Jiachen Liu

Wei Li, Can Gao, Guocheng Niu et al. · baidu

Vision-Language Pre-training (VLP) has achieved impressive performance on various cross-modal downstream tasks. However, most existing methods can only learn from aligned image-caption data and rely heavily on expensive regional features, which greatly limits their scalability and performance. In this paper, we propose an end-to-end unified-modal pre-training framework, namely UNIMO-2, for joint learning on both aligned image-caption data and unaligned image-only and text-only corpus. We build a unified Transformer model to jointly learn visual representations, textual representations and semantic alignment between images and texts. In particular, we propose to conduct grounded learning on both images and texts via a sharing grounded space, which helps bridge unaligned images and texts, and align the visual and textual semantic spaces on different types of corpora. The experiments show that our grounded learning method can improve textual and visual semantic alignment for improving performance on various cross-modal tasks. Moreover, benefiting from effective joint modeling of different types of corpora, our model also achieves impressive performance on single-modal visual and textual tasks. Our code and models are public at the UNIMO project page https://unimo-ptm.github.io/.

Shuaike Li, Kai Zhang, Xianquan Wang et al.

While Knowledge Editing (KE) enables efficient updates, its dominant Static Fact Overwriting paradigm treats LLMs as discrete databases, forcibly injecting isolated facts. Fracturing pre-trained logical topologies, this triggers Epistemic Dissonance -- a pathology where un-evolved legacy priors force the model to explicitly negate the injected update. Idealized interventions reveal that this is an inherent structural flaw rather than mere algorithmic noise, with a zero-distortion proxy yielding a catastrophic 95.6% self-refutation rate. Given the causally driven nature of real-world knowledge, grounding updates in explicit causal narratives effectively collapses this conflict rate to just 6.6%, underscoring the imperative for a paradigm shift toward Causal Editing. To internalize this evolution, we propose CODE (Causal On-policy Distillation for Editing). By coupling causal bootstrapping with asymmetric on-policy distillation, CODE engraves causal transition logic directly into parametric memory. Experiments on LLaMA-3.1 and Qwen-2.5 show CODE drastically suppresses self-refutation to 1.8% while securing robust multi-hop accuracy (up to 83.5%), seamlessly transforming discrete fact injection into coherent knowledge evolution. Code is available at https://github.com/CrashBugger/CODE.

Jiachen Liu, Pan Ji, Nitin Bansal et al.

We present a novel framework named PlaneMVS for 3D plane reconstruction from multiple input views with known camera poses. Most previous learning-based plane reconstruction methods reconstruct 3D planes from single images, which highly rely on single-view regression and suffer from depth scale ambiguity. In contrast, we reconstruct 3D planes with a multi-view-stereo (MVS) pipeline that takes advantage of multi-view geometry. We decouple plane reconstruction into a semantic plane detection branch and a plane MVS branch. The semantic plane detection branch is based on a single-view plane detection framework but with differences. The plane MVS branch adopts a set of slanted plane hypotheses to replace conventional depth hypotheses to perform plane sweeping strategy and finally learns pixel-level plane parameters and its planar depth map. We present how the two branches are learned in a balanced way, and propose a soft-pooling loss to associate the outputs of the two branches and make them benefit from each other. Extensive experiments on various indoor datasets show that PlaneMVS significantly outperforms state-of-the-art (SOTA) single-view plane reconstruction methods on both plane detection and 3D geometry metrics. Our method even outperforms a set of SOTA learning-based MVS methods thanks to the learned plane priors. To the best of our knowledge, this is the first work on 3D plane reconstruction within an end-to-end MVS framework. Source code: https://github.com/oppo-us-research/PlaneMVS.

Peng Di, Jianguo Li, Hang Yu et al.

Code Large Language Models (Code LLMs) have gained significant attention in the industry due to their wide applications in the full lifecycle of software engineering. However, the effectiveness of existing models in understanding non-English inputs for multi-lingual code-related tasks is still far from well studied. This paper introduces CodeFuse-13B, an open-sourced pre-trained code LLM. It is specifically designed for code-related tasks with both English and Chinese prompts and supports over 40 programming languages. CodeFuse achieves its effectiveness by utilizing a high quality pre-training dataset that is carefully filtered by program analyzers and optimized during the training process. Extensive experiments are conducted using real-world usage scenarios, the industry-standard benchmark HumanEval-x, and the specially designed CodeFuseEval for Chinese prompts. To assess the effectiveness of CodeFuse, we actively collected valuable human feedback from the AntGroup's software development process where CodeFuse has been successfully deployed. The results demonstrate that CodeFuse-13B achieves a HumanEval pass@1 score of 37.10%, positioning it as one of the top multi-lingual code LLMs with similar parameter sizes. In practical scenarios, such as code generation, code translation, code comments, and testcase generation, CodeFuse performs better than other models when confronted with Chinese prompts.

Wei Li, Wenhao Wu, Moye Chen et al. · baidu

Natural Language Generation (NLG) has made great progress in recent years due to the development of deep learning techniques such as pre-trained language models. This advancement has resulted in more fluent, coherent and even properties controllable (e.g. stylistic, sentiment, length etc.) generation, naturally leading to development in downstream tasks such as abstractive summarization, dialogue generation, machine translation, and data-to-text generation. However, the faithfulness problem that the generated text usually contains unfaithful or non-factual information has become the biggest challenge, which makes the performance of text generation unsatisfactory for practical applications in many real-world scenarios. Many studies on analysis, evaluation, and optimization methods for faithfulness problems have been proposed for various tasks, but have not been organized, compared and discussed in a combined manner. In this survey, we provide a systematic overview of the research progress on the faithfulness problem of NLG, including problem analysis, evaluation metrics and optimization methods. We organize the evaluation and optimization methods for different tasks into a unified taxonomy to facilitate comparison and learning across tasks. Several research trends are discussed further.

Wei Li, Xue Xu, Xinyan Xiao et al. · baidu

Diffusion generative models have recently greatly improved the power of text-conditioned image generation. Existing image generation models mainly include text conditional diffusion model and cross-modal guided diffusion model, which are good at small scene image generation and complex scene image generation respectively. In this work, we propose a simple yet effective approach, namely UPainting, to unify simple and complex scene image generation, as shown in Figure 1. Based on architecture improvements and diverse guidance schedules, UPainting effectively integrates cross-modal guidance from a pretrained image-text matching model into a text conditional diffusion model that utilizes a pretrained Transformer language model as the text encoder. Our key findings is that combining the power of large-scale Transformer language model in understanding language and image-text matching model in capturing cross-modal semantics and style, is effective to improve sample fidelity and image-text alignment of image generation. In this way, UPainting has a more general image generation capability, which can generate images of both simple and complex scenes more effectively. To comprehensively compare text-to-image models, we further create a more general benchmark, UniBench, with well-written Chinese and English prompts in both simple and complex scenes. We compare UPainting with recent models and find that UPainting greatly outperforms other models in terms of caption similarity and image fidelity in both simple and complex scenes. UPainting project page \url{https://upainting.github.io/}.

Wenhao Wu, Wei Li, Jiachen Liu et al. · baidu

Despite being able to generate fluent and grammatical text, current Seq2Seq summarization models still suffering from the unfaithful generation problem. In this paper, we study the faithfulness of existing systems from a new perspective of factual robustness which is the ability to correctly generate factual information over adversarial unfaithful information. We first measure a model's factual robustness by its success rate to defend against adversarial attacks when generating factual information. The factual robustness analysis on a wide range of current systems shows its good consistency with human judgments on faithfulness. Inspired by these findings, we propose to improve the faithfulness of a model by enhancing its factual robustness. Specifically, we propose a novel training strategy, namely FRSUM, which teaches the model to defend against both explicit adversarial samples and implicit factual adversarial perturbations. Extensive automatic and human evaluation results show that FRSUM consistently improves the faithfulness of various Seq2Seq models, such as T5, BART.

Jiachen Liu, Fan Lai, Yinwei Dai et al.

Federated learning (FL) is an emerging machine learning (ML) paradigm that enables heterogeneous edge devices to collaboratively train ML models without revealing their raw data to a logically centralized server. However, beyond the heterogeneous device capacity, FL participants often exhibit differences in their data distributions, which are not independent and identically distributed (Non-IID). Many existing works present point solutions to address issues like slow convergence, low final accuracy, and bias in FL, all stemming from client heterogeneity. In this paper, we explore an additional layer of complexity to mitigate such heterogeneity by grouping clients with statistically similar data distributions (cohorts). We propose Auxo to gradually identify such cohorts in large-scale, low-availability, and resource-constrained FL populations. Auxo then adaptively determines how to train cohort-specific models in order to achieve better model performance and ensure resource efficiency. Our extensive evaluations show that, by identifying cohorts with smaller heterogeneity and performing efficient cohort-based training, Auxo boosts various existing FL solutions in terms of final accuracy (2.1% - 8.2%), convergence time (up to 2.2x), and model bias (4.8% - 53.8%).

Wenhao Wu, Wei Li, Jiachen Liu et al. · baidu

Though model robustness has been extensively studied in language understanding, the robustness of Seq2Seq generation remains understudied. In this paper, we conduct the first quantitative analysis on the robustness of pre-trained Seq2Seq models. We find that even current SOTA pre-trained Seq2Seq model (BART) is still vulnerable, which leads to significant degeneration in faithfulness and informativeness for text generation tasks. This motivated us to further propose a novel adversarial augmentation framework, namely AdvSeq, for generally improving faithfulness and informativeness of Seq2Seq models via enhancing their robustness. AdvSeq automatically constructs two types of adversarial augmentations during training, including implicit adversarial samples by perturbing word representations and explicit adversarial samples by word swapping, both of which effectively improve Seq2Seq robustness. Extensive experiments on three popular text generation tasks demonstrate that AdvSeq significantly improves both the faithfulness and informativeness of Seq2Seq generation under both automatic and human evaluation settings.

Ziheng Geng, Jiachen Liu, Ian Franklin et al.

Recent advances in large language models (LLMs) have shown the promise to significantly accelerate the workflow by automating structural modeling and analysis. However, existing studies primarily focus on enabling LLMs to operate a single structural analysis software platform. In practice, structural engineers often rely on multiple finite element analysis (FEA) tools, such as ETABS, SAP2000, and OpenSees, depending on project needs, user preferences, and company constraints. This limitation restricts the practical deployment of LLM-assisted engineering workflows. To address this gap, this study develops LLMs capable of automating frame structural analysis across multiple software platforms. The LLMs adopt a two-stage multi-agent architecture. In Stage 1, a cohort of agents collaboratively interpret user input and perform structured reasoning to infer geometric, material, boundary, and load information required for finite element modeling. The outputs of these agents are compiled into a unified JSON representation. In Stage 2, code translation agents operate in parallel to convert the JSON file into executable scripts across multiple structural analysis platforms. Each agent is prompted with the syntax rules and modeling workflows of its target software. The LLMs are evaluated using 20 representative frame problems across three widely used platforms: ETABS, SAP2000, and OpenSees. Results from ten repeated trials demonstrate consistently reliable performance, achieving accuracy exceeding 90% across all cases.

Zhe Hu, Hou Pong Chan, Jiachen Liu et al.

Despite recent progress of pre-trained language models on generating fluent text, existing methods still suffer from incoherence problems in long-form text generation tasks that require proper content control and planning to form a coherent high-level logical flow. In this work, we propose PLANET, a novel generation framework leveraging autoregressive self-attention mechanism to conduct content planning and surface realization dynamically. To guide the generation of output sentences, our framework enriches the Transformer decoder with latent representations to maintain sentence-level semantic plans grounded by bag-of-words. Moreover, we introduce a new coherence-based contrastive learning objective to further improve the coherence of output. Extensive experiments are conducted on two challenging long-form text generation tasks including counterargument generation and opinion article generation. Both automatic and human evaluations show that our method significantly outperforms strong baselines and generates more coherent texts with richer contents.

Ziheng Geng, Jiachen Liu, Ran Cao et al.

Flood insurance is an effective strategy for individuals to mitigate disaster-related losses. However, participation rates among at-risk populations in the United States remain strikingly low. This gap underscores the need to understand and model the behavioral mechanisms underlying insurance decisions. Large language models (LLMs) have recently exhibited human-like intelligence across wide-ranging tasks, offering promising tools for simulating human decision-making. This study constructs a benchmark dataset to capture insurance purchase probabilities across factors. Using this dataset, the capacity of LLMs is evaluated: while LLMs exhibit a qualitative understanding of factors, they fall short in estimating quantitative probabilities. To address this limitation, InsurAgent, an LLM-empowered agent comprising five modules including perception, retrieval, reasoning, action, and memory, is proposed. The retrieval module leverages retrieval-augmented generation (RAG) to ground decisions in empirical survey data, achieving accurate estimation of marginal and bivariate probabilities. The reasoning module leverages LLM common sense to extrapolate beyond survey data, capturing contextual information that is intractable for traditional models. The memory module supports the simulation of temporal decision evolutions, illustrated through a roller coaster life trajectory. Overall, InsurAgent provides a valuable tool for behavioral modeling and policy analysis.

Jiachen Liu, Yuan Xue, Jose Duarte et al.

The automatic generation of floorplans given user inputs has great potential in architectural design and has recently been explored in the computer vision community. However, the majority of existing methods synthesize floorplans in the format of rasterized images, which are difficult to edit or customize. In this paper, we aim to synthesize floorplans as sequences of 1-D vectors, which eases user interaction and design customization. To generate high fidelity vectorized floorplans, we propose a novel two-stage framework, including a draft stage and a multi-round refining stage. In the first stage, we encode the room connectivity graph input by users with a graph convolutional network (GCN), then apply an autoregressive transformer network to generate an initial floorplan sequence. To polish the initial design and generate more visually appealing floorplans, we further propose a novel panoptic refinement network(PRN) composed of a GCN and a transformer network. The PRN takes the initial generated sequence as input and refines the floorplan design while encouraging the correct room connectivity with our proposed geometric loss. We have conducted extensive experiments on a real-world floorplan dataset, and the results show that our method achieves state-of-the-art performance under different settings and evaluation metrics.

Wenhao Wu, Wei Li, Xinyan Xiao et al.

A crucial issue of current text generation models is that they often uncontrollably generate factually inconsistent text with respective of their inputs. Limited by the lack of annotated data, existing works in evaluating factual consistency directly transfer the reasoning ability of models trained on other data-rich upstream tasks like question answering (QA) and natural language inference (NLI) without any further adaptation. As a result, they perform poorly on the real generated text and are biased heavily by their single-source upstream tasks. To alleviate this problem, we propose a weakly supervised framework that aggregates multiple resources to train a precise and efficient factual metric, namely WeCheck. WeCheck first utilizes a generative model to accurately label a real generated sample by aggregating its weak labels, which are inferred from multiple resources. Then, we train the target metric model with the weak supervision while taking noises into consideration. Comprehensive experiments on a variety of tasks demonstrate the strong performance of WeCheck, which achieves a 3.4\% absolute improvement over previous state-of-the-art methods on TRUE benchmark on average.

Luyang Huang, Guocheng Niu, Jiachen Liu et al.

Due to the limitations of the model structure and pre-training objectives, existing vision-and-language generation models cannot utilize pair-wise images and text through bi-directional generation. In this paper, we propose DU-VLG, a framework which unifies vision-and-language generation as sequence generation problems. DU-VLG is trained with novel dual pre-training tasks: multi-modal denoising autoencoder tasks and modality translation tasks. To bridge the gap between image understanding and generation, we further design a novel commitment loss. We compare pre-training objectives on image captioning and text-to-image generation datasets. Results show that DU-VLG yields better performance than variants trained with uni-directional generation objectives or the variant without the commitment loss. We also obtain higher scores compared to previous state-of-the-art systems on three vision-and-language generation tasks. In addition, human judges further confirm that our model generates real and relevant images as well as faithful and informative captions.

Zhongwei Wan, Xin Wang, Che Liu et al.

Large Language Models (LLMs) have demonstrated remarkable capabilities in important tasks such as natural language understanding and language generation, and thus have the potential to make a substantial impact on our society. Such capabilities, however, come with the considerable resources they demand, highlighting the strong need to develop effective techniques for addressing their efficiency challenges. In this survey, we provide a systematic and comprehensive review of efficient LLMs research. We organize the literature in a taxonomy consisting of three main categories, covering distinct yet interconnected efficient LLMs topics from model-centric, data-centric, and framework-centric perspective, respectively. We have also created a GitHub repository where we organize the papers featured in this survey at https://github.com/AIoT-MLSys-Lab/Efficient-LLMs-Survey. We will actively maintain the repository and incorporate new research as it emerges. We hope our survey can serve as a valuable resource to help researchers and practitioners gain a systematic understanding of efficient LLMs research and inspire them to contribute to this important and exciting field.

Haomiao Ni, Jiachen Liu, Yuan Xue et al.

Motion transfer of talking-head videos involves generating a new video with the appearance of a subject video and the motion pattern of a driving video. Current methodologies primarily depend on a limited number of subject images and 2D representations, thereby neglecting to fully utilize the multi-view appearance features inherent in the subject video. In this paper, we propose a novel 3D-aware talking-head video motion transfer network, Head3D, which fully exploits the subject appearance information by generating a visually-interpretable 3D canonical head from the 2D subject frames with a recurrent network. A key component of our approach is a self-supervised 3D head geometry learning module, designed to predict head poses and depth maps from 2D subject video frames. This module facilitates the estimation of a 3D head in canonical space, which can then be transformed to align with driving video frames. Additionally, we employ an attention-based fusion network to combine the background and other details from subject frames with the 3D subject head to produce the synthetic target video. Our extensive experiments on two public talking-head video datasets demonstrate that Head3D outperforms both 2D and 3D prior arts in the practical cross-identity setting, with evidence showing it can be readily adapted to the pose-controllable novel view synthesis task.

Rui Yu, Jiachen Liu, Zihan Zhou et al.

In various applications, such as robotic navigation and remote visual assistance, expanding the field of view (FOV) of the camera proves beneficial for enhancing environmental perception. Unlike image outpainting techniques aimed solely at generating aesthetically pleasing visuals, these applications demand an extended view that faithfully represents the scene. To achieve this, we formulate a new problem of faithful FOV extrapolation that utilizes a set of pre-captured images as prior knowledge of the scene. To address this problem, we present a simple yet effective solution called NeRF-Enhanced Outpainting (NEO) that uses extended-FOV images generated through NeRF to train a scene-specific image outpainting model. To assess the performance of NEO, we conduct comprehensive evaluations on three photorealistic datasets and one real-world dataset. Extensive experiments on the benchmark datasets showcase the robustness and potential of our method in addressing this challenge. We believe our work lays a strong foundation for future exploration within the research community.

Patrick Tser Jern Kon, Jiachen Liu, Qiuyi Ding et al.

Scientific experimentation, a cornerstone of human progress, demands rigor in reliability, methodical control, and interpretability to yield meaningful results. Despite the growing capabilities of large language models (LLMs) in automating different aspects of the scientific process, automating rigorous experimentation remains a significant challenge. To address this gap, we propose Curie, an AI agent framework designed to embed rigor into the experimentation process through three key components: an intra-agent rigor module to enhance reliability, an inter-agent rigor module to maintain methodical control, and an experiment knowledge module to enhance interpretability. To evaluate Curie, we design a novel experimental benchmark composed of 46 questions across four computer science domains, derived from influential research papers, and widely adopted open-source projects. Compared to the strongest baseline tested, we achieve a 3.4$\times$ improvement in correctly answering experimental questions. Curie is open-sourced at https://github.com/Just-Curieous/Curie.

Lingdong Kong, Xian Sun, Wei Chow et al.

AI-assisted research is crossing a threshold: fully automated systems can now generate research papers for as little as $15, while long-horizon agents can execute experiments, draft manuscripts, and simulate critique with minimal human input. Yet this productivity frontier exposes a deeper integrity problem: under scientific pressure, even frontier LLMs still fabricate results, miss hidden errors, and fail to judge novelty reliably. Studying developments through April 2026, we present an end-to-end analysis of AI across the complete research lifecycle, organized into four epistemological phases: Creation (idea generation, literature review, coding & experiments, tables & figures), Writing (paper writing), Validation (peer review, rebuttal & revision), and Dissemination (posters, slides, videos, social media, project pages, and interactive agents). We identify a sharp, stage-dependent boundary between reliable assistance and unreliable autonomy: AI excels at structured, retrieval-grounded, and tool-mediated tasks, but remains fragile for genuinely novel ideas, research-level experiments, and scientific judgment. Generated ideas often degrade after implementation, research code lags far behind pattern-matching benchmarks, and end-to-end autonomous systems have not yet consistently reached major-venue acceptance standards. We further show that greater automation can obscure rather than eliminate failure modes, making human-governed collaboration the most credible deployment paradigm. Finally, we provide a structured taxonomy, benchmark suite, and tool inventory, cross-stage design principles, and a practitioner-oriented playbook, with resources maintained at our project page.

Jiachen Liu, Rui Yu, Sili Chen et al.

3D plane reconstruction from a single image is a crucial yet challenging topic in 3D computer vision. Previous state-of-the-art (SOTA) methods have focused on training their system on a single dataset from either indoor or outdoor domain, limiting their generalizability across diverse testing data. In this work, we introduce a novel framework dubbed ZeroPlane, a Transformer-based model targeting zero-shot 3D plane detection and reconstruction from a single image, over diverse domains and environments. To enable data-driven models across multiple domains, we have curated a large-scale planar benchmark, comprising over 14 datasets and 560,000 high-resolution, dense planar annotations for diverse indoor and outdoor scenes. To address the challenge of achieving desirable planar geometry on multi-dataset training, we propose to disentangle the representation of plane normal and offset, and employ an exemplar-guided, classification-then-regression paradigm to learn plane and offset respectively. Additionally, we employ advanced backbones as image encoder, and present an effective pixel-geometry-enhanced plane embedding module to further facilitate planar reconstruction. Extensive experiments across multiple zero-shot evaluation datasets have demonstrated that our approach significantly outperforms previous methods on both reconstruction accuracy and generalizability, especially over in-the-wild data. Our code and data are available at: https://github.com/jcliu0428/ZeroPlane.

Patrick Tser Jern Kon, Jiachen Liu, Xinyi Zhu et al.

Automating AI research holds immense potential for accelerating scientific progress, yet current AI agents struggle with the complexities of rigorous, end-to-end experimentation. We introduce EXP-Bench, a novel benchmark designed to systematically evaluate AI agents on complete research experiments sourced from influential AI publications. Given a research question and incomplete starter code, EXP-Bench challenges AI agents to formulate hypotheses, design and implement experimental procedures, execute them, and analyze results. To enable the creation of such intricate and authentic tasks with high-fidelity, we design a semi-autonomous pipeline to extract and structure crucial experimental details from these research papers and their associated open-source code. With the pipeline, EXP-Bench curated 461 AI research tasks from 51 top-tier AI research papers. Evaluations of leading LLM-based agents, such as OpenHands and IterativeAgent on EXP-Bench demonstrate partial capabilities: while scores on individual experimental aspects such as design or implementation correctness occasionally reach 20-35%, the success rate for complete, executable experiments was a mere 0.5%. By identifying these bottlenecks and providing realistic step-by-step experiment procedures, EXP-Bench serves as a vital tool for future AI agents to improve their ability to conduct AI research experiments. EXP-Bench is open-sourced at https://github.com/Just-Curieous/Curie/tree/main/benchmark/exp_bench.

Jae-Won Chung, Jeff J. Ma, Ruofan Wu et al.

As the adoption of Generative AI in real-world services grow explosively, energy has emerged as a critical bottleneck resource. However, energy remains a metric that is often overlooked, under-explored, or poorly understood in the context of building ML systems. We present the ML$.$ENERGY Benchmark, a benchmark suite and tool for measuring inference energy consumption under realistic service environments, and the corresponding ML$.$ENERGY Leaderboard, which have served as a valuable resource for those hoping to understand and optimize the energy consumption of their generative AI services. In this paper, we explain four key design principles for benchmarking ML energy we have acquired over time, and then describe how they are implemented in the ML$.$ENERGY Benchmark. We then highlight results from the early 2025 iteration of the benchmark, including energy measurements of 40 widely used model architectures across 6 different tasks, case studies of how ML design choices impact energy consumption, and how automated optimization recommendations can lead to significant (sometimes more than 40%) energy savings without changing what is being computed by the model. The ML$.$ENERGY Benchmark is open-source and can be easily extended to various customized models and application scenarios.

Jiachen Liu, Fan Lai, Ding Ding et al.

In recent years, collaborative learning (CL) has emerged as a promising approach for machine learning (ML) and data science across distributed edge devices. As the deployment of CL jobs increases, they inevitably contend for limited resources. However, efficient resource scheduling in this context is challenging because of the ephemeral nature and resource heterogeneity of devices, coupled with the overlapping resource requirements of diverse CL jobs. Existing resource managers often assign devices to CL jobs randomly for simplicity and scalability, but this approach compromises job efficiency. In this paper, we present Venn, a CL resource manager that efficiently schedules ephemeral, heterogeneous devices among multiple CL jobs to reduce the average job completion time (JCT). Venn formulates the Intersection Resource Scheduling (IRS) problem to identify complex resource contention among multiple CL jobs. It then proposes a contention-aware scheduling heuristic to minimize the average scheduling delay. Furthermore, it proposes a resource-aware device-to-job matching heuristic to optimize response collection time by mitigating stragglers. Our evaluation shows that, compared to the state-of-the-art CL resource managers, Venn improves the average JCT by up to 1.88x. The code is available at https://github.com/SymbioticLab/Venn.

Long Phan, Alice Gatti, Ziwen Han et al. · amazon-science, apple-ml

Benchmarks are important tools for tracking the rapid advancements in large language model (LLM) capabilities. However, benchmarks are not keeping pace in difficulty: LLMs now achieve over 90\% accuracy on popular benchmarks like MMLU, limiting informed measurement of state-of-the-art LLM capabilities. In response, we introduce Humanity's Last Exam (HLE), a multi-modal benchmark at the frontier of human knowledge, designed to be the final closed-ended academic benchmark of its kind with broad subject coverage. HLE consists of 2,500 questions across dozens of subjects, including mathematics, humanities, and the natural sciences. HLE is developed globally by subject-matter experts and consists of multiple-choice and short-answer questions suitable for automated grading. Each question has a known solution that is unambiguous and easily verifiable, but cannot be quickly answered via internet retrieval. State-of-the-art LLMs demonstrate low accuracy and calibration on HLE, highlighting a significant gap between current LLM capabilities and the expert human frontier on closed-ended academic questions. To inform research and policymaking upon a clear understanding of model capabilities, we publicly release HLE at https://lastexam.ai.

Jiachen Liu, Maestro Harmon, Zechen Zhang

While AI innovation accelerates rapidly, the intellectual process behind breakthroughs -- how researchers identify gaps, synthesize prior work, and generate insights -- remains poorly understood. The lack of structured data on scientific reasoning hinders systematic analysis and development of AI research agents. We introduce Sci-Reasoning, the first dataset capturing the intellectual synthesis behind high-quality AI research. Using community-validated quality signals and an LLM-accelerated, human-verified pipeline, we trace Oral and Spotlight papers across NeurIPS, ICML, and ICLR (2023-2025) to its key predecessors, articulating specific reasoning links in a structured format. Our analysis identifies 15 distinct thinking patterns, with three dominant strategies accounting for 52.7%: Gap-Driven Reframing (24.2%), Cross-Domain Synthesis (18.0%), and Representation Shift (10.5%). The most powerful innovation recipes combine multiple patterns: Gap-Driven Reframing + Representation Shift, Cross-Domain Synthesis + Representation Shift, and Gap-Driven Reframing + Cross-Domain Synthesis. This dataset enables quantitative studies of scientific progress and provides structured reasoning trajectories for training the next generation AI research agents.

Wang Zhao, Jiachen Liu, Sheng Zhang et al.

This paper presents a generalizable 3D plane detection and reconstruction framework named MonoPlane. Unlike previous robust estimator-based works (which require multiple images or RGB-D input) and learning-based works (which suffer from domain shift), MonoPlane combines the best of two worlds and establishes a plane reconstruction pipeline based on monocular geometric cues, resulting in accurate, robust and scalable 3D plane detection and reconstruction in the wild. Specifically, we first leverage large-scale pre-trained neural networks to obtain the depth and surface normals from a single image. These monocular geometric cues are then incorporated into a proximity-guided RANSAC framework to sequentially fit each plane instance. We exploit effective 3D point proximity and model such proximity via a graph within RANSAC to guide the plane fitting from noisy monocular depths, followed by image-level multi-plane joint optimization to improve the consistency among all plane instances. We further design a simple but effective pipeline to extend this single-view solution to sparse-view 3D plane reconstruction. Extensive experiments on a list of datasets demonstrate our superior zero-shot generalizability over baselines, achieving state-of-the-art plane reconstruction performance in a transferring setting. Our code is available at https://github.com/thuzhaowang/MonoPlane .

Moye Chen, Wei Li, Jiachen Liu et al.

Most of existing extractive multi-document summarization (MDS) methods score each sentence individually and extract salient sentences one by one to compose a summary, which have two main drawbacks: (1) neglecting both the intra and cross-document relations between sentences; (2) neglecting the coherence and conciseness of the whole summary. In this paper, we propose a novel MDS framework (SgSum) to formulate the MDS task as a sub-graph selection problem, in which source documents are regarded as a relation graph of sentences (e.g., similarity graph or discourse graph) and the candidate summaries are its sub-graphs. Instead of selecting salient sentences, SgSum selects a salient sub-graph from the relation graph as the summary. Comparing with traditional methods, our method has two main advantages: (1) the relations between sentences are captured by modeling both the graph structure of the whole document set and the candidate sub-graphs; (2) directly outputs an integrate summary in the form of sub-graph which is more informative and coherent. Extensive experiments on MultiNews and DUC datasets show that our proposed method brings substantial improvements over several strong baselines. Human evaluation results also demonstrate that our model can produce significantly more coherent and informative summaries compared with traditional MDS methods. Moreover, the proposed architecture has strong transfer ability from single to multi-document input, which can reduce the resource bottleneck in MDS tasks. Our code and results are available at: \url{https://github.com/PaddlePaddle/Research/tree/master/NLP/EMNLP2021-SgSum}.

Fan Lai, Yinwei Dai, Sanjay S. Singapuram et al.

We present FedScale, a federated learning (FL) benchmarking suite with realistic datasets and a scalable runtime to enable reproducible FL research. FedScale datasets encompass a wide range of critical FL tasks, ranging from image classification and object detection to language modeling and speech recognition. Each dataset comes with a unified evaluation protocol using real-world data splits and evaluation metrics. To reproduce realistic FL behavior, FedScale contains a scalable and extensible runtime. It provides high-level APIs to implement FL algorithms, deploy them at scale across diverse hardware and software backends, and evaluate them at scale, all with minimal developer efforts. We combine the two to perform systematic benchmarking experiments and highlight potential opportunities for heterogeneity-aware co-optimizations in FL. FedScale is open-source and actively maintained by contributors from different institutions at http://fedscale.ai. We welcome feedback and contributions from the community.

Jiachen Liu, Jae-Won Chung, Zhiyu Wu et al.

Large language models (LLMs) are now at the core of conversational AI services such as real-time translation and chatbots, which provide live user interaction by incrementally streaming text to the user. However, existing LLM serving systems fail to provide good user experience because their optimization metrics are not always aligned with user experience. In this paper, we first introduce and define the notion of Quality-of-Experience (QoE) for text streaming services by considering each user's end-to-end interaction timeline. Based on this, we propose Andes, a QoE-aware LLM serving system that enhances user experience by ensuring that users receive the first token promptly and subsequent tokens at a smooth, digestible pace, even during surge periods. This is enabled by Andes's preemptive request scheduler that dynamically prioritizes requests at the token granularity based on each request's expected QoE gain and GPU resource usage. Our evaluations demonstrate that, compared to state-of-the-art LLM serving systems, Andes improves the average QoE by up to $4.7\times$ given the same GPU resource, or saves up to 61% GPU resources while maintaining the same high QoE.

Jiachen Liu, Jiaxin Pei, Jintao Huang et al.

Scientific publication compresses a branching, iterative research process into a linear narrative, discarding the majority of what was discovered along the way. This compilation imposes two structural costs: a Storytelling Tax, where failed experiments, rejected hypotheses, and the branching exploration process are discarded to fit a linear narrative; and an Engineering Tax, where the gap between reviewer-sufficient prose and agent-sufficient specification leaves critical implementation details unwritten. Tolerable for human readers, these costs become critical when AI agents must understand, reproduce, and extend published work. We introduce the Agent-Native Research Artifact (Ara), a protocol that replaces the narrative paper with a machine-executable research package structured around four layers: scientific logic, executable code with full specifications, an exploration graph that preserves the failures compilation discards, and evidence grounding every claim in raw outputs. Three mechanisms support the ecosystem: a Live Research Manager that captures decisions and dead ends during ordinary development; an Ara Compiler that translates legacy PDFs and repos into Aras; and an Ara-native review system that automates objective checks so human reviewers can focus on significance, novelty, and taste. On PaperBench and RE-Bench, Ara raises question-answering accuracy from 72.4% to 93.7% and reproduction success from 57.4% to 64.4%. On RE-Bench's five open-ended extension tasks, preserved failure traces in Ara accelerate progress, but can also constrain a capable agent from stepping outside the prior-run box depending on the agent's capabilities.

Jiachen Liu, Ziheng Geng, Ran Cao et al.

Large language models (LLMs) have exhibited remarkable capabilities across diverse open-domain tasks, yet their application in specialized domains such as civil engineering remains largely unexplored. This paper starts bridging this gap by evaluating and enhancing the reliability and robustness of LLMs in structural analysis of beams. Reliability is assessed through the accuracy of correct outputs under repetitive runs of the same problems, whereas robustness is evaluated via the performance across varying load and boundary conditions. A benchmark dataset, comprising eight beam analysis problems, is created to test the Llama-3.3 70B Instruct model. Results show that, despite a qualitative understanding of structural mechanics, the LLM lacks the quantitative reliability and robustness for engineering applications. To address these limitations, a shift is proposed that reframes the structural analysis as code generation tasks. Accordingly, an LLM-empowered agent is developed that (a) integrates chain-of-thought and few-shot prompting to generate accurate OpeeSeesPy code, and (b) automatically executes the code to produce structural analysis results. Experimental results demonstrate that the agent achieves accuracy exceeding 99.0% on the benchmark dataset, exhibiting reliable and robust performance across diverse conditions. Ablation studies highlight the complete example and function usage examples as the primary contributors to the agent's enhanced performance.

Yu Zeng, Yang Zhang, Jiachen Liu et al.

Hair editing is a critical image synthesis task that aims to edit hair color and hairstyle using text descriptions or reference images, while preserving irrelevant attributes (e.g., identity, background, cloth). Many existing methods are based on StyleGAN to address this task. However, due to the limited spatial distribution of StyleGAN, it struggles with multiple hair color editing and facial preservation. Considering the advancements in diffusion models, we utilize Latent Diffusion Models (LDMs) for hairstyle editing. Our approach introduces Multi-stage Hairstyle Blend (MHB), effectively separating control of hair color and hairstyle in diffusion latent space. Additionally, we train a warping module to align the hair color with the target region. To further enhance multi-color hairstyle editing, we fine-tuned a CLIP model using a multi-color hairstyle dataset. Our method not only tackles the complexity of multi-color hairstyles but also addresses the challenge of preserving original colors during diffusion editing. Extensive experiments showcase the superiority of our method in editing multi-color hairstyles while preserving facial attributes given textual descriptions and reference images.

Qixin Wan, Zilong Wang, Jingwen Zhou et al.

Foundation models have shown remarkable capabilities in various domains, but their performance on complex, multimodal engineering problems remains largely unexplored. We introduce SoM-1K, the first large-scale multimodal benchmark dataset dedicated to evaluating foundation models on problems in the strength of materials (SoM). The dataset, which contains 1,065 annotated SoM problems, mirrors real-world engineering tasks by including both textual problem statements and schematic diagrams. Due to the limited capabilities of current foundation models in understanding complicated visual information, we propose a novel prompting strategy called Descriptions of Images (DoI), which provides rigorous expert-generated text descriptions of the visual diagrams as the context. We evaluate eight representative foundation models, including both large language models (LLMs) and vision language models (VLMs). Our results show that current foundation models struggle significantly with these engineering problems, with the best-performing model achieving only 56.6% accuracy. Interestingly, we found that LLMs, when provided with DoI, often outperform VLMs provided with visual diagrams. A detailed error analysis reveals that DoI plays a crucial role in mitigating visual misinterpretation errors, suggesting that accurate text-based descriptions can be more effective than direct image input for current foundation models. This work establishes a rigorous benchmark for engineering AI and highlights a critical need for developing more robust multimodal reasoning capabilities in foundation models, particularly in scientific and engineering contexts.

Ziheng Geng, Jiachen Liu, Ran Cao et al.

Large language models (LLMs) such as GPT and Gemini have demonstrated remarkable capabilities in contextual understanding and reasoning. The strong performance of LLMs has sparked growing interest in leveraging them to automate tasks traditionally dependent on human expertise. Recently, LLMs have been integrated into intelligent agents capable of operating structural analysis software (e.g., OpenSees) to construct structural models and perform analyses. However, existing LLMs are limited in handling multi-step structural modeling due to frequent hallucinations and error accumulation during long-sequence operations. To this end, this study presents a novel multi-agent architecture to automate the structural modeling and analysis using OpenSeesPy. First, problem analysis and construction planning agents extract key parameters from user descriptions and formulate a stepwise modeling plan. Node and element agents then operate in parallel to assemble the frame geometry, followed by a load assignment agent. The resulting geometric and load information is translated into executable OpenSeesPy scripts by code translation agents. The proposed architecture is evaluated on a benchmark of 20 frame problems over ten repeated trials, achieving 100% accuracy in 18 cases and 90% in the remaining two. The architecture also significantly improves computational efficiency and demonstrates scalability to larger structural systems.

Ziheng Geng, Jiachen Liu, Ran Cao et al.

Large language models (LLMs) have recently been used to empower autonomous agents in engineering, significantly improving automation and efficiency in labor-intensive workflows. However, their potential remains underexplored in structural engineering, particularly for finite element modeling tasks requiring geometric modeling, complex reasoning, and domain knowledge. To bridge this gap, this paper develops a LLM-based multi-agent system to automate finite element modeling of 2D frames. The system decomposes structural analysis into subtasks, each managed by a specialized agent powered by the lightweight Llama-3.3 70B Instruct model. The workflow begins with a Problem Analysis Agent, which extracts geometry, boundary, and material parameters from the user input. Next, a Geometry Agent incrementally derives node coordinates and element connectivity by applying expert-defined rules. These structured outputs are converted into executable OpenSeesPy code by a Translation Agent and refined by a Model Validation Agent through consistency checks. Then, a Load Agent applies load conditions into the assembled structural model. Experimental evaluations on 20 benchmark problems demonstrate that the system achieves accuracy over 80% in most cases across 10 repeated trials, outperforming Gemini-2.5 Pro and ChatGPT-4o models.

Yuxuan Zhu, Jiachen Liu, Mosharaf Chowdhury et al.

Federated learning (FL) aims to train machine learning (ML) models across potentially millions of edge client devices. Yet, training and customizing models for FL clients is notoriously challenging due to the heterogeneity of client data, device capabilities, and the massive scale of clients, making individualized model exploration prohibitively expensive. State-of-the-art FL solutions personalize a globally trained model or concurrently train multiple models, but they often incur suboptimal model accuracy and huge training costs. In this paper, we introduce FedTrans, a multi-model FL training framework that automatically produces and trains high-accuracy, hardware-compatible models for individual clients at scale. FedTrans begins with a basic global model, identifies accuracy bottlenecks in model architectures during training, and then employs model transformation to derive new models for heterogeneous clients on the fly. It judiciously assigns models to individual clients while performing soft aggregation on multi-model updates to minimize total training costs. Our evaluations using realistic settings show that FedTrans improves individual client model accuracy by 14% - 72% while slashing training costs by 1.6X - 20X over state-of-the-art solutions.

Sarah Jabbour, Trenton Chang, Anindya Das Antar et al.

Generative AI (GenAI) models have become vital across industries, yet current evaluation methods have not adapted to their widespread use. Traditional evaluations often rely on benchmarks and fixed datasets, frequently failing to reflect real-world performance, which creates a gap between lab-tested outcomes and practical applications. This white paper proposes a comprehensive framework for how we should evaluate real-world GenAI systems, emphasizing diverse, evolving inputs and holistic, dynamic, and ongoing assessment approaches. The paper offers guidance for practitioners on how to design evaluation methods that accurately reflect real-time capabilities, and provides policymakers with recommendations for crafting GenAI policies focused on societal impacts, rather than fixed performance numbers or parameter sizes. We advocate for holistic frameworks that integrate performance, fairness, and ethics and the use of continuous, outcome-oriented methods that combine human and automated assessments while also being transparent to foster trust among stakeholders. Implementing these strategies ensures GenAI models are not only technically proficient but also ethically responsible and impactful.

Jiachen Liu, Yuan Xue, Haomiao Ni et al.

Generating realistic building layouts for automatic building design has been studied in both the computer vision and architecture domains. Traditional approaches from the architecture domain, which are based on optimization techniques or heuristic design guidelines, can synthesize desirable layouts, but usually require post-processing and involve human interaction in the design pipeline, making them costly and timeconsuming. The advent of deep generative models has significantly improved the fidelity and diversity of the generated architecture layouts, reducing the workload by designers and making the process much more efficient. In this paper, we conduct a comprehensive review of three major research topics of architecture layout design and generation: floorplan layout generation, scene layout synthesis, and generation of some other formats of building layouts. For each topic, we present an overview of the leading paradigms, categorized either by research domains (architecture or machine learning) or by user input conditions or constraints. We then introduce the commonly-adopted benchmark datasets that are used to verify the effectiveness of the methods, as well as the corresponding evaluation metrics. Finally, we identify the well-solved problems and limitations of existing approaches, then propose new perspectives as promising directions for future research in this important research area. A project associated with this survey to maintain the resources is available at awesome-building-layout-generation.

Changkun Liu, Bin Tan, Zeran Ke et al.

This paper addresses metric 3D reconstruction of indoor scenes by exploiting their inherent geometric regularities with compact representations. Using planar 3D primitives - a well-suited representation for man-made environments - we introduce PLANA3R, a pose-free framework for metric Planar 3D Reconstruction from unposed two-view images. Our approach employs Vision Transformers to extract a set of sparse planar primitives, estimate relative camera poses, and supervise geometry learning via planar splatting, where gradients are propagated through high-resolution rendered depth and normal maps of primitives. Unlike prior feedforward methods that require 3D plane annotations during training, PLANA3R learns planar 3D structures without explicit plane supervision, enabling scalable training on large-scale stereo datasets using only depth and normal annotations. We validate PLANA3R on multiple indoor-scene datasets with metric supervision and demonstrate strong generalization to out-of-domain indoor environments across diverse tasks under metric evaluation protocols, including 3D surface reconstruction, depth estimation, and relative pose estimation. Furthermore, by formulating with planar 3D representation, our method emerges with the ability for accurate plane segmentation. The project page is available at https://lck666666.github.io/plana3r

Wei Li, Xue Xu, Jiachen Liu et al.

Existing text-to-image diffusion models primarily generate images from text prompts. However, the inherent conciseness of textual descriptions poses challenges in faithfully synthesizing images with intricate details, such as specific entities or scenes. This paper presents UNIMO-G, a simple multimodal conditional diffusion framework that operates on multimodal prompts with interleaved textual and visual inputs, which demonstrates a unified ability for both text-driven and subject-driven image generation. UNIMO-G comprises two core components: a Multimodal Large Language Model (MLLM) for encoding multimodal prompts, and a conditional denoising diffusion network for generating images based on the encoded multimodal input. We leverage a two-stage training strategy to effectively train the framework: firstly pre-training on large-scale text-image pairs to develop conditional image generation capabilities, and then instruction tuning with multimodal prompts to achieve unified image generation proficiency. A well-designed data processing pipeline involving language grounding and image segmentation is employed to construct multi-modal prompts. UNIMO-G excels in both text-to-image generation and zero-shot subject-driven synthesis, and is notably effective in generating high-fidelity images from complex multimodal prompts involving multiple image entities.

Zhe Hu, Zhiwei Cao, Hou Pong Chan et al.

Controllable text generation is an appealing but challenging task, which allows users to specify particular attributes of the generated outputs. In this paper, we propose a controllable dialogue generation model to steer response generation under multi-attribute constraints. Specifically, we define and categorize the commonly used control attributes into global and local ones, which possess different granularities of effects on response generation. Then, we significantly extend the conventional seq2seq framework by introducing a novel two-stage decoder, which first uses a multi-grained style specification layer to impose the stylistic constraints and determine word-level control states of responses based on the attributes, and then employs a response generation layer to generate final responses maintaining both semantic relevancy to the contexts and fidelity to the attributes. Furthermore, we train our model with an attribute consistency reward to promote response control with explicit supervision signals. Extensive experiments and in-depth analyses on two datasets indicate that our model can significantly outperform competitive baselines in terms of response quality, content diversity and controllability.

Wenhao Wu, Wei Li, Xinyan Xiao et al.

Abstractive summarization for long-document or multi-document remains challenging for the Seq2Seq architecture, as Seq2Seq is not good at analyzing long-distance relations in text. In this paper, we present BASS, a novel framework for Boosting Abstractive Summarization based on a unified Semantic graph, which aggregates co-referent phrases distributing across a long range of context and conveys rich relations between phrases. Further, a graph-based encoder-decoder model is proposed to improve both the document representation and summary generation process by leveraging the graph structure. Specifically, several graph augmentation methods are designed to encode both the explicit and implicit relations in the text while the graph-propagation attention mechanism is developed in the decoder to select salient content into the summary. Empirical results show that the proposed architecture brings substantial improvements for both long-document and multi-document summarization tasks.

Wei Li, Can Gao, Guocheng Niu et al.

Existed pre-training methods either focus on single-modal tasks or multi-modal tasks, and cannot effectively adapt to each other. They can only utilize single-modal data (i.e. text or image) or limited multi-modal data (i.e. image-text pairs). In this work, we propose a unified-modal pre-training architecture, namely UNIMO, which can effectively adapt to both single-modal and multi-modal understanding and generation tasks. Large scale of free text corpus and image collections can be utilized to improve the capability of visual and textual understanding, and cross-modal contrastive learning (CMCL) is leveraged to align the textual and visual information into a unified semantic space over a corpus of image-text pairs. As the non-paired single-modal data is very rich, our model can utilize much larger scale of data to learn more generalizable representations. Moreover, the textual knowledge and visual knowledge can enhance each other in the unified semantic space. The experimental results show that UNIMO significantly improves the performance of several single-modal and multi-modal downstream tasks. Our code and pre-trained models are public at the UNIMO project page https://unimo-ptm.github.io/

Wei Li, Xinyan Xiao, Jiachen Liu et al.

Graphs that capture relations between textual units have great benefits for detecting salient information from multiple documents and generating overall coherent summaries. In this paper, we develop a neural abstractive multi-document summarization (MDS) model which can leverage well-known graph representations of documents such as similarity graph and discourse graph, to more effectively process multiple input documents and produce abstractive summaries. Our model utilizes graphs to encode documents in order to capture cross-document relations, which is crucial to summarizing long documents. Our model can also take advantage of graphs to guide the summary generation process, which is beneficial for generating coherent and concise summaries. Furthermore, pre-trained language models can be easily combined with our model, which further improve the summarization performance significantly. Empirical results on the WikiSum and MultiNews dataset show that the proposed architecture brings substantial improvements over several strong baselines.

Chulun Zhou, Liangyu Chen, Jiachen Liu et al.

Unsupervised style transfer aims to change the style of an input sentence while preserving its original content without using parallel training data. In current dominant approaches, owing to the lack of fine-grained control on the influence from the target style,they are unable to yield desirable output sentences. In this paper, we propose a novel attentional sequence-to-sequence (Seq2seq) model that dynamically exploits the relevance of each output word to the target style for unsupervised style transfer. Specifically, we first pretrain a style classifier, where the relevance of each input word to the original style can be quantified via layer-wise relevance propagation. In a denoising auto-encoding manner, we train an attentional Seq2seq model to reconstruct input sentences and repredict word-level previously-quantified style relevance simultaneously. In this way, this model is endowed with the ability to automatically predict the style relevance of each output word. Then, we equip the decoder of this model with a neural style component to exploit the predicted wordlevel style relevance for better style transfer. Particularly, we fine-tune this model using a carefully-designed objective function involving style transfer, style relevance consistency, content preservation and fluency modeling loss terms. Experimental results show that our proposed model achieves state-of-the-art performance in terms of both transfer accuracy and content preservation.

Hengkai Guo, Wenji Wang, Guanjun Guo et al.

While propagation-based approaches have achieved state-of-the-art performance for video object segmentation, the literature lacks a fair comparison of different methods using the same settings. In this paper, we carry out an empirical study for propagation-based methods. We view these approaches from a unified perspective and conduct detailed ablation study for core methods, input cues, multi-object combination and training strategies. With careful designs, our improved end-to-end memory networks achieve a global mean of 76.1 on DAVIS 2017 val set.

Zhen Wang, Jiachen Liu, Xinyan Xiao et al.

While sophisticated neural-based techniques have been developed in reading comprehension, most approaches model the answer in an independent manner, ignoring its relations with other answer candidates. This problem can be even worse in open-domain scenarios, where candidates from multiple passages should be combined to answer a single question. In this paper, we formulate reading comprehension as an extract-then-select two-stage procedure. We first extract answer candidates from passages, then select the final answer by combining information from all the candidates. Furthermore, we regard candidate extraction as a latent variable and train the two-stage process jointly with reinforcement learning. As a result, our approach has improved the state-of-the-art performance significantly on two challenging open-domain reading comprehension datasets. Further analysis demonstrates the effectiveness of our model components, especially the information fusion of all the candidates and the joint training of the extract-then-select procedure.