Zhiyi Shi

Pengcheng Jiang, Zhiyi Shi, Kelly Hong et al.

Search agents are often trained as policies over growing transcripts: the model must decide how to search while also remembering what it has seen, which evidence is useful, which constraints remain open, and which claims have actually been checked. We argue that this formulation puts too much routine state management inside the policy: reinforcement learning is forced to optimize both semantic search decisions and recoverable bookkeeping that the environment can maintain more reliably. We introduce Harness-1, a 20B search agent (retrieval subagent) trained with reinforcement learning inside a stateful search harness. The harness maintains environment-side working memory, including a candidate pool, an importance-tagged curated set, compact evidence links, verification records, compressed and deduplicated observations, and budget-aware context rendering. The policy retains the semantic decisions: what to search, which documents to keep or discard, what to verify, and when to stop. Across eight retrieval benchmarks spanning web, finance, patents, and multi-hop QA, Harness-1 achieves 0.730 average curated recall, outperforming the next strongest open search subagent by +11.4 points and remaining competitive with much larger frontier-model searchers. Its gains are especially strong on held-out transfer benchmarks, suggesting that reinforcement learning over explicit search state can produce retrieval behaviors that generalize beyond the training domains. Our code is available at https://github.com/pat-jj/harness-1.

Pengcheng Jiang, Jiacheng Lin, Zhiyi Shi et al. · stanford

Cutting-edge agentic AI systems are built on foundation models that can be adapted to plan, reason, and interact with external tools to perform increasingly complex and specialized tasks. As these systems grow in capability and scope, adaptation becomes a central mechanism for improving performance, reliability, and generalization. In this paper, we unify the rapidly expanding research landscape into a systematic framework that spans both agent adaptations and tool adaptations. We further decompose these into tool-execution-signaled and agent-output-signaled forms of agent adaptation, as well as agent-agnostic and agent-supervised forms of tool adaptation. We demonstrate that this framework helps clarify the design space of adaptation strategies in agentic AI, makes their trade-offs explicit, and provides practical guidance for selecting or switching among strategies during system design. We then review the representative approaches in each category, analyze their strengths and limitations, and highlight key open challenges and future opportunities. Overall, this paper aims to offer a conceptual foundation and practical roadmap for researchers and practitioners seeking to build more capable, efficient, and reliable agentic AI systems.

Jiashuo Sun, Jimeng Shi, Yixuan Xie et al.

Retrieval-Augmented Generation (RAG) has become a standard approach for knowledge-intensive question answering, but existing systems remain brittle on multi-hop questions, where solving the task requires chaining multiple retrieval and reasoning steps. Key challenges are that current methods represent reasoning through free-form natural language, where intermediate states are implicit, retrieval queries can drift from intended entities, and errors are detected by the same model that produces them making self-reflection an unreliable, ungrounded signal. We observe that multi-hop question answering is a typical form of step-by-step computation, and that this structured process aligns closely with how code-specialized language models are trained to operate. Motivated by this, we introduce \pyrag, a framework that reformulates multi-hop RAG as program synthesis and execution. Instead of free-form reasoning trajectories, \pyrag represents the reasoning process as an executable Python program over retrieval and QA tools, exposing intermediate states as variables, producing deterministic feedback through execution, and yielding an inspectable trace of the entire reasoning process. This formulation further enables compiler-grounded self-repair and execution-driven adaptive retrieval without any additional training. Experiments on five QA benchmarks (PopQA, HotpotQA, 2WikiMultihopQA, MuSiQue, and Bamboogle) show that \pyrag consistently outperforms strong baselines under both training-free and RL-trained settings, with especially large gains on compositional multi-hop datasets. Our code, data and models are publicly available at https://github.com/GasolSun36/PyRAG.

Zhiyi Shi, Junsik Kim, Wanhua Li et al.

Multi-modal pre-trained models efficiently extract and fuse features from different modalities with low memory requirements for fine-tuning. Despite this efficiency, their application in disease diagnosis is under-explored. A significant challenge is the frequent occurrence of missing modalities, which impairs performance. Additionally, fine-tuning the entire pre-trained model demands substantial computational resources. To address these issues, we introduce Modality-aware Low-Rank Adaptation (MoRA), a computationally efficient method. MoRA projects each input to a low intrinsic dimension but uses different modality-aware up-projections for modality-specific adaptation in cases of missing modalities. Practically, MoRA integrates into the first block of the model, significantly improving performance when a modality is missing. It requires minimal computational resources, with less than 1.6% of the trainable parameters needed compared to training the entire model. Experimental results show that MoRA outperforms existing techniques in disease diagnosis, demonstrating superior performance, robustness, and training efficiency.

Chongjie Si, Zhiyi Shi, Shifan Zhang et al.

Large language models demonstrate impressive performance on downstream tasks, yet they require extensive resource consumption when fully fine-tuning all parameters. To mitigate this, Parameter Efficient Fine-Tuning (PEFT) strategies, such as LoRA, have been developed. In this paper, we delve into the concept of task-specific directions (TSDs), which are critical for transitioning large models from pretrained states to task-specific enhancements in PEFT. We propose a framework to clearly define these directions and explore their properties and practical utilization challenges. We then introduce a novel approach, LoRA-Dash, which aims to maximize the impact of TSDs during the fine-tuning process, thereby enhancing model performance on targeted tasks. Additionally, based on our exploration of TSD, we focus on an important issue in PEFT: the initialization of LoRA. While some works have pointed out the significance of initialization for LoRA's performance and proposed various strategies, these methods are often empirical and not task-specific. To address this issue, we propose LoRA-Init. Starting from TSD, we identify the directions that require the most adjustment during fine-tuning for downstream tasks. By initializing the matrices in LoRA with these directions, LoRA-Init significantly enhances LoRA's performance. Moreover, we can combine LoRA-Dash and LoRA-Init to create the final version of LoRA based on TSDs, which we refer to as LoRA-TSD. Extensive experiments have conclusively demonstrated the effectiveness of these methods, and in-depth analyses further reveal the underlying mechanisms behind their success.

Heng Wang, Pengcheng Jiang, Jiashuo Sun et al.

Large language models (LLMs) are increasingly used to assist ideation in research, but evaluating the quality of LLM-generated research proposals remains difficult: novelty and soundness are hard to measure automatically, and large-scale human evaluation is costly. We propose a verifiable alternative by reframing proposal generation as a time-sliced scientific forecasting problem. Given a research question and inspiring papers available before a cutoff time, the model generates a structured proposal and is evaluated by whether it anticipates research directions that appear in papers published after the time. We operationalize this objective with the Future Alignment Score (FAS), computed via retrieval and LLM-based semantic scoring against a held-out future corpus. To train models, we build a time-consistent dataset of 17,771 papers from targets and their pre-cutoff citations, and synthesize reasoning traces that teach gap identification and inspiration borrowing. Across Llama-3.1 and Qwen2.5 models, future-aligned tuning improves future alignment over unaligned baselines (up to +10.6% overall FAS), and domain-expert human evaluation corroborates improved proposal quality. Finally, we demonstrate practical impact by implementing two model-generated proposals with a code agent, obtaining 4.17% accuracy gain on MATH from a new prompting strategy and consistent improvements for a novel model-merging method.

Zhiyi Shi, Binjie Wang, Chongjie Si et al.

Model editing aims to efficiently update a pre-trained model's knowledge without the need for time-consuming full retraining. While existing pioneering editing methods achieve promising results, they primarily focus on editing single-modal language models (LLMs). However, for vision-language models (VLMs), which involve multiple modalities, the role and impact of each modality on editing performance remain largely unexplored. To address this gap, we explore the impact of textual and visual modalities on model editing and find that: (1) textual and visual representations reach peak sensitivity at different layers, reflecting their varying importance; and (2) editing both modalities can efficiently update knowledge, but this comes at the cost of compromising the model's original capabilities. Based on our findings, we propose DualEdit, an editor that modifies both textual and visual modalities at their respective key layers. Additionally, we introduce a gating module within the more sensitive textual modality, allowing DualEdit to efficiently update new knowledge while preserving the model's original information. We evaluate DualEdit across multiple VLM backbones and benchmark datasets, demonstrating its superiority over state-of-the-art VLM editing baselines as well as adapted LLM editing methods on different evaluation metrics. Codes are available at https://github.com/zhiyiscs/DualEdit

Kaiwen Geng, Zhiyi Shi, Xiaoyan Zhao et al.

Abstract Background: Pulmonary function tests (PFTs) and computed tomography (CT) imaging are vital in diagnosing, managing, and monitoring lung diseases. A common issue in practice is the lack of access to recorded pulmonary functions despite available chest CT scans. Purpose: To develop and validate a deep learning algorithm for predicting pulmonary function directly from chest CT scans. Methods: The development cohort came from the Pittsburgh Lung Screening Study (PLuSS) (n=3619). The validation cohort came from the Specialized Centers of Clinically Oriented Research (SCCOR) in COPD (n=662). A deep learning model called BeyondCT, combining a three-dimensional (3D) convolutional neural network (CNN) and Vision Transformer (ViT) architecture, was used to predict forced vital capacity (FVC) and forced expiratory volume in one second (FEV1) from non-contrasted inspiratory chest CT scans. A 3D CNN model without ViT was used for comparison. Subject demographics (age, gender, smoking status) were also incorporated into the model. Performance was compared to actual PFTs using mean absolute error (MAE, L), percentage error, and R square. Results: The 3D-CNN model achieved MAEs of 0.395 L and 0.383 L, percentage errors of 13.84% and 18.85%, and R square of 0.665 and 0.679 for FVC and FEV1, respectively. The BeyondCT model without demographics had MAEs of 0.362 L and 0.371 L, percentage errors of 10.89% and 14.96%, and R square of 0.719 and 0.727, respectively. Including demographics improved performance (p<0.05), with MAEs of 0.356 L and 0.353 L, percentage errors of 10.79% and 14.82%, and R square of 0.77 and 0.739 for FVC and FEV1 in the test set. Conclusion: The BeyondCT model showed robust performance in predicting lung function from non-contrast inspiratory chest CT scans.

Junsik Kim, Zhiyi Shi, Davin Jeong et al.

In clinical In-Vitro Fertilization (IVF), identifying the most viable embryo for transfer is important to increasing the likelihood of a successful pregnancy. Traditionally, this process involves embryologists manually assessing embryos' static morphological features at specific intervals using light microscopy. This manual evaluation is not only time-intensive and costly, due to the need for expert analysis, but also inherently subjective, leading to variability in the selection process. To address these challenges, we develop a multimodal model that leverages both time-lapse video data and Electronic Health Records (EHRs) to predict embryo viability. One of the primary challenges of our research is to effectively combine time-lapse video and EHR data, owing to their inherent differences in modality. We comprehensively analyze our multimodal model with various modality inputs and integration approaches. Our approach will enable fast and automated embryo viability predictions in scale for clinical IVF.

Zhiyi Shi, Junsik Kim, Helen Y. Yang et al.

Automating embryo viability prediction for in vitro fertilization (IVF) is important but challenging due to the limited availability of labeled pregnancy outcome data, as only a small fraction of embryos are labeled after transfer. Self-supervised learning (SSL) can leverage both labeled and unlabeled data to improve prediction. However, existing SSL methods for videos are not directly applicable to embryo development videos due to two challenges: (1) embryo time-lapse videos contain hundreds of frames, requiring significant GPU memory for conventional SSL; (2) the dataset contains videos with varying lengths and many outlier frames, causing traditional video alignment methods to struggle with semantic misalignment. We propose Spatial-Temporal Pre-Training (STPT) to address these challenges. STPT includes two stages: spatial and temporal. In each stage, only one encoder is trained while the other is frozen, reducing memory demands. To handle temporal misalignment, STPT avoids frame-by-frame alignment across videos. The spatial stage learns from alignments within each video and its temporally consistent augmentations. The temporal stage then models relationships between video embeddings. Our method efficiently handles long videos and temporal variability. On 23,027 time-lapse videos (3,286 labeled), STPT achieves the highest AUC of 0.635 (95% CI: 0.632-0.638) compared to baselines, with limited computational resources.

Chongjie Si, Zhiyi Shi, Yadao Wang et al.

The rapid development of large language models has revolutionized natural language processing, but their fine-tuning remains computationally expensive, hindering broad deployment. Parameter-efficient fine-tuning (PEFT) methods, such as LoRA, have emerged as solutions. Recent work like DoRA attempts to further decompose weight adaptation into direction and magnitude components. However, existing formulations often define direction heuristically at the column level, lacking a principled geometric foundation. In this paper, we propose MAP, a novel framework that reformulates weight matrices as high-dimensional vectors and decouples their adaptation into direction and magnitude in a rigorous manner. MAP normalizes the pre-trained weights, learns a directional update, and introduces two scalar coefficients to independently scale the magnitude of the base and update vectors. This design enables more interpretable and flexible adaptation, and can be seamlessly integrated into existing PEFT methods. Extensive experiments show that MAP significantly improves performance when coupling with existing methods, offering a simple yet powerful enhancement to existing PEFT methods. Given the universality and simplicity of MAP, we hope it can serve as a default setting for designing future PEFT methods.

Chongjie Si, Zhiyi Shi, Xuehui Wang et al.

Adapting pre-trained foundation models for diverse downstream tasks is a core practice in artificial intelligence. However, the wide range of tasks and high computational costs make full fine-tuning impractical. To overcome this, parameter-efficient fine-tuning (PEFT) methods like LoRA have emerged and are becoming a growing research focus. Despite the success of these methods, they are primarily designed for linear layers, focusing on two-dimensional matrices while largely ignoring higher-dimensional parameter spaces like convolutional kernels. Moreover, directly applying these methods to higher-dimensional parameter spaces often disrupts their structural relationships. Given the rapid advancements in matrix-based PEFT methods, rather than designing a specialized strategy, we propose a generalization that extends matrix-based PEFT methods to higher-dimensional parameter spaces without compromising their structural properties. Specifically, we treat parameters as elements of a Lie group, with updates modeled as perturbations in the corresponding Lie algebra. These perturbations are mapped back to the Lie group through the exponential map, ensuring smooth, consistent updates that preserve the inherent structure of the parameter space. Extensive experiments on computer vision and natural language processing validate the effectiveness and versatility of our approach, demonstrating clear improvements over existing methods.

Hyun-Jic Oh, Junsik Kim, Zhiyi Shi et al.

Multiplex imaging is revolutionizing pathology by enabling the simultaneous visualization of multiple biomarkers within tissue samples, providing molecular-level insights that traditional hematoxylin and eosin (H&E) staining cannot provide. However, the complexity and cost of multiplex data acquisition have hindered its widespread adoption. Additionally, most existing large repositories of H&E images lack corresponding multiplex images, limiting opportunities for multimodal analysis. To address these challenges, we leverage recent advances in latent diffusion models (LDMs), which excel at modeling complex data distributions by utilizing their powerful priors for fine-tuning to a target domain. In this paper, we introduce a novel framework for virtual multiplex staining that utilizes pretrained LDM parameters to generate multiplex images from H&E images using a conditional diffusion model. Our approach enables marker-by-marker generation by conditioning the diffusion model on each marker, while sharing the same architecture across all markers. To tackle the challenge of varying pixel value distributions across different marker stains and to improve inference speed, we fine-tune the model for single-step sampling, enhancing both color contrast fidelity and inference efficiency through pixel-level loss functions. We validate our framework on two publicly available datasets, notably demonstrating its effectiveness in generating up to 18 different marker types with improved accuracy, a substantial increase over the 2-3 marker types achieved in previous approaches. This validation highlights the potential of our framework, pioneering virtual multiplex staining. Finally, this paper bridges the gap between H&E and multiplex imaging, potentially enabling retrospective studies and large-scale analyses of existing H&E image repositories.

Shangsi Ren, Cameron A. Beeche, Zhiyi Shi et al.

This study aims to establish the causal relationship network between various factors leading to workday loss in underground coal mines using a novel causal artificial intelligence (AI) method. The analysis utilizes data obtained from the National Institute for Occupational Safety and Health (NIOSH). A total of 101,010 injury records from 3,982 unique underground coal mines spanning the years from 1990 to 2020 were extracted from the NIOSH database. Causal relationships were analyzed and visualized using a novel causal AI method called Grouped Greedy Equivalence Search (GGES). The impact of each variable on workday loss was assessed through intervention do-calculus adjustment (IDA) scores. Model training and validation were performed using the 10-fold cross-validation technique. Performance metrics, including adjacency precision (AP), adjacency recall (AR), arrowhead precision (AHP), and arrowhead recall (AHR), were utilized to evaluate the models. Findings revealed that after 2006, key direct causes of workday loss among mining employees included total mining experience, mean office employees, mean underground employees, county, and total mining experience (years). Total mining experience emerged as the most influential factor, whereas mean employees per mine exhibited the least influence. The analyses emphasized the significant role of total mining experience in determining workday loss. The models achieved optimal performance, with AP, AR, AHP, and AHR values measuring 0.694, 0.653, 0.386, and 0.345, respectively. This study demonstrates the feasibility of utilizing the new GGES method to clarify the causal factors behind the workday loss by analyzing employment demographics and injury records and establish their causal relationship network.