Xiangyu Peng

Ziheng Qin, Kai Wang, Zangwei Zheng et al.

Data pruning aims to obtain lossless performances with less overall cost. A common approach is to filter out samples that make less contribution to the training. This could lead to gradient expectation bias compared to the original data. To solve this problem, we propose \textbf{InfoBatch}, a novel framework aiming to achieve lossless training acceleration by unbiased dynamic data pruning. Specifically, InfoBatch randomly prunes a portion of less informative samples based on the loss distribution and rescales the gradients of the remaining samples to approximate the original gradient. As a plug-and-play and architecture-agnostic framework, InfoBatch consistently obtains lossless training results on classification, semantic segmentation, vision pertaining, and instruction fine-tuning tasks. On CIFAR10/100, ImageNet-1K, and ADE20K, InfoBatch losslessly saves 40\% overall cost. For pertaining MAE and diffusion model, InfoBatch can respectively save 24.8\% and 27\% cost. For LLaMA instruction fine-tuning, InfoBatch is also able to save 20\% cost and is compatible with coreset selection methods. The code is publicly available at \href{https://github.com/henryqin1997/InfoBatch}{github.com/NUS-HPC-AI-Lab/InfoBatch}.

Wei Zhou, Xiangyu Peng, Mark Riedl · gatech

One major challenge in reinforcement learning (RL) is the large amount of steps for the RL agent needs to converge in the training process and learn the optimal policy, especially in text-based game environments where the action space is extensive. However, non-player characters (NPCs) sometimes hold some key information about the game, which can potentially help to train RL agents faster. Thus, this paper explores how to interact and converse with NPC agents to get the key information using large language models (LLMs), as well as incorporate this information to speed up RL agent's training using knowledge graphs (KGs) and Story Shaping.

Xiangyu Peng, Christopher Cui, Wei Zhou et al. · gatech

Reward design for reinforcement learning agents can be difficult in situations where one not only wants the agent to achieve some effect in the world but where one also cares about how that effect is achieved. For example, we might wish for an agent to adhere to a tacit understanding of commonsense, align itself to a preference for how to behave for purposes of safety, or taking on a particular role in an interactive game. Storytelling is a mode for communicating tacit procedural knowledge. We introduce a technique, Story Shaping, in which a reinforcement learning agent infers tacit knowledge from an exemplar story of how to accomplish a task and intrinsically rewards itself for performing actions that make its current environment adhere to that of the inferred story world. Specifically, Story Shaping infers a knowledge graph representation of the world state from observations, and also infers a knowledge graph from the exemplar story. An intrinsic reward is generated based on the similarity between the agent's inferred world state graph and the inferred story world graph. We conducted experiments in text-based games requiring commonsense reasoning and shaping the behaviors of agents as virtual game characters.

Zexin Chen, Eric Zhou, Kenneth Eaton et al. · gatech

Imaginative play is an area of creativity that could allow robots to engage with the world around them in a much more personified way. Imaginary play can be seen as taking real objects and locations and using them as imaginary objects and locations in virtual scenarios. We adopted the story generation capability of large language models (LLMs) to obtain the stories used for imaginary play with human-written prompts. Those generated stories will be simplified and mapped into action sequences that can guide the agent in imaginary play. To evaluate whether the agent can successfully finish the imaginary play, we also designed a text adventure game to simulate a house as the playground for the agent to interact.

Xiangyu Peng, Chen Xing, Prafulla Kumar Choubey et al. · gatech, salesforce

Prompt tuning approaches, which learn task-specific soft prompts for a downstream task conditioning on frozen pre-trained models, have attracted growing interest due to its parameter efficiency. With large language models and sufficient training data, prompt tuning performs comparably to full-model tuning. However, with limited training samples in few-shot settings, prompt tuning fails to match the performance of full-model fine-tuning. In this work, we focus on improving the few-shot performance of prompt tuning by transferring knowledge from soft prompts of source tasks. Recognizing the good generalization capabilities of ensemble methods in low-data regime, we first experiment and show that a simple ensemble of model predictions based on different source prompts, outperforms existing multi-prompt knowledge transfer approaches such as source prompt fusion in the few-shot setting. Motivated by this observation, we further investigate model ensembles and propose Sample-specific Ensemble of Source Models (SESoM). SESoM learns to adjust the contribution of each source model for each target sample separately when ensembling source model outputs. Through this way, SESoM inherits the superior generalization of model ensemble approaches and simultaneously captures the sample-specific competence of each source prompt. We conduct experiments across a diverse set of eight NLP tasks using models of different scales (T5-{base, large, XL}) and find that SESoM consistently outperforms the existing models of the same as well as larger parametric scale by a large margin.

Daquan Zhou, Kai Wang, Jianyang Gu et al.

State-of-the-art deep neural networks are trained with large amounts (millions or even billions) of data. The expensive computation and memory costs make it difficult to train them on limited hardware resources, especially for recent popular large language models (LLM) and computer vision models (CV). Recent popular dataset distillation methods are thus developed, aiming to reduce the number of training samples via synthesizing small-scale datasets via gradient matching. However, as the gradient calculation is coupled with the specific network architecture, the synthesized dataset is biased and performs poorly when used for training unseen architectures. To address these limitations, we present dataset quantization (DQ), a new framework to compress large-scale datasets into small subsets which can be used for training any neural network architectures. Extensive experiments demonstrate that DQ is able to generate condensed small datasets for training unseen network architectures with state-of-the-art compression ratios for lossless model training. To the best of our knowledge, DQ is the first method that can successfully distill large-scale datasets such as ImageNet-1k with a state-of-the-art compression ratio. Notably, with 60% data from ImageNet and 20% data from Alpaca's instruction tuning data, the models can be trained with negligible or no performance drop for both vision tasks (including classification, semantic segmentation, and object detection) as well as language tasks (including instruction tuning tasks such as BBH and DROP).

Jiaxin Zhang, Xiangyu Peng, Qinglin Chen et al.

On-policy distillation (OPD) is an increasingly important paradigm for post-training language models. However, we identify a pervasive Scaling Law of Miscalibration: while OPD effectively improves task accuracy, it systematically traps models in severe overconfidence. We trace this failure to an information mismatch: teacher supervision is formed under privileged context available during training, whereas the deployed model must report confidence using only deployment-time information. We formalize this perspective theoretically, showing that teacher-conditioned success is generally not a valid target for deployment-time confidence and that helpful privileged context induces entropy collapse and a systematic optimism bias. To address this, we propose a calibration-aware OPD framework, CaOPD, that estimates empirical confidence from model rollouts, replaces self-reported confidence with this student-grounded target, and distills the revised response through the same self-distillation pipeline. Experiments across various models and domains show that CaOPD achieves Pareto-optimal calibration while maintaining competitive capability, generalizing robustly under out-of-distribution and continual learning. Our findings highlight that capability distillation does not imply calibrated confidence, and that confidence should be treated as an essential objective in post-training. Code: https://github.com/SalesforceAIResearch/CaOPD

Jianfei Yang, Xiangyu Peng, Kai Wang et al.

Domain Adaptation of Black-box Predictors (DABP) aims to learn a model on an unlabeled target domain supervised by a black-box predictor trained on a source domain. It does not require access to both the source-domain data and the predictor parameters, thus addressing the data privacy and portability issues of standard domain adaptation. Existing DABP approaches mostly rely on model distillation from the black-box predictor, \emph{i.e.}, training the model with its noisy target-domain predictions, which however inevitably introduces the confirmation bias accumulated from the prediction noises. To mitigate such bias, we propose a new method, named BETA, to incorporate knowledge distillation and noisy label learning into one coherent framework. This is enabled by a new divide-to-adapt strategy. BETA divides the target domain into an easy-to-adapt subdomain with less noise and a hard-to-adapt subdomain. Then it deploys mutually-teaching twin networks to filter the predictor errors for each other and improve them progressively, from the easy to hard subdomains. As such, BETA effectively purifies the noisy labels and reduces error accumulation. We theoretically show that the target error of BETA is minimized by decreasing the noise ratio of the subdomains. Extensive experiments demonstrate BETA outperforms existing methods on all DABP benchmarks, and is even comparable with the standard domain adaptation methods that use the source-domain data.

Kai Wang, Bo Zhao, Xiangyu Peng et al.

Face recognition, as one of the most successful applications in artificial intelligence, has been widely used in security, administration, advertising, and healthcare. However, the privacy issues of public face datasets have attracted increasing attention in recent years. Previous works simply mask most areas of faces or synthesize samples using generative models to construct privacy-preserving face datasets, which overlooks the trade-off between privacy protection and data utility. In this paper, we propose a novel framework FaceMAE, where the face privacy and recognition performance are considered simultaneously. Firstly, randomly masked face images are used to train the reconstruction module in FaceMAE. We tailor the instance relation matching (IRM) module to minimize the distribution gap between real faces and FaceMAE reconstructed ones. During the deployment phase, we use trained FaceMAE to reconstruct images from masked faces of unseen identities without extra training. The risk of privacy leakage is measured based on face retrieval between reconstructed and original datasets. Experiments prove that the identities of reconstructed images are difficult to be retrieved. We also perform sufficient privacy-preserving face recognition on several public face datasets (i.e. CASIA-WebFace and WebFace260M). Compared to previous state of the arts, FaceMAE consistently \textbf{reduces at least 50\% error rate} on LFW, CFP-FP and AgeDB.

SII-OpenMOSS Team, Donghua Yu, Mingshu Chen et al.

Audio is indispensable for real-world video, yet generation models have largely overlooked audio components. Current approaches to producing audio-visual content often rely on cascaded pipelines, which increase cost, accumulate errors, and degrade overall quality. While systems such as Veo 3 and Sora 2 emphasize the value of simultaneous generation, joint multimodal modeling introduces unique challenges in architecture, data, and training. Moreover, the closed-source nature of existing systems limits progress in the field. In this work, we introduce MOVA (MOSS Video and Audio), an open-source model capable of generating high-quality, synchronized audio-visual content, including realistic lip-synced speech, environment-aware sound effects, and content-aligned music. MOVA employs a Mixture-of-Experts (MoE) architecture, with a total of 32B parameters, of which 18B are active during inference. It supports IT2VA (Image-Text to Video-Audio) generation task. By releasing the model weights and code, we aim to advance research and foster a vibrant community of creators. The released codebase features comprehensive support for efficient inference, LoRA fine-tuning, and prompt enhancement.

Kai Wang, Xiangyu Peng, Shuo Yang et al.

Learning with noisy labels has aroused much research interest since data annotations, especially for large-scale datasets, may be inevitably imperfect. Recent approaches resort to a semi-supervised learning problem by dividing training samples into clean and noisy sets. This paradigm, however, is prone to significant degeneration under heavy label noise, as the number of clean samples is too small for conventional methods to behave well. In this paper, we introduce a novel framework, termed as LC-Booster, to explicitly tackle learning under extreme noise. The core idea of LC-Booster is to incorporate label correction into the sample selection, so that more purified samples, through the reliable label correction, can be utilized for training, thereby alleviating the confirmation bias. Experiments show that LC-Booster advances state-of-the-art results on several noisy-label benchmarks, including CIFAR-10, CIFAR-100, Clothing1M and WebVision. Remarkably, under the extreme 90\% noise ratio, LC-Booster achieves 92.9\% and 48.4\% accuracy on CIFAR-10 and CIFAR-100, surpassing state-of-the-art methods by a large margin.

Jonathan Balloch, Zhiyu Lin, Mustafa Hussain et al. · gatech

A robust body of reinforcement learning techniques have been developed to solve complex sequential decision making problems. However, these methods assume that train and evaluation tasks come from similarly or identically distributed environments. This assumption does not hold in real life where small novel changes to the environment can make a previously learned policy fail or introduce simpler solutions that might never be found. To that end we explore the concept of {\em novelty}, defined in this work as the sudden change to the mechanics or properties of environment. We provide an ontology of for novelties most relevant to sequential decision making, which distinguishes between novelties that affect objects versus actions, unary properties versus non-unary relations, and the distribution of solutions to a task. We introduce NovGrid, a novelty generation framework built on MiniGrid, acting as a toolkit for rapidly developing and evaluating novelty-adaptation-enabled reinforcement learning techniques. Along with the core NovGrid we provide exemplar novelties aligned with our ontology and instantiate them as novelty templates that can be applied to many MiniGrid-compliant environments. Finally, we present a set of metrics built into our framework for the evaluation of novelty-adaptation-enabled machine-learning techniques, and show characteristics of a baseline RL model using these metrics.

Jonathan Balloch, Zhiyu Lin, Robert Wright et al. · gatech

Open-world novelty--a sudden change in the mechanics or properties of an environment--is a common occurrence in the real world. Novelty adaptation is an agent's ability to improve its policy performance post-novelty. Most reinforcement learning (RL) methods assume that the world is a closed, fixed process. Consequentially, RL policies adapt inefficiently to novelties. To address this, we introduce WorldCloner, an end-to-end trainable neuro-symbolic world model for rapid novelty adaptation. WorldCloner learns an efficient symbolic representation of the pre-novelty environment transitions, and uses this transition model to detect novelty and efficiently adapt to novelty in a single-shot fashion. Additionally, WorldCloner augments the policy learning process using imagination-based adaptation, where the world model simulates transitions of the post-novelty environment to help the policy adapt. By blending ''imagined'' transitions with interactions in the post-novelty environment, performance can be recovered with fewer total environment interactions. Using environments designed for studying novelty in sequential decision-making problems, we show that the symbolic world model helps its neural policy adapt more efficiently than model-based and model-based neural-only reinforcement learning methods.

Christopher Cui, Xiangyu Peng, Mark Riedl · gatech

Text-adventure games and text role-playing games are grand challenges for reinforcement learning game playing agents. Text role-playing games are open-ended environments where an agent must faithfully play a particular character. We consider the distinction between characters and actors, where an actor agent has the ability to play multiple characters. We present a framework we call a thespian agent that can learn to emulate multiple characters along with a soft prompt that can be used to direct it as to which character to play at any time. We further describe an attention mechanism that allows the agent to learn new characters that are based on previously learned characters in a few-shot fashion. We show that our agent outperforms the state of the art agent framework in multi-character learning and few-shot learning.

Xiangyu Peng, Can Qin, An Yan et al.

Multimodal large language models (MLLMs) have demonstrated strong capabilities in visual understanding, yet they remain limited in complex, multi-step reasoning that requires deep searching and integrating visual evidence with external knowledge. In this work, we address this challenge by constructing high-quality, verified multi-hop vision-language training data for multimodal deep-search agents. We propose a Multi-hop Tool-Augmented Agent for Evidence-based QA Synthesis (MTA-Agent), which automatically selects tools and their parameters to retrieve and validate evidence from both visual and textual sources and generates structured multi-hop question-answer trajectories. Starting from diverse VQA seed datasets, our pipeline produces a large-scale training dataset, MTA-Vision-DeepSearch, containing 21K high-quality multi-hop examples. The data is filtered through a multi-stage verification process to ensure factual consistency and answer uniqueness. Using MTA-Vision-DeepSearch, a 32B open-source multimodal search agent achieves state-of-the-art performance, reaching an average of 54.63\% across six challenging benchmarks, outperforming GPT-5 (51.86\%), Gemini-2.5-Pro (50.98\%), and Gemini-3-Pro (54.46\%) under the same tool settings. We further show that training on our data improves both reasoning depth and tool-use behavior, increasing the average number of steps from 2.27 to 4.28, and leading to more systematic and persistent search strategies. Additionally, we demonstrate that training can be performed without real-time tool calls by replaying cached interactions, significantly reducing training cost. Importantly, we present MTA-Agent as a fully open recipe for multimodal deep search: we release the entire dataset, training trajectories, and implementation details to enable reproducibility and future research on open multimodal search agents.

Zangwei Zheng, Xiangyu Peng, Tianji Yang et al.

Vision and language are the two foundational senses for humans, and they build up our cognitive ability and intelligence. While significant breakthroughs have been made in AI language ability, artificial visual intelligence, especially the ability to generate and simulate the world we see, is far lagging behind. To facilitate the development and accessibility of artificial visual intelligence, we created Open-Sora, an open-source video generation model designed to produce high-fidelity video content. Open-Sora supports a wide spectrum of visual generation tasks, including text-to-image generation, text-to-video generation, and image-to-video generation. The model leverages advanced deep learning architectures and training/inference techniques to enable flexible video synthesis, which could generate video content of up to 15 seconds, up to 720p resolution, and arbitrary aspect ratios. Specifically, we introduce Spatial-Temporal Diffusion Transformer (STDiT), an efficient diffusion framework for videos that decouples spatial and temporal attention. We also introduce a highly compressive 3D autoencoder to make representations compact and further accelerate training with an ad hoc training strategy. Through this initiative, we aim to foster innovation, creativity, and inclusivity within the community of AI content creation. By embracing the open-source principle, Open-Sora democratizes full access to all the training/inference/data preparation codes as well as model weights. All resources are publicly available at: https://github.com/hpcaitech/Open-Sora.

Prafulla Kumar Choubey, Kung-Hsiang Huang, Pranav Narayanan Venkit et al.

Enterprise deep research often fails to produce decision-ready reports due to uneven information coverage, context explosion, and premature stopping. We propose a scalable Enterprise Deep Research (EDR) architecture to address these failures. Our system (i) decomposes requests into coverage-driven objectives via outline generation with reflection, (ii) localizes context with dependency-guided execution and explicit information sharing, and (iii) enforces evidence-based completion criteria so agents iteratively collect information until sufficiency conditions are met. We evaluate on an internal sales enablement task and the public DeepResearch Bench benchmark, where our proposed system design achieves the strongest overall performance compared with competitive deep-research baselines. The results show that dependency-controlled context and explicit evidence sufficiency criteria reduce premature stopping and improve the consistency and depth of enterprise research outputs.

Xiangyu Peng, Zangwei Zheng, Chenhui Shen et al.

Video generation models have achieved remarkable progress in the past year. The quality of AI video continues to improve, but at the cost of larger model size, increased data quantity, and greater demand for training compute. In this report, we present Open-Sora 2.0, a commercial-level video generation model trained for only $200k. With this model, we demonstrate that the cost of training a top-performing video generation model is highly controllable. We detail all techniques that contribute to this efficiency breakthrough, including data curation, model architecture, training strategy, and system optimization. According to human evaluation results and VBench scores, Open-Sora 2.0 is comparable to global leading video generation models including the open-source HunyuanVideo and the closed-source Runway Gen-3 Alpha. By making Open-Sora 2.0 fully open-source, we aim to democratize access to advanced video generation technology, fostering broader innovation and creativity in content creation. All resources are publicly available at: https://github.com/hpcaitech/Open-Sora.

Xiangyu Peng, Kai Wang, Jianfei Yang et al.

With the advance of deep learning, much progress has been made in building powerful artificial intelligence (AI) systems for automatic Chest X-ray (CXR) analysis. Most existing AI models are trained to be a binary classifier with the aim of distinguishing positive and negative cases. However, a large gap exists between the simple binary setting and complicated real-world medical scenarios. In this work, we reinvestigate the problem of automatic radiology diagnosis. We first observe that there is considerable diversity among cases within the positive class, which means simply classifying them as positive loses many important details. This motivates us to build AI models that can communicate fine-grained knowledge from medical images like human experts. To this end, we first propose a new benchmark on fine granularity learning from medical images. Specifically, we devise a division rule based on medical knowledge to divide positive cases into two subcategories, namely atypical positive and typical positive. Then, we propose a new metric termed AUC$^\text{FG}$ on the two subcategories for evaluation of the ability to separate them apart. With the proposed benchmark, we encourage the community to develop AI diagnosis systems that could better learn fine granularity from medical images. Last, we propose a simple risk modulation approach to this problem by only using coarse labels in training. Empirical results show that despite its simplicity, the proposed method achieves superior performance and thus serves as a strong baseline.

Kai Wang, Bo Zhao, Xiangyu Peng et al.

Dataset condensation aims at reducing the network training effort through condensing a cumbersome training set into a compact synthetic one. State-of-the-art approaches largely rely on learning the synthetic data by matching the gradients between the real and synthetic data batches. Despite the intuitive motivation and promising results, such gradient-based methods, by nature, easily overfit to a biased set of samples that produce dominant gradients, and thus lack global supervision of data distribution. In this paper, we propose a novel scheme to Condense dataset by Aligning FEatures (CAFE), which explicitly attempts to preserve the real-feature distribution as well as the discriminant power of the resulting synthetic set, lending itself to strong generalization capability to various architectures. At the heart of our approach is an effective strategy to align features from the real and synthetic data across various scales, while accounting for the classification of real samples. Our scheme is further backed up by a novel dynamic bi-level optimization, which adaptively adjusts parameter updates to prevent over-/under-fitting. We validate the proposed CAFE across various datasets, and demonstrate that it generally outperforms the state of the art: on the SVHN dataset, for example, the performance gain is up to 11%. Extensive experiments and analyses verify the effectiveness and necessity of proposed designs.

Xiangyu Peng, Jessica Quaye, Sudha Rao et al.

We explore how interaction with large language models (LLMs) can give rise to emergent behaviors, empowering players to participate in the evolution of game narratives. Our testbed is a text-adventure game in which players attempt to solve a mystery under a fixed narrative premise, but can freely interact with non-player characters generated by GPT-4, a large language model. We recruit 28 gamers to play the game and use GPT-4 to automatically convert the game logs into a node-graph representing the narrative in the player's gameplay. We find that through their interactions with the non-deterministic behavior of the LLM, players are able to discover interesting new emergent nodes that were not a part of the original narrative but have potential for being fun and engaging. Players that created the most emergent nodes tended to be those that often enjoy games that facilitate discovery, exploration and experimentation.

Fan Yin, Philippe Laban, Xiangyu Peng et al. · microsoft-research

Malicious content generated by large language models (LLMs) can pose varying degrees of harm. Although existing LLM-based moderators can detect harmful content, they struggle to assess risk levels and may miss lower-risk outputs. Accurate risk assessment allows platforms with different safety thresholds to tailor content filtering and rejection. In this paper, we introduce per-topic severity rubrics for 11 harmful topics and build BingoGuard, an LLM-based moderation system designed to predict both binary safety labels and severity levels. To address the lack of annotations on levels of severity, we propose a scalable generate-then-filter framework that first generates responses across different severity levels and then filters out low-quality responses. Using this framework, we create BingoGuardTrain, a training dataset with 54,897 examples covering a variety of topics, response severity, styles, and BingoGuardTest, a test set with 988 examples explicitly labeled based on our severity rubrics that enables fine-grained analysis on model behaviors on different severity levels. Our BingoGuard-8B, trained on BingoGuardTrain, achieves the state-of-the-art performance on several moderation benchmarks, including WildGuardTest and HarmBench, as well as BingoGuardTest, outperforming best public models, WildGuard, by 4.3\%. Our analysis demonstrates that incorporating severity levels into training significantly enhances detection performance and enables the model to effectively gauge the severity of harmful responses.

Prafulla Kumar Choubey, Xin Su, Man Luo et al.

Knowledge graphs (KGs) generated by large language models (LLMs) are becoming increasingly valuable for Retrieval-Augmented Generation (RAG) applications that require knowledge-intensive reasoning. However, existing KG extraction methods predominantly rely on prompt-based approaches, which are inefficient for processing large-scale corpora. These approaches often suffer from information loss, particularly with long documents, due to the lack of specialized design for KG construction. Additionally, there is a gap in evaluation datasets and methodologies for ontology-free KG construction. To overcome these limitations, we propose SynthKG, a multi-step, document-level ontology-free KG synthesis workflow based on LLMs. By fine-tuning a smaller LLM on the synthesized document-KG pairs, we streamline the multi-step process into a single-step KG generation approach called Distill-SynthKG, substantially reducing the number of LLM inference calls. Furthermore, we re-purpose existing question-answering datasets to establish KG evaluation datasets and introduce new evaluation metrics. Using KGs produced by Distill-SynthKG, we also design a novel graph-based retrieval framework for RAG. Experimental results demonstrate that Distill-SynthKG not only surpasses all baseline models in KG quality -- including models up to eight times larger -- but also consistently excels in retrieval and question-answering tasks. Our proposed graph retrieval framework also outperforms all KG-retrieval methods across multiple benchmark datasets. We release the SynthKG dataset and Distill-SynthKG model publicly to support further research and development.

Prafulla Kumar Choubey, Xiangyu Peng, Shilpa Bhagavath et al.

We present a new benchmark for evaluating Deep Search--a realistic and complex form of retrieval-augmented generation (RAG) that requires source-aware, multi-hop reasoning over diverse, sparsed, but related sources. These include documents, meeting transcripts, Slack messages, GitHub, and URLs, which vary in structure and often contain human-to-human interactions. We build it using a synthetic data pipeline that simulates business workflows across product planning, development, and support stages, generating interconnected content with realistic noise and multi-hop questions with guaranteed ground-truth answers. We release our benchmark with both answerable and unanswerable queries, and retrieval pool of 39,190 enterprise artifacts, enabling fine-grained evaluation of long-context LLM and RAG systems. Our experiments reveal that even the best-performing agentic RAG methods achieve an average performance score of 32.96 on our benchmark. With further analysis, we highlight retrieval as the main bottleneck: existing methods struggle to conduct deep searches and retrieve all necessary evidence. Consequently, they often reason over partial context, leading to significant performance degradation.

Prafulla Kumar Choubey, Xiangyu Peng, Shilpa Bhagavath et al.

Automated service agents require well-structured workflows to provide consistent and accurate responses to customer queries. However, these workflows are often undocumented, and their automatic extraction from conversations remains unexplored. In this work, we present a novel framework for extracting and evaluating dialog workflows from historical interactions. Our extraction process consists of two key stages: (1) a retrieval step to select relevant conversations based on key procedural elements, and (2) a structured workflow generation process using a question-answer-based chain-of-thought (QA-CoT) prompting. To comprehensively assess the quality of extracted workflows, we introduce an automated agent and customer bots simulation framework that measures their effectiveness in resolving customer issues. Extensive experiments on the ABCD and SynthABCD datasets demonstrate that our QA-CoT technique improves workflow extraction by 12.16\% in average macro accuracy over the baseline. Moreover, our evaluation method closely aligns with human assessments, providing a reliable and scalable framework for future research.

Xiangyu Peng, Prafulla Kumar Choubey, Caiming Xiong et al.

Existing evaluation frameworks for retrieval-augmented generation (RAG) systems focus on answerable queries, but they overlook the importance of appropriately rejecting unanswerable requests. In this paper, we introduce UAEval4RAG, a framework designed to evaluate whether RAG systems can handle unanswerable queries effectively. We define a taxonomy with six unanswerable categories, and UAEval4RAG automatically synthesizes diverse and challenging queries for any given knowledge base with unanswered ratio and acceptable ratio metrics. We conduct experiments with various RAG components, including retrieval models, rewriting methods, rerankers, language models, and prompting strategies, and reveal hidden trade-offs in performance of RAG systems. Our findings highlight the critical role of component selection and prompt design in optimizing RAG systems to balance the accuracy of answerable queries with high rejection rates of unanswerable ones. UAEval4RAG provides valuable insights and tools for developing more robust and reliable RAG systems.

Xiangyu Peng, Can Qin, Zeyuan Chen et al.

Multimodal retrieval-augmented generation (MM-RAG) is a key approach for applying large language models (LLMs) and agents to real-world knowledge bases, yet current evaluations are fragmented, focusing on either text or images in isolation or on simplified multimodal setups that fail to capture document-centric multimodal use cases. In this paper, we introduce UniDoc-Bench, the first large-scale, realistic benchmark for MM-RAG built from 70k real-world PDF pages across eight domains. Our pipeline extracts and links evidence from text, tables, and figures, then generates 1,600 multimodal QA pairs spanning factual retrieval, comparison, summarization, and logical reasoning queries. To ensure reliability, 20% of QA pairs are validated by multiple annotators and expert adjudication. UniDoc-Bench supports apples-to-apples comparison across four paradigms: (1) text-only, (2) image-only, (3) multimodal text-image fusion, and (4) multimodal joint retrieval -- under a unified protocol with standardized candidate pools, prompts, and evaluation metrics. Our experiments show that multimodal text-image fusion RAG systems consistently outperform both unimodal and jointly multimodal embedding-based retrieval, indicating that neither text nor images alone are sufficient and that current multimodal embeddings remain inadequate. Beyond benchmarking, our analysis reveals when and how visual context complements textual evidence, uncovers systematic failure modes, and offers actionable guidance for developing more robust MM-RAG pipelines.

Honglu Zhou, Xiangyu Peng, Shrikant Kendre et al. · salesforce, stanford

Next-generation AI companions must go beyond general video understanding to resolve spatial and temporal references in dynamic, real-world environments. Existing Video Large Language Models (Video LLMs), while capable of coarse-level comprehension, struggle with fine-grained, spatiotemporal reasoning, especially when user queries rely on time-based event references for temporal anchoring, or gestural cues for spatial anchoring to clarify object references and positions. To bridge this critical gap, we introduce Strefer, a synthetic instruction data generation framework designed to equip Video LLMs with spatiotemporal referring and reasoning capabilities. Strefer produces diverse instruction-tuning data using a data engine that pseudo-annotates temporally dense, fine-grained video metadata, capturing rich spatial and temporal information in a structured manner, including subjects, objects, their locations as masklets, and their action descriptions and timelines. Our approach enhances the ability of Video LLMs to interpret spatial and temporal references, fostering more versatile, space-time-aware reasoning essential for real-world AI companions. Without using proprietary models, costly human annotation, or the need to annotate large volumes of new videos, experimental evaluations show that models trained with data produced by Strefer outperform baselines on tasks requiring spatial and temporal disambiguation. Additionally, these models exhibit enhanced space-time-aware reasoning, establishing a new foundation for perceptually grounded, instruction-tuned Video LLMs.

Claire Jin, Sudha Rao, Xiangyu Peng et al.

Advancements in large language models (LLMs) are revolutionizing interactive game design, enabling dynamic plotlines and interactions between players and non-player characters (NPCs). However, LLMs may exhibit flaws such as hallucinations, forgetfulness, or misinterpretations of prompts, causing logical inconsistencies and unexpected deviations from intended designs. Automated techniques for detecting such game bugs are still lacking. To address this, we propose a systematic LLM-based method for automatically identifying such bugs from player game logs, eliminating the need for collecting additional data such as post-play surveys. Applied to a text-based game DejaBoom!, our approach effectively identifies bugs inherent in LLM-powered interactive games, surpassing unstructured LLM-powered bug-catching methods and filling the gap in automated detection of logical and design flaws.

Christopher Z. Cui, Xiangyu Peng, Mark O. Riedl

Open-ended worlds are those in which there are no pre-specified goals or environmental reward signal. As a consequence, an agent must know how to perform a multitude of tasks. However, when a new task is presented to an agent, we expect it to be able to reuse some of what it knows from previous tasks to rapidly learn that new task. We introduce a novel technique whereby policies for different a priori known tasks are combined into a Mixture-of-Experts model with an attention mechanism across a mix of frozen and unfrozen experts. The model learns when to attend to frozen task-specific experts when appropriate and learns new experts to handle novel situations. We work in an open-ended text-based environment in which the agent is tasked with behaving like different types of character roles and must rapidly learn behaviors associated with new character role types. We show that our agent both obtains more rewards in the zero-shot setting, and discovers these rewards with greater sample efficiency in the few-shot learning settings.

Xiangyu Peng, Michael Sollami

Multimodal conditionality in transformer-based natural language models has demonstrated state-of-the-art performance in the task of product description generation. Recent approaches condition a language model on one or more images and other textual metadata to achieve near-human performance for describing products from e-commerce stores. However, generated descriptions may exhibit degrees of inaccuracy or even contradictory claims relative to the inputs of a given product. In this paper, we propose a controllable language generation framework called Extract-Finetune-Boost (XFBoost), which addresses the problem of inaccurate low-quality inference. By using visual semantic attributes as constraints at the decoding stage of the generation process and finetuning the language model with policy gradient techniques, the XFBoost framework is found to produce significantly more descriptive text with higher image relevancy, outperforming baselines and lowering the frequency of factually inaccurate descriptions. We further demonstrate the application of XFBoost to online learning wherein human-in-the-loop critics improve language models with active feedback.

Xiangyu Peng, Kai Wang, Zheng Zhu et al.

Recent self-supervised contrastive learning methods greatly benefit from the Siamese structure that aims at minimizing distances between positive pairs. For high performance Siamese representation learning, one of the keys is to design good contrastive pairs. Most previous works simply apply random sampling to make different crops of the same image, which overlooks the semantic information that may degrade the quality of views. In this work, we propose ContrastiveCrop, which could effectively generate better crops for Siamese representation learning. Firstly, a semantic-aware object localization strategy is proposed within the training process in a fully unsupervised manner. This guides us to generate contrastive views which could avoid most false positives (i.e., object vs. background). Moreover, we empirically find that views with similar appearances are trivial for the Siamese model training. Thus, a center-suppressed sampling is further designed to enlarge the variance of crops. Remarkably, our method takes a careful consideration of positive pairs for contrastive learning with negligible extra training overhead. As a plug-and-play and framework-agnostic module, ContrastiveCrop consistently improves SimCLR, MoCo, BYOL, SimSiam by 0.4% ~ 2.0% classification accuracy on CIFAR-10, CIFAR-100, Tiny ImageNet and STL-10. Superior results are also achieved on downstream detection and segmentation tasks when pre-trained on ImageNet-1K.

Xiangyu Peng, Mark O. Riedl, Prithviraj Ammanabrolu

We focus on the task of creating a reinforcement learning agent that is inherently explainable -- with the ability to produce immediate local explanations by thinking out loud while performing a task and analyzing entire trajectories post-hoc to produce causal explanations. This Hierarchically Explainable Reinforcement Learning agent (HEX-RL), operates in Interactive Fictions, text-based game environments in which an agent perceives and acts upon the world using textual natural language. These games are usually structured as puzzles or quests with long-term dependencies in which an agent must complete a sequence of actions to succeed -- providing ideal environments in which to test an agent's ability to explain its actions. Our agent is designed to treat explainability as a first-class citizen, using an extracted symbolic knowledge graph-based state representation coupled with a Hierarchical Graph Attention mechanism that points to the facts in the internal graph representation that most influenced the choice of actions. Experiments show that this agent provides significantly improved explanations over strong baselines, as rated by human participants generally unfamiliar with the environment, while also matching state-of-the-art task performance.

Xiangyu Peng, Kaige Xie, Amal Alabdulkarim et al.

Automated storytelling has long captured the attention of researchers for the ubiquity of narratives in everyday life. However, it is challenging to maintain coherence and stay on-topic toward a specific ending when generating narratives with neural language models. In this paper, we introduce Story generation with Reader Models (StoRM), a framework in which a reader model is used to reason about the story should progress. A reader model infers what a human reader believes about the concepts, entities, and relations about the fictional story world. We show how an explicit reader model represented as a knowledge graph affords story coherence and provides controllability in the form of achieving a given story world state goal. Experiments show that our model produces significantly more coherent and on-topic stories, outperforming baselines in dimensions including plot plausibility and staying on topic.

Xiangyu Peng, Jonathan C. Balloch, Mark O. Riedl

Open-world novelty occurs when the rules of an environment can change abruptly, such as when a game player encounters "house rules". To address open-world novelty, game playing agents must be able to detect when novelty is injected, and to quickly adapt to the new rules. We propose a model-based reinforcement learning approach where game state and rules are represented as knowledge graphs. The knowledge graph representation of the state and rules allows novelty to be detected as changes in the knowledge graph, assists with the training of deep reinforcement learners, and enables imagination-based re-training where the agent uses the knowledge graph to perform look-ahead.

Xiangyu Peng, Siyan Li, Sarah Wiegreffe et al.

Transformer-based language model approaches to automated story generation currently provide state-of-the-art results. However, they still suffer from plot incoherence when generating narratives over time, and critically lack basic commonsense reasoning. Furthermore, existing methods generally focus only on single-character stories, or fail to track characters at all. To improve the coherence of generated narratives and to expand the scope of character-centric narrative generation, we introduce Commonsense-inference Augmented neural StoryTelling (CAST), a framework for introducing commonsense reasoning into the generation process with the option to model the interaction between multiple characters. We find that our CAST method produces significantly more coherent, on-topic, enjoyable and fluent stories than existing models in both the single-character and two-character settings in three storytelling domains.

Amal Alabdulkarim, Siyan Li, Xiangyu Peng

The scope of this survey paper is to explore the challenges in automatic story generation. We hope to contribute in the following ways: 1. Explore how previous research in story generation addressed those challenges. 2. Discuss future research directions and new technologies that may aid more advancements. 3. Shed light on emerging and often overlooked challenges such as creativity and discourse.

Xiangyu Peng, Siyan Li, Spencer Frazier et al.

Large-scale, transformer-based language models such as GPT-2 are pretrained on diverse corpora scraped from the internet. Consequently, they are prone to generating non-normative text (i.e. in violation of social norms). We introduce a technique for fine-tuning GPT-2, using a policy gradient reinforcement learning technique and a normative text classifier to produce reward and punishment values. We evaluate our technique on five data sets using automated and human participant experiments. The normative text classifier is 81-90% accurate when compared to gold-standard human judgments of normative and non-normative generated text. Our normative fine-tuning technique is able to reduce non-normative text by 27-61%, depending on the data set.