Peng Qi

Ethan A. Chi, Ashwin Paranjape, Abigail See et al. · meta-ai, stanford

We present Chirpy Cardinal, an open-domain social chatbot. Aiming to be both informative and conversational, our bot chats with users in an authentic, emotionally intelligent way. By integrating controlled neural generation with scaffolded, hand-written dialogue, we let both the user and bot take turns driving the conversation, producing an engaging and socially fluent experience. Deployed in the fourth iteration of the Alexa Prize Socialbot Grand Challenge, Chirpy Cardinal handled thousands of conversations per day, placing second out of nine bots with an average user rating of 3.58/5.

Kaiser Sun, Peng Qi, Yuhao Zhang et al. · stanford

Generative models have been widely applied to solve extractive tasks, where parts of the input is extracted to form the desired output, and achieved significant success. For example, in extractive question answering (QA), generative models have constantly yielded state-of-the-art results. In this work, we identify the issue of tokenization inconsistency that is commonly neglected in training these models. This issue damages the extractive nature of these tasks after the input and output are tokenized inconsistently by the tokenizer, and thus leads to performance drop as well as hallucination. We propose a simple yet effective fix to this issue and conduct a case study on extractive QA. We show that, with consistent tokenization, the model performs better in both in-domain and out-of-domain datasets, with a notable average of +1.7 F2 gain when a BART model is trained on SQuAD and evaluated on 8 QA datasets. Further, the model converges faster, and becomes less likely to generate out-of-context answers. With these findings, we would like to call for more attention on how tokenization should be done when solving extractive tasks and recommend applying consistent tokenization during training.

Yuyan Bu, Qiang Sheng, Juan Cao et al.

With information consumption via online video streaming becoming increasingly popular, misinformation video poses a new threat to the health of the online information ecosystem. Though previous studies have made much progress in detecting misinformation in text and image formats, video-based misinformation brings new and unique challenges to automatic detection systems: 1) high information heterogeneity brought by various modalities, 2) blurred distinction between misleading video manipulation and nonmalicious artistic video editing, and 3) new patterns of misinformation propagation due to the dominant role of recommendation systems on online video platforms. To facilitate research on this challenging task, we conduct this survey to present advances in misinformation video detection. We first analyze and characterize the misinformation video from three levels including signals, semantics, and intents. Based on the characterization, we systematically review existing works for detection from features of various modalities to techniques for clue integration. We also introduce existing resources including representative datasets and useful tools. Besides summarizing existing studies, we discuss related areas and outline open issues and future directions to encourage and guide more research on misinformation video detection. The corresponding repository is at https://github.com/ICTMCG/Awesome-Misinfo-Video-Detection.

Rujun Han, Yuhao Zhang, Peng Qi et al. · stanford

Question answering based on retrieval augmented generation (RAG-QA) is an important research topic in NLP and has a wide range of real-world applications. However, most existing datasets for this task are either constructed using a single source corpus or consist of short extractive answers, which fall short of evaluating large language model (LLM) based RAG-QA systems on cross-domain generalization. To address these limitations, we create Long-form RobustQA (LFRQA), a new dataset comprising human-written long-form answers that integrate short extractive answers from multiple documents into a single, coherent narrative, covering 26K queries and large corpora across seven different domains. We further propose RAG-QA Arena by directly comparing model-generated answers against LFRQA's answers using LLMs as evaluators. We show via extensive experiments that RAG-QA Arena and human judgments on answer quality are highly correlated. Moreover, only 41.3% of the most competitive LLM's answers are preferred to LFRQA's answers, demonstrating RAG-QA Arena as a challenging evaluation platform for future research.

Xiyang Hu, Xinchi Chen, Peng Qi et al.

Multilingual information retrieval (IR) is challenging since annotated training data is costly to obtain in many languages. We present an effective method to train multilingual IR systems when only English IR training data and some parallel corpora between English and other languages are available. We leverage parallel and non-parallel corpora to improve the pretrained multilingual language models' cross-lingual transfer ability. We design a semantic contrastive loss to align representations of parallel sentences that share the same semantics in different languages, and a new language contrastive loss to leverage parallel sentence pairs to remove language-specific information in sentence representations from non-parallel corpora. When trained on English IR data with these losses and evaluated zero-shot on non-English data, our model demonstrates significant improvement to prior work on retrieval performance, while it requires much less computational effort. We also demonstrate the value of our model for a practical setting when a parallel corpus is only available for a few languages, but a lack of parallel corpora resources persists for many other low-resource languages. Our model can work well even with a small number of parallel sentences, and be used as an add-on module to any backbones and other tasks.

Beizhe Hu, Qiang Sheng, Juan Cao et al.

Detecting fake news requires both a delicate sense of diverse clues and a profound understanding of the real-world background, which remains challenging for detectors based on small language models (SLMs) due to their knowledge and capability limitations. Recent advances in large language models (LLMs) have shown remarkable performance in various tasks, but whether and how LLMs could help with fake news detection remains underexplored. In this paper, we investigate the potential of LLMs in fake news detection. First, we conduct an empirical study and find that a sophisticated LLM such as GPT 3.5 could generally expose fake news and provide desirable multi-perspective rationales but still underperforms the basic SLM, fine-tuned BERT. Our subsequent analysis attributes such a gap to the LLM's inability to select and integrate rationales properly to conclude. Based on these findings, we propose that current LLMs may not substitute fine-tuned SLMs in fake news detection but can be a good advisor for SLMs by providing multi-perspective instructive rationales. To instantiate this proposal, we design an adaptive rationale guidance network for fake news detection (ARG), in which SLMs selectively acquire insights on news analysis from the LLMs' rationales. We further derive a rationale-free version of ARG by distillation, namely ARG-D, which services cost-sensitive scenarios without querying LLMs. Experiments on two real-world datasets demonstrate that ARG and ARG-D outperform three types of baseline methods, including SLM-based, LLM-based, and combinations of small and large language models.

Chao Shang, Guangtao Wang, Peng Qi et al.

Question answering over temporal knowledge graphs (KGs) efficiently uses facts contained in a temporal KG, which records entity relations and when they occur in time, to answer natural language questions (e.g., "Who was the president of the US before Obama?"). These questions often involve three time-related challenges that previous work fail to adequately address: 1) questions often do not specify exact timestamps of interest (e.g., "Obama" instead of 2000); 2) subtle lexical differences in time relations (e.g., "before" vs "after"); 3) off-the-shelf temporal KG embeddings that previous work builds on ignore the temporal order of timestamps, which is crucial for answering temporal-order related questions. In this paper, we propose a time-sensitive question answering (TSQA) framework to tackle these problems. TSQA features a timestamp estimation module to infer the unwritten timestamp from the question. We also employ a time-sensitive KG encoder to inject ordering information into the temporal KG embeddings that TSQA is based on. With the help of techniques to reduce the search space for potential answers, TSQA significantly outperforms the previous state of the art on a new benchmark for question answering over temporal KGs, especially achieving a 32% (absolute) error reduction on complex questions that require multiple steps of reasoning over facts in the temporal KG.

Zhengxuan Wu, Yuhao Zhang, Peng Qi et al. · stanford

Modern language models (LMs) need to follow human instructions while being faithful; yet, they often fail to achieve both. Here, we provide concrete evidence of a trade-off between instruction following (i.e., follow open-ended instructions) and faithfulness (i.e., ground responses in given context) when training LMs with these objectives. For instance, fine-tuning LLaMA-7B on instruction following datasets renders it less faithful. Conversely, instruction-tuned Vicuna-7B shows degraded performance at following instructions when further optimized on tasks that require contextual grounding. One common remedy is multi-task learning (MTL) with data mixing, yet it remains far from achieving a synergic outcome. We propose a simple yet effective method that relies on Rejection Sampling for Continued Self-instruction Tuning (ReSet), which significantly outperforms vanilla MTL. Surprisingly, we find that less is more, as training ReSet with high-quality, yet substantially smaller data (three-fold less) yields superior results. Our findings offer a better understanding of objective discrepancies in alignment training of LMs.

Guang Yang, Juan Cao, Qiang Sheng et al.

The daily practice of sharing images on social media raises a severe issue about privacy leakage. To address the issue, privacy-leaking image detection is studied recently, with the goal to automatically identify images that may leak privacy. Recent advance on this task benefits from focusing on crucial objects via pretrained object detectors and modeling their correlation. However, these methods have two limitations: 1) they neglect other important elements like scenes, textures, and objects beyond the capacity of pretrained object detectors; 2) the correlation among objects is fixed, but a fixed correlation is not appropriate for all the images. To overcome the limitations, we propose the Dynamic Region-Aware Graph Convolutional Network (DRAG) that dynamically finds out crucial regions including objects and other important elements, and models their correlation adaptively for each input image. To find out crucial regions, we cluster spatially-correlated feature channels into several region-aware feature maps. Further, we dynamically model the correlation with the self-attention mechanism and explore the interaction among the regions with a graph convolutional network. The DRAG achieved an accuracy of 87% on the largest dataset for privacy-leaking image detection, which is 10 percentage points higher than the state of the art. The further case study demonstrates that it found out crucial regions containing not only objects but other important elements like textures.

Yuyan Bu, Qiang Sheng, Juan Cao et al.

As short-form video-sharing platforms become a significant channel for news consumption, fake news in short videos has emerged as a serious threat in the online information ecosystem, making developing detection methods for this new scenario an urgent need. Compared with that in text and image formats, fake news on short video platforms contains rich but heterogeneous information in various modalities, posing a challenge to effective feature utilization. Unlike existing works mostly focusing on analyzing what is presented, we introduce a novel perspective that considers how it might be created. Through the lens of the creative process behind news video production, our empirical analysis uncovers the unique characteristics of fake news videos in material selection and editing. Based on the obtained insights, we design FakingRecipe, a creative process-aware model for detecting fake news short videos. It captures the fake news preferences in material selection from sentimental and semantic aspects and considers the traits of material editing from spatial and temporal aspects. To improve evaluation comprehensiveness, we first construct FakeTT, an English dataset for this task, and conduct experiments on both FakeTT and the existing Chinese FakeSV dataset. The results show FakingRecipe's superiority in detecting fake news on short video platforms.

Xiaofei Wen, Wenjie Jacky Mo, Xingyu Fu et al.

Despite rapid progress in video-capable MLLMs, we find that their apparent audio understanding in videos is often vision-driven: models rely on visual cues to infer or hallucinate acoustic information, rather than verifying the audio stream. This issue appears across both state-of-the-art open-source omni models and leading closed-source models from providers such as Google and OpenAI. We characterize this failure mode as an audio-visual Clever Hans effect, in which models appear (falsely) audio-grounded, but actually exploit visual-acoustic correlations without verifying whether the audio and visual streams are truly aligned. To systematically study this behavior, we introduce Thud, an intervention-driven probing framework based on three counterfactual audio edits: Shift, which tests temporal synchronization; Mute, which tests sound existence; and Swap, which tests audio-visual consistency. Beyond diagnosis, we further study a two-stage alignment recipe: intervention-derived preference pairs teach audio verification, while event-level general video preferences regularize the model against over-specialization. Our best 10K-sample recipe improves average performance across the three intervention dimensions by 28 percentage points, while slightly improving performance on general video and audio-visual QA benchmarks.

Hendrik Schuff, Heike Adel, Peng Qi et al.

While much research focused on producing explanations, it is still unclear how the produced explanations' quality can be evaluated in a meaningful way. Today's predominant approach is to quantify explanations using proxy scores which compare explanations to (human-annotated) gold explanations. This approach assumes that explanations which reach higher proxy scores will also provide a greater benefit to human users. In this paper, we present problems of this approach. Concretely, we (i) formulate desired characteristics of explanation quality, (ii) describe how current evaluation practices violate them, and (iii) support our argumentation with initial evidence from a crowdsourcing case study in which we investigate the explanation quality of state-of-the-art explainable question answering systems. We find that proxy scores correlate poorly with human quality ratings and, additionally, become less expressive the more often they are used (i.e. following Goodhart's law). Finally, we propose guidelines to enable a meaningful evaluation of explanations to drive the development of systems that provide tangible benefits to human users.

Tiehua Zhang, Yuze Liu, Zhishu Shen et al.

Graph neural network (GNN) has gained increasing popularity in recent years owing to its capability and flexibility in modeling complex graph structure data. Among all graph learning methods, hypergraph learning is a technique for exploring the implicit higher-order correlations when training the embedding space of the graph. In this paper, we propose a hypergraph learning framework named LFH that is capable of dynamic hyperedge construction and attentive embedding update utilizing the heterogeneity attributes of the graph. Specifically, in our framework, the high-quality features are first generated by the pairwise fusion strategy that utilizes explicit graph structure information when generating initial node embedding. Afterwards, a hypergraph is constructed through the dynamic grouping of implicit hyperedges, followed by the type-specific hypergraph learning process. To evaluate the effectiveness of our proposed framework, we conduct comprehensive experiments on several popular datasets with eleven state-of-the-art models on both node classification and link prediction tasks, which fall into categories of homogeneous pairwise graph learning, heterogeneous pairwise graph learning, and hypergraph learning. The experiment results demonstrate a significant performance gain (average 12.5% in node classification and 13.3% in link prediction) compared with recent state-of-the-art methods.

Peng Qi, Guangtao Wang, Jing Huang

Distilling supervision signal from a long sequence to make predictions is a challenging task in machine learning, especially when not all elements in the input sequence contribute equally to the desired output. In this paper, we propose SpanDrop, a simple and effective data augmentation technique that helps models identify the true supervision signal in a long sequence with very few examples. By directly manipulating the input sequence, SpanDrop randomly ablates parts of the sequence at a time and ask the model to perform the same task to emulate counterfactual learning and achieve input attribution. Based on theoretical analysis of its properties, we also propose a variant of SpanDrop based on the beta-Bernoulli distribution, which yields diverse augmented sequences while providing a learning objective that is more consistent with the original dataset. We demonstrate the effectiveness of SpanDrop on a set of carefully designed toy tasks, as well as various natural language processing tasks that require reasoning over long sequences to arrive at the correct answer, and show that it helps models improve performance both when data is scarce and abundant.

Sanjari Srivastava, Gang Li, Cheng Chang et al.

Training web agents to navigate complex, real-world websites requires them to master $\textit{subtasks}$ - short-horizon interactions on multiple UI components (e.g., choosing the correct date in a date picker, or scrolling in a container to extract information). We introduce WARC-Bench (Web Archive Benchmark), a novel web navigation benchmark featuring 438 tasks designed to evaluate multimodal AI agents on subtasks. WARC-Bench enables sandboxed interactions with dynamic and realistic webpages using Web ARChive files. We show that WARC-Bench is challenging for leading computer-use models, with the highest observed success rate being 64.8%. To improve open source models on subtask, we explore two common training techniques: supervised fine-tuning (SFT) and reinforcement learning with verifiable rewards (RLVR). Experiments show that SFT models obtain a 48.8% success rate on the benchmark. Training with RLVR over SFT checkpoints, even in data-scarce settings, improves the score to 52.8% on WARC-Bench, outperforming many frontier models. Our analysis concludes that mastering these subtasks is essential for robust web planning and navigation, and is a capability not extensively evaluated by existing benchmarks.

Peng Qi, Yuhao Zhang, Yuhui Zhang et al.

We introduce Stanza, an open-source Python natural language processing toolkit supporting 66 human languages. Compared to existing widely used toolkits, Stanza features a language-agnostic fully neural pipeline for text analysis, including tokenization, multi-word token expansion, lemmatization, part-of-speech and morphological feature tagging, dependency parsing, and named entity recognition. We have trained Stanza on a total of 112 datasets, including the Universal Dependencies treebanks and other multilingual corpora, and show that the same neural architecture generalizes well and achieves competitive performance on all languages tested. Additionally, Stanza includes a native Python interface to the widely used Java Stanford CoreNLP software, which further extends its functionality to cover other tasks such as coreference resolution and relation extraction. Source code, documentation, and pretrained models for 66 languages are available at https://stanfordnlp.github.io/stanza.

Peng Qi, Zehong Yan, Wynne Hsu et al.

Misinformation is a prevalent societal issue due to its potential high risks. Out-of-context (OOC) misinformation, where authentic images are repurposed with false text, is one of the easiest and most effective ways to mislead audiences. Current methods focus on assessing image-text consistency but lack convincing explanations for their judgments, which is essential for debunking misinformation. While Multimodal Large Language Models (MLLMs) have rich knowledge and innate capability for visual reasoning and explanation generation, they still lack sophistication in understanding and discovering the subtle crossmodal differences. In this paper, we introduce SNIFFER, a novel multimodal large language model specifically engineered for OOC misinformation detection and explanation. SNIFFER employs two-stage instruction tuning on InstructBLIP. The first stage refines the model's concept alignment of generic objects with news-domain entities and the second stage leverages language-only GPT-4 generated OOC-specific instruction data to fine-tune the model's discriminatory powers. Enhanced by external tools and retrieval, SNIFFER not only detects inconsistencies between text and image but also utilizes external knowledge for contextual verification. Our experiments show that SNIFFER surpasses the original MLLM by over 40% and outperforms state-of-the-art methods in detection accuracy. SNIFFER also provides accurate and persuasive explanations as validated by quantitative and human evaluations.

Boyu Zhu, Xiaofei Wen, Wenjie Jacky Mo et al.

Omni-modal Large Language Models (OLLMs) that process text, images, videos, and audio introduce new challenges for safety and value guardrails in human-AI interaction. Prior guardrail research largely targets unimodal settings and typically frames safeguarding as binary classification, which limits robustness across diverse modalities and tasks. To address this gap, we propose OmniGuard, the first family of omni-modal guardrails that performs safeguarding across all modalities with deliberate reasoning ability. To support the training of OMNIGUARD, we curate a large, comprehensive omni-modal safety dataset comprising over 210K diverse samples, with inputs that cover all modalities through both unimodal and cross-modal samples. Each sample is annotated with structured safety labels and carefully curated safety critiques from expert models through targeted distillation. Extensive experiments on 15 benchmarks show that OmniGuard achieves strong effectiveness and generalization across a wide range of multimodal safety scenarios. Importantly, OmniGuard provides a unified framework that enforces policies and mitigates risks in omni-modalities, paving the way toward building more robust and capable omnimodal safeguarding systems.

Yu Gu, Kai Zhang, Yuting Ning et al. · microsoft-research

Language agents based on large language models (LLMs) have demonstrated great promise in automating web-based tasks. Recent work has shown that incorporating advanced planning algorithms, e.g., tree search, is advantageous over reactive planning for web agents. However, unlike simulated sandbox environments, real-world environments such as the web are rife with irreversible actions. This undermines the feasibility of backtracking, a cornerstone of (tree) search. Overly relying on test-time search also hurts efficiency. We advocate model-based planning for web agents that employs a world model to simulate and deliberate over the outcome of each candidate action before committing to one. We systematically explore this paradigm by (1) Proposing a model-based planning framework, WebDreamer, which employs LLMs to serve as both world models and value functions; (2) Training specialized LLMs as world models with a scalable data synthesis pipeline. Empirical results demonstrate that WebDreamer achieves substantial performance improvements over reactive baselines. It is competitive, while being 4-5 times more efficient, with tree search in sandbox environments (VisualWebArena) and also works effectively on real-world websites (Online-Mind2Web and Mind2Web-Live). Furthermore, our trained world model, Dreamer-7B, performs comparable to GPT-4o, highlighting the potential of specialized world models for efficient and effective planning in complex web environments.

Xiaoran Cai, Wang Yang, Xiyu Ren et al.

Sustainability or ESG rating agencies use company disclosures and external data to produce scores or ratings that assess the environmental, social, and governance performance of a company. However, sustainability ratings across agencies for a single company vary widely, limiting their comparability, credibility, and relevance to decision-making. To harmonize the rating results, we propose adopting a universal human-AI collaboration framework to generate trustworthy benchmark datasets for evaluating sustainability rating methodologies. The framework comprises two complementary parts: STRIDE (Sustainability Trust Rating & Integrity Data Equation) provides principled criteria and a scoring system that guide the construction of firm-level benchmark datasets using large language models (LLMs), and SR-Delta, a discrepancy-analysis procedural framework that surfaces insights for potential adjustments. The framework enables scalable and comparable assessment of sustainability rating methodologies. We call on the broader AI community to adopt AI-powered approaches to strengthen and advance sustainability rating methodologies that support and enforce urgent sustainability agendas.

Tianliang Yao, Bo Lu, Markus Kowarschik et al.

Endovascular procedures have revolutionized the treatment of vascular diseases thanks to minimally invasive solutions that significantly reduce patient recovery time and enhance clinical outcomes. However, the precision and dexterity required during these procedures poses considerable challenges for interventionists. Robotic systems have emerged offering transformative solutions, addressing issues such as operator fatigue, radiation exposure, and the inherent limitations of human precision. The integration of Embodied Intelligence (EI) into these systems signifies a paradigm shift, enabling robots to navigate complex vascular networks and adapt to dynamic physiological conditions. Data-driven approaches, advanced computer vision, medical image analysis, and machine learning techniques, are at the forefront of this evolution. These methods augment procedural intelligence by facilitating real-time vessel segmentation, device tracking, and anatomical landmark detection. Reinforcement learning and imitation learning further refine navigation strategies and replicate experts' techniques. This review systematically examines the integration of EI principles into robotic technologies, in relation to endovascular procedures. We discuss recent advancements in intelligent perception and data-driven control, and their practical applications in robot-assisted endovascular procedures. By critically evaluating current limitations and emerging opportunities, this review establishes a framework for future developments, emphasizing the potential for greater autonomy and improved clinical outcomes. Emerging trends and specific areas of research, such as federated learning for medical data sharing, explainable AI for clinical decision support, and advanced human-robot collaboration paradigms, are also explored, offering insights into the future direction of this rapidly evolving field.

Yumo Xu, Peng Qi, Jifan Chen et al. · amazon-science

Citation quality is crucial in information-seeking systems, directly influencing trust and the effectiveness of information access. Current evaluation frameworks, both human and automatic, mainly rely on Natural Language Inference (NLI) to assess binary or ternary supportiveness from cited sources, which we argue is a suboptimal proxy for citation evaluation. In this work we introduce CiteEval, a citation evaluation framework driven by principles focusing on fine-grained citation assessment within a broad context, encompassing not only the cited sources but the full retrieval context, user query, and generated text. Guided by the proposed framework, we construct CiteBench, a multi-domain benchmark with high-quality human annotations on citation quality. To enable efficient evaluation, we further develop CiteEval-Auto, a suite of model-based metrics that exhibit strong correlation with human judgments. Experiments across diverse systems demonstrate CiteEval-Auto's superior ability to capture the multifaceted nature of citations compared to existing metrics, offering a principled and scalable approach to evaluate and improve model-generated citations.

Zehong Yan, Peng Qi, Wynne Hsu et al.

Multimodal misinformation, encompassing textual, visual, and cross-modal distortions, poses an increasing societal threat that is amplified by generative AI. Existing methods typically focus on a single type of distortion and struggle to generalize to unseen scenarios. In this work, we observe that different distortion types share common reasoning capabilities while also requiring task-specific skills. We hypothesize that joint training across distortion types facilitates knowledge sharing and enhances the model's ability to generalize. To this end, we introduce TRUST-VL, a unified and explainable vision-language model for general multimodal misinformation detection. TRUST-VL incorporates a novel Question-Aware Visual Amplifier module, designed to extract task-specific visual features. To support training, we also construct TRUST-Instruct, a large-scale instruction dataset containing 198K samples featuring structured reasoning chains aligned with human fact-checking workflows. Extensive experiments on both in-domain and zero-shot benchmarks demonstrate that TRUST-VL achieves state-of-the-art performance, while also offering strong generalization and interpretability.

Tianliang Yao, Zhiqiang Pei, Yong Li et al.

An ever-growing incorporation of AI solutions into clinical practices enhances the efficiency and effectiveness of healthcare services. This paper focuses on guidewire tip tracking tasks during image-guided therapy for cardiovascular diseases, aiding physicians in improving diagnostic and therapeutic quality. A novel tracking framework based on a Siamese network with dual attention mechanisms combines self- and cross-attention strategies for robust guidewire tip tracking. This design handles visual ambiguities, tissue deformations, and imaging artifacts through enhanced spatial-temporal feature learning. Validation occurred on 3 randomly selected clinical digital subtraction angiography (DSA) sequences from a dataset of 15 sequences, covering multiple interventional scenarios. The results indicate a mean localization error of 0.421 $\pm$ 0.138 mm, with a maximum error of 1.736 mm, and a mean Intersection over Union (IoU) of 0.782. The framework maintains an average processing speed of 57.2 frames per second, meeting the temporal demands of endovascular imaging. Further validations with robotic platforms for automating diagnostics and therapies in clinical routines yielded tracking errors of 0.708 $\pm$ 0.695 mm and 0.148 $\pm$ 0.057 mm in two distinct experimental scenarios.

Yuchen Mao, Hongwei Li, Yinyi Lai et al.

Large-scale vision models like SAM have extensive visual knowledge, yet their general nature and computational demands limit their use in specialized tasks like medical image segmentation. In contrast, task-specific models such as U-Net++ often underperform due to sparse labeled data. This study introduces a strategic knowledge mining method that leverages SAM's broad understanding to boost the performance of small, locally hosted deep learning models. In our approach, we trained a U-Net++ model on a limited labeled dataset and extend its capabilities by converting SAM's output infered on unlabeled images into prompts. This process not only harnesses SAM's generalized visual knowledge but also iteratively improves SAM's prediction to cater specialized medical segmentation tasks via U-Net++. The mined knowledge, serving as "pseudo labels", enriches the training dataset, enabling the fine-tuning of the local network. Applied to the Kvasir SEG and COVID-QU-Ex datasets which consist of gastrointestinal polyp and lung X-ray images respectively, our proposed method consistently enhanced the segmentation performance on Dice by 3% and 1% respectively over the baseline U-Net++ model, when the same amount of labelled data were used during training (75% and 50% of labelled data). Remarkably, our proposed method surpassed the baseline U-Net++ model even when the latter was trained exclusively on labeled data (100% of labelled data). These results underscore the potential of knowledge mining to overcome data limitations in specialized models by leveraging the broad, albeit general, knowledge of large-scale models like SAM, all while maintaining operational efficiency essential for clinical applications.

Simon Yu, Gang Li, Weiyan Shi et al.

Large language models (LLMs) are moving beyond static uses and are now powering agents that learn continually during their interaction with external environments. For example, agents can learn reusable skills while navigating web pages or toggling new tools. However, existing methods for skill learning often create skills that are over-specialized to a single website and fail to generalize. We introduce PolySkill, a new framework that enables agents to learn generalizable and compositional skills. The core idea, inspired by polymorphism in software engineering, is to decouple a skill's abstract goal (what it accomplishes) and its concrete implementation (how it is executed). Experiments show that our method (1) improves skill reuse by 1.7x on seen websites and (2) boosts success rates by up to 9.4% on Mind2Web and 13.9% on unseen websites, while reducing steps by over 20%. (3) In self-exploration settings without specified tasks, our framework improves the quality of proposed tasks and enables agents to learn generalizable skills that work across different sites. By enabling the agent to identify and refine its own goals, the PolySkill enhances the agent's ability to learn a better curriculum, leading to the acquisition of more generalizable skills compared to baseline methods. This work provides a practical path toward building agents capable of continual learning in adaptive environments. Our findings show that separating a skill's goal from its execution is a crucial step toward developing autonomous agents that can learn and generalize across the open web continuously.

Mengzhao Jia, Zhihan Zhang, Ignacio Cases et al.

Multimodal large language models (MLLMs) have rapidly advanced from perception tasks to complex multi-step reasoning, yet reinforcement learning with verifiable rewards (RLVR) often leads to spurious reasoning since only the final-answer correctness is rewarded. To address this limitation, we propose AutoRubric-R1V, a framework that integrates RLVR with process-level supervision through automatically collected rubric-based generative rewards. Our key innovation lies in a scalable self-aggregation method that distills consistent reasoning checkpoints from successful trajectories, enabling problem-specific rubric construction without human annotation or stronger teacher models. By jointly leveraging rubric-based and outcome rewards, AutoRubric-R1V achieves state-of-the-art performance on six multimodal reasoning benchmarks and substantially improves reasoning faithfulness in dedicated evaluations.

Yijin Ni, Peng Qi

Frontier large language models such as GPT-5 and Claude Sonnet remain prone to overconfidence even after alignment through Reinforcement Learning with Human Feedback (RLHF) and Direct Preference Optimization (DPO). For instance, they tend to offer the same conservative answer "No" to both questions "Can I eat the [food / potato chips] that has been left out overnight?" despite the latter requiring no refridgeration for safe consumption. We find that this failure is potentially attributed to a limitation of existing preference learning: it emphasizes selecting the correct response for a given prompt, while neglecting counterfactual prompts that should alter the response. To address this limitation, we propose abductive preference learning, a fine-tuning paradigm that reverses the conventional conditioning by learning preferences over prompts given a response. To validate this idea, we construct an abductive dataset derived from the HaluEval QA benchmark with 1,001 entries, implementing abductive DPO and its variant DPOP. Experiments reveal complementary strengths: standard methods improve response selection, abductive methods improve prompt discrimination, while a multitask objective unifies both. On the abductive dataset, multitask DPOP boosts accuracy from $90.0\%$ to $99.5\%$ in response selection and $54.7\%$ to $85.0\%$ in prompt discrimination, with qualitative evidence highlighting improved sensitivity to prompt differences. Finally, evaluation on AlpacaEval shows multitask DPOP improves win rate (from $5.26\%$ to $6.17\%$), confirming that abductive preference learning preserves the benefits of conventional preference optimization while addressing the overlooked challenge of counterfactual prompts.

Chengzhi Liu, Yuzhe Yang, Kaiwen Zhou et al.

The promotion of academic papers has become an important means of enhancing research visibility. However, existing automated methods struggle limited storytelling, insufficient aesthetic quality, and constrained self-adjustment, making it difficult to achieve efficient and engaging dissemination. At the heart of those challenges is a simple principle: \emph{there is no way to improve it when you cannot evaluate it right}. To address this, we introduce \textbf{EvoPresent}, a self-improvement agent framework that unifies coherent narratives, aesthetic-aware designs, and realistic presentation delivery via virtual characters. Central to EvoPresent is \textbf{PresAesth}, a multi-task reinforcement learning (RL) aesthetic model that provides reliable aesthetic scoring, defect adjustment, and comparative feedback, enabling iterative self-improvement even under limited aesthetic training data. To systematically evaluate the methods, we introduce \textbf{EvoPresent Benchmark}, a comprehensive benchmark comprising: \textit{Presentation Generation Quality}, built on 650 top-tier AI conference papers with multimodal resources (slides, videos and scripts) to assess both content and design; and \textit{Aesthetic Awareness}, consisting of 2,000 slide pairs with varying aesthetic levels, supporting joint training and evaluation on scoring, defect adjustment, and comparison. Our findings highlight that (i) High-quality feedback is essential for agent self-improvement, while initial capability alone does not guarantee effective self-correction. (ii) Automated generation pipelines exhibit a trade-off between visual design and content construction. (iii) Multi-task RL training shows stronger generalization in aesthetic awareness tasks.

Yuyan Bu, Qiang Sheng, Juan Cao et al.

The emergence of fake news on short video platforms has become a new significant societal concern, necessitating automatic video-news-specific detection. Current detectors primarily rely on pattern-based features to separate fake news videos from real ones. However, limited and less diversified training data lead to biased patterns and hinder their performance. This weakness stems from the complex many-to-many relationships between video material segments and fabricated news events in real-world scenarios: a single video clip can be utilized in multiple ways to create different fake narratives, while a single fabricated event often combines multiple distinct video segments. However, existing datasets do not adequately reflect such relationships due to the difficulty of collecting and annotating large-scale real-world data, resulting in sparse coverage and non-comprehensive learning of the characteristics of potential fake news video creation. To address this issue, we propose a data augmentation framework, AgentAug, that generates diverse fake news videos by simulating typical creative processes. AgentAug implements multiple LLM-driven pipelines of four fabrication categories for news video creation, combined with an active learning strategy based on uncertainty sampling to select the potentially useful augmented samples during training. Experimental results on two benchmark datasets demonstrate that AgentAug consistently improves the performance of short video fake news detectors.

Xiaofei Wen, Wenjie Jacky Mo, Yanan Xie et al.

Autonomous web agents need to operate under externally imposed or human-specified policies while generating long-horizon trajectories. However, little work has examined whether these trajectories comply with such policies, or whether policy violations persist across different contexts such as domains (e.g., shopping or coding websites) and subdomains (e.g., product search and order management in shopping). To address this gap, we introduce PolicyGuardBench, a benchmark of about 60k examples for detecting policy violations in agent trajectories. From diverse agent runs, we generate a broad set of policies and create both within subdomain and cross subdomain pairings with violation labels. In addition to full-trajectory evaluation, PolicyGuardBench also includes a prefix-based violation detection task where models must anticipate policy violations from truncated trajectory prefixes rather than complete sequences. Using this dataset, we train PolicyGuard-4B, a lightweight guardrail model that delivers strong detection accuracy across all tasks while keeping inference efficient. Notably, PolicyGuard-4B generalizes across domains and preserves high accuracy on unseen settings. Together, PolicyGuardBench and PolicyGuard-4B provide the first comprehensive framework for studying policy compliance in web agent trajectories, and show that accurate and generalizable guardrails are feasible at small scales.

Peng Qi, Wenxi Qu, Tianliang Yao et al.

Percutaneous Coronary Intervention (PCI) is a minimally invasive procedure that improves coronary blood flow and treats coronary artery disease. Although PCI typically requires 2D X-ray angiography (XRA) to guide catheter placement at real-time, computed tomography angiography (CTA) may substantially improve PCI by providing precise information of 3D vascular anatomy and status. To leverage real-time XRA and detailed 3D CTA anatomy for PCI, accurate multimodal image registration of XRA and CTA is required, to guide the procedure and avoid complications. This is a challenging process as it requires registration of images from different geometrical modalities (2D -> 3D and vice versa), with variations in contrast and noise levels. In this paper, we propose a novel multimodal coronary artery image registration method based on a swarm optimization algorithm, which effectively addresses challenges such as large deformations, low contrast, and noise across these imaging modalities. Our algorithm consists of two main modules: 1) preprocessing of XRA and CTA images separately, and 2) a registration module based on feature extraction using the Steger and Superpixel Particle Swarm Optimization algorithms. Our technique was evaluated on a pilot dataset of 28 pairs of XRA and CTA images from 10 patients who underwent PCI. The algorithm was compared with four state-of-the-art (SOTA) methods in terms of registration accuracy, robustness, and efficiency. Our method outperformed the selected SOTA baselines in all aspects. Experimental results demonstrate the significant effectiveness of our algorithm, surpassing the previous benchmarks and proposes a novel clinical approach that can potentially have merit for improving patient outcomes in coronary artery disease.

Zehong Yan, Peng Qi, Wynne Hsu et al.

While large generative artificial intelligence (GenAI) models have achieved significant success, they also raise growing concerns about online information security due to their potential misuse for generating deceptive content. Out-of-context (OOC) multimodal misinformation detection, which often retrieves Web evidence to identify the repurposing of images in false contexts, faces the issue of reasoning over GenAI-polluted evidence to derive accurate predictions. Existing works simulate GenAI-powered pollution at the claim level with stylistic rewriting to conceal linguistic cues, and ignore evidence-level pollution for such information-seeking applications. In this work, we investigate how polluted evidence affects the performance of existing OOC detectors, revealing a performance degradation of more than 9 percentage points. We propose two strategies, cross-modal evidence reranking and cross-modal claim-evidence reasoning, to address the challenges posed by polluted evidence. Extensive experiments on two benchmark datasets show that these strategies can effectively enhance the robustness of existing out-of-context detectors amidst polluted evidence.

Bo Li, Peng Qi, Bo Liu et al.

The rapid development of Artificial Intelligence (AI) technology has enabled the deployment of various systems based on it. However, many current AI systems are found vulnerable to imperceptible attacks, biased against underrepresented groups, lacking in user privacy protection. These shortcomings degrade user experience and erode people's trust in all AI systems. In this review, we provide AI practitioners with a comprehensive guide for building trustworthy AI systems. We first introduce the theoretical framework of important aspects of AI trustworthiness, including robustness, generalization, explainability, transparency, reproducibility, fairness, privacy preservation, and accountability. To unify currently available but fragmented approaches toward trustworthy AI, we organize them in a systematic approach that considers the entire lifecycle of AI systems, ranging from data acquisition to model development, to system development and deployment, finally to continuous monitoring and governance. In this framework, we offer concrete action items for practitioners and societal stakeholders (e.g., researchers, engineers, and regulators) to improve AI trustworthiness. Finally, we identify key opportunities and challenges for the future development of trustworthy AI systems, where we identify the need for a paradigm shift toward comprehensively trustworthy AI systems.

Peng Qi, Juan Cao, Xirong Li et al.

Recently, fake news with text and images have achieved more effective diffusion than text-only fake news, raising a severe issue of multimodal fake news detection. Current studies on this issue have made significant contributions to developing multimodal models, but they are defective in modeling the multimodal content sufficiently. Most of them only preliminarily model the basic semantics of the images as a supplement to the text, which limits their performance on detection. In this paper, we find three valuable text-image correlations in multimodal fake news: entity inconsistency, mutual enhancement, and text complementation. To effectively capture these multimodal clues, we innovatively extract visual entities (such as celebrities and landmarks) to understand the news-related high-level semantics of images, and then model the multimodal entity inconsistency and mutual enhancement with the help of visual entities. Moreover, we extract the embedded text in images as the complementation of the original text. All things considered, we propose a novel entity-enhanced multimodal fusion framework, which simultaneously models three cross-modal correlations to detect diverse multimodal fake news. Extensive experiments demonstrate the superiority of our model compared to the state of the art.

Peng Qi, Jing Huang, Youzheng Wu et al.

Conversational artificial intelligence (ConvAI) systems have attracted much academic and commercial attention recently, making significant progress on both fronts. However, little existing work discusses how these systems can be developed and deployed for social good in real-world applications, with comprehensive case studies and analyses of pros and cons. In this paper, we briefly review the progress the community has made towards better ConvAI systems and reflect on how existing technologies can help advance social good initiatives from various angles that are unique for ConvAI, or not yet become common knowledge in the community. We further discuss about the challenges ahead for ConvAI systems to better help us achieve these goals and highlight the risks involved in their development and deployment in the real world.

Xiaochen Hou, Peng Qi, Guangtao Wang et al.

Recent work on aspect-level sentiment classification has demonstrated the efficacy of incorporating syntactic structures such as dependency trees with graph neural networks(GNN), but these approaches are usually vulnerable to parsing errors. To better leverage syntactic information in the face of unavoidable errors, we propose a simple yet effective graph ensemble technique, GraphMerge, to make use of the predictions from differ-ent parsers. Instead of assigning one set of model parameters to each dependency tree, we first combine the dependency relations from different parses before applying GNNs over the resulting graph. This allows GNN mod-els to be robust to parse errors at no additional computational cost, and helps avoid overparameterization and overfitting from GNN layer stacking by introducing more connectivity into the ensemble graph. Our experiments on the SemEval 2014 Task 4 and ACL 14 Twitter datasets show that our GraphMerge model not only outperforms models with single dependency tree, but also beats other ensemble mod-els without adding model parameters.

Peng Qi, Haejun Lee, Oghenetegiri "TG" Sido et al.

We develop a unified system to answer directly from text open-domain questions that may require a varying number of retrieval steps. We employ a single multi-task transformer model to perform all the necessary subtasks -- retrieving supporting facts, reranking them, and predicting the answer from all retrieved documents -- in an iterative fashion. We avoid crucial assumptions of previous work that do not transfer well to real-world settings, including exploiting knowledge of the fixed number of retrieval steps required to answer each question or using structured metadata like knowledge bases or web links that have limited availability. Instead, we design a system that can answer open-domain questions on any text collection without prior knowledge of reasoning complexity. To emulate this setting, we construct a new benchmark, called BeerQA, by combining existing one- and two-step datasets with a new collection of 530 questions that require three Wikipedia pages to answer, unifying Wikipedia corpora versions in the process. We show that our model demonstrates competitive performance on both existing benchmarks and this new benchmark. We make the new benchmark available at https://beerqa.github.io/.

Ashwin Paranjape, Abigail See, Kathleen Kenealy et al.

We present Chirpy Cardinal, an open-domain dialogue agent, as a research platform for the 2019 Alexa Prize competition. Building an open-domain socialbot that talks to real people is challenging - such a system must meet multiple user expectations such as broad world knowledge, conversational style, and emotional connection. Our socialbot engages users on their terms - prioritizing their interests, feelings and autonomy. As a result, our socialbot provides a responsive, personalized user experience, capable of talking knowledgeably about a wide variety of topics, as well as chatting empathetically about ordinary life. Neural generation plays a key role in achieving these goals, providing the backbone for our conversational and emotional tone. At the end of the competition, Chirpy Cardinal progressed to the finals with an average rating of 3.6/5.0, a median conversation duration of 2 minutes 16 seconds, and a 90th percentile duration of over 12 minutes.

Devendra Singh Sachan, Yuhao Zhang, Peng Qi et al.

Much recent work suggests that incorporating syntax information from dependency trees can improve task-specific transformer models. However, the effect of incorporating dependency tree information into pre-trained transformer models (e.g., BERT) remains unclear, especially given recent studies highlighting how these models implicitly encode syntax. In this work, we systematically study the utility of incorporating dependency trees into pre-trained transformers on three representative information extraction tasks: semantic role labeling (SRL), named entity recognition, and relation extraction. We propose and investigate two distinct strategies for incorporating dependency structure: a late fusion approach, which applies a graph neural network on the output of a transformer, and a joint fusion approach, which infuses syntax structure into the transformer attention layers. These strategies are representative of prior work, but we introduce additional model design elements that are necessary for obtaining improved performance. Our empirical analysis demonstrates that these syntax-infused transformers obtain state-of-the-art results on SRL and relation extraction tasks. However, our analysis also reveals a critical shortcoming of these models: we find that their performance gains are highly contingent on the availability of human-annotated dependency parses, which raises important questions regarding the viability of syntax-augmented transformers in real-world applications.

Yuhao Zhang, Yuhui Zhang, Peng Qi et al.

We introduce biomedical and clinical English model packages for the Stanza Python NLP library. These packages offer accurate syntactic analysis and named entity recognition capabilities for biomedical and clinical text, by combining Stanza's fully neural architecture with a wide variety of open datasets as well as large-scale unsupervised biomedical and clinical text data. We show via extensive experiments that our packages achieve syntactic analysis and named entity recognition performance that is on par with or surpasses state-of-the-art results. We further show that these models do not compromise speed compared to existing toolkits when GPU acceleration is available, and are made easy to download and use with Stanza's Python interface. A demonstration of our packages is available at: http://stanza.run/bio.

Peng Qi, Yuhao Zhang, Christopher D. Manning

We investigate the problem of generating informative questions in information-asymmetric conversations. Unlike previous work on question generation which largely assumes knowledge of what the answer might be, we are interested in the scenario where the questioner is not given the context from which answers are drawn, but must reason pragmatically about how to acquire new information, given the shared conversation history. We identify two core challenges: (1) formally defining the informativeness of potential questions, and (2) exploring the prohibitively large space of potential questions to find the good candidates. To generate pragmatic questions, we use reinforcement learning to optimize an informativeness metric we propose, combined with a reward function designed to promote more specific questions. We demonstrate that the resulting pragmatic questioner substantially improves the informativeness and specificity of questions generated over a baseline model, as evaluated by our metrics as well as humans.

Juan Cao, Peng Qi, Qiang Sheng et al.

The increasing popularity of social media promotes the proliferation of fake news, which has caused significant negative societal effects. Therefore, fake news detection on social media has recently become an emerging research area of great concern. With the development of multimedia technology, fake news attempts to utilize multimedia content with images or videos to attract and mislead consumers for rapid dissemination, which makes visual content an important part of fake news. Despite the importance of visual content, our understanding of the role of visual content in fake news detection is still limited. This chapter presents a comprehensive review of the visual content in fake news, including the basic concepts, effective visual features, representative detection methods and challenging issues of multimedia fake news detection. This chapter can help readers to understand the role of visual content in fake news detection, and effectively utilize visual content to assist in detecting multimedia fake news.

Peng Qi, Xiaowen Lin, Leo Mehr et al.

It is challenging for current one-step retrieve-and-read question answering (QA) systems to answer questions like "Which novel by the author of 'Armada' will be adapted as a feature film by Steven Spielberg?" because the question seldom contains retrievable clues about the missing entity (here, the author). Answering such a question requires multi-hop reasoning where one must gather information about the missing entity (or facts) to proceed with further reasoning. We present GoldEn (Gold Entity) Retriever, which iterates between reading context and retrieving more supporting documents to answer open-domain multi-hop questions. Instead of using opaque and computationally expensive neural retrieval models, GoldEn Retriever generates natural language search queries given the question and available context, and leverages off-the-shelf information retrieval systems to query for missing entities. This allows GoldEn Retriever to scale up efficiently for open-domain multi-hop reasoning while maintaining interpretability. We evaluate GoldEn Retriever on the recently proposed open-domain multi-hop QA dataset, HotpotQA, and demonstrate that it outperforms the best previously published model despite not using pretrained language models such as BERT.

Juan Cao, Qiang Sheng, Peng Qi et al.

Social media has become a major information platform where people consume and share news. However, it has also enabled the wide dissemination of false news, i.e., news posts published on social media that are verifiably false, causing significant negative effects on society. In order to help prevent further propagation of false news on social media, we set up this competition to motivate the development of automated real-time false news detection approaches. Specifically, this competition includes three sub-tasks: false-news text detection, false-news image detection and false-news multi-modal detetcion, which aims to motivate participants to further explore the efficiency of multiple modalities in detecting false news and reasonable fusion approaches of multi-modal contents. To better support this competition, we also construct and publicize a multi-modal data repository about False News on Weibo Social platform(MCG-FNeWS}) to help evaluate the performance of different approaches from participants.

Peng Qi, Juan Cao, Tianyun Yang et al.

The increasing popularity of social media promotes the proliferation of fake news. With the development of multimedia technology, fake news attempts to utilize multimedia contents with images or videos to attract and mislead readers for rapid dissemination, which makes visual contents an important part of fake news. Fake-news images, images attached in fake news posts,include not only fake images which are maliciously tampered but also real images which are wrongly used to represent irrelevant events. Hence, how to fully exploit the inherent characteristics of fake-news images is an important but challenging problem for fake news detection. In the real world, fake-news images may have significantly different characteristics from real-news images at both physical and semantic levels, which can be clearly reflected in the frequency and pixel domain, respectively. Therefore, we propose a novel framework Multi-domain Visual Neural Network (MVNN) to fuse the visual information of frequency and pixel domains for detecting fake news. Specifically, we design a CNN-based network to automatically capture the complex patterns of fake-news images in the frequency domain; and utilize a multi-branch CNN-RNN model to extract visual features from different semantic levels in the pixel domain. An attention mechanism is utilized to fuse the feature representations of frequency and pixel domains dynamically. Extensive experiments conducted on a real-world dataset demonstrate that MVNN outperforms existing methods with at least 9.2% in accuracy, and can help improve the performance of multimodal fake news detection by over 5.2%.

Peng Qi, Timothy Dozat, Yuhao Zhang et al.

This paper describes Stanford's system at the CoNLL 2018 UD Shared Task. We introduce a complete neural pipeline system that takes raw text as input, and performs all tasks required by the shared task, ranging from tokenization and sentence segmentation, to POS tagging and dependency parsing. Our single system submission achieved very competitive performance on big treebanks. Moreover, after fixing an unfortunate bug, our corrected system would have placed the 2nd, 1st, and 3rd on the official evaluation metrics LAS,MLAS, and BLEX, and would have outperformed all submission systems on low-resource treebank categories on all metrics by a large margin. We further show the effectiveness of different model components through extensive ablation studies.

Yuhao Zhang, Peng Qi, Christopher D. Manning

Dependency trees help relation extraction models capture long-range relations between words. However, existing dependency-based models either neglect crucial information (e.g., negation) by pruning the dependency trees too aggressively, or are computationally inefficient because it is difficult to parallelize over different tree structures. We propose an extension of graph convolutional networks that is tailored for relation extraction, which pools information over arbitrary dependency structures efficiently in parallel. To incorporate relevant information while maximally removing irrelevant content, we further apply a novel pruning strategy to the input trees by keeping words immediately around the shortest path between the two entities among which a relation might hold. The resulting model achieves state-of-the-art performance on the large-scale TACRED dataset, outperforming existing sequence and dependency-based neural models. We also show through detailed analysis that this model has complementary strengths to sequence models, and combining them further improves the state of the art.

Zhilin Yang, Peng Qi, Saizheng Zhang et al.

Existing question answering (QA) datasets fail to train QA systems to perform complex reasoning and provide explanations for answers. We introduce HotpotQA, a new dataset with 113k Wikipedia-based question-answer pairs with four key features: (1) the questions require finding and reasoning over multiple supporting documents to answer; (2) the questions are diverse and not constrained to any pre-existing knowledge bases or knowledge schemas; (3) we provide sentence-level supporting facts required for reasoning, allowing QA systems to reason with strong supervision and explain the predictions; (4) we offer a new type of factoid comparison questions to test QA systems' ability to extract relevant facts and perform necessary comparison. We show that HotpotQA is challenging for the latest QA systems, and the supporting facts enable models to improve performance and make explainable predictions.

Urvashi Khandelwal, He He, Peng Qi et al.

We know very little about how neural language models (LM) use prior linguistic context. In this paper, we investigate the role of context in an LSTM LM, through ablation studies. Specifically, we analyze the increase in perplexity when prior context words are shuffled, replaced, or dropped. On two standard datasets, Penn Treebank and WikiText-2, we find that the model is capable of using about 200 tokens of context on average, but sharply distinguishes nearby context (recent 50 tokens) from the distant history. The model is highly sensitive to the order of words within the most recent sentence, but ignores word order in the long-range context (beyond 50 tokens), suggesting the distant past is modeled only as a rough semantic field or topic. We further find that the neural caching model (Grave et al., 2017b) especially helps the LSTM to copy words from within this distant context. Overall, our analysis not only provides a better understanding of how neural LMs use their context, but also sheds light on recent success from cache-based models.