Search

The Chain-of-Thought (CoT) paradigm has emerged as a critical approach for enhancing the reasoning capabilities of large language models (LLMs). However, despite their widespread adoption and success, CoT methods often exhibit instability due to their inability to consistently ensure the quality of generated reasoning paths, leading to sub-optimal reasoning performance. To address this challenge, we propose the \textbf{Strategic Chain-of-Thought} (SCoT), a novel methodology designed to refine LLM performance by integrating strategic knowledge prior to generating intermediate reasoning steps. SCoT employs a two-stage approach within a single prompt: first eliciting an effective problem-solving strategy, which is then used to guide the generation of high-quality CoT paths and final answers. Our experiments across eight challenging reasoning datasets demonstrate significant improvements, including a 21.05\% increase on the GSM8K dataset and 24.13\% on the Tracking\_Objects dataset, respectively, using the Llama3-8b model. Additionally, we extend the SCoT framework to develop a few-shot method with automatically matched demonstrations, yielding even stronger results. These findings underscore the efficacy of SCoT, highlighting its potential to substantially enhance LLM performance in complex reasoning tasks.

Real-world decision-making tasks are usually partially observable Markov decision processes (POMDPs), where the state is not fully observable. Recent progress has demonstrated that recurrent reinforcement learning (RL), which consists of a context encoder based on recurrent neural networks (RNNs) for unobservable state prediction and a multilayer perceptron (MLP) policy for decision making, can mitigate partial observability and serve as a robust baseline for POMDP tasks. However, previous recurrent RL methods face training stability issues due to the gradient instability of RNNs. In this paper, we propose Recurrent Off-policy RL with Context-Encoder-Specific Learning Rate (RESeL) to tackle this issue. Specifically, RESeL uses a lower learning rate for context encoder than other MLP layers to ensure the stability of the former while maintaining the training efficiency of the latter. We integrate this technique into existing off-policy RL methods, resulting in the RESeL algorithm. We evaluated RESeL in 18 POMDP tasks, including classic, meta-RL, and credit assignment scenarios, as well as five MDP locomotion tasks. The experiments demonstrate significant improvements in training stability with RESeL. Comparative results show that RESeL achieves notable performance improvements over previous recurrent RL baselines in POMDP tasks, and is competitive with or even surpasses state-of-the-art methods in MDP tasks. Further ablation studies highlight the necessity of applying a distinct learning rate for the context encoder.

Earth system predictability is challenged by the complexity of environmental dynamics and the multitude of variables involved. Current AI foundation models, although advanced by leveraging large and heterogeneous data, are often constrained by their size and data integration, limiting their effectiveness in addressing the full range of Earth system prediction challenges. To overcome these limitations, we introduce the Oak Ridge Base Foundation Model for Earth System Predictability (ORBIT), an advanced vision transformer model that scales up to 113 billion parameters using a novel hybrid tensor-data orthogonal parallelism technique. As the largest model of its kind, ORBIT surpasses the current climate AI foundation model size by a thousandfold. Performance scaling tests conducted on the Frontier supercomputer have demonstrated that ORBIT achieves 684 petaFLOPS to 1.6 exaFLOPS sustained throughput, with scaling efficiency maintained at 41% to 85% across 49,152 AMD GPUs. These breakthroughs establish new advances in AI-driven climate modeling and demonstrate promise to significantly improve the Earth system predictability.

Vision Transformers (ViTs) are pivotal for foundational models in scientific imagery, including Earth science applications, due to their capability to process large sequence lengths. While transformers for text has inspired scaling sequence lengths in ViTs, yet adapting these for ViTs introduces unique challenges. We develop distributed sequence parallelism for ViTs, enabling them to handle up to 1M tokens. Our approach, leveraging DeepSpeed-Ulysses and Long-Sequence-Segmentation with model sharding, is the first to apply sequence parallelism in ViT training, achieving a 94% batch scaling efficiency on 2,048 AMD-MI250X GPUs. Evaluating sequence parallelism in ViTs, particularly in models up to 10B parameters, highlighted substantial bottlenecks. We countered these with hybrid sequence, pipeline, tensor parallelism, and flash attention strategies, to scale beyond single GPU memory limits. Our method significantly enhances climate modeling accuracy by 20% in temperature predictions, marking the first training of a transformer model on a full-attention matrix over 188K sequence length.

We introduce a conditional pseudo-reversible normalizing flow for constructing surrogate models of a physical model polluted by additive noise to efficiently quantify forward and inverse uncertainty propagation. Existing surrogate modeling approaches usually focus on approximating the deterministic component of physical model. However, this strategy necessitates knowledge of noise and resorts to auxiliary sampling methods for quantifying inverse uncertainty propagation. In this work, we develop the conditional pseudo-reversible normalizing flow model to directly learn and efficiently generate samples from the conditional probability density functions. The training process utilizes dataset consisting of input-output pairs without requiring prior knowledge about the noise and the function. Our model, once trained, can generate samples from any conditional probability density functions whose high probability regions are covered by the training set. Moreover, the pseudo-reversibility feature allows for the use of fully-connected neural network architectures, which simplifies the implementation and enables theoretical analysis. We provide a rigorous convergence analysis of the conditional pseudo-reversible normalizing flow model, showing its ability to converge to the target conditional probability density function using the Kullback-Leibler divergence. To demonstrate the effectiveness of our method, we apply it to several benchmark tests and a real-world geologic carbon storage problem.

Binary function similarity detection plays an important role in a wide range of security applications. Existing works usually assume that the query function and target function share equal semantics and compare their full semantics to obtain the similarity. However, we find that the function mapping is more complex, especially when function inlining happens. In this paper, we will systematically investigate cross-inlining binary function similarity detection. We first construct a cross-inlining dataset by compiling 51 projects using 9 compilers, with 4 optimizations, to 6 architectures, with 2 inlining flags, which results in two datasets both with 216 combinations. Then we construct the cross-inlining function mappings by linking the common source functions in these two datasets. Through analysis of this dataset, we find that three cross-inlining patterns widely exist while existing work suffers when detecting cross-inlining binary function similarity. Next, we propose a pattern-based model named CI-Detector for cross-inlining matching. CI-Detector uses the attributed CFG to represent the semantics of binary functions and GNN to embed binary functions into vectors. CI-Detector respectively trains a model for these three cross-inlining patterns. Finally, the testing pairs are input to these three models and all the produced similarities are aggregated to produce the final similarity. We conduct several experiments to evaluate CI-Detector. Results show that CI-Detector can detect cross-inlining pairs with a precision of 81% and a recall of 97%, which exceeds all state-of-the-art works., Comment: Accepted at ICSE 2024 (Second Cycle). Camera-ready version

Learning a precise dynamics model can be crucial for offline reinforcement learning, which, unfortunately, has been found to be quite challenging. Dynamics models that are learned by fitting historical transitions often struggle to generalize to unseen transitions. In this study, we identify a hidden but pivotal factor termed dynamics reward that remains consistent across transitions, offering a pathway to better generalization. Therefore, we propose the idea of reward-consistent dynamics models: any trajectory generated by the dynamics model should maximize the dynamics reward derived from the data. We implement this idea as the MOREC (Model-based Offline reinforcement learning with Reward Consistency) method, which can be seamlessly integrated into previous offline model-based reinforcement learning (MBRL) methods. MOREC learns a generalizable dynamics reward function from offline data, which is subsequently employed as a transition filter in any offline MBRL method: when generating transitions, the dynamics model generates a batch of transitions and selects the one with the highest dynamics reward value. On a synthetic task, we visualize that MOREC has a strong generalization ability and can surprisingly recover some distant unseen transitions. On 21 offline tasks in D4RL and NeoRL benchmarks, MOREC improves the previous state-of-the-art performance by a significant margin, i.e., 4.6% on D4RL tasks and 25.9% on NeoRL tasks. Notably, MOREC is the first method that can achieve above 95% online RL performance in 6 out of 12 D4RL tasks and 3 out of 9 NeoRL tasks.

In the upcoming decade, deep learning may revolutionize the natural sciences, enhancing our capacity to model and predict natural occurrences. This could herald a new era of scientific exploration, bringing significant advancements across sectors from drug development to renewable energy. To answer this call, we present DeepSpeed4Science initiative (deepspeed4science.ai) which aims to build unique capabilities through AI system technology innovations to help domain experts to unlock today's biggest science mysteries. By leveraging DeepSpeed's current technology pillars (training, inference and compression) as base technology enablers, DeepSpeed4Science will create a new set of AI system technologies tailored for accelerating scientific discoveries by addressing their unique complexity beyond the common technical approaches used for accelerating generic large language models (LLMs). In this paper, we showcase the early progress we made with DeepSpeed4Science in addressing two of the critical system challenges in structural biology research.

Mini-app is an emerging form of mobile application that combines web technology with native capabilities. Its features, e.g., no need to download and no installation, have made it popular rapidly. However, privacy issues that violate the laws or regulations are breeding in the swiftly expanding mini-app ecosystem. The consistency between what the mini-app does about the data in the program code and what it declares in its privacy policy description is important. But no work has systematically investigated the privacy problem of the mini-app before. In this paper, to our best knowledge, we are the first to conduct the compliance detection of data practice and policy description in mini-apps. In this paper, we first customize a taint analysis method based on data entity dependency network to adapt to the characteristics of the JavaScript language in the mini-apps. Then, we transform data types and data operations to data practices in program codes and privacy policies, so as to finish a fine-grained consistency matching model.We crawl 100,000 mini-apps on WeChat client in the wild and extract 2,998 with a privacy policy. Among them, only 318 meet the consistency requirements, 2,680 are inconsistent, and the proportion of inconsistencies is as high as 89.4%. The inconsistency in the mini-app is very serious. Based on 6 real-world cases analyzed, in order to reduce this potential data leakage risk, we suggest that the developer should reduce the collection of irrelevant information and the straightforward use of templates, and the platform should provide data flow detection tools and privacy policy writing support.

Deep neural networks (DNNs) and natural language processing (NLP) systems have developed rapidly and have been widely used in various real-world fields. However, they have been shown to be vulnerable to backdoor attacks. Specifically, the adversary injects a backdoor into the model during the training phase, so that input samples with backdoor triggers are classified as the target class. Some attacks have achieved high attack success rates on the pre-trained language models (LMs), but there have yet to be effective defense methods. In this work, we propose a defense method based on deep model mutation testing. Our main justification is that backdoor samples are much more robust than clean samples if we impose random mutations on the LMs and that backdoors are generalizable. We first confirm the effectiveness of model mutation testing in detecting backdoor samples and select the most appropriate mutation operators. We then systematically defend against three extensively studied backdoor attack levels (i.e., char-level, word-level, and sentence-level) by detecting backdoor samples. We also make the first attempt to defend against the latest style-level backdoor attacks. We evaluate our approach on three benchmark datasets (i.e., IMDB, Yelp, and AG news) and three style transfer datasets (i.e., SST-2, Hate-speech, and AG news). The extensive experimental results demonstrate that our approach can detect backdoor samples more efficiently and accurately than the three state-of-the-art defense approaches.

Binary2source function matching is a fundamental task for many security applications, including Software Component Analysis (SCA). The "1-to-1" mechanism has been applied in existing binary2source matching works, in which one binary function is matched against one source function. However, we discovered that such mapping could be "1-to-n" (one query binary function maps multiple source functions), due to the existence of function inlining. To help conduct binary2source function matching under function inlining, we propose a method named O2NMatcher to generate Source Function Sets (SFSs) as the matching target for binary functions with inlining. We first propose a model named ECOCCJ48 for inlined call site prediction. To train this model, we leverage the compilable OSS to generate a dataset with labeled call sites (inlined or not), extract several features from the call sites, and design a compiler-opt-based multi-label classifier by inspecting the inlining correlations between different compilations. Then, we use this model to predict the labels of call sites in the uncompilable OSS projects without compilation and obtain the labeled function call graphs of these projects. Next, we regard the construction of SFSs as a sub-tree generation problem and design root node selection and edge extension rules to construct SFSs automatically. Finally, these SFSs will be added to the corpus of source functions and compared with binary functions with inlining. We conduct several experiments to evaluate the effectiveness of O2NMatcher and results show our method increases the performance of existing works by 6% and exceeds all the state-of-the-art works.

This paper revisits Deep Mutual Learning (DML), a simple yet effective computing paradigm. We propose using R\'{e}nyi divergence instead of the KL divergence, which is more flexible and tunable, to improve vanilla DML. This modification is able to consistently improve performance over vanilla DML with limited additional complexity. The convergence properties of the proposed paradigm are analyzed theoretically, and Stochastic Gradient Descent with a constant learning rate is shown to converge with $\mathcal{O}(1)$-bias in the worst case scenario for nonconvex optimization tasks. That is, learning will reach nearby local optima but continue searching within a bounded scope, which may help mitigate overfitting. Finally, our extensive empirical results demonstrate the advantage of combining DML and R\'{e}nyi divergence, leading to further improvement in model generalization.

This paper addresses policy learning in non-stationary environments and games with continuous actions. Rather than the classical reward maximization mechanism, inspired by the ideas of follow-the-regularized-leader (FTRL) and mirror descent (MD) update, we propose a no-regret style reinforcement learning algorithm PORL for continuous action tasks. We prove that PORL has a last-iterate convergence guarantee, which is important for adversarial and cooperative games. Empirical studies show that, in stationary environments such as MuJoCo locomotion controlling tasks, PORL performs equally well as, if not better than, the soft actor-critic (SAC) algorithm; in non-stationary environments including dynamical environments, adversarial training, and competitive games, PORL is superior to SAC in both a better final policy performance and a more stable training process.

Multimodel ensembling has been widely used to improve climate model predictions, and the improvement strongly depends on the ensembling scheme. In this work, we propose a Bayesian neural network (BNN) ensembling method, which combines climate models within a Bayesian model averaging framework, to improve the predictive capability of model ensembles. Our proposed BNN approach calculates spatiotemporally varying model weights and biases by leveraging individual models' simulation skill, calibrates the ensemble prediction against observations by considering observation data uncertainty, and quantifies epistemic uncertainty when extrapolating to new conditions. More importantly, the BNN method provides interpretability about which climate model contributes more to the ensemble prediction at which locations and times. Thus, beyond its predictive capability, the method also brings insights and understanding of the models to guide further model and data development. In this study, we apply the BNN weighting scheme to an ensemble of CMIP6 climate models for monthly precipitation prediction over the conterminous United States. In both synthetic and real case studies, we demonstrate that BNN produces predictions of monthly precipitation with higher accuracy than three baseline ensembling methods. BNN can correctly assign a larger weight to the regions and seasons where the individual model fits the observation better. Moreover, its offered interpretability is consistent with our understanding of localized climate model performance. Additionally, BNN shows an increasing uncertainty when the prediction is farther away from the period with constrained data, which appropriately reflects our predictive confidence and trustworthiness of the models in the changing climate.

Relative permeability is commonly used to model immiscible fluid flow through porous materials. In this work we derive the relative permeability relationship from conservation of energy, assuming that the system to be non-ergodic at large length scales and relying on averaging in both space and time to homogenize the behavior. Explicit criteria are obtained to define stationary conditions: (1) there can be no net change for extensive measures of the system state over the time averaging interval; (2) the net energy inputs into the system are zero, meaning that the net rate of work done on the system must balance with the heat removed; and (3) there is no net work performed due to the contribution of internal energy fluctuations. Results are then evaluated based on direct numerical simulation. Dynamic connectivity is observed during steady-state flow, which is quantitatively assessed based the Euler characteristic. We show that even during steady-state flow at low capillary number ($\mathsf{Ca}\sim1\times10^5$), typical flow processes will explore multiple connectivity states. The residence time for each connectivity state is captured based on the time-and-space average. The distribution for energy fluctuations is shown to be multi-modal and non-Gaussian when terms are considered independently. However, we demonstrate that their sum is zero. Given an appropriate choice of the thermodynamic driving force, we show that the conventional relative permeability relationship is sufficient to model the energy dissipation in systems with complex pore-scale dynamics that routinely alter the structure of fluid connected pathways.

Reinforcement learning (RL) solves sequential decision-making problems via a trial-and-error process interacting with the environment. While RL achieves outstanding success in playing complex video games that allow huge trial-and-error, making errors is always undesired in the real world. To improve the sample efficiency and thus reduce the errors, model-based reinforcement learning (MBRL) is believed to be a promising direction, which builds environment models in which the trial-and-errors can take place without real costs. In this survey, we take a review of MBRL with a focus on the recent progress in deep RL. For non-tabular environments, there is always a generalization error between the learned environment model and the real environment. As such, it is of great importance to analyze the discrepancy between policy training in the environment model and that in the real environment, which in turn guides the algorithm design for better model learning, model usage, and policy training. Besides, we also discuss the recent advances of model-based techniques in other forms of RL, including offline RL, goal-conditioned RL, multi-agent RL, and meta-RL. Moreover, we discuss the applicability and advantages of MBRL in real-world tasks. Finally, we end this survey by discussing the promising prospects for the future development of MBRL. We think that MBRL has great potential and advantages in real-world applications that were overlooked, and we hope this survey could attract more research on MBRL.

Recovering reward function from expert demonstrations is a fundamental problem in reinforcement learning. The recovered reward function captures the motivation of the expert. Agents can imitate experts by following these reward functions in their environment, which is known as apprentice learning. However, the agents may face environments different from the demonstrations, and therefore, desire transferable reward functions. Classical reward learning methods such as inverse reinforcement learning (IRL) or, equivalently, adversarial imitation learning (AIL), recover reward functions coupled with training dynamics, which are hard to be transferable. Previous dynamics-agnostic reward learning methods rely on assumptions such as that the reward function has to be state-only, restricting their applicability. In this work, we present a dynamics-agnostic discriminator-ensemble reward learning method (DARL) within the AIL framework, capable of learning both state-action and state-only reward functions. DARL achieves this by decoupling the reward function from training dynamics, employing a dynamics-agnostic discriminator on a latent space derived from the original state-action space. This latent space is optimized to minimize information on the dynamics. We moreover discover the policy-dependency issue of the AIL framework that reduces the transferability. DARL represents the reward function as an ensemble of discriminators during training to eliminate policy dependencies. Empirical studies on MuJoCo tasks with changed dynamics show that DARL better recovers the reward function and results in better imitation performance in transferred environments, handling both state-only and state-action reward scenarios.

The fairness characteristic is a critical attribute of trusted AI systems. A plethora of research has proposed diverse methods for individual fairness testing. However, they are suffering from three major limitations, i.e., low efficiency, low effectiveness, and model-specificity. This work proposes ExpGA, an explanationguided fairness testing approach through a genetic algorithm (GA). ExpGA employs the explanation results generated by interpretable methods to collect high-quality initial seeds, which are prone to derive discriminatory samples by slightly modifying feature values. ExpGA then adopts GA to search discriminatory sample candidates by optimizing a fitness value. Benefiting from this combination of explanation results and GA, ExpGA is both efficient and effective to detect discriminatory individuals. Moreover, ExpGA only requires prediction probabilities of the tested model, resulting in a better generalization capability to various models. Experiments on multiple real-world benchmarks, including tabular and text datasets, show that ExpGA presents higher efficiency and effectiveness than four state-of-the-art approaches.

Binary similarity analysis is critical to many code-reuse-related issues and "1-to-1" mechanism is widely applied, where one function in a binary file is matched against one function in a source file or binary file. However, we discover that function mapping is a more complex problem of "1-to-n" or even "n-to-n" due to the existence of function inlining. In this paper, we investigate the effect of function inlining on binary similarity analysis. We first construct 4 inlining-oriented datasets for four similarity analysis tasks, including code search, OSS reuse detection, vulnerability detection, and patch presence test. Then, we further study the extent of function inlining, the performance of existing works under function inlining, and the effectiveness of existing inlining-simulation strategies. Results show that the proportion of function inlining can reach nearly 70%, while most existing works neglect it and use "1-to-1" mechanism. The mismatches cause a 30% loss in performance during code search and a 40% loss during vulnerability detection. Moreover, two existing inlining-simulation strategies can only recover 60% of the inlined functions. We discover that inlining is usually cumulative when optimization increases. Conditional inlining and incremental inlining are suggested to design low-cost and high-coverage inlining-simulation strategies.