Fei Cheng

Jiahao Huang, Fei Cheng, Junfeng Jiang et al.

Self-evolving agents improve over time by reflecting on past failures, but existing evaluation is limited in two ways: it measures only task scores, leaving reflection quality unknown, and it relies on agents' own episode runs, offering no mechanism to target specific failure patterns. We present \textbf{BenchTrace}, a benchmark for evaluating self-evolution ability in LLM agents. BenchTrace is built on a snapshot-reflection dataset of 1,821 annotated episodes spanning six diverse tasks, and comprises a \textbf{Reflection Evaluation} that probes failure identification through targeted QA tasks, and an \textbf{Evolution Evaluation} that tests whether past failure experience translates into avoidance behavior in a controlled self-evolution simulation. Building on BenchTrace, we propose \textbf{failure avoidance rate (FAR)}, a new evaluation metric measuring the fraction of test cases in which the agent successfully avoids the target failure instance. Experiments with Qwen3-32B and GPT-4.1 reveal that both models fall below a 30\% end-to-end pass rate on reflection evaluation, with diagnosis as the primary bottleneck. Evolution evaluation shows that self-evolution methods generally improve FAR over the non-evolving baseline, but agents forget early lessons as noise episodes accumulate, and agents fail to generalize their reflections beyond the specific context, causing negative transfer across task contexts. Our correlation analysis further reveals that only a fully correct reflection is strongly associated with higher FAR. BenchTrace exposes concrete limits of current self-evolution approaches and provides a controlled, model-agnostic framework for targeted evaluation.

Jiahao Huang, Fei Cheng, Junfeng Jiang et al.

Reasoning distillation transfers complex reasoning abilities from large language models (LLMs) to smaller ones, yet its success depends on how well the training data align with the student model. This paper introduces the Data-Model Compatibility (DMC) metric, which can be used to assess the suitability of a dataset for reasoning distillation on a student model. DMC provides an assessment by jointly considering data quality, relative difficulty, and student capability. We validated the effectiveness of DMC from two perspectives: (1) DMC exhibits a strong correlation with reasoning distillation performance; and (2) using DMC as the criterion for data selection leads to improved reasoning distillation performance. Both findings are consistently demonstrated across multiple student models and tasks. Moreover, since the DMC of each dataset dynamically changes during training, our experiments demonstrate that dynamically selecting datasets based on DMC can further enhance performance.

LLM-jp, Akiko Aizawa, Eiji Aramaki et al.

This paper introduces LLM-jp, a cross-organizational project for the research and development of Japanese large language models (LLMs). LLM-jp aims to develop open-source and strong Japanese LLMs, and as of this writing, more than 1,500 participants from academia and industry are working together for this purpose. This paper presents the background of the establishment of LLM-jp, summaries of its activities, and technical reports on the LLMs developed by LLM-jp. For the latest activities, visit https://llm-jp.nii.ac.jp/en/.

Zhen Wan, Fei Cheng, Qianying Liu et al.

Contrastive pre-training on distant supervision has shown remarkable effectiveness in improving supervised relation extraction tasks. However, the existing methods ignore the intrinsic noise of distant supervision during the pre-training stage. In this paper, we propose a weighted contrastive learning method by leveraging the supervised data to estimate the reliability of pre-training instances and explicitly reduce the effect of noise. Experimental results on three supervised datasets demonstrate the advantages of our proposed weighted contrastive learning approach compared to two state-of-the-art non-weighted baselines.Our code and models are available at: https://github.com/YukinoWan/WCL

Yibin Shen, Qianying Liu, Zhuoyuan Mao et al.

Solving math word problems is the task that analyses the relation of quantities and requires an accurate understanding of contextual natural language information. Recent studies show that current models rely on shallow heuristics to predict solutions and could be easily misled by small textual perturbations. To address this problem, we propose a Textual Enhanced Contrastive Learning framework, which enforces the models to distinguish semantically similar examples while holding different mathematical logic. We adopt a self-supervised manner strategy to enrich examples with subtle textual variance by textual reordering or problem re-construction. We then retrieve the hardest to differentiate samples from both equation and textual perspectives and guide the model to learn their representations. Experimental results show that our method achieves state-of-the-art on both widely used benchmark datasets and also exquisitely designed challenge datasets in English and Chinese. \footnote{Our code and data is available at \url{https://github.com/yiyunya/Textual_CL_MWP}

Zhen Wan, Qianying Liu, Zhuoyuan Mao et al.

Relation extraction (RE) has achieved remarkable progress with the help of pre-trained language models. However, existing RE models are usually incapable of handling two situations: implicit expressions and long-tail relation types, caused by language complexity and data sparsity. In this paper, we introduce a simple enhancement of RE using $k$ nearest neighbors ($k$NN-RE). $k$NN-RE allows the model to consult training relations at test time through a nearest-neighbor search and provides a simple yet effective means to tackle the two issues above. Additionally, we observe that $k$NN-RE serves as an effective way to leverage distant supervision (DS) data for RE. Experimental results show that the proposed $k$NN-RE achieves state-of-the-art performances on a variety of supervised RE datasets, i.e., ACE05, SciERC, and Wiki80, along with outperforming the best model to date on the i2b2 and Wiki80 datasets in the setting of allowing using DS. Our code and models are available at: https://github.com/YukinoWan/kNN-RE.

Chuang Ma, Qianying Liu, Tomoyuki Obuchi et al.

Multimodal large language models (MLLMs) remain unreliable on spatial multiple-choice questions, and their failures are often attributed to poorly attended visual information. In this work, we identify a complementary failure mode, spatial lexical bias: adding a spatial relation word to the answer options can attract the model's decision and make the newly added option likely to be selected. Using nine open-weight MLLMs, we show that this phenomenon is widely observed. In particular, models can answer a binary spatial question correctly, yet consistently select an incorrect third spatial option once it is added to the answer set. We isolate such binary-stable but ternary-fragile cases as diagnostic examples and leverage mechanistic interpretability tools, revealing that a substantial part of the failure instead originates on the language side rather than the visual side: visual attention analyses and residual-stream probes show the correct spatial relation remains internally available on these failures, while irrelevant-option controls, activation patching, and sparse component interventions trace the bias to specific LLM-side channels and neurons. Based on this finding, we show that a lightweight LLM-only DPO update on tiny single-object-pair synthetic data mitigates the bias, lifting four-way robust accuracy by up to 100 points on synthetic data, and by 68.0, 32.6, and 20.1 points on broader evaluation datasets WhatsUp, SpatialMQA-Direct, and VSR.

Xiaofei Sun, Linfeng Dong, Xiaoya Li et al.

Despite the success of ChatGPT, its performances on most NLP tasks are still well below the supervised baselines. In this work, we looked into the causes, and discovered that its subpar performance was caused by the following factors: (1) token limit in the prompt does not allow for the full utilization of the supervised datasets; (2) mismatch between the generation nature of ChatGPT and NLP tasks; (3) intrinsic pitfalls of LLMs models, e.g., hallucination, overly focus on certain keywords, etc. In this work, we propose a collection of general modules to address these issues, in an attempt to push the limits of ChatGPT on NLP tasks. Our proposed modules include (1) a one-input-multiple-prompts strategy that employs multiple prompts for one input to accommodate more demonstrations; (2) using fine-tuned models for better demonstration retrieval; (3) transforming tasks to formats that are more tailored to the generation nature; (4) employing reasoning strategies that are tailored to addressing the task-specific complexity; (5) the self-verification strategy to address the hallucination issue of LLMs; (6) the paraphrase strategy to improve the robustness of model predictions. We conduct experiments on 21 datasets of 10 representative NLP tasks, including question answering, commonsense reasoning, natural language inference, sentiment analysis, named entity recognition, entity-relation extraction, event extraction, dependency parsing, semantic role labeling, and part-of-speech tagging. Using the proposed assemble of techniques, we are able to significantly boost the performance of ChatGPT on the selected NLP tasks, achieving performances comparable to or better than supervised baselines, or even existing SOTA performances.

Chengzhi Zhong, Fei Cheng, Qianying Liu et al.

In this study, we investigate whether non-English-centric LLMs, despite their strong performance, `think' in their respective dominant language: more precisely, `think' refers to how the representations of intermediate layers, when un-embedded into the vocabulary space, exhibit higher probabilities for certain dominant languages during generation. We term such languages as internal $\textbf{latent languages}$. We examine the latent language of three typical categories of models for Japanese processing: Llama2, an English-centric model; Swallow, an English-centric model with continued pre-training in Japanese; and LLM-jp, a model pre-trained on balanced English and Japanese corpora. Our empirical findings reveal that, unlike Llama2 which relies exclusively on English as the internal latent language, Japanese-specific Swallow and LLM-jp employ both Japanese and English, exhibiting dual internal latent languages. For any given target language, the model preferentially activates the latent language most closely related to it. In addition, we explore how intermediate layers respond to questions involving cultural conflicts between latent internal and target output languages. We further explore how the language identity shifts across layers while keeping consistent semantic meaning reflected in the intermediate layer representations. This study deepens the understanding of non-English-centric large language models, highlighting the intricate dynamics of language representation within their intermediate layers.

Shanshan Liu, Noriki Nishida, Fei Cheng et al.

Generalization to unseen concepts is a central challenge due to the scarcity of human annotations in Mention-agnostic Biomedical Concept Recognition (MA-BCR). This work makes two key contributions to systematically address this issue. First, we propose an evaluation framework built on hierarchical concept indices and novel metrics to measure generalization. Second, we explore LLM-based Auto-Labeled Data (ALD) as a scalable resource, creating a task-specific pipeline for its generation. Our research unequivocally shows that while LLM-generated ALD cannot fully substitute for manual annotations, it is a valuable resource for improving generalization, successfully providing models with the broader coverage and structural knowledge needed to approach recognizing unseen concepts. Code and datasets are available at https://github.com/bio-ie-tool/hi-ald.

Fei Cheng, Masayuki Asahara, Ichiro Kobayashi et al.

Temporal relation classification is a pair-wise task for identifying the relation of a temporal link (TLINK) between two mentions, i.e. event, time, and document creation time (DCT). It leads to two crucial limits: 1) Two TLINKs involving a common mention do not share information. 2) Existing models with independent classifiers for each TLINK category (E2E, E2T, and E2D) hinder from using the whole data. This paper presents an event centric model that allows to manage dynamic event representations across multiple TLINKs. Our model deals with three TLINK categories with multi-task learning to leverage the full size of data. The experimental results show that our proposal outperforms state-of-the-art models and two transfer learning baselines on both the English and Japanese data.

Zhen wan, Yating Zhang, Yexiang Wang et al.

While large language models (LLMs) like GPT-4 have recently demonstrated astonishing zero-shot capabilities in general domain tasks, they often generate content with hallucinations in specific domains such as Chinese law, hindering their application in these areas. This is typically due to the absence of training data that encompasses such a specific domain, preventing GPT-4 from acquiring in-domain knowledge. A pressing challenge is that it's not plausible to continue training LLMs of such scale on in-domain data. This paper introduces a simple and effective domain adaptation framework for GPT-4 by reformulating generation as an \textbf{adapt-retrieve-revise} process. The initial step is to \textbf{adapt} an affordable 7B LLM to the target domain by continuing learning on in-domain data. When solving a task, we leverage the adapted LLM to generate a draft answer given a task query. Then, the draft answer will be used to \textbf{retrieve} supporting evidence candidates from an external in-domain knowledge base. Finally, the draft answer and retrieved evidence are concatenated into a whole prompt to let GPT-4 assess the evidence and \textbf{revise} the draft answer to generate the final answer. Our proposal combines the advantages of the efficiency of adapting a smaller 7B model with the evidence-assessing capability of GPT-4 and effectively prevents GPT-4 from generating hallucinatory content. In the zero-shot setting of four Chinese legal tasks, our method improves accuracy by 33.3\% compared to the direct generation by GPT-4. When compared to two stronger retrieval-based baselines, our method outperforms them by 15.4\% and 23.9\%. Our code will be released

Jivnesh Sandhan, Fei Cheng, Tushar Sandhan et al.

Large Language Models (LLMs) are increasingly deployed in domains such as education, mental health and customer support, where stable and consistent personas are critical for reliability. Yet, existing studies focus on narrative or role-playing tasks and overlook how adversarial conversational history alone can reshape induced personas. Black-box persona manipulation remains unexplored, raising concerns for robustness in realistic interactions. In response, we introduce the task of persona editing, which adversarially steers LLM traits through user-side inputs under a black-box, inference-only setting. To this end, we propose PHISH (Persona Hijacking via Implicit Steering in History), the first framework to expose a new vulnerability in LLM safety that embeds semantically loaded cues into user queries to gradually induce reverse personas. We also define a metric to quantify attack success. Across 3 benchmarks and 8 LLMs, PHISH predictably shifts personas, triggers collateral changes in correlated traits, and exhibits stronger effects in multi-turn settings. In high-risk domains mental health, tutoring, and customer support, PHISH reliably manipulates personas, validated by both human and LLM-as-Judge evaluations. Importantly, PHISH causes only a small reduction in reasoning benchmark performance, leaving overall utility largely intact while still enabling significant persona manipulation. While current guardrails offer partial protection, they remain brittle under sustained attack. Our findings expose new vulnerabilities in personas and highlight the need for context-resilient persona in LLMs. Our codebase and dataset is available at: https://github.com/Jivnesh/PHISH

Jivnesh Sandhan, Harshit Jaiswal, Fei Cheng et al.

The rapid adoption of LLMs has increased the need for reliable AI text detection, yet existing detectors often fail outside controlled benchmarks. We systematically evaluate 2 dominant paradigms (training-free and supervised) and show that both are brittle under distribution shift, unseen generators, and simple stylistic perturbations. To address these limitations, we propose a supervised contrastive learning (SCL) framework that learns discriminative style embeddings. Experiments show that while supervised detectors excel in-domain, they degrade sharply out-of-domain, and training-free methods remain highly sensitive to proxy choice. Overall, our results expose fundamental challenges in building domain-agnostic detectors. Our code is available at: https://github.com/HARSHITJAIS14/DetectAI

Qianying Liu, Zhuo Gong, Zhengdong Yang et al.

Low-resource speech recognition has been long-suffering from insufficient training data. In this paper, we propose an approach that leverages neighboring languages to improve low-resource scenario performance, founded on the hypothesis that similar linguistic units in neighboring languages exhibit comparable term frequency distributions, which enables us to construct a Huffman tree for performing multilingual hierarchical Softmax decoding. This hierarchical structure enables cross-lingual knowledge sharing among similar tokens, thereby enhancing low-resource training outcomes. Empirical analyses demonstrate that our method is effective in improving the accuracy and efficiency of low-resource speech recognition.

Yibin Shen, Qianying Liu, Zhuoyuan Mao et al.

To solve Math Word Problems, human students leverage diverse reasoning logic that reaches different possible equation solutions. However, the mainstream sequence-to-sequence approach of automatic solvers aims to decode a fixed solution equation supervised by human annotation. In this paper, we propose a controlled equation generation solver by leveraging a set of control codes to guide the model to consider certain reasoning logic and decode the corresponding equations expressions transformed from the human reference. The empirical results suggest that our method universally improves the performance on single-unknown (Math23K) and multiple-unknown (DRAW1K, HMWP) benchmarks, with substantial improvements up to 13.2% accuracy on the challenging multiple-unknown datasets.

Yahan Yu, Noa Nakanishi, Fei Cheng

Large Language Models (LLMs) often produce explicit reflective traces during complex reasoning, accompanied by anthropomorphic markers such as wait, hmm, and alternatively. Although these markers are commonly used as visible indicators of reflection, their mechanisms remain unclear, which leaves the risk of overthinking associated with redundant and repetitive reflection markers. In this work, we revisit anthropomorphic reflection markers, examining their necessity for reasoning and role in the reflection. We suppress these markers through prompt-level and token-level interventions, and analyze their effects on task performance across four benchmarks and two model scales. Our results show that anthropomorphic markers are not uniformly necessary for reasoning performance: suppressing them can preserve or improve performance in several settings, especially under larger sampling budgets. Meanwhile, marker suppression does not necessarily remove reflection behavior, as models can still perform marker-free verification. These suggest that anthropomorphic markers tend to be surface cues rather than reliable proxies for reflection itself, and motivate future research on reasoning mechanisms beyond explicit marker patterns.

Yuxin Wen, Qing Shuai, Di Kang et al.

We present HY-Motion 1.0, a series of state-of-the-art, large-scale, motion generation models capable of generating 3D human motions from textual descriptions. HY-Motion 1.0 represents the first successful attempt to scale up Diffusion Transformer (DiT)-based flow matching models to the billion-parameter scale within the motion generation domain, delivering instruction-following capabilities that significantly outperform current open-source benchmarks. Uniquely, we introduce a comprehensive, full-stage training paradigm -- including large-scale pretraining on over 3,000 hours of motion data, high-quality fine-tuning on 400 hours of curated data, and reinforcement learning from both human feedback and reward models -- to ensure precise alignment with the text instruction and high motion quality. This framework is supported by our meticulous data processing pipeline, which performs rigorous motion cleaning and captioning. Consequently, our model achieves the most extensive coverage, spanning over 200 motion categories across 6 major classes. We release HY-Motion 1.0 to the open-source community to foster future research and accelerate the transition of 3D human motion generation models towards commercial maturity.

Yihua Zhu, Qianying Liu, Fei Cheng et al.

Reinforcement learning with verifiable rewards (RLVR) has become central to post-training reasoning models, yet a key limitation of existing studies is their narrow view of the reasoning space: difficulty is treated as reasoning depth alone, and reward is concentrated on forward deductive state tracking. We instead characterize the reasoning space along two dimensions. Difficulty. Beyond reasoning depth, we study environment complexity, where models must identify the correct path amid distractors and interacting structures. Rewarded reasoning form. We consider four abilities core to real-world reasoning: deductive state tracking, abductive recovery of hidden events or facts, inductive rule induction, and analogical transfer. To disentangle these factors, we construct a synthetic knowledge-graph environment with controlled pre- and post-training distributions, where each instance varies along depth, complexity, and task family. Three findings emerge: joint depth-complexity coverage outperforms single-axis recipes; reasoning families respond non-uniformly, with abductive reasoning degrading outside the RL-covered region and task correlations clustering into deductive-abductive and inductive-analogy pairs; and uniform mixing outperforms staged curricula under a fixed budget. We also find that recent off-the-shelf models exhibit the same deductive-over-abductive asymmetry, suggesting that this gap is not merely an artifact of our controlled setup.

Qianying Liu, Wenyu Guan, Jianhao Shen et al.

Previous studies have introduced a weakly-supervised paradigm for solving math word problems requiring only the answer value annotation. While these methods search for correct value equation candidates as pseudo labels, they search among a narrow sub-space of the enormous equation space. To address this problem, we propose a novel search algorithm with combinatorial strategy \textbf{ComSearch}, which can compress the search space by excluding mathematically equivalent equations. The compression allows the searching algorithm to enumerate all possible equations and obtain high-quality data. We investigate the noise in the pseudo labels that hold wrong mathematical logic, which we refer to as the \textit{false-matching} problem, and propose a ranking model to denoise the pseudo labels. Our approach holds a flexible framework to utilize two existing supervised math word problem solvers to train pseudo labels, and both achieve state-of-the-art performance in the weak supervision task.

Ming Li, Ariunaa Enkhtur, Beverley Anne Yamamoto et al.

Purpose:Generative Artificial Intelligence (GAI) models, such as ChatGPT, may inherit or amplify societal biases due to their training on extensive datasets. With the increasing usage of GAI by students, faculty, and staff in higher education institutions (HEIs), it is urgent to examine the ethical issues and potential biases associated with these technologies. Design/Approach/Methods:This scoping review aims to elucidate how biases related to GAI in HEIs have been researched and discussed in recent academic publications. We categorized the potential societal biases that GAI might cause in the field of higher education. Our review includes articles written in English, Chinese, and Japanese across four main databases, focusing on GAI usage in higher education and bias. Findings:Our findings reveal that while there is meaningful scholarly discussion around bias and discrimination concerning LLMs in the AI field, most articles addressing higher education approach the issue superficially. Few articles identify specific types of bias under different circumstances, and there is a notable lack of empirical research. Most papers in our review focus primarily on educational and research fields related to medicine and engineering, with some addressing English education. However, there is almost no discussion regarding the humanities and social sciences. Additionally, a significant portion of the current discourse is in English and primarily addresses English-speaking contexts. Originality/Value:To the best of our knowledge, our study is the first to summarize the potential societal biases in higher education. This review highlights the need for more in-depth studies and empirical work to understand the specific biases that GAI might introduce or amplify in educational settings, guiding the development of more ethical AI applications in higher education.

Ming Li, Ariunaa Enkhtur, Fei Cheng et al.

This scoping review explores the ethical challenges of using ChatGPT in higher education. By reviewing recent academic articles in English, Chinese, and Japanese, we aimed to provide a deep dive review and identify gaps in the literature. Drawing on Arksey and O'Malley's (2005) scoping review framework, we defined search terms and identified relevant publications from four databases in the three target languages. The research results showed that the majority of the papers were discussion papers, but there was some early empirical work. The ethical issues highlighted in these works mainly concern academic integrity, assessment issues, and data protection. Given the rapid deployment of generative artificial intelligence, it is imperative for educators to conduct more empirical studies to develop sound ethical policies for its use.

Yihua Zhu, Qianying Liu, Jiaxin Wang et al.

Autoregressive LLMs perform well on relational tasks that require linking entities via relational words (e.g., father/son, friend), but it is unclear whether they learn the logical semantics of such relations (e.g., symmetry and inversion logic) and, if so, whether reversal-type failures arise from missing relational semantics or left-to-right order bias. We propose a controlled Knowledge Graph-based synthetic framework that generates text from symmetric/inverse triples, train GPT-style autoregressive models from scratch, and evaluate memorization, logical inference, and in-context generalization to unseen entities to address these questions. We find a sharp phase transition in which relational semantics emerge with sufficient logic-bearing supervision, even in shallow (2-3 layer) models, and that successful generalization aligns with stable intermediate-layer signals. Finally, order-matched forward/reverse tests and a diffusion baseline indicate that reversal failures are primarily driven by autoregressive order bias rather than deficient inversion semantics.

Wan Jou She, Lis Kanashiro Pereira, Fei Cheng et al.

This paper introduces EmplifAI, a Japanese empathetic dialogue dataset designed to support patients coping with chronic medical conditions. They often experience a wide range of positive and negative emotions (e.g., hope and despair) that shift across different stages of disease management. EmplifAI addresses this complexity by providing situation-based dialogues grounded in 28 fine-grained emotion categories, adapted and validated from the GoEmotions taxonomy. The dataset includes 280 medically contextualized situations and 4125 two-turn dialogues, collected through crowdsourcing and expert review. To evaluate emotional alignment in empathetic dialogues, we assessed model predictions on situation--dialogue pairs using BERTScore across multiple large language models (LLMs), achieving F1 scores of 0.83. Fine-tuning a baseline Japanese LLM (LLM-jp-3.1-13b-instruct4) with EmplifAI resulted in notable improvements in fluency, general empathy, and emotion-specific empathy. Furthermore, we compared the scores assigned by LLM-as-a-Judge and human raters on dialogues generated by multiple LLMs to validate our evaluation pipeline and discuss the insights and potential risks derived from the correlation analysis.

Pharath Sathya, Yin Jou Huang, Fei Cheng

Evaluating creative text generation remains a challenge because existing reference-based metrics fail to capture the subjective nature of creativity. We propose a structured evaluation framework for AI story generation comprising four components (Novelty, Value, Adherence, and Resonance) and eleven sub-components. Using controlled story generation via ``Spike Prompting'' and a crowdsourced study of 115 readers, we examine how different creative components shape both immediate and reflective human creativity judgments. Our findings show that creativity is evaluated hierarchically rather than cumulatively, with different dimensions becoming salient at different stages of judgment, and that reflective evaluation substantially alters both ratings and inter-rater agreement. Together, these results support the effectiveness of our framework in revealing dimensions of creativity that are obscured by reference-based evaluation.

Shiho Matta, Lis Kanashiro Pereira, Peitao Han et al.

Modern vision-language models (VLMs) excel at many multimodal tasks, yet their grasp of temporal information in video remains weak and, crucially, under-evaluated. We probe this gap with a deceptively simple but revealing challenge: judging the arrow of time (AoT)-whether a short clip is played forward or backward. We introduce AoT-PsyPhyBENCH, a psychophysically validated benchmark that tests whether VLMs can infer temporal direction in natural videos using the same stimuli and behavioral baselines established for humans. Our comprehensive evaluation of open-weight and proprietary, reasoning and non-reasoning VLMs reveals that most models perform near chance, and even the best lag far behind human accuracy on physically irreversible processes (e.g., free fall, diffusion/explosion) and causal manual actions (division/addition) that humans recognize almost instantly. These results highlight a fundamental gap in current multimodal systems: while they capture rich visual-semantic correlations, they lack the inductive biases required for temporal continuity and causal understanding. We release the code and data for AoT-PsyPhyBENCH to encourage further progress in the physical and temporal reasoning capabilities of VLMs.

Jivnesh Sandhan, Fei Cheng, Tushar Sandhan et al.

Psychometric tests, traditionally used to assess humans, are now being applied to Large Language Models (LLMs) to evaluate their behavioral traits. However, existing studies follow a context-free approach, answering each question in isolation to avoid contextual influence. We term this the Disney World test, an artificial setting that ignores real-world applications, where conversational history shapes responses. To bridge this gap, we propose the first Context-Aware Personality Evaluation (CAPE) framework for LLMs, incorporating prior conversational interactions. To thoroughly analyze the influence of context, we introduce novel metrics to quantify the consistency of LLM responses, a fundamental trait in human behavior. Our exhaustive experiments on 7 LLMs reveal that conversational history enhances response consistency via in-context learning but also induces personality shifts, with GPT-3.5-Turbo and GPT-4-Turbo exhibiting extreme deviations. While GPT models are robust to question ordering, Gemini-1.5-Flash and Llama-8B display significant sensitivity. Moreover, GPT models response stem from their intrinsic personality traits as well as prior interactions, whereas Gemini-1.5-Flash and Llama--8B heavily depend on prior interactions. Finally, applying our framework to Role Playing Agents (RPAs) shows context-dependent personality shifts improve response consistency and better align with human judgments. Our code and datasets are publicly available at: https://github.com/jivnesh/CAPE

Qianying Liu, Katrina Qiyao Wang, Fei Cheng et al.

Large Language Models have garnered significant attention for their capabilities in multilingual natural language processing, while studies on risks associated with cross biases are limited to immediate context preferences. Cross-language disparities in reasoning-based recommendations remain largely unexplored, with a lack of even descriptive analysis. This study is the first to address this gap. We test LLM's applicability and capability in providing personalized advice across three key scenarios: university applications, travel, and relocation. We investigate multilingual bias in state-of-the-art LLMs by analyzing their responses to decision-making tasks across multiple languages. We quantify bias in model-generated scores and assess the impact of demographic factors and reasoning strategies (e.g., Chain-of-Thought prompting) on bias patterns. Our findings reveal that local language bias is prevalent across different tasks, with GPT-4 and Sonnet reducing bias for English-speaking countries compared to GPT-3.5 but failing to achieve robust multilingual alignment, highlighting broader implications for multilingual AI agents and applications such as education. \footnote{Code available at: https://github.com/yiyunya/assess_agentic_national_bias

Qianying Liu, Wenyu Guan, Sujian Li et al.

Automatically solving math word problems is a critical task in the field of natural language processing. Recent models have reached their performance bottleneck and require more high-quality data for training. We propose a novel data augmentation method that reverses the mathematical logic of math word problems to produce new high-quality math problems and introduce new knowledge points that can benefit learning the mathematical reasoning logic. We apply the augmented data on two SOTA math word problem solving models and compare our results with a strong data augmentation baseline. Experimental results show the effectiveness of our approach. We release our code and data at https://github.com/yiyunya/RODA.

Ranran Haoran Zhang, Qianying Liu, Aysa Xuemo Fan et al.

Joint entity and relation extraction aims to extract relation triplets from plain text directly. Prior work leverages Sequence-to-Sequence (Seq2Seq) models for triplet sequence generation. However, Seq2Seq enforces an unnecessary order on the unordered triplets and involves a large decoding length associated with error accumulation. These introduce exposure bias, which may cause the models overfit to the frequent label combination, thus deteriorating the generalization. We propose a novel Sequence-to-Unordered-Multi-Tree (Seq2UMTree) model to minimize the effects of exposure bias by limiting the decoding length to three within a triplet and removing the order among triplets. We evaluate our model on two datasets, DuIE and NYT, and systematically study how exposure bias alters the performance of Seq2Seq models. Experiments show that the state-of-the-art Seq2Seq model overfits to both datasets while Seq2UMTree shows significantly better generalization. Our code is available at https://github.com/WindChimeRan/OpenJERE .

Peitao Han, Fei Cheng, Lis K. Pereira et al.

The Arrow-of-Time (AoT) task, determining whether a video plays forward or backward by recognizing temporal irreversibility, is one humans solve with near-perfect accuracy, yet frontier Video Large Language Models (Video-LLMs) perform only modestly above chance. This gap raises a key question: do visual backbones fail to encode temporal information, or does information bottleneck lie elsewhere in the Video-LLM architecture? We address this question by isolating the vision encoder from the Video-LLM and tracing temporal information across the encoder, projector, and LLM. We find that video-centric encoders with explicit temporal modeling encode strong temporal signals, whereas frame-centric encoders do not. However, when video-centric representations are passed through a standard Video-LLM architecture, performance often collapses, revealing a bottleneck of temporal information flow. We identify projector design as a key factor: Q-Former disrupts temporal information, while a time-preserved MLP projection substantially improves the LLM's access to such information. Our layer-wise analysis further shows temporal representation dynamics across encoder layers. Guided by these findings, we build a Video-LLM with temporal-aware video-centric encoder, time-preserved projector, and AoT supervision, surpassing human performance on AoT$_{PPB}$ with 98.1\% accuracy, and improving broader temporal reasoning tasks by up to 6.0 points on VITATECS-Direction and 1.3 points on TVBench. Our results show that temporal reasoning in Video-LLMs requires both effective temporal encoding and reliable transfer of this information to the LLM.

Zeqing Wang, Fei Cheng, Kangye Ji et al.

Continual learning (CL) is a technique that enables neural networks to constantly adapt to their dynamic surroundings. Despite being overlooked for a long time, this technology can considerably address the customized needs of users in edge devices. Actually, most CL methods require huge resource consumption by the training behavior to acquire generalizability among all tasks for delaying forgetting regardless of edge scenarios. Therefore, this paper proposes a compact algorithm called LightCL, which evaluates and compresses the redundancy of already generalized components in structures of the neural network. Specifically, we consider two factors of generalizability, learning plasticity and memory stability, and design metrics of both to quantitatively assess generalizability of neural networks during CL. This evaluation shows that generalizability of different layers in a neural network exhibits a significant variation. Thus, we $\textit{Maintain Generalizability}$ by freezing generalized parts without the resource-intensive training process and $\textit{Memorize Feature Patterns}$ by stabilizing feature extracting of previous tasks to enhance generalizability for less-generalized parts with a little extra memory, which is far less than the reduction by freezing. Experiments illustrate that LightCL outperforms other state-of-the-art methods and reduces at most $\textbf{6.16$\times$}$ memory footprint. We also verify the effectiveness of LightCL on the edge device.

Zelin Zhang, Fei Cheng, Chenhui Chu

Although outcome-based reinforcement learning (RL) significantly advances the mathematical reasoning capabilities of Large Language Models (LLMs), its reliance on computationally expensive ground-truth annotations imposes a severe scalability bottleneck. Unsupervised RL guided by intrinsic rewards offers a scalable alternative, yet it suffers from opaque training dynamics and catastrophic instability, such as policy collapse and reward hacking. In this paper, we first design and evaluate a suite of intrinsic rewards that explicitly enforce concise and certain generation. Second, to discover the boundaries of this approach, we test base models across a spectrum of intrinsic reasoning capabilities, revealing how a model's foundational logical prior dictates its success or failure. Finally, to demystify why certain configurations stabilize while others collapse, we introduce a novel geometric diagnostic lens, showing that successful cases are enveloped by manifolds. Ultimately, our work goes beyond merely demonstrating that enforcing concise and certain responses successfully boosts mathematical reasoning; we reveal when this unsupervised approach breaks down and geometrically diagnose why.

Yikun Sun, Zhen Wan, Nobuhiro Ueda et al.

The creation of instruction data and evaluation benchmarks for serving Large language models often involves enormous human annotation. This issue becomes particularly pronounced when rapidly developing such resources for a non-English language like Japanese. Instead of following the popular practice of directly translating existing English resources into Japanese (e.g., Japanese-Alpaca), we propose an efficient self-instruct method based on GPT-4. We first translate a small amount of English instructions into Japanese and post-edit them to obtain native-level quality. GPT-4 then utilizes them as demonstrations to automatically generate Japanese instruction data. We also construct an evaluation benchmark containing 80 questions across 8 categories, using GPT-4 to automatically assess the response quality of LLMs without human references. The empirical results suggest that the models fine-tuned on our GPT-4 self-instruct data significantly outperformed the Japanese-Alpaca across all three base pre-trained models. Our GPT-4 self-instruct data allowed the LLaMA 13B model to defeat GPT-3.5 (Davinci-003) with a 54.37\% win-rate. The human evaluation exhibits the consistency between GPT-4's assessments and human preference. Our high-quality instruction data and evaluation benchmark have been released here.

Zeqing Wang, Kangye Ji, Di Wang et al.

Segment Anything Model (SAM) struggles with segmenting objects in the open world, especially across diverse and dynamic domains. Continual segmentation (CS) is a potential technique to solve this issue, but a significant obstacle is the intractable balance between previous domains (stability) and new domains (plasticity) during CS. Furthermore, how to utilize two kinds of features of SAM, images and prompts, in an efficient and effective CS manner remains a significant hurdle. In this work, we propose a novel CS method, termed SAMCL, to address these challenges. It is the first study to empower SAM with the CS ability across dynamic domains. SAMCL decouples stability and plasticity during CS by two components: $\textit{AugModule}$ and $\textit{Module Selector}$. Specifically, SAMCL leverages individual $\textit{AugModule}$ to effectively and efficiently learn new relationships between images and prompts in each domain. $\textit{Module Selector}$ selects the appropriate module during testing, based on the inherent ability of SAM to distinguish between different domains. These two components enable SAMCL to realize a task-agnostic method without any interference across different domains. Experimental results demonstrate that SAMCL outperforms state-of-the-art methods, achieving an exceptionally low average forgetting of just $0.5$%, along with at least a $2.5$% improvement in transferring to unseen domains. Moreover, the tunable parameter consumption in AugModule is about $0.236$MB, marking at least a $23.3$% reduction compared to other fine-tuning methods.

Ruifeng Luo, Zhengjie Liu, Tianxiao Cheng et al.

Recognizing symbols in architectural CAD drawings is critical for various advanced engineering applications. In this paper, we propose a novel CAD data annotation engine that leverages intrinsic attributes from systematically archived CAD drawings to automatically generate high-quality annotations, thus significantly reducing manual labeling efforts. Utilizing this engine, we construct ArchCAD-400K, a large-scale CAD dataset consisting of 413,062 chunks from 5538 highly standardized drawings, making it over 26 times larger than the largest existing CAD dataset. ArchCAD-400K boasts an extended drawing diversity and broader categories, offering line-grained annotations. Furthermore, we present a new baseline model for panoptic symbol spotting, termed Dual-Pathway Symbol Spotter (DPSS). It incorporates an adaptive fusion module to enhance primitive features with complementary image features, achieving state-of-the-art performance and enhanced robustness. Extensive experiments validate the effectiveness of DPSS, demonstrating the value of ArchCAD-400K and its potential to drive innovation in architectural design and construction.

Fuya Nakamori, Yin Jou Huang, Fei Cheng

This study proposes a method to improve the performance of Werewolf agents by switching between predefined strategies based on the attitudes of other players and the context of conversations. While prior works of Werewolf agents using prompt engineering have employed methods where effective strategies are implicitly defined, they cannot adapt to changing situations. In this research, we propose a method that explicitly selects an appropriate strategy based on the game context and the estimated roles of other players. We compare the strategy adaptation Werewolf agents with baseline agents using implicit or fixed strategies and verify the effectiveness of our proposed method.

Fei Cheng, Ribeka Tanaka, Sadao Kurohashi

Clinical information extraction (e.g., 2010 i2b2/VA challenge) usually presents tasks of concept recognition, assertion classification, and relation extraction. Jointly modeling the multi-stage tasks in the clinical domain is an underexplored topic. The existing independent task setting (reference inputs given in each stage) makes the joint models not directly comparable to the existing pipeline work. To address these issues, we define a joint task setting and propose a novel end-to-end system to jointly optimize three-stage tasks. We empirically investigate the joint evaluation of our proposal and the pipeline baseline with various embedding techniques: word, contextual, and in-domain contextual embeddings. The proposed joint system substantially outperforms the pipeline baseline by +0.3, +1.4, +3.1 for the concept, assertion, and relation F1. This work bridges joint approaches and clinical information extraction. The proposed approach could serve as a strong joint baseline for future research. The code is publicly available.

Chengzhi Zhong, Fei Cheng, Qianying Liu et al.

Large language models exhibit strong multilingual capabilities despite limited exposure to non-English data. Prior studies show that English-centric large language models map multilingual content into English-aligned representations at intermediate layers and then project them back into target-language token spaces in the final layer. From this observation, we hypothesize that this cross-lingual transition is governed by a small and sparse set of dimensions, which occur at consistent indices across the intermediate to final layers. Building on this insight, we introduce a simple, training-free method to identify and manipulate these dimensions, requiring only as few as 50 sentences of either parallel or monolingual data. Experiments on a multilingual generation control task reveal the interpretability of these dimensions, demonstrating that the interventions in these dimensions can switch the output language while preserving semantic content, and that it surpasses the performance of prior neuron-based approaches at a substantially lower cost.

Zhen Wan, Chao-Han Huck Yang, Yahan Yu et al.

We introduce Speech-based Intelligence Quotient (SIQ) as a new form of human cognition-inspired evaluation pipeline for voice understanding large language models, LLM Voice, designed to assess their voice understanding ability. Moving beyond popular voice understanding metrics such as word error rate (WER), SIQ examines LLM Voice across three cognitive levels motivated by Bloom's Taxonomy: (1) Remembering (i.e., WER for verbatim accuracy); (2) Understanding (i.e., similarity of LLM's interpretations); and (3) Application (i.e., QA accuracy for simulating downstream tasks). We demonstrate that SIQ not only quantifies voice understanding abilities but also provides unified comparisons between cascaded methods (e.g., ASR LLM) and end-to-end models, identifies annotation errors in existing benchmarks, and detects hallucinations in LLM Voice. Our framework represents a first-of-its-kind intelligence examination that bridges cognitive principles with voice-oriented benchmarks, while exposing overlooked challenges in multi-modal training.

Sakiko Yahata, Zhen Wan, Fei Cheng et al.

Extracting causal relationships from a medical case report is essential for comprehending the case, particularly its diagnostic process. Since the diagnostic process is regarded as a bottom-up inference, causal relationships in cases naturally form a multi-layered tree structure. The existing tasks, such as medical relation extraction, are insufficient for capturing the causal relationships of an entire case, as they treat all relations equally without considering the hierarchical structure inherent in the diagnostic process. Thus, we propose a novel task, Causal Tree Extraction (CTE), which receives a case report and generates a causal tree with the primary disease as the root, providing an intuitive understanding of a case's diagnostic process. Subsequently, we construct a Japanese case report CTE dataset, J-Casemap, propose a generation-based CTE method that outperforms the baseline by 20.2 points in the human evaluation, and introduce evaluation metrics that reflect clinician preferences. Further experiments also show that J-Casemap enhances the performance of solving other medical tasks, such as question answering.

Zhengdong Yang, Qianying Liu, Sheng Li et al.

We present a novel approach centered on the decoding stage of Automatic Speech Recognition (ASR) that enhances multilingual performance, especially for low-resource languages. It utilizes a cross-lingual embedding clustering method to construct a hierarchical Softmax (H-Softmax) decoder, which enables similar tokens across different languages to share similar decoder representations. It addresses the limitations of the previous Huffman-based H-Softmax method, which relied on shallow features in token similarity assessments. Through experiments on a downsampled dataset of 15 languages, we demonstrate the effectiveness of our approach in improving low-resource multilingual ASR accuracy.

Peitao Han, Lis Kanashiro Pereira, Fei Cheng et al.

Existing in-context learning (ICL) methods for relation extraction (RE) often prioritize language similarity over structural similarity, which can lead to overlooking entity relationships. To address this, we propose an AMR-enhanced retrieval-based ICL method for RE. Our model retrieves in-context examples based on semantic structure similarity between task inputs and training samples. Evaluations on four standard English RE datasets show that our model outperforms baselines in the unsupervised setting across all datasets. In the supervised setting, it achieves state-of-the-art results on three datasets and competitive results on the fourth.

Haochuan Jiang, Guanyu Yang, Fei Cheng et al.

Synthesizing Chinese characters with consistent style using few stylized examples is challenging. Existing models struggle to generate arbitrary style characters with limited examples. In this paper, we propose the Generalized W-Net, a novel class of W-shaped architectures that addresses this. By incorporating Adaptive Instance Normalization and introducing multi-content, our approach can synthesize Chinese characters in any desired style, even with limited examples. It handles seen and unseen styles during training and can generate new character contents. Experimental results demonstrate the effectiveness of our approach.

Felix Virgo, Fei Cheng, Lis Kanashiro Pereira et al.

We propose a voting-driven semi-supervised approach to automatically acquire the typical duration of an event and use it as pseudo-labeled data. The human evaluation demonstrates that our pseudo labels exhibit surprisingly high accuracy and balanced coverage. In the temporal commonsense QA task, experimental results show that using only pseudo examples of 400 events, we achieve performance comparable to the existing BERT-based weakly supervised approaches that require a significant amount of training examples. When compared to the RoBERTa baselines, our best approach establishes state-of-the-art performance with a 7% improvement in Exact Match.

Tatsuro Inaba, Hirokazu Kiyomaru, Fei Cheng et al.

Large language models (LLMs) have achieved impressive performance on various reasoning tasks. To further improve the performance, we propose MultiTool-CoT, a novel framework that leverages chain-of-thought (CoT) prompting to incorporate multiple external tools, such as a calculator and a knowledge retriever, during the reasoning process. We apply MultiTool-CoT to the Task 2 dataset of NumGLUE, which requires both numerical reasoning and domain-specific knowledge. The experiments show that our method significantly outperforms strong baselines and achieves state-of-the-art performance.

Qianying Liu, Dongsheng Yang, Wenjie Zhong et al.

Numerical reasoning over table-and-text hybrid passages, such as financial reports, poses significant challenges and has numerous potential applications. Noise and irrelevant variables in the model input have been a hindrance to its performance. Additionally, coarse-grained supervision of the whole solution program has impeded the model's ability to learn the underlying numerical reasoning process. In this paper, we propose three pretraining tasks that operate at both the whole program and sub-program level: Variable Integrity Ranking, which guides the model to focus on useful variables; Variable Operator Prediction, which decomposes the supervision into fine-grained single operator prediction; and Variable Keyphrase Masking, which encourages the model to identify key evidence that sub-programs are derived from. Experimental results demonstrate the effectiveness of our proposed methods, surpassing transformer-based model baselines.

Zhen Wan, Fei Cheng, Zhuoyuan Mao et al.

In spite of the potential for ground-breaking achievements offered by large language models (LLMs) (e.g., GPT-3), they still lag significantly behind fully-supervised baselines (e.g., fine-tuned BERT) in relation extraction (RE). This is due to the two major shortcomings of LLMs in RE: (1) low relevance regarding entity and relation in retrieved demonstrations for in-context learning; and (2) the strong inclination to wrongly classify NULL examples into other pre-defined labels. In this paper, we propose GPT-RE to bridge the gap between LLMs and fully-supervised baselines. GPT-RE successfully addresses the aforementioned issues by (1) incorporating task-specific entity representations in demonstration retrieval; and (2) enriching the demonstrations with gold label-induced reasoning logic. We evaluate GPT-RE on four widely-used RE datasets, and observe that GPT-RE achieves improvements over not only existing GPT-3 baselines, but also fully-supervised baselines. Specifically, GPT-RE achieves SOTA performances on the Semeval and SciERC datasets, and competitive performances on the TACRED and ACE05 datasets.

Qianying Liu, Fei Cheng, Sadao Kurohashi

Meta learning with auxiliary languages has demonstrated promising improvements for cross-lingual natural language processing. However, previous studies sample the meta-training and meta-testing data from the same language, which limits the ability of the model for cross-lingual transfer. In this paper, we propose XLA-MAML, which performs direct cross-lingual adaption in the meta-learning stage. We conduct zero-shot and few-shot experiments on Natural Language Inference and Question Answering. The experimental results demonstrate the effectiveness of our method across different languages, tasks, and pretrained models. We also give analysis on various cross-lingual specific settings for meta-learning including sampling strategy and parallelism.

Fei Cheng, Shuntaro Yada, Ribeka Tanaka et al.

We present an open-access natural language processing toolkit for Japanese medical information extraction. We first propose a novel relation annotation schema for investigating the medical and temporal relations between medical entities in Japanese medical reports. We experiment with the practical annotation scenarios by separately annotating two different types of reports. We design a pipeline system with three components for recognizing medical entities, classifying entity modalities, and extracting relations. The empirical results show accurate analyzing performance and suggest the satisfactory annotation quality, the effective annotation strategy for targeting report types, and the superiority of the latest contextual embedding models.