Yicheng Wang

Yicheng Wang, Jiayi Yuan, Yu-Neng Chuang et al. · amazon-science, microsoft-research

Large Language Models (LLMs) are increasingly serving as evaluators in Natural Language Generation (NLG) tasks; this is often referred to as ``LLM-as-a-judge'' paradigm. However, the capabilities of LLMs in evaluating NLG quality remain underexplored. Current studies depend on human assessments and simple metrics that fail to capture the discernment of LLMs across diverse NLG tasks. To address this gap, we propose the Discernment of Hierarchical Perturbation (DHP) benchmarking framework, which provides quantitative discernment scores for LLMs. This framework leverages hierarchically perturbed text data and statistical tests to systematically measure the NLG evaluation capabilities of LLMs. We re-established six evaluation datasets for this benchmark, covering four NLG tasks: Summarization, Story Completion, Question Answering, and Translation. Our comprehensive benchmarking of five major LLM families provides critical insight into their strengths and limitations as NLG evaluators. Our dataset is available at https://huggingface.co/datasets/YCWANGVINCE/DHP_Benchmark.

Andrey Ignatov, Grigory Malivenko, Radu Timofte et al. · tencent-ai

Various depth estimation models are now widely used on many mobile and IoT devices for image segmentation, bokeh effect rendering, object tracking and many other mobile tasks. Thus, it is very crucial to have efficient and accurate depth estimation models that can run fast on low-power mobile chipsets. In this Mobile AI challenge, the target was to develop deep learning-based single image depth estimation solutions that can show a real-time performance on IoT platforms and smartphones. For this, the participants used a large-scale RGB-to-depth dataset that was collected with the ZED stereo camera capable to generated depth maps for objects located at up to 50 meters. The runtime of all models was evaluated on the Raspberry Pi 4 platform, where the developed solutions were able to generate VGA resolution depth maps at up to 27 FPS while achieving high fidelity results. All models developed in the challenge are also compatible with any Android or Linux-based mobile devices, their detailed description is provided in this paper.

Zhuoer Wang, Yicheng Wang, Ziwei Zhu et al. · amazon-science, microsoft-research

Question generation is a widely used data augmentation approach with extensive applications, and extracting qualified candidate answers from context passages is a critical step for most question generation systems. However, existing methods for candidate answer extraction are reliant on linguistic rules or annotated data that face the partial annotation issue and challenges in generalization. To overcome these limitations, we propose a novel unsupervised candidate answer extraction approach that leverages the inherent structure of context passages through a Differentiable Masker-Reconstructor (DMR) Model with the enforcement of self-consistency for picking up salient information tokens. We curated two datasets with exhaustively-annotated answers and benchmark a comprehensive set of supervised and unsupervised candidate answer extraction methods. We demonstrate the effectiveness of the DMR model by showing its performance is superior among unsupervised methods and comparable to supervised methods.

Yicheng Wang, Xiaotian Han, Chia-Yuan Chang et al.

Physics-informed neural networks (PINNs) are revolutionizing science and engineering practice by bringing together the power of deep learning to bear on scientific computation. In forward modeling problems, PINNs are meshless partial differential equation (PDE) solvers that can handle irregular, high-dimensional physical domains. Naturally, the neural architecture hyperparameters have a large impact on the efficiency and accuracy of the PINN solver. However, this remains an open and challenging problem because of the large search space and the difficulty of identifying a proper search objective for PDEs. Here, we propose Auto-PINN, the first systematic, automated hyperparameter optimization approach for PINNs, which employs Neural Architecture Search (NAS) techniques to PINN design. Auto-PINN avoids manually or exhaustively searching the hyperparameter space associated with PINNs. A comprehensive set of pre-experiments using standard PDE benchmarks allows us to probe the structure-performance relationship in PINNs. We find that the different hyperparameters can be decoupled, and that the training loss function of PINNs is a good search objective. Comparison experiments with baseline methods demonstrate that Auto-PINN produces neural architectures with superior stability and accuracy over alternative baselines.

Tianshi Zheng, Jiaxin Bai, Yicheng Wang et al. · tencent-ai

While large language models (LLMs) have demonstrated impressive capabilities across various natural language processing tasks by acquiring rich factual knowledge from their broad training data, their ability to synthesize and logically reason with this knowledge in complex ways remains underexplored. In this work, we present a systematic evaluation of state-of-the-art LLMs' complex logical reasoning abilities through a novel benchmark of automatically generated complex reasoning questions over general domain and biomedical knowledge graphs. Our extensive experiments, employing diverse in-context learning techniques, reveal that LLMs excel at reasoning over general world knowledge but face significant challenges with specialized domain-specific knowledge. We find that prompting with explicit Chain-of-Thought demonstrations can substantially improve LLM performance on complex logical reasoning tasks with diverse logical operations. Interestingly, our controlled evaluations uncover an asymmetry where LLMs display proficiency at set union operations, but struggle considerably with set intersections - a key building block of logical reasoning. To foster further work, we will publicly release our evaluation benchmark and code.

Hongtao Wang, Jiangshe Zhang, Xiaoli Wei et al.

Picking the first arrival times of prestack gathers is called First Arrival Time (FAT) picking, which is an indispensable step in seismic data processing, and is mainly solved manually in the past. With the current increasing density of seismic data collection, the efficiency of manual picking has been unable to meet the actual needs. Therefore, automatic picking methods have been greatly developed in recent decades, especially those based on deep learning. However, few of the current supervised deep learning-based method can avoid the dependence on labeled samples. Besides, since the gather data is a set of signals which are greatly different from the natural images, it is difficult for the current method to solve the FAT picking problem in case of a low Signal to Noise Ratio (SNR). In this paper, for hard rock seismic gather data, we propose a Multi-Stage Segmentation Pickup Network (MSSPN), which solves the generalization problem across worksites and the picking problem in the case of low SNR. In MSSPN, there are four sub-models to simulate the manually picking processing, which is assumed to four stages from coarse to fine. Experiments on seven field datasets with different qualities show that our MSSPN outperforms benchmarks by a large margin.Particularly, our method can achieve more than 90\% accurate picking across worksites in the case of medium and high SNRs, and even fine-tuned model can achieve 88\% accurate picking of the dataset with low SNR.

Yicheng Wang, Xiaotian Han, Leisheng Yu et al.

Parkinson's disease (PD), a neurodegenerative disorder, often manifests as speech and voice dysfunction. While utilizing voice data for PD detection has great potential in clinical applications, the widely used deep learning models currently have fairness issues regarding different ages of onset. These deep models perform well for the elderly group (age $>$ 55) but are less accurate for the young group (age $\leq$ 55). Through our investigation, the discrepancy between the elderly and the young arises due to 1) an imbalanced dataset and 2) the milder symptoms often seen in early-onset patients. However, traditional debiasing methods are impractical as they typically impair the prediction accuracy for the majority group while minimizing the discrepancy. To address this issue, we present a new debiasing method using GradCAM-based feature masking combined with ensemble models, ensuring that neither fairness nor accuracy is compromised. Specifically, the GradCAM-based feature masking selectively obscures age-related features in the input voice data while preserving essential information for PD detection. The ensemble models further improve the prediction accuracy for the minority (young group). Our approach effectively improves detection accuracy for early-onset patients without sacrificing performance for the elderly group. Additionally, we propose a two-step detection strategy for the young group, offering a practical risk assessment for potential early-onset PD patients.

Gemini Team, Petko Georgiev, Ving Ian Lei et al. · deepmind, mila

In this report, we introduce the Gemini 1.5 family of models, representing the next generation of highly compute-efficient multimodal models capable of recalling and reasoning over fine-grained information from millions of tokens of context, including multiple long documents and hours of video and audio. The family includes two new models: (1) an updated Gemini 1.5 Pro, which exceeds the February version on the great majority of capabilities and benchmarks; (2) Gemini 1.5 Flash, a more lightweight variant designed for efficiency with minimal regression in quality. Gemini 1.5 models achieve near-perfect recall on long-context retrieval tasks across modalities, improve the state-of-the-art in long-document QA, long-video QA and long-context ASR, and match or surpass Gemini 1.0 Ultra's state-of-the-art performance across a broad set of benchmarks. Studying the limits of Gemini 1.5's long-context ability, we find continued improvement in next-token prediction and near-perfect retrieval (>99%) up to at least 10M tokens, a generational leap over existing models such as Claude 3.0 (200k) and GPT-4 Turbo (128k). Finally, we highlight real-world use cases, such as Gemini 1.5 collaborating with professionals on completing their tasks achieving 26 to 75% time savings across 10 different job categories, as well as surprising new capabilities of large language models at the frontier; when given a grammar manual for Kalamang, a language with fewer than 200 speakers worldwide, the model learns to translate English to Kalamang at a similar level to a person who learned from the same content.

Gheorghe Comanici, Eric Bieber, Mike Schaekermann et al. · amazon-science, baidu

In this report, we introduce the Gemini 2.X model family: Gemini 2.5 Pro and Gemini 2.5 Flash, as well as our earlier Gemini 2.0 Flash and Flash-Lite models. Gemini 2.5 Pro is our most capable model yet, achieving SoTA performance on frontier coding and reasoning benchmarks. In addition to its incredible coding and reasoning skills, Gemini 2.5 Pro is a thinking model that excels at multimodal understanding and it is now able to process up to 3 hours of video content. Its unique combination of long context, multimodal and reasoning capabilities can be combined to unlock new agentic workflows. Gemini 2.5 Flash provides excellent reasoning abilities at a fraction of the compute and latency requirements and Gemini 2.0 Flash and Flash-Lite provide high performance at low latency and cost. Taken together, the Gemini 2.X model generation spans the full Pareto frontier of model capability vs cost, allowing users to explore the boundaries of what is possible with complex agentic problem solving.

Daxiang Dong, Mingming Zheng, Dong Xu et al.

We present Qianfan-OCR, a 4B-parameter end-to-end vision-language model that unifies document parsing, layout analysis, and document understanding within a single architecture. It performs direct image-to-Markdown conversion and supports diverse prompt-driven tasks including table extraction, chart understanding, document QA, and key information extraction. To address the loss of explicit layout analysis in end-to-end OCR, we propose Layout-as-Thought, an optional thinking phase triggered by special think tokens that generates structured layout representations -- bounding boxes, element types, and reading order -- before producing final outputs, recovering layout grounding capabilities while improving accuracy on complex layouts. Qianfan-OCR ranks first among end-to-end models on OmniDocBench v1.5 (93.12) and OlmOCR Bench (79.8), achieves competitive results on OCRBench, CCOCR, DocVQA, and ChartQA against general VLMs of comparable scale, and attains the highest average score on public key information extraction benchmarks, surpassing Gemini-3.1-Pro, Seed-2.0, and Qwen3-VL-235B. The model is publicly accessible via the Baidu AI Cloud Qianfan platform.

Yicheng Wang, Mark Cusick, Mohamed Laila et al.

Automatic speech recognition (ASR) techniques have become powerful tools, enhancing efficiency in law enforcement scenarios. To ensure fairness for demographic groups in different acoustic environments, ASR engines must be tested across a variety of speakers in realistic settings. However, describing the fairness discrepancies between models with confidence remains a challenge. Meanwhile, most public ASR datasets are insufficient to perform a satisfying fairness evaluation. To address the limitations, we built FairLENS - a systematic fairness evaluation framework. We propose a novel and adaptable evaluation method to examine the fairness disparity between different models. We also collected a fairness evaluation dataset covering multiple scenarios and demographic dimensions. Leveraging this framework, we conducted fairness assessments on 1 open-source and 11 commercially available state-of-the-art ASR models. Our results reveal that certain models exhibit more biases than others, serving as a fairness guideline for users to make informed choices when selecting ASR models for a given real-world scenario. We further explored model biases towards specific demographic groups and observed that shifts in the acoustic domain can lead to the emergence of new biases.

Daochen Zha, Zaid Pervaiz Bhat, Yi-Wei Chen et al.

Action recognition is an important task for video understanding with broad applications. However, developing an effective action recognition solution often requires extensive engineering efforts in building and testing different combinations of the modules and their hyperparameters. In this demo, we present AutoVideo, a Python system for automated video action recognition. AutoVideo is featured for 1) highly modular and extendable infrastructure following the standard pipeline language, 2) an exhaustive list of primitives for pipeline construction, 3) data-driven tuners to save the efforts of pipeline tuning, and 4) easy-to-use Graphical User Interface (GUI). AutoVideo is released under MIT license at https://github.com/datamllab/autovideo

Jiaxin Bai, Yicheng Wang, Tianshi Zheng et al.

Abductive reasoning is the process of making educated guesses to provide explanations for observations. Although many applications require the use of knowledge for explanations, the utilization of abductive reasoning in conjunction with structured knowledge, such as a knowledge graph, remains largely unexplored. To fill this gap, this paper introduces the task of complex logical hypothesis generation, as an initial step towards abductive logical reasoning with KG. In this task, we aim to generate a complex logical hypothesis so that it can explain a set of observations. We find that the supervised trained generative model can generate logical hypotheses that are structurally closer to the reference hypothesis. However, when generalized to unseen observations, this training objective does not guarantee better hypothesis generation. To address this, we introduce the Reinforcement Learning from Knowledge Graph (RLF-KG) method, which minimizes differences between observations and conclusions drawn from generated hypotheses according to the KG. Experiments show that, with RLF-KG's assistance, the generated hypotheses provide better explanations, and achieve state-of-the-art results on three widely used KGs.

Junru Zhou, Yicheng Wang, Pan Li

Graph Neural Networks (GNNs) have emerged as a promising approach for ``learning to branch'' in Mixed-Integer Linear Programming (MILP). While standard Message-Passing GNNs (MPNNs) are efficient, they theoretically lack the expressive power to fully represent MILP structures. Conversely, higher-order GNNs (like 2-FGNNs) are expressive but computationally prohibitive. In this work, we investigate Subgraph GNNs as a theoretical middle ground. Crucially, while previous work [Chen et al., 2025] demonstrated that GNNs with 3-WL expressive power can approximate Strong Branching, we prove a sharper result: node-anchored Subgraph GNNs whose expressive power is strictly lower than 3-WL [Zhang et al., 2023] are sufficient to approximate Strong Branching scores. However, our extensive empirical evaluation on four benchmark datasets reveals a stark contrast between theory and practice. While node-anchored Subgraph GNNs theoretically offer superior branching decisions, their $O(n)$ complexity overhead results in significant memory bottlenecks and slower solving times than MPNNs and heuristics. Our results indicate that for MILP branching, the computational cost of expressive GNNs currently outweighs their gains in decision quality, suggesting that future research must focus on efficiency-preserving expressivity.

Chengli Tan, Jiangshe Zhang, Junmin Liu et al.

Recently, sharpness-aware minimization (SAM) has attracted much attention because of its surprising effectiveness in improving generalization performance. However, compared to stochastic gradient descent (SGD), it is more prone to getting stuck at the saddle points, which as a result may lead to performance degradation. To address this issue, we propose a simple renormalization strategy, dubbed Stable SAM (SSAM), so that the gradient norm of the descent step maintains the same as that of the ascent step. Our strategy is easy to implement and flexible enough to integrate with SAM and its variants, almost at no computational cost. With elementary tools from convex optimization and learning theory, we also conduct a theoretical analysis of sharpness-aware training, revealing that compared to SGD, the effectiveness of SAM is only assured in a limited regime of learning rate. In contrast, we show how SSAM extends this regime of learning rate and then it can consistently perform better than SAM with the minor modification. Finally, we demonstrate the improved performance of SSAM on several representative data sets and tasks.

Yicheng Wang, Zhikang Zhang, Jue Wang et al. · amazon-science

In various video-language learning tasks, the challenge of achieving cross-modality alignment with multi-grained data persists. We propose a method to tackle this challenge from two crucial perspectives: data and modeling. Given the absence of a multi-grained video-text pretraining dataset, we introduce a Granularity EXpansion (GEX) method with Integration and Compression operations to expand the granularity of a single-grained dataset. To better model multi-grained data, we introduce an Iterative Approximation Module (IAM), which embeds multi-grained videos and texts into a unified, low-dimensional semantic space while preserving essential information for cross-modal alignment. Furthermore, GEXIA is highly scalable with no restrictions on the number of video-text granularities for alignment. We evaluate our work on three categories of video tasks across seven benchmark datasets, showcasing state-of-the-art or comparable performance. Remarkably, our model excels in tasks involving long-form video understanding, even though the pretraining dataset only contains short video clips.

Yicheng Wang, Feng Liu, Junmin Liu et al.

As a promising field in open-world learning, \textit{Novel Class Discovery} (NCD) is usually a task to cluster unseen novel classes in an unlabeled set based on the prior knowledge of labeled data within the same domain. However, the performance of existing NCD methods could be severely compromised when novel classes are sampled from a different distribution with the labeled ones. In this paper, we explore and establish the solvability of NCD with cross domain setting under the necessary condition that the style information needs to be removed. Based on the theoretical analysis, we introduce an exclusive style removal module for extracting style information that is distinctive from the baseline features, thereby facilitating inference. Moreover, this module is easy to integrate with other NCD methods, acting as a plug-in to improve performance on novel classes with different distributions compared to the labeled set. Additionally, recognizing the non-negligible influence of different backbones and pre-training strategies on the performance of the NCD methods, we build a fair benchmark for future NCD research. Extensive experiments on three common datasets demonstrate the effectiveness of our proposed style removal strategy.

Zimo Wang, Yicheng Wang, Sensen Wu

Confronted with the spatial heterogeneity of real estate market, some traditional research utilized Geographically Weighted Regression (GWR) to estimate the house price. However, its kernel function is non-linear, elusive, and complex to opt bandwidth, the predictive power could also be improved. Consequently, a novel technique, Geographical Neural Network Weighted Regression (GNNWR), has been applied to improve the accuracy of real estate appraisal with the help of neural networks. Based on Shenzhen house price dataset, this work conspicuously captures the weight distribution of different variants at Shenzhen real estate market, which GWR is difficult to materialize. Moreover, we focus on the performance of GNNWR, verify its robustness and superiority, refine the experiment process with 10-fold cross-validation, extend its application area from natural to socioeconomic geospatial data. It's a practical and trenchant way to assess house price, and we demonstrate the effectiveness of GNNWR on a complex socioeconomic dataset.

Andrey Ignatov, Grigory Malivenko, David Plowman et al.

Depth estimation is an important computer vision problem with many practical applications to mobile devices. While many solutions have been proposed for this task, they are usually very computationally expensive and thus are not applicable for on-device inference. To address this problem, we introduce the first Mobile AI challenge, where the target is to develop an end-to-end deep learning-based depth estimation solutions that can demonstrate a nearly real-time performance on smartphones and IoT platforms. For this, the participants were provided with a new large-scale dataset containing RGB-depth image pairs obtained with a dedicated stereo ZED camera producing high-resolution depth maps for objects located at up to 50 meters. The runtime of all models was evaluated on the popular Raspberry Pi 4 platform with a mobile ARM-based Broadcom chipset. The proposed solutions can generate VGA resolution depth maps at up to 10 FPS on the Raspberry Pi 4 while achieving high fidelity results, and are compatible with any Android or Linux-based mobile devices. A detailed description of all models developed in the challenge is provided in this paper.

Yicheng Wang, Shuang Xu, Junmin Liu et al.

Multi-Focus Image Fusion (MFIF) is a promising image enhancement technique to obtain all-in-focus images meeting visual needs and it is a precondition of other computer vision tasks. One of the research trends of MFIF is to avoid the defocus spread effect (DSE) around the focus/defocus boundary (FDB). In this paper,we propose a network termed MFIF-GAN to attenuate the DSE by generating focus maps in which the foreground region are correctly larger than the corresponding objects. The Squeeze and Excitation Residual module is employed in the network. By combining the prior knowledge of training condition, this network is trained on a synthetic dataset based on an α-matte model. In addition, the reconstruction and gradient regularization terms are combined in the loss functions to enhance the boundary details and improve the quality of fused images. Extensive experiments demonstrate that the MFIF-GAN outperforms several state-of-the-art (SOTA) methods in visual perception, quantitative analysis as well as efficiency. Moreover, the edge diffusion and contraction module is firstly proposed to verify that focus maps generated by our method are accurate at the pixel level.

Shuang Xu, Zixiang Zhao, Yicheng Wang et al.

Image fusion is a significant problem in many fields including digital photography, computational imaging and remote sensing, to name but a few. Recently, deep learning has emerged as an important tool for image fusion. This paper presents three deep convolutional sparse coding (CSC) networks for three kinds of image fusion tasks (i.e., infrared and visible image fusion, multi-exposure image fusion, and multi-modal image fusion). The CSC model and the iterative shrinkage and thresholding algorithm are generalized into dictionary convolution units. As a result, all hyper-parameters are learned from data. Our extensive experiments and comprehensive comparisons reveal the superiority of the proposed networks with regard to quantitative evaluation and visual inspection.

Xiangyu Zhang, Ramin Bashizade, Yicheng Wang et al.

Statistical machine learning often uses probabilistic algorithms, such as Markov Chain Monte Carlo (MCMC), to solve a wide range of problems. Probabilistic computations, often considered too slow on conventional processors, can be accelerated with specialized hardware by exploiting parallelism and optimizing the design using various approximation techniques. Current methodologies for evaluating correctness of probabilistic accelerators are often incomplete, mostly focusing only on end-point result quality ("accuracy"). It is important for hardware designers and domain experts to look beyond end-point "accuracy" and be aware of the hardware optimizations impact on other statistical properties. This work takes a first step towards defining metrics and a methodology for quantitatively evaluating correctness of probabilistic accelerators beyond end-point result quality. We propose three pillars of statistical robustness: 1) sampling quality, 2) convergence diagnostic, and 3) goodness of fit. We apply our framework to a representative MCMC accelerator and surface design issues that cannot be exposed using only application end-point result quality. Applying the framework to guide design space exploration shows that statistical robustness comparable to floating-point software can be achieved by slightly increasing the bit representation, without floating-point hardware requirements.

Yixin Nie, Yicheng Wang, Mohit Bansal

Success in natural language inference (NLI) should require a model to understand both lexical and compositional semantics. However, through adversarial evaluation, we find that several state-of-the-art models with diverse architectures are over-relying on the former and fail to use the latter. Further, this compositionality unawareness is not reflected via standard evaluation on current datasets. We show that removing RNNs in existing models or shuffling input words during training does not induce large performance loss despite the explicit removal of compositional information. Therefore, we propose a compositionality-sensitivity testing setup that analyzes models on natural examples from existing datasets that cannot be solved via lexical features alone (i.e., on which a bag-of-words model gives a high probability to one wrong label), hence revealing the models' actual compositionality awareness. We show that this setup not only highlights the limited compositional ability of current NLI models, but also differentiates model performance based on design, e.g., separating shallow bag-of-words models from deeper, linguistically-grounded tree-based models. Our evaluation setup is an important analysis tool: complementing currently existing adversarial and linguistically driven diagnostic evaluations, and exposing opportunities for future work on evaluating models' compositional understanding.

Lisa Bauer, Yicheng Wang, Mohit Bansal

Reading comprehension QA tasks have seen a recent surge in popularity, yet most works have focused on fact-finding extractive QA. We instead focus on a more challenging multi-hop generative task (NarrativeQA), which requires the model to reason, gather, and synthesize disjoint pieces of information within the context to generate an answer. This type of multi-step reasoning also often requires understanding implicit relations, which humans resolve via external, background commonsense knowledge. We first present a strong generative baseline that uses a multi-attention mechanism to perform multiple hops of reasoning and a pointer-generator decoder to synthesize the answer. This model performs substantially better than previous generative models, and is competitive with current state-of-the-art span prediction models. We next introduce a novel system for selecting grounded multi-hop relational commonsense information from ConceptNet via a pointwise mutual information and term-frequency based scoring function. Finally, we effectively use this extracted commonsense information to fill in gaps of reasoning between context hops, using a selectively-gated attention mechanism. This boosts the model's performance significantly (also verified via human evaluation), establishing a new state-of-the-art for the task. We also show promising initial results of the generalizability of our background knowledge enhancements by demonstrating some improvement on QAngaroo-WikiHop, another multi-hop reasoning dataset.

Xiaoteng Zhang, Yixin Bao, Feiyun Zhang et al.

In this paper, we introduce our submissions for the tasks of trimmed activity recognition (Kinetics) and trimmed event recognition (Moments in Time) for Activitynet Challenge 2018. In the two tasks, non-local neural networks and temporal segment networks are implemented as our base models. Multi-modal cues such as RGB image, optical flow and acoustic signal have also been used in our method. We also propose new non-local-based models for further improvement on the recognition accuracy. The final submissions after ensembling the models achieve 83.5% top-1 accuracy and 96.8% top-5 accuracy on the Kinetics validation set, 35.81% top-1 accuracy and 62.59% top-5 accuracy on the MIT validation set.

Yicheng Wang, Mohit Bansal

It is shown that many published models for the Stanford Question Answering Dataset (Rajpurkar et al., 2016) lack robustness, suffering an over 50% decrease in F1 score during adversarial evaluation based on the AddSent (Jia and Liang, 2017) algorithm. It has also been shown that retraining models on data generated by AddSent has limited effect on their robustness. We propose a novel alternative adversary-generation algorithm, AddSentDiverse, that significantly increases the variance within the adversarial training data by providing effective examples that punish the model for making certain superficial assumptions. Further, in order to improve robustness to AddSent's semantic perturbations (e.g., antonyms), we jointly improve the model's semantic-relationship learning capabilities in addition to our AddSentDiverse-based adversarial training data augmentation. With these additions, we show that we can make a state-of-the-art model significantly more robust, achieving a 36.5% increase in F1 score under many different types of adversarial evaluation while maintaining performance on the regular SQuAD task.