Bowen Jiang

Bowen Jiang, Zhijun Zhuang, Shreyas S. Shivakumar et al.

This work introduces an enhanced approach to generating scene graphs by incorporating both a relationship hierarchy and commonsense knowledge. Specifically, we begin by proposing a hierarchical relation head that exploits an informative hierarchical structure. It jointly predicts the relation super-category between object pairs in an image, along with detailed relations under each super-category. Following this, we implement a robust commonsense validation pipeline that harnesses foundation models to critique the results from the scene graph prediction system, removing nonsensical predicates even with a small language-only model. Extensive experiments on Visual Genome and OpenImage V6 datasets demonstrate that the proposed modules can be seamlessly integrated as plug-and-play enhancements to existing scene graph generation algorithms. The results show significant improvements with an extensive set of reasonable predictions beyond dataset annotations. Codes are available at https://github.com/bowen-upenn/scene_graph_commonsense.

Bowen Jiang, Camillo J. Taylor

This paper presents a finding that leveraging the hierarchical structures among labels for relationships and objects can substantially improve the performance of scene graph generation systems. The focus of this work is to create an informative hierarchical structure that can divide object and relationship categories into disjoint super-categories in a systematic way. Specifically, we introduce a Bayesian prediction head to jointly predict the super-category of relationships between a pair of object instances, as well as the detailed relationship within that super-category simultaneously, facilitating more informative predictions. The resulting model exhibits the capability to produce a more extensive set of predicates beyond the dataset annotations, and to tackle the prevalent issue of low annotation quality. While our paper presents preliminary findings, experiments on the Visual Genome dataset show its strong performance, particularly in predicate classifications and zero-shot settings, that demonstrates the promise of our approach.

Bowen Jiang, Yuan Yuan, Maohao Shen et al. · uw

Personalization is one of the next milestones in advancing AI capability and alignment. We introduce PersonaMem-v2, the state-of-the-art dataset for LLM personalization that simulates 1,000 realistic user-chatbot interactions on 300+ scenarios, 20,000+ user preferences, and 128k-token context windows, where most user preferences are implicitly revealed to reflect real-world interactions. Using this data, we investigate how reinforcement fine-tuning enables a model to improve its long-context reasoning capabilities for user understanding and personalization. We also develop a framework for training an agentic memory system, which maintains a single, human-readable memory that grows with each user over time. In our experiments, frontier LLMs still struggle with implicit personalization, achieving only 37-48% accuracy. While they support long context windows, reasoning remains the bottleneck for implicit personalization tasks. Using reinforcement fine-tuning, we successfully train Qwen3-4B to outperforms GPT-5, reaching 53% accuracy in implicit personalization. Moreover, our agentic memory framework achieves state-of-the-art 55% accuracy while using 16x fewer input tokens, relying on a 2k-token memory instead of full 32k conversation histories. These results underscore the impact of our dataset and demonstrate agentic memory as a scalable path toward real-world personalized intelligence.

Maohao Shen, Bowen Jiang, Jacky Yibo Zhang et al.

Active learning enables efficient model training by leveraging interactions between machine learning agents and human annotators. We study and propose a novel framework that formulates batch active learning from the sparse approximation's perspective. Our active learning method aims to find an informative subset from the unlabeled data pool such that the corresponding training loss function approximates its full data pool counterpart. We realize the framework as sparsity-constrained discontinuous optimization problems, which explicitly balance uncertainty and representation for large-scale applications and could be solved by greedy or proximal iterative hard thresholding algorithms. The proposed method can adapt to various settings, including both Bayesian and non-Bayesian neural networks. Numerical experiments show that our work achieves competitive performance across different settings with lower computational complexity.

Yiming Gao, Jieming Yin, Yuxiang Wang et al.

Existing Point Cloud Networks (PCNs) have proven to achieve great success in many point cloud tasks such as object part segmentation, shape classification, and so on. The most popular point-based PCNs are usually composed of two sequential steps: Data Structuring (DS) and Feature Computation (FC). In this paper, we first describe an important characteristic of the PCN-specific DS step that has not been addressed in existing PCN accelerators: the spatial locality resulting from overlapping points of the gathered point subsets. Using algorithm-hardware co-design, L-PCN (Locality-aware PCN) proposes two novel techniques to exploit this characteristic to reduce the large amount of repetitive operations in the overall PCN. The first of which is a point cloud partitioning technique, Octree-based Islandization. Using Octree-based adjacency gathering, a point cloud is partitioned into islands in L-PCN, where the point subsets inside the same island exhibit a strong spatial correlation. After partitioning, L-PCN performs the rest of PCN steps at the granularity of islands. The second method of L-PCN is scheduling the intra-island computation with a Hub-based Scheduling to exploit the intra-island data reuse by dynamically caching, updating, and reusing the repeated data. The two methods are implemented in an Islandization Unit, which can be seamlessly integrated into standard PCN workflow. Our evaluation shows that based on our methods for exploiting spatial locality, L-PCN achieves a theoretical reduction in feature fetching ranging from 55.2% to 93.8% and in feature computation ranging from 45.4% to 80.6% during the PCN process. For experimentation, prototype L-PCN accelerators are implemented on the Intel Arria 10 GX FPGA. Experimental results prove that with the Islandization Unit as a plug-in, state-of-the-art PCN accelerators can achieve an additional speedup ranging from 1.2x to 3.2x.

Carl Qi, Yilin Wu, Lifan Yu et al.

Autonomous systems that efficiently utilize tools can assist humans in completing many common tasks such as cooking and cleaning. However, current systems fall short of matching human-level of intelligence in terms of adapting to novel tools. Prior works based on affordance often make strong assumptions about the environments and cannot scale to more complex, contact-rich tasks. In this work, we tackle this challenge and explore how agents can learn to use previously unseen tools to manipulate deformable objects. We propose to learn a generative model of the tool-use trajectories as a sequence of tool point clouds, which generalizes to different tool shapes. Given any novel tool, we first generate a tool-use trajectory and then optimize the sequence of tool poses to align with the generated trajectory. We train a single model on four different challenging deformable object manipulation tasks, using demonstration data from only one tool per task. The model generalizes to various novel tools, significantly outperforming baselines. We further test our trained policy in the real world with unseen tools, where it achieves the performance comparable to human. Additional materials can be found on our project website: https://sites.google.com/view/toolgen.

Ruiying Li, Yunlang Zhou, YuYao Zhu et al.

Vision-Language-Action (VLA) systems have shown strong potential for language-driven robotic manipulation. However, scaling them to long-horizon tasks remains challenging. Existing pipelines typically separate data collection, policy learning, and deployment, resulting in heavy reliance on manual environment resets and brittle multi-policy execution. We present RoboClaw, an agentic robotics framework that unifies data collection, policy learning, and task execution under a single VLM-driven controller. At the policy level, RoboClaw introduces Entangled Action Pairs (EAP), which couple forward manipulation behaviors with inverse recovery actions to form self-resetting loops for autonomous data collection. This mechanism enables continuous on-policy data acquisition and iterative policy refinement with minimal human intervention. During deployment, the same agent performs high-level reasoning and dynamically orchestrates learned policy primitives to accomplish long-horizon tasks. By maintaining consistent contextual semantics across collection and execution, RoboClaw reduces mismatch between the two phases and improves multi-policy robustness. Experiments in real-world manipulation tasks demonstrate improved stability and scalability compared to conventional open-loop pipelines, while significantly reducing human effort throughout the robot lifecycle, achieving a 25% improvement in success rate over baseline methods on long-horizon tasks and reducing human time investment by 53.7%.

Mengti Sun, Bowen Jiang, Bibit Bianchini et al.

This work presents an instance-agnostic learning framework that fuses vision with dynamics to simultaneously learn shape, pose trajectories, and physical properties via the use of geometry as a shared representation. Unlike many contact learning approaches that assume motion capture input and a known shape prior for the collision model, our proposed framework learns an object's geometric and dynamic properties from RGBD video, without requiring either category-level or instance-level shape priors. We integrate a vision system, BundleSDF, with a dynamics system, ContactNets, and propose a cyclic training pipeline to use the output from the dynamics module to refine the poses and the geometry from the vision module, using perspective reprojection. Experiments demonstrate our framework's ability to learn the geometry and dynamics of rigid and convex objects and improve upon the current tracking framework.

Weixing Zhang, Bowen Jiang, Rahul Sharma et al.

In model-driven engineering, metamodel evolution leads to the need to adapt corresponding grammars to maintain consistency, which typically requires tedious manual work. Existing rule-based methods can achieve partial automation but have limitations when handling complex grammar scenarios. This paper proposes a Large Language Model-based approach that automatically applies adaptations to new grammars after evolution by learning grammar adaptations from previous versions. We evaluated this approach on six real-world Xtext domain-specific languages, using four DSLs as a training set to develop prompting strategies, two DSLs as a test set for validation, and conducting a longitudinal case study on QVTo. The evaluation used three Large Language Models (Claude Sonnet 4.5, ChatGPT 5.1, Gemini 3) and measured grammar adaptation quality from three dimensions: grammar rule-level adaptation consistency, output similarity, and metamodel conformance. Results show that on the test set, all three LLMs achieved 100% adaptation consistency and output similarity, while the rule-based approach achieved only 84.21% on DOT and 62.50% on Xcore. In the QVTo longitudinal study, the LLM-based approach successfully reused learned adaptations across all three evolution steps without manual grammar editing, while the rule-based approach required manual adjustments in two of three transitions. However, on large-scale grammars (EAST-ADL, 297 rules), LLMs' adaptation consistency was far below 90%. This study demonstrates the advantages of LLM-based approaches in handling complex grammar scenarios, while revealing their limitations in large-scale grammar adaptation.

Bowen Jiang, Zhuoqun Hao, Young-Min Cho et al.

Large Language Models (LLMs) have emerged as personalized assistants for users across a wide range of tasks -- from offering writing support to delivering tailored recommendations or consultations. Over time, the interaction history between a user and an LLM can provide extensive information about an individual's traits and preferences. However, open questions remain on how well LLMs today can effectively leverage such history to (1) internalize the user's inherent traits and preferences, (2) track how the user profiling and preferences evolve over time, and (3) generate personalized responses accordingly in new scenarios. In this work, we introduce the PERSONAMEM benchmark. PERSONAMEM features curated user profiles with over 180 simulated user-LLM interaction histories, each containing up to 60 sessions of multi-turn conversations across 15 real-world tasks that require personalization. Given an in-situ user query, i.e. query issued by the user from the first-person perspective, we evaluate LLM chatbots' ability to identify the most suitable response according to the current state of the user's profile. We observe that current LLMs still struggle to recognize the dynamic evolution in users' profiles over time through direct prompting approaches. As a consequence, LLMs often fail to deliver responses that align with users' current situations and preferences, with frontier models such as GPT-4.1, o4-mini, GPT-4.5, o1, or Gemini-2.0 achieving only around 50% overall accuracy, suggesting room for improvement. We hope that PERSONAMEM, along with the user profile and conversation simulation pipeline, can facilitate future research in the development of truly user-aware chatbots. Code and data are available at github.com/bowen-upenn/PersonaMem.

Bowen Jiang, Yilin Wu, Wenxuan Zhou et al.

Although end-to-end robot learning has shown some success for robot manipulation, the learned policies are often not sufficiently robust to variations in object pose or geometry. To improve the policy generalization, we introduce spatially-grounded parameterized motion primitives in our method HACMan++. Specifically, we propose an action representation consisting of three components: what primitive type (such as grasp or push) to execute, where the primitive will be grounded (e.g. where the gripper will make contact with the world), and how the primitive motion is executed, such as parameters specifying the push direction or grasp orientation. These three components define a novel discrete-continuous action space for reinforcement learning. Our framework enables robot agents to learn to chain diverse motion primitives together and select appropriate primitive parameters to complete long-horizon manipulation tasks. By grounding the primitives on a spatial location in the environment, our method is able to effectively generalize across object shape and pose variations. Our approach significantly outperforms existing methods, particularly in complex scenarios demanding both high-level sequential reasoning and object generalization. With zero-shot sim-to-real transfer, our policy succeeds in challenging real-world manipulation tasks, with generalization to unseen objects. Videos can be found on the project website: https://sgmp-rss2024.github.io.

Bowen Jiang, Taiwei Shi, Ryo Kamoi et al.

This paper introduces OMAR: One Model, All Roles, a reinforcement learning framework that enables AI to develop social intelligence through multi-turn, multi-agent conversational self-play. Unlike traditional paradigms that rely on static, single-turn optimizations, OMAR allows a single model to role-play all participants in a conversation simultaneously, learning to achieve long-term goals and complex social norms directly from dynamic social interaction. To ensure training stability across long dialogues, we implement a hierarchical advantage estimation that calculates turn-level and token-level advantages. Evaluations in the SOTOPIA social environment and Werewolf strategy games show that our trained models develop fine-grained, emergent social intelligence, such as empathy, persuasion, and compromise seeking, demonstrating the effectiveness of learning collaboration even under competitive scenarios. While we identify practical challenges like reward hacking, our results show that rich social intelligence can emerge without human supervision. We hope this work incentivizes further research on AI social intelligence in group conversations.

Weixing Zhang, Mario Herb, Martin Armbruster et al.

Despite Domain-Driven Design's proven value in managing complex business logic, a fundamental semantic expressiveness gap persists between generic modeling languages and tactical DDD patterns, causing continuous divergence between design intent and implementation. We envision a constraint-based tactical modeling environment that transforms abstract architectural principles into explicit, tool-enforced engineering constraints. At its core is a DDD-native metamodel where tactical patterns are first-class modeling primitives, coupled with a real-time constraint verification engine that prevents architectural violations during modeling, and bidirectional synchronization mechanisms that maintain model-code consistency through round-trip engineering. This approach aims to democratize tactical DDD by embedding expert-level architectural knowledge directly into modeling constraints, enabling small teams and junior developers to build complex business systems without sacrificing long-term maintainability. By lowering the technical barriers to DDD adoption, we envision transforming tactical DDD from an elite practice requiring continuous expert oversight into an accessible engineering discipline with tool-supported verification.

Jiahe Song, Chuang Wang, Yinfan Wang et al.

Reaction diagram parsing (RxnDP) is critical for extracting chemical synthesis information from literature. Although recent Vision-Language Models (VLMs) have emerged as a promising paradigm to automate this complex visual reasoning task, their application is fundamentally bottlenecked by the inability to align visual chemical entities with pre-trained knowledge, alongside the inherent discrepancy between token-level training and reaction-level evaluation. To address these dual challenges, this work enhances VLM-based RxnDP from two complementary perspectives: prompting representation and learning paradigms. First, we propose Identifier as Visual Prompting (IdtVP), which leverages naturally occurring molecule identifiers (e.g., bold numerals like 1a) to activate the chemical knowledge acquired during VLM pre-training. IdtVP enables powerful zero-shot and out-of-distribution capabilities, outperforming existing prompting strategies. Second, to further optimize performance within fine-tuning paradigms, we introduce Re3-DAPO, a reinforcement learning algorithm that leverages verifiable rewards to directly optimize reaction-level metrics, thereby achieving consistent gains over standard supervised fine-tuning. Additionally, we release the ScannedRxn benchmark, comprising scanned historical reaction diagrams with real-world artifacts, to rigorously assess model robustness and out-of-distribution ability. Our contributions advance the accuracy and generalization of VLM-based reaction diagram parsing. We will release data, models, and code on GitHub.

Bowen Jiang, Yuan Yuan, Xinyi Bai et al.

This work demonstrates that diffusion models can achieve font-controllable multilingual text rendering using just raw images without font label annotations.Visual text rendering remains a significant challenge. While recent methods condition diffusion on glyphs, it is impossible to retrieve exact font annotations from large-scale, real-world datasets, which prevents user-specified font control. To address this, we propose a data-driven solution that integrates the conditional diffusion model with a text segmentation model, utilizing segmentation masks to capture and represent fonts in pixel space in a self-supervised manner, thereby eliminating the need for any ground-truth labels and enabling users to customize text rendering with any multilingual font of their choice. The experiment provides a proof of concept of our algorithm in zero-shot text and font editing across diverse fonts and languages, providing valuable insights for the community and industry toward achieving generalized visual text rendering. Code is available at github.com/bowen-upenn/ControlText.

Yuzhen Xiao, Jiahe Song, Yongxin Xu et al.

In-Context Learning (ICL) technique based on Large Language Models (LLMs) has gained prominence in Named Entity Recognition (NER) tasks for its lower computing resource consumption, less manual labeling overhead, and stronger generalizability. Nevertheless, most ICL-based NER methods depend on large-parameter LLMs: the open-source models demand substantial computational resources for deployment and inference, while the closed-source ones incur high API costs, raise data-privacy concerns, and hinder community collaboration. To address this question, we propose an Ensemble Learning Method for Named Entity Recognition (EL4NER), which aims at aggregating the ICL outputs of multiple open-source, small-parameter LLMs to enhance overall performance in NER tasks at less deployment and inference cost. Specifically, our method comprises three key components. First, we design a task decomposition-based pipeline that facilitates deep, multi-stage ensemble learning. Second, we introduce a novel span-level sentence similarity algorithm to establish an ICL demonstration retrieval mechanism better suited for NER tasks. Third, we incorporate a self-validation mechanism to mitigate the noise introduced during the ensemble process. We evaluated EL4NER on multiple widely adopted NER datasets from diverse domains. Our experimental results indicate that EL4NER surpasses most closed-source, large-parameter LLM-based methods at a lower parameter cost and even attains state-of-the-art (SOTA) performance among ICL-based methods on certain datasets. These results show the parameter efficiency of EL4NER and underscore the feasibility of employing open-source, small-parameter LLMs within the ICL paradigm for NER tasks.

Bowen Jiang, Runchuan Zhu, Jiang Wu et al.

We introduce KoLasSimpleQA, the first benchmark evaluating the multilingual factual ability of Large Language Models (LLMs). Inspired by existing research, we created the question set with features such as single knowledge point coverage, absolute objectivity, unique answers, and temporal stability. These questions enable efficient evaluation using the LLM-as-judge paradigm, testing both the LLMs' factual memory and self-awareness ("know what they don't know"). KoLasSimpleQA expands existing research in two key dimensions: (1) Breadth (Multilingual Coverage): It includes 9 languages, supporting global applicability evaluation. (2) Depth (Dual Domain Design): It covers both the general domain (global facts) and the language-specific domain (such as history, culture, and regional traditions) for a comprehensive assessment of multilingual capabilities. We evaluated mainstream LLMs, including traditional LLM and emerging Large Reasoning Models. Results show significant performance differences between the two domains, particularly in performance metrics, ranking, calibration, and robustness. This highlights the need for targeted evaluation and optimization in multilingual contexts. We hope KoLasSimpleQA will help the research community better identify LLM capability boundaries in multilingual contexts and provide guidance for model optimization. We will release KoLasSimpleQA at https://github.com/opendatalab/KoLasSimpleQA .

Haote Yang, Xingjian Wei, Jiang Wu et al.

We introduce OpenHuEval, the first benchmark for LLMs focusing on the Hungarian language and specifics. OpenHuEval is constructed from a vast collection of Hungarian-specific materials sourced from multiple origins. In the construction, we incorporated the latest design principles for evaluating LLMs, such as using real user queries from the internet, emphasizing the assessment of LLMs' generative capabilities, and employing LLM-as-judge to enhance the multidimensionality and accuracy of evaluations. Ultimately, OpenHuEval encompasses eight Hungarian-specific dimensions, featuring five tasks and 3953 questions. Consequently, OpenHuEval provides the comprehensive, in-depth, and scientifically accurate assessment of LLM performance in the context of the Hungarian language and its specifics. We evaluated current mainstream LLMs, including both traditional LLMs and recently developed Large Reasoning Models. The results demonstrate the significant necessity for evaluation and model optimization tailored to the Hungarian language and specifics. We also established the framework for analyzing the thinking processes of LRMs with OpenHuEval, revealing intrinsic patterns and mechanisms of these models in non-English languages, with Hungarian serving as a representative example. We will release OpenHuEval at https://github.com/opendatalab/OpenHuEval .

Jiahe Song, Chuang Wang, Bowen Jiang et al.

Large-scale chemical reaction datasets are crucial for AI research in chemistry. However, existing chemical reaction data often exist as images within papers, making them not machine-readable and unusable for training machine learning models. In response to this challenge, we propose the RxnCaption framework for the task of chemical Reaction Diagram Parsing (RxnDP). Our framework reformulates the traditional coordinate prediction driven parsing process into an image captioning problem, which Large Vision-Language Models (LVLMs) handle naturally. We introduce a strategy termed "BBox and Index as Visual Prompt" (BIVP), which uses our state-of-the-art molecular detector, MolYOLO, to pre-draw molecular bounding boxes and indices directly onto the input image. This turns the downstream parsing into a natural-language description problem. Extensive experiments show that the BIVP strategy significantly improves structural extraction quality while simplifying model design. We further construct the RxnCaption-11k dataset, an order of magnitude larger than prior real-world literature benchmarks, with a balanced test subset across four layout archetypes. Experiments demonstrate that RxnCaption-VL achieves state-of-the-art performance on multiple metrics. We believe our method, dataset, and models will advance structured information extraction from chemical literature and catalyze broader AI applications in chemistry. We will release data, models, and code on GitHub.

Weixing Zhang, Bowen Jiang, Yuhong Fu et al.

Software languages evolve over time for reasons such as feature additions. When grammars evolve, textual instances that originally conformed to them may become outdated. While model-driven engineering provides many techniques for co-evolving models with metamodel changes, these approaches are not designed for textual DSLs and may lose human-relevant information such as layout and comments. This study systematically evaluates the potential of large language models (LLMs) for co-evolving grammars and instances of textual DSLs. Using Claude Sonnet 4.5 and GPT-5.2 across ten case languages with ten runs each, we assess both correctness and preservation of human-oriented information. Results show strong performance on small-scale cases ($\geq$94% precision and recall for instances requiring fewer than 20 modified lines), but performance degraded with scale: Claude maintains 85% recall at 40 lines, while GPT fails on the largest instances. Response time increases substantially with instance size, and grammar evolution complexity and deletion granularity affect performance more than change type. These findings clarify when LLM-based co-evolution is effective and where current limitations remain.

Runchuan Zhu, Bowen Jiang, Lingrui Mei et al.

Recent advances in large language models (LLMs) have sparked growing interest in agentic workflows, which are structured sequences of LLM invocations intended to solve complex tasks. However, existing approaches often rely on static templates or manually designed workflows, which limit adaptability to diverse tasks and hinder scalability. We propose AdaptFlow, a natural language-based meta-learning framework inspired by model-agnostic meta-learning (MAML). AdaptFlow learns a generalizable workflow initialization that enables rapid subtask-level adaptation. It employs a bi-level optimization scheme: the inner loop refines the workflow for a specific subtask using LLM-generated feedback, while the outer loop updates the shared initialization to perform well across tasks. This setup allows AdaptFlow to generalize effectively to unseen tasks by adapting the initialized workflow through language-guided modifications. Evaluated across question answering, code generation, and mathematical reasoning benchmarks, AdaptFlow consistently outperforms both manually crafted and automatically searched baselines, achieving state-of-the-art results with strong generalization across tasks and models. The source code and data are available at https://github.com/microsoft/DKI_LLM/tree/AdaptFlow/AdaptFlow.

Bowen Jiang, Yangxinyu Xie, Zhuoqun Hao et al.

This study introduces a hypothesis-testing framework to assess whether large language models (LLMs) possess genuine reasoning abilities or primarily depend on token bias. We go beyond evaluating LLMs on accuracy; rather, we aim to investigate their token bias in solving logical reasoning tasks. Specifically, we develop carefully controlled synthetic datasets, featuring conjunction fallacy and syllogistic problems. Our framework outlines a list of hypotheses where token biases are readily identifiable, with all null hypotheses assuming genuine reasoning capabilities of LLMs. The findings in this study suggest, with statistical guarantee, that most LLMs still struggle with logical reasoning. While they may perform well on classic problems, their success largely depends on recognizing superficial patterns with strong token bias, thereby raising concerns about their actual reasoning and generalization abilities. Codes and data are open-sourced at https://github.com/bowen-upenn/llm_token_bias.

Bowen Jiang, Yangxinyu Xie, Xiaomeng Wang et al.

This work discusses how to build more rational language and multimodal agents and what criteria define rationality in intelligent systems. Rationality is the quality of being guided by reason, characterized by decision-making that aligns with evidence and logical principles. It plays a crucial role in reliable problem-solving by ensuring well-grounded and consistent solutions. Despite their progress, large language models (LLMs) often fall short of rationality due to their bounded knowledge space and inconsistent outputs. In response, recent efforts have shifted toward developing multimodal and multi-agent systems, as well as integrating modules like external tools, programming codes, symbolic reasoners, utility function, and conformal risk controls rather than relying solely on a single LLM for decision-making. This paper surveys state-of-the-art advancements in language and multimodal agents, assesses their role in enhancing rationality, and outlines open challenges and future research directions. We maintain an open repository at https://github.com/bowen-upenn/Agent_Rationality.

Yangxinyu Xie, Bowen Jiang, Tanwi Mallick et al.

Recent advancement of large language models (LLMs) represents a transformational capability at the frontier of artificial intelligence. However, LLMs are generalized models, trained on extensive text corpus, and often struggle to provide context-specific information, particularly in areas requiring specialized knowledge, such as wildfire details within the broader context of climate change. For decision-makers focused on wildfire resilience and adaptation, it is crucial to obtain responses that are not only precise but also domain-specific. To that end, we developed WildfireGPT, a prototype LLM agent designed to transform user queries into actionable insights on wildfire risks. We enrich WildfireGPT by providing additional context, such as climate projections and scientific literature, to ensure its information is current, relevant, and scientifically accurate. This enables WildfireGPT to be an effective tool for delivering detailed, user-specific insights on wildfire risks to support a diverse set of end users, including but not limited to researchers and engineers, for making positive impact and decision making.

Jiashu He, Jinxuan Fan, Bowen Jiang et al.

When addressing complex questions that require new information, people often associate the question with existing knowledge to derive a sensible answer. For instance, when evaluating whether melatonin aids insomnia, one might associate "hormones helping mental disorders" with "melatonin being a hormone and insomnia a mental disorder" to complete the reasoning. Large Language Models (LLMs) also require such associative thinking, particularly in resolving scientific inquiries when retrieved knowledge is insufficient and does not directly answer the question. Graph Inspired Veracity Extrapolation (GIVE) addresses this by using a knowledge graph (KG) to extrapolate structured knowledge. However, it involves the construction and pruning of many hypothetical triplets, which limits efficiency and generalizability. We propose Self-GIVE, a retrieve-RL framework that enhances LLMs with automatic associative thinking through reinforcement learning. Self-GIVE extracts structured information and entity sets to assist the model in linking to the queried concepts. We address GIVE's key limitations: (1) extensive LLM calls and token overhead for knowledge extrapolation, (2) difficulty in deploying on smaller LLMs (3B or 7B) due to complex instructions, and (3) inaccurate knowledge from LLM pruning. Specifically, after fine-tuning using self-GIVE with a 135 node UMLS KG, it improves the performance of the Qwen2.5 3B and 7B models by up to $\textbf{28.5%$\rightarrow$71.4%}$ and $\textbf{78.6$\rightarrow$90.5%}$ in samples $\textbf{unseen}$ in challenging biomedical QA tasks. In particular, Self-GIVE allows the 7B model to match or outperform GPT3.5 turbo with GIVE, while cutting token usage by over 90%. Self-GIVE enhances the scalable integration of structured retrieval and reasoning with associative thinking.

Bowen Jiang, Zhijun Zhuang, Shreyas S. Shivakumar et al.

This work explores the zero-shot capabilities of foundation models in Visual Question Answering (VQA) tasks. We propose an adaptive multi-agent system, named Multi-Agent VQA, to overcome the limitations of foundation models in object detection and counting by using specialized agents as tools. Unlike existing approaches, our study focuses on the system's performance without fine-tuning it on specific VQA datasets, making it more practical and robust in the open world. We present preliminary experimental results under zero-shot scenarios and highlight some failure cases, offering new directions for future research.

Yangxinyu Xie, Bowen Jiang, Tanwi Mallick et al.

Large language models (LLMs) are a transformational capability at the frontier of artificial intelligence and machine learning that can support decision-makers in addressing pressing societal challenges such as extreme natural hazard events. As generalized models, LLMs often struggle to provide context-specific information, particularly in areas requiring specialized knowledge. In this work we propose a retrieval-augmented generation (RAG)-based multi-agent LLM system to support analysis and decision-making in the context of natural hazards and extreme weather events. As a proof of concept, we present WildfireGPT, a specialized system focused on wildfire hazards. The architecture employs a user-centered, multi-agent design to deliver tailored risk insights across diverse stakeholder groups. By integrating natural hazard and extreme weather projection data, observational datasets, and scientific literature through an RAG framework, the system ensures both the accuracy and contextual relevance of the information it provides. Evaluation across ten expert-led case studies demonstrates that WildfireGPT significantly outperforms existing LLM-based solutions for decision support.

Hanbo Huang, Yihan Li, Bowen Jiang et al.

Privacy-sensitive users require deploying large language models (LLMs) within their own infrastructure (on-premises) to safeguard private data and enable customization. However, vulnerabilities in local environments can lead to unauthorized access and potential model theft. To address this, prior research on small models has explored securing only the output layer within hardware-secured devices to balance model confidentiality and customization. Yet this approach fails to protect LLMs effectively. In this paper, we discover that (1) query-based distillation attacks targeting the secured top layer can produce a functionally equivalent replica of the victim model; (2) securing the same number of layers, bottom layers before a transition layer provide stronger protection against distillation attacks than top layers, with comparable effects on customization performance; and (3) the number of secured layers creates a trade-off between protection and customization flexibility. Based on these insights, we propose SOLID, a novel deployment framework that secures a few bottom layers in a secure environment and introduces an efficient metric to optimize the trade-off by determining the ideal number of hidden layers. Extensive experiments on five models (1.3B to 70B parameters) demonstrate that SOLID outperforms baselines, achieving a better balance between protection and downstream customization.

Taiwei Shi, Sihao Chen, Bowen Jiang et al.

Reinforcement learning has become the central approach for language models (LMs) to learn from environmental reward or feedback. In practice, the environmental feedback is usually sparse and delayed. Learning from such signals is challenging, as LMs must implicitly infer how observed failures should translate into behavioral changes for future iterations. We introduce Experiential Reinforcement Learning (ERL), a training paradigm that embeds an explicit experience-reflection-consolidation loop into the reinforcement learning process. Given a task, the model generates an initial attempt, receives environmental feedback, and produces a reflection that guides a refined second attempt, whose success is reinforced and internalized into the base policy. This process converts feedback into structured behavioral revision, improving exploration and stabilizing optimization while preserving gains at deployment without additional inference cost. Across sparse-reward control environments and agentic reasoning benchmarks, ERL consistently improves learning efficiency and final performance over strong reinforcement learning baselines, achieving gains of up to +81% in complex multi-step environments and up to +11% in tool-using reasoning tasks. These results suggest that integrating explicit self-reflection into policy training provides a practical mechanism for transforming feedback into durable behavioral improvement.

Zhaoyang Wei, Chenhui Qiang, Bowen Jiang et al.

Chain-of-Thought (CoT) reasoning has emerged as a powerful approach to enhance the structured, multi-step decision-making capabilities of Multi-Modal Large Models (MLLMs), is particularly crucial for autonomous driving with adverse weather conditions and complex traffic environments. However, existing benchmarks have largely overlooked the need for rigorous evaluation of CoT processes in these specific and challenging scenarios. To address this critical gap, we introduce AD^2-Bench, the first Chain-of-Thought benchmark specifically designed for autonomous driving with adverse weather and complex scenes. AD^2-Bench is meticulously constructed to fulfill three key criteria: comprehensive data coverage across diverse adverse environments, fine-grained annotations that support multi-step reasoning, and a dedicated evaluation framework tailored for assessing CoT performance. The core contribution of AD^2-Bench is its extensive collection of over 5.4k high-quality, manually annotated CoT instances. Each intermediate reasoning step in these annotations is treated as an atomic unit with explicit ground truth, enabling unprecedented fine-grained analysis of MLLMs' inferential processes under text-level, point-level, and region-level visual prompts. Our comprehensive evaluation of state-of-the-art MLLMs on AD^2-Bench reveals accuracy below 60%, highlighting the benchmark's difficulty and the need to advance robust, interpretable end-to-end autonomous driving systems. AD^2-Bench thus provides a standardized evaluation platform, driving research forward by improving MLLMs' reasoning in autonomous driving, making it an invaluable resource.

Bowen Jiang, Yangxinyu Xie, Xiaomeng Wang et al.

We present GeoGrid-Bench, a benchmark designed to evaluate the ability of foundation models to understand geo-spatial data in the grid structure. Geo-spatial datasets pose distinct challenges due to their dense numerical values, strong spatial and temporal dependencies, and unique multimodal representations including tabular data, heatmaps, and geographic visualizations. To assess how foundation models can support scientific research in this domain, GeoGrid-Bench features large-scale, real-world data covering 16 climate variables across 150 locations and extended time frames. The benchmark includes approximately 3,200 question-answer pairs, systematically generated from 8 domain expert-curated templates to reflect practical tasks encountered by human scientists. These range from basic queries at a single location and time to complex spatiotemporal comparisons across regions and periods. Our evaluation reveals that vision-language models perform best overall, and we provide a fine-grained analysis of the strengths and limitations of different foundation models in different geo-spatial tasks. This benchmark offers clearer insights into how foundation models can be effectively applied to geo-spatial data analysis and used to support scientific research.

Jiaxiang Li, Junwei Feng, Jie Luo et al.

Machine learning interatomic potentials have revolutionized complex materials design by enabling rapid exploration of material configurational spaces via crystal structure prediction with ab initio accuracy. However, critical challenges persist in ensuring robust generalization to unknown structures and minimizing the requirement for substantial expert knowledge and time-consuming manual interventions. Here, we propose an automated crystal structure prediction framework built upon the attention-coupled neural networks potential to address these limitations. The generalizability of the potential is achieved by sampling regions across the local minima of the potential energy surface, where the self-evolving pipeline autonomously refines the potential iteratively while minimizing human intervention. The workflow is validated on Mg-Ca-H ternary and Be-P-N-O quaternary systems by exploring nearly 10 million configurations, demonstrating substantial speedup compared to first-principles calculations. These results underscore the effectiveness of our approach in accelerating the exploration and discovery of complex multi-component functional materials.

Ziyi Liu, Priyanka Dey, Jen-tse Huang et al.

Cultural Intelligence (CQ) refers to the ability to understand unfamiliar cultural contexts, a crucial skill for large language models (LLMs) to effectively engage with globally diverse users. Existing studies often focus on explicitly stated cultural norms, but fail to capture the subtle, implicit values that are common in daily conversation. To address this gap, we introduce CQBench, a benchmark specifically designed to assess LLMs' capability to infer implicit cultural values from natural conversational contexts. CQBench consists of multi character conversation based stories using values from the World Value Survey and the GlobalOpinions, with topics including ethical, religious, social, etc. Our automatic dataset construction pipeline integrates rigorous validation procedures (incorporation, consistency, and implicitness checks), achieving a 94.5% human model agreement in the final validation. To leverage CQBench data, we design three tasks of increasing complexity: attitude detection, value selection, and value extraction. These tasks evaluate whether models can detect attitude and recognize values embedded within natural dialogues rather than relying on explicit cultural knowledge. We find that while frontier models like o1 reach human level performance in value selection (0.809 F1), they still fall short in nuanced attitude detection (0.622 F1). Notably, finetuning a smaller LLaMA-3.2-3B on only 500 culturally rich examples improves performance by over 10%, even outperforming o3-mini in some cases. Using CQ-Bench, we provide insights into the current challenges in LLMs' CQ research and suggest practical pathways for enhancing LLMs' cross-cultural reasoning abilities.

Wenxuan Zhou, Bowen Jiang, Fan Yang et al.

Manipulating objects without grasping them is an essential component of human dexterity, referred to as non-prehensile manipulation. Non-prehensile manipulation may enable more complex interactions with the objects, but also presents challenges in reasoning about gripper-object interactions. In this work, we introduce Hybrid Actor-Critic Maps for Manipulation (HACMan), a reinforcement learning approach for 6D non-prehensile manipulation of objects using point cloud observations. HACMan proposes a temporally-abstracted and spatially-grounded object-centric action representation that consists of selecting a contact location from the object point cloud and a set of motion parameters describing how the robot will move after making contact. We modify an existing off-policy RL algorithm to learn in this hybrid discrete-continuous action representation. We evaluate HACMan on a 6D object pose alignment task in both simulation and in the real world. On the hardest version of our task, with randomized initial poses, randomized 6D goals, and diverse object categories, our policy demonstrates strong generalization to unseen object categories without a performance drop, achieving an 89% success rate on unseen objects in simulation and 50% success rate with zero-shot transfer in the real world. Compared to alternative action representations, HACMan achieves a success rate more than three times higher than the best baseline. With zero-shot sim2real transfer, our policy can successfully manipulate unseen objects in the real world for challenging non-planar goals, using dynamic and contact-rich non-prehensile skills. Videos can be found on the project website: https://hacman-2023.github.io.

Bowen Jiang, Maohao Shen

When performing classification tasks, raw high dimensional features often contain redundant information, and lead to increased computational complexity and overfitting. In this paper, we assume the data samples lie on a single underlying smooth manifold, and define intra-class and inter-class similarities using pairwise local kernel distances. We aim to find a linear projection to maximize the intra-class similarities and minimize the inter-class similarities simultaneously, so that the projected low dimensional data has optimized pairwise distances based on the label information, which is more suitable for a Diffusion Map to do further dimensionality reduction. Numerical experiments on several benchmark datasets show that our proposed approaches are able to extract low dimensional discriminate features that could help us achieve higher classification accuracy.