Yi Dong

Boxin Wang, Wei Ping, Peng Xu et al. · nvidia

Large decoder-only language models (LMs) can be largely improved in terms of perplexity by retrieval (e.g., RETRO), but its impact on text generation quality and downstream task accuracy is unclear. Thus, it is still an open question: shall we pretrain large autoregressive LMs with retrieval? To answer it, we perform a comprehensive study on a scalable pre-trained retrieval-augmented LM (i.e., RETRO) compared with standard GPT and retrieval-augmented GPT incorporated at fine-tuning or inference stages. We first provide the recipe to reproduce RETRO up to 9.5B parameters while retrieving a text corpus with 330B tokens. Based on that, we have the following novel findings: i) RETRO outperforms GPT on text generation with much less degeneration (i.e., repetition), moderately higher factual accuracy, and slightly lower toxicity with a nontoxic retrieval database. ii) On the LM Evaluation Harness benchmark, RETRO largely outperforms GPT on knowledge-intensive tasks, but is on par with GPT on other tasks. Furthermore, we introduce a simple variant of the model, RETRO++, which largely improves open-domain QA results of original RETRO (e.g., EM score +8.6 on Natural Question) and significantly outperforms retrieval-augmented GPT in both fine-tuning and zero-shot evaluation settings. Our findings highlight the promising direction of pretraining autoregressive LMs with retrieval as future foundation models. We release our code and model at: https://github.com/NVIDIA/Megatron-LM/blob/main/tools/retro/README.md

Zhilin Wang, Yi Dong, Jiaqi Zeng et al. · nvidia

Existing open-source helpfulness preference datasets do not specify what makes some responses more helpful and others less so. Models trained on these datasets can incidentally learn to model dataset artifacts (e.g. preferring longer but unhelpful responses only due to their length). To alleviate this problem, we collect HelpSteer, a multi-attribute helpfulness dataset annotated for the various aspects that make responses helpful. Specifically, our 37k-sample dataset has annotations for correctness, coherence, complexity, and verbosity in addition to overall helpfulness of responses. Training Llama 2 70B using the HelpSteer dataset with SteerLM technique produces a model that scores 7.54 on MT Bench, which is currently the highest score for open models that do not require training data from more powerful models (e.g. GPT4). We release this dataset with CC-BY-4.0 license at https://huggingface.co/datasets/nvidia/HelpSteer

Yi Dong, Zhilin Wang, Makesh Narsimhan Sreedhar et al. · nvidia

Model alignment with human preferences is an essential step in making Large Language Models (LLMs) helpful and consistent with human values. It typically consists of supervised fine-tuning (SFT) and reinforcement learning from human feedback (RLHF) stages. However, RLHF faces inherent limitations stemming from a complex training setup and its tendency to align the model with implicit values that end users cannot control at run-time. Moreover, reward models in RLHF stage commonly rely on single-dimensional feedback as opposed to explicit, multifaceted signals that indicate attributes such as helpfulness, humor, and toxicity. To address these limitations, we propose SteerLM, a supervised fine-tuning method that empowers end-users to control responses during inference. SteerLM conditions responses to conform to an explicitly defined multi-dimensional set of attributes, thereby empowering a steerable AI capable of generating helpful and high-quality responses while maintaining customizability. Experiments show that SteerLM trained on open source datasets generates responses that are preferred by human and automatic evaluators to many state-of-the-art baselines trained with RLHF while being much easier to train. Try SteerLM at https://huggingface.co/nvidia/SteerLM-llama2-13B

Yi Dong, Xingyu Zhao, Sen Wang et al.

Deep Reinforcement Learning (DRL) has achieved impressive performance in robotics and autonomous systems (RAS). A key challenge to its deployment in real-life operations is the presence of spuriously unsafe DRL policies. Unexplored states may lead the agent to make wrong decisions that could result in hazards, especially in applications where DRL-trained end-to-end controllers govern the behaviour of RAS. This paper proposes a novel quantitative reliability assessment framework for DRL-controlled RAS, leveraging verification evidence generated from formal reliability analysis of neural networks. A two-level verification framework is introduced to check the safety property with respect to inaccurate observations that are due to, e.g., environmental noise and state changes. Reachability verification tools are leveraged locally to generate safety evidence of trajectories. In contrast, at the global level, we quantify the overall reliability as an aggregated metric of local safety evidence, corresponding to a set of distinct tasks and their occurrence probabilities. The effectiveness of the proposed verification framework is demonstrated and validated via experiments on real RAS.

Yi Dong, Yang Chen, Xingyu Zhao et al.

With the employment of smart meters, massive data on consumer behaviour can be collected by retailers. From the collected data, the retailers may obtain the household profile information and implement demand response. While retailers prefer to acquire a model as accurate as possible among different customers, there are two major challenges. First, different retailers in the retail market do not share their consumer's electricity consumption data as these data are regarded as their assets, which has led to the problem of data island. Second, the electricity load data are highly heterogeneous since different retailers may serve various consumers. To this end, a fully distributed short-term load forecasting framework based on a consensus algorithm and Long Short-Term Memory (LSTM) is proposed, which may protect the customer's privacy and satisfy the accurate load forecasting requirement. Specifically, a fully distributed learning framework is exploited for distributed training, and a consensus technique is applied to meet confidential privacy. Case studies show that the proposed method has comparable performance with centralised methods regarding the accuracy, but the proposed method shows advantages in training speed and data privacy.

Hongjin Su, Shizhe Diao, Ximing Lu et al.

Large language models are powerful generalists, yet solving deep and complex problems such as those of the Humanity's Last Exam (HLE) remains both conceptually challenging and computationally expensive. We show that small orchestrators managing other models and a variety of tools can both push the upper bound of intelligence and improve efficiency in solving difficult agentic tasks. We introduce ToolOrchestra, a method for training small orchestrators that coordinate intelligent tools. ToolOrchestra explicitly uses reinforcement learning with outcome-, efficiency-, and user-preference-aware rewards. Using ToolOrchestra, we produce Orchestrator, an 8B model that achieves higher accuracy at lower cost than previous tool-use agents while aligning with user preferences on which tools are to be used for a given query. On HLE, Orchestrator achieves a score of 37.1%, outperforming GPT-5 (35.1%) while being 2.5x more efficient. On tau2-Bench and FRAMES, Orchestrator surpasses GPT-5 by a wide margin while using only about 30% of the cost. Extensive analysis shows that Orchestrator achieves the best trade-off between performance and cost under multiple metrics, and generalizes robustly to unseen tools. These results demonstrate that composing diverse tools with a lightweight orchestration model is both more efficient and more effective than existing methods, paving the way for practical and scalable tool-augmented reasoning systems.

Jinwei Hu, Yi Dong, Zhengtao Ding et al.

This paper presents a defense framework for enhancing the safety of large language model (LLM) empowered multi-agent systems (MAS) in safety-critical domains such as aerospace. We apply randomized smoothing, a statistical robustness certification technique, to the MAS consensus context, enabling probabilistic guarantees on agent decisions under adversarial influence. Unlike traditional verification methods, our approach operates in black-box settings and employs a two-stage adaptive sampling mechanism to balance robustness and computational efficiency. Simulation results demonstrate that our method effectively prevents the propagation of adversarial behaviors and hallucinations while maintaining consensus performance. This work provides a practical and scalable path toward safe deployment of LLM-based MAS in real-world, high-stakes environments.

Yi Dong, Yingjie Wang, Mariana Gama et al.

In the realm of power systems, the increasing involvement of residential users in load forecasting applications has heightened concerns about data privacy. Specifically, the load data can inadvertently reveal the daily routines of residential users, thereby posing a risk to their property security. While federated learning (FL) has been employed to safeguard user privacy by enabling model training without the exchange of raw data, these FL models have shown vulnerabilities to emerging attack techniques, such as Deep Leakage from Gradients and poisoning attacks. To counteract these, we initially employ a Secure-Aggregation (SecAgg) algorithm that leverages multiparty computation cryptographic techniques to mitigate the risk of gradient leakage. However, the introduction of SecAgg necessitates the deployment of additional sub-center servers for executing the multiparty computation protocol, thereby escalating computational complexity and reducing system robustness, especially in scenarios where one or more sub-centers are unavailable. To address these challenges, we introduce a Markovian Switching-based distributed training framework, the convergence of which is substantiated through rigorous theoretical analysis. The Distributed Markovian Switching (DMS) topology shows strong robustness towards the poisoning attacks as well. Case studies employing real-world power system load data validate the efficacy of our proposed algorithm. It not only significantly minimizes communication complexity but also maintains accuracy levels comparable to traditional FL methods, thereby enhancing the scalability of our load forecasting algorithm.

Haoran Chen, Dong Yi, Moyan Cao et al.

Anomaly analysis in surveillance videos is a crucial topic in computer vision. In recent years, multimodal large language models (MLLMs) have outperformed task-specific models in various domains. Although MLLMs are particularly versatile, their abilities to understand anomalous concepts and details are insufficiently studied because of the outdated benchmarks of this field not providing MLLM-style QAs and efficient algorithms to assess the model's open-ended text responses. To fill this gap, we propose a benchmark for crime surveillance video analysis with large models denoted as UCVL, including 1,829 videos and reorganized annotations from the UCF-Crime and UCF-Crime Annotation datasets. We design six types of questions and generate diverse QA pairs. Then we develop detailed instructions and use OpenAI's GPT-4o for accurate assessment. We benchmark eight prevailing MLLMs ranging from 0.5B to 40B parameters, and the results demonstrate the reliability of this bench. Moreover, we finetune LLaVA-OneVision on UCVL's training set. The improvement validates our data's high quality for video anomaly analysis.

Xiaowei Huang, Wenjie Ruan, Wei Huang et al.

Large Language Models (LLMs) have exploded a new heatwave of AI for their ability to engage end-users in human-level conversations with detailed and articulate answers across many knowledge domains. In response to their fast adoption in many industrial applications, this survey concerns their safety and trustworthiness. First, we review known vulnerabilities and limitations of the LLMs, categorising them into inherent issues, attacks, and unintended bugs. Then, we consider if and how the Verification and Validation (V&V) techniques, which have been widely developed for traditional software and deep learning models such as convolutional neural networks as independent processes to check the alignment of their implementations against the specifications, can be integrated and further extended throughout the lifecycle of the LLMs to provide rigorous analysis to the safety and trustworthiness of LLMs and their applications. Specifically, we consider four complementary techniques: falsification and evaluation, verification, runtime monitoring, and regulations and ethical use. In total, 370+ references are considered to support the quick understanding of the safety and trustworthiness issues from the perspective of V&V. While intensive research has been conducted to identify the safety and trustworthiness issues, rigorous yet practical methods are called for to ensure the alignment of LLMs with safety and trustworthiness requirements.

Yi Dong, Wei Huang, Vibhav Bharti et al.

The increasing use of Machine Learning (ML) components embedded in autonomous systems -- so-called Learning-Enabled Systems (LESs) -- has resulted in the pressing need to assure their functional safety. As for traditional functional safety, the emerging consensus within both, industry and academia, is to use assurance cases for this purpose. Typically assurance cases support claims of reliability in support of safety, and can be viewed as a structured way of organising arguments and evidence generated from safety analysis and reliability modelling activities. While such assurance activities are traditionally guided by consensus-based standards developed from vast engineering experience, LESs pose new challenges in safety-critical application due to the characteristics and design of ML models. In this article, we first present an overall assurance framework for LESs with an emphasis on quantitative aspects, e.g., breaking down system-level safety targets to component-level requirements and supporting claims stated in reliability metrics. We then introduce a novel model-agnostic Reliability Assessment Model (RAM) for ML classifiers that utilises the operational profile and robustness verification evidence. We discuss the model assumptions and the inherent challenges of assessing ML reliability uncovered by our RAM and propose solutions to practical use. Probabilistic safety argument templates at the lower ML component-level are also developed based on the RAM. Finally, to evaluate and demonstrate our methods, we not only conduct experiments on synthetic/benchmark datasets but also scope our methods with case studies on simulated Autonomous Underwater Vehicles and physical Unmanned Ground Vehicles.

Yi Dong, Xingyu Zhao, Xiaowei Huang

While Deep Reinforcement Learning (DRL) provides transformational capabilities to the control of Robotics and Autonomous Systems (RAS), the black-box nature of DRL and uncertain deployment environments of RAS pose new challenges on its dependability. Although existing works impose constraints on the DRL policy to ensure successful completion of the mission, it is far from adequate to assess the DRL-driven RAS in a holistic way considering all dependability properties. In this paper, we formally define a set of dependability properties in temporal logic and construct a Discrete-Time Markov Chain (DTMC) to model the dynamics of risk/failures of a DRL-driven RAS interacting with the stochastic environment. We then conduct Probabilistic Model Checking (PMC) on the designed DTMC to verify those properties. Our experimental results show that the proposed method is effective as a holistic assessment framework while uncovering conflicts between the properties that may need trade-offs in training. Moreover, we find that the standard DRL training cannot improve dependability properties, thus requiring bespoke optimisation objectives. Finally, our method offers sensitivity analysis of dependability properties to disturbance levels from environments, providing insights for the assurance of real RAS.

Xingyu Zhao, Wei Huang, Sven Schewe et al.

The utilisation of Deep Learning (DL) raises new challenges regarding its dependability in critical applications. Sound verification and validation methods are needed to assure the safe and reliable use of DL. However, state-of-the-art debug testing methods on DL that aim at detecting adversarial examples (AEs) ignore the operational profile, which statistically depicts the software's future operational use. This may lead to very modest effectiveness on improving the software's delivered reliability, as the testing budget is likely to be wasted on detecting AEs that are unrealistic or encountered very rarely in real-life operation. In this paper, we first present the novel notion of "operational AEs" which are AEs that have relatively high chance to be seen in future operation. Then an initial design of a new DL testing method to efficiently detect "operational AEs" is provided, as well as some insights on our prospective research plan.