of large speech-text models (LSMs), outperforming previous state-of-the-art approaches. Our findings highlight the importance of inner-layer speech-text alignment in LLMs and provide new insights into enhancing cross-modal learning.

Qiyuan Deng, Xuefeng Bai, Kehai Chen et al.

Reinforcement Learning (RL) algorithms for safety alignment of Large Language Models (LLMs), such as Direct Preference Optimization (DPO), encounter the challenge of distribution shift. Current approaches typically address this issue through online sampling from the target policy, which requires significant computational resources. In this paper, we hypothesize that during off-policy training, while the ranking order of output generated by policy changes, their overall distribution remains relatively stable. This stability allows the conversion of the sampling process from the target policy into a computationally efficient re-ranking of preference data. Building on this hypothesis, we propose a new framework that leverages the model's intrinsic safety judgment capability to extract reward signals, which are then used to calculate label confidence for preference reordering. Extensive experiments and theoretical analysis demonstrate that the proposed method effectively addresses the distribution shift issue, remarkably enhancing the safety performance while avoiding about 300x computational overheads.

Xingzuo Li, Kehai Chen, Yunfei Long et al.

Large language model (LLM) agents typically adopt a step-by-step reasoning framework, in which they interleave the processes of thinking and acting to accomplish the given task. However, this paradigm faces a deep-rooted one-pass issue whereby each generated intermediate thought is plugged into the trajectory regardless of its correctness, which can cause irreversible error propagation. To address the issue, this paper proposes a novel framework called Generator-Assistant Stepwise Rollback (GA-Rollback) to induce better decision-making for LLM agents. Particularly, GA-Rollback utilizes a generator to interact with the environment and an assistant to examine each action produced by the generator, where the assistant triggers a rollback operation upon detection of incorrect actions. Moreover, we introduce two additional strategies tailored for the rollback scenario to further improve its effectiveness. Extensive experiments show that GA-Rollback achieves significant improvements over several strong baselines on three widely used benchmarks. Our analysis further reveals that GA-Rollback can function as a robust plug-and-play module, integrating seamlessly with other methods.

Hongbin Zhang, Kehai Chen, Xuefeng Bai et al.

Large language models (LLMs) have exhibited impressive multilingual reasoning capabilities, driven by extensive multilingual pre-training corpora and instruction fine-tuning data. However, a performance gap exists between high- and low-resource language reasoning tasks due to the language imbalance in the pre-training corpus, which is exacerbated by evaluation bias in existing reasoning benchmarks lacking low-resource language coverage. To alleviate this issue, we propose LinguaLIFT, a two-stage instruction tuning framework for advancing low-resource language reasoning. LinguaLIFT employs a language alignment layer to capture multilingual alignment in a code-switched tuning way without requiring multilingual instruction or parallel data, thereby transferring the cross-lingual reasoning capabilities to low-resource languages through English-only instruction tuning data. To comprehensively evaluate the multilingual reasoning capabilities, we introduce the Multilingual Math World Problem (MMWP) benchmark, which spans 21 low-resource, 17 medium-resource, and 10 high-resource languages. Experimental results show that LinguaLIFT outperforms several competitive baselines across MMWP and four widely used benchmarks.

Jinchao Hu, Meizhi Zhong, Kehai Chen et al.

Large language models are increasingly being deployed as autonomous agents yet their real world effectiveness depends on reliable tools for information retrieval, computation and external action. Existing studies remain fragmented across tasks, tool types, and training settings, lacking a unified view of how tool-use methods differ and evolve. This paper organizes the literature into three paradigms: prompting as plug-and-play, supervised tool learning and reward-driven tool policy learning, analyzes their methods, strengths and failure modes, reviews the evaluation landscape and highlights key challenges, aiming to address this fragmentation and provide a more structured evolutionary view of agentic tool use.

Yingjie Zhu, Xuefeng Bai, Kehai Chen et al.

Large Vision-Language Models (LVLMs) have achieved remarkable success across a wide range of multimodal tasks, yet their robustness to spatial variations remains insufficiently understood. In this work, we present a systematic study of the spatial bias of LVLMs, focusing on how models respond when identical key visual information is placed at different locations within an image. Through a carefully designed probing dataset, we demonstrate that current LVLMs often produce inconsistent outputs under such spatial shifts, revealing a fundamental limitation in their spatial-semantic understanding. Further analysis shows that this phenomenon originates not from the vision encoder, which reliably perceives and interprets visual content across positions, but from the unbalanced design of position embeddings in the language model component. In particular, the widely adopted position embedding strategies, such as RoPE, introduce imbalance during cross-modal interaction, leading image tokens at different positions to exert unequal influence on semantic understanding. To mitigate this issue, we introduce Balanced Position Assignment (BaPA), a simple yet effective mechanism that assigns identical position embeddings to all image tokens, promoting a more balanced integration of visual information. Extensive experiments show that BaPA enhances the spatial robustness of LVLMs without retraining and further boosts their performance across diverse multimodal benchmarks when combined with lightweight fine-tuning. Further analysis of information flow reveals that BaPA yields balanced attention, enabling more holistic visual understanding.

Hongbin Zhang, Kehai Chen, Xuefeng Bai et al.

Reward models (RMs) are crucial for aligning large language models (LLMs) with diverse cultures. Consequently, evaluating their cultural awareness is essential for further advancing global alignment of LLMs. However, existing RM evaluations fall short in assessing cultural awareness due to the scarcity of culturally relevant evaluation datasets. To fill this gap, we propose Cultural Awareness Reward modeling Benchmark (CARB), covering 10 distinct cultures across 4 cultural domains. Our extensive evaluation of state-of-the-art RMs reveals their deficiencies in modeling cultural awareness and demonstrates a positive correlation between performance on CARB and downstream multilingual cultural alignment tasks. Further analysis identifies the spurious correlations within culture-aware reward modeling, wherein RM's scoring relies predominantly on surface-level features rather than authentic cultural nuance understanding. To address these, we propose Think-as-Locals to elicit deeper culturally grounded reasoning from generative RMs via reinforcement learning from verifiable rewards (RLVR) and employ well-designed rewards to ensure accurate preference judgments and high-quality structured evaluation criteria generation. Experimental results validate its efficacy in mitigating spurious features interference and advancing culture-aware reward modeling.

Xuchen Wei, Yangxin Wu, Yaoyin Zhang et al.

This paper presents HITSZ's submission for the IWSLT 2025 Indic track, focusing on speech-to-text translation (ST) for English-to-Indic and Indic-to-English language pairs. To enhance translation quality in this low-resource scenario, we propose an end-to-end system integrating the pre-trained Whisper automated speech recognition (ASR) model with Krutrim, an Indic-specialized large language model (LLM). Experimental results demonstrate that our end-to-end system achieved average BLEU scores of $28.88$ for English-to-Indic directions and $27.86$ for Indic-to-English directions. Furthermore, we investigated the Chain-of-Thought (CoT) method. While this method showed potential for significant translation quality improvements on successfully parsed outputs (e.g. a $13.84$ BLEU increase for Tamil-to-English), we observed challenges in ensuring the model consistently adheres to the required CoT output format.

Zuhong Lin, Daoyuan Ren, Kai Ran et al.

Accurately identifying the synthesis conditions of metal-organic frameworks (MOFs) is essential for guiding experimental design, yet remains challenging because relevant information in the literature is often scattered, inconsistent, and difficult to interpret. We present MOFh6, a large language model driven system that reads raw articles or crystal codes and converts them into standardized synthesis tables. It links related descriptions across paragraphs, unifies ligand abbreviations with full names, and outputs structured parameters ready for use. MOFh6 achieved 99% extraction accuracy, resolved 94.1% of abbreviation cases across five major publishers, and maintained a precision of 0.93 +/- 0.01. Processing a full text takes 9.6 s, locating synthesis descriptions 36 s, with 100 papers processed for USD 4.24. By replacing static database lookups with real-time extraction, MOFh6 reshapes MOF synthesis research, accelerating the conversion of literature knowledge into practical synthesis protocols and enabling scalable, data-driven materials discovery.

Zhengsheng Guo, Linwei Zheng, Xinyang Chen et al.

While human cognition inherently retrieves information from diverse and specialized knowledge sources during decision-making processes, current Retrieval-Augmented Generation (RAG) systems typically operate through single-source knowledge retrieval, leading to a cognitive-algorithmic discrepancy. To bridge this gap, we introduce MoK-RAG, a novel multi-source RAG framework that implements a mixture of knowledge paths enhanced retrieval mechanism through functional partitioning of a large language model (LLM) corpus into distinct sections, enabling retrieval from multiple specialized knowledge paths. Applied to the generation of 3D simulated environments, our proposed MoK-RAG3D enhances this paradigm by partitioning 3D assets into distinct sections and organizing them based on a hierarchical knowledge tree structure. Different from previous methods that only use manual evaluation, we pioneered the introduction of automated evaluation methods for 3D scenes. Both automatic and human evaluations in our experiments demonstrate that MoK-RAG3D can assist Embodied AI agents in generating diverse scenes.

Zhiyu Yin, Kehai Chen, Xuefeng Bai et al.

Video generation, by leveraging a dynamic visual generation method, pushes the boundaries of Artificial Intelligence Generated Content (AIGC). Video generation presents unique challenges beyond static image generation, requiring both high-quality individual frames and temporal coherence to maintain consistency across the spatiotemporal sequence. Recent works have aimed at addressing the spatiotemporal consistency issue in video generation, while few literature review has been organized from this perspective. This gap hinders a deeper understanding of the underlying mechanisms for high-quality video generation. In this survey, we systematically review the recent advances in video generation, covering five key aspects: foundation models, information representations, generation schemes, post-processing techniques, and evaluation metrics. We particularly focus on their contributions to maintaining spatiotemporal consistency. Finally, we discuss the future directions and challenges in this field, hoping to inspire further efforts to advance the development of video generation.

Ruili Jiang, Kehai Chen, Xuefeng Bai et al.

The recent surge of versatile large language models (LLMs) largely depends on aligning increasingly capable foundation models with human intentions by preference learning, enhancing LLMs with excellent applicability and effectiveness in a wide range of contexts. Despite the numerous related studies conducted, a perspective on how human preferences are introduced into LLMs remains limited, which may prevent a deeper comprehension of the relationships between human preferences and LLMs as well as the realization of their limitations. In this survey, we review the progress in exploring human preference learning for LLMs from a preference-centered perspective, covering the sources and formats of preference feedback, the modeling and usage of preference signals, as well as the evaluation of the aligned LLMs. We first categorize the human feedback according to data sources and formats. We then summarize techniques for human preferences modeling and compare the advantages and disadvantages of different schools of models. Moreover, we present various preference usage methods sorted by the objectives to utilize human preference signals. Finally, we summarize some prevailing approaches to evaluate LLMs in terms of alignment with human intentions and discuss our outlooks on the human intention alignment for LLMs.

Andong Chen, Lianzhang Lou, Kehai Chen et al.

Recently, large language models (LLMs) enhanced by self-reflection have achieved promising performance on machine translation. The key idea is guiding LLMs to generate translation with human-like feedback. However, existing self-reflection methods lack effective feedback information, limiting the translation performance. To address this, we introduce a DUAL-REFLECT framework, leveraging the dual learning of translation tasks to provide effective feedback, thereby enhancing the models' self-reflective abilities and improving translation performance. The application of this method across various translation tasks has proven its effectiveness in improving translation accuracy and eliminating ambiguities, especially in translation tasks with low-resource language pairs.

Hongbin Zhang, Kehai Chen, Xuefeng Bai et al.

Large language models (LLMs) have showcased impressive multilingual machine translation ability. However, unlike encoder-decoder style models, decoder-only LLMs lack an explicit alignment between source and target contexts. Analyzing contribution scores during generation processes revealed that LLMs can be biased towards previously generated tokens over corresponding source tokens, leading to unfaithful translations. To address this issue, we propose to encourage LLMs to pay more attention to the source context from both source and target perspectives in zeroshot prompting: 1) adjust source context attention weights; 2) suppress irrelevant target prefix influence; Additionally, we propose 3) avoiding over-reliance on the target prefix in instruction tuning. Experimental results from both human-collected unfaithfulness test sets focusing on LLM-generated unfaithful translations and general test sets, verify our methods' effectiveness across multiple language pairs. Further human evaluation shows our method's efficacy in reducing hallucinatory translations and facilitating faithful translation generation.

Cunxiang Wang, Zhikun Xu, Qipeng Guo et al.

The Open-Domain Question Answering (ODQA) task involves retrieving and subsequently generating answers from fine-grained relevant passages within a database. Current systems leverage Pretrained Language Models (PLMs) to model the relationship between questions and passages. However, the diversity in surface form expressions can hinder the model's ability to capture accurate correlations, especially within complex contexts. Therefore, we utilize Abstract Meaning Representation (AMR) graphs to assist the model in understanding complex semantic information. We introduce a method known as Graph-as-Token (GST) to incorporate AMRs into PLMs. Results from Natural Questions (NQ) and TriviaQA (TQ) demonstrate that our GST method can significantly improve performance, resulting in up to 2.44/3.17 Exact Match score improvements on NQ/TQ respectively. Furthermore, our method enhances robustness and outperforms alternative Graph Neural Network (GNN) methods for integrating AMRs. To the best of our knowledge, we are the first to employ semantic graphs in ODQA.

Yun Luo, Zhen Yang, Xuefeng Bai et al.

Representation forgetting refers to the drift of contextualized representations during continual training. Intuitively, the representation forgetting can influence the general knowledge stored in pre-trained language models (LMs), but the concrete effect is still unclear. In this paper, we study the effect of representation forgetting on the generality of pre-trained language models, i.e. the potential capability for tackling future downstream tasks. Specifically, we design three metrics, including overall generality destruction (GD), syntactic knowledge forgetting (SynF), and semantic knowledge forgetting (SemF), to measure the evolution of general knowledge in continual learning. With extensive experiments, we find that the generality is destructed in various pre-trained LMs, and syntactic and semantic knowledge is forgotten through continual learning. Based on our experiments and analysis, we further get two insights into alleviating general knowledge forgetting: 1) training on general linguistic tasks at first can mitigate general knowledge forgetting; 2) the hybrid continual learning method can mitigate the generality destruction and maintain more general knowledge compared with those only considering rehearsal or regularization.

Xuefeng Bai, Yulong Chen, Linfeng Song et al.

Although neural models have achieved competitive results in dialogue systems, they have shown limited ability in representing core semantics, such as ignoring important entities. To this end, we exploit Abstract Meaning Representation (AMR) to help dialogue modeling. Compared with the textual input, AMR explicitly provides core semantic knowledge and reduces data sparsity. We develop an algorithm to construct dialogue-level AMR graphs from sentence-level AMRs and explore two ways to incorporate AMRs into dialogue systems. Experimental results on both dialogue understanding and response generation tasks show the superiority of our model. To our knowledge, we are the first to leverage a formal semantic representation into neural dialogue modeling.

Xuefeng Bai, Linfeng Song, Yue Zhang

AMR-to-text generation aims to recover a text containing the same meaning as an input AMR graph. Current research develops increasingly powerful graph encoders to better represent AMR graphs, with decoders based on standard language modeling being used to generate outputs. We propose a decoder that back predicts projected AMR graphs on the target sentence during text generation. As the result, our outputs can better preserve the input meaning than standard decoders. Experiments on two AMR benchmarks show the superiority of our model over the previous state-of-the-art system based on graph Transformer.

Xuefeng Bai, Pengbo Liu, Yue Zhang

Targeted sentiment classification predicts the sentiment polarity on given target mentions in input texts. Dominant methods employ neural networks for encoding the input sentence and extracting relations between target mentions and their contexts. Recently, graph neural network has been investigated for integrating dependency syntax for the task, achieving the state-of-the-art results. However, existing methods do not consider dependency label information, which can be intuitively useful. To solve the problem, we investigate a novel relational graph attention network that integrates typed syntactic dependency information. Results on standard benchmarks show that our method can effectively leverage label information for improving targeted sentiment classification performances. Our final model significantly outperforms state-of-the-art syntax-based approaches.