Huan Wang

Devansh Arpit, Matthew Fernandez, Itai Feigenbaum et al. · salesforce, stanford

We introduce the Salesforce CausalAI Library, an open-source library for causal analysis using observational data. It supports causal discovery and causal inference for tabular and time series data, of discrete, continuous and heterogeneous types. This library includes algorithms that handle linear and non-linear causal relationships between variables, and uses multi-processing for speed-up. We also include a data generator capable of generating synthetic data with specified structural equation model for the aforementioned data formats and types, that helps users control the ground-truth causal process while investigating various algorithms. Finally, we provide a user interface (UI) that allows users to perform causal analysis on data without coding. The goal of this library is to provide a fast and flexible solution for a variety of problems in the domain of causality. This technical report describes the Salesforce CausalAI API along with its capabilities, the implementations of the supported algorithms, and experiments demonstrating their performance and speed. Our library is available at \url{https://github.com/salesforce/causalai}.

Aadyot Bhatnagar, Huan Wang, Caiming Xiong et al. · salesforce

We study the problem of uncertainty quantification via prediction sets, in an online setting where the data distribution may vary arbitrarily over time. Recent work develops online conformal prediction techniques that leverage regret minimization algorithms from the online learning literature to learn prediction sets with approximately valid coverage and small regret. However, standard regret minimization could be insufficient for handling changing environments, where performance guarantees may be desired not only over the full time horizon but also in all (sub-)intervals of time. We develop new online conformal prediction methods that minimize the strongly adaptive regret, which measures the worst-case regret over all intervals of a fixed length. We prove that our methods achieve near-optimal strongly adaptive regret for all interval lengths simultaneously, and approximately valid coverage. Experiments show that our methods consistently obtain better coverage and smaller prediction sets than existing methods on real-world tasks, such as time series forecasting and image classification under distribution shift.

Runyu Zhang, Qinghua Liu, Huan Wang et al. · salesforce

This paper studies policy optimization algorithms for multi-agent reinforcement learning. We begin by proposing an algorithm framework for two-player zero-sum Markov Games in the full-information setting, where each iteration consists of a policy update step at each state using a certain matrix game algorithm, and a value update step with a certain learning rate. This framework unifies many existing and new policy optimization algorithms. We show that the state-wise average policy of this algorithm converges to an approximate Nash equilibrium (NE) of the game, as long as the matrix game algorithms achieve low weighted regret at each state, with respect to weights determined by the speed of the value updates. Next, we show that this framework instantiated with the Optimistic Follow-The-Regularized-Leader (OFTRL) algorithm at each state (and smooth value updates) can find an $\mathcal{\widetilde{O}}(T^{-5/6})$ approximate NE in $T$ iterations, and a similar algorithm with slightly modified value update rule achieves a faster $\mathcal{\widetilde{O}}(T^{-1})$ convergence rate. These improve over the current best $\mathcal{\widetilde{O}}(T^{-1/2})$ rate of symmetric policy optimization type algorithms. We also extend this algorithm to multi-player general-sum Markov Games and show an $\mathcal{\widetilde{O}}(T^{-3/4})$ convergence rate to Coarse Correlated Equilibria (CCE). Finally, we provide a numerical example to verify our theory and investigate the importance of smooth value updates, and find that using "eager" value updates instead (equivalent to the independent natural policy gradient algorithm) may significantly slow down the convergence, even on a simple game with $H=2$ layers.

Itai Feigenbaum, Huan Wang, Shelby Heinecke et al. · salesforce, stanford

Causal discovery aims to recover a causal graph from data generated by it; constraint based methods do so by searching for a d-separating conditioning set of nodes in the graph via an oracle. In this paper, we provide analytic evidence that on large graphs, d-separation is a rare phenomenon, even when guaranteed to exist, unless the graph is extremely sparse. We then provide an analytic average case analysis of the PC Algorithm for causal discovery, as well as a variant of the SGS Algorithm we call UniformSGS. We consider a set $V=\{v_1,\ldots,v_n\}$ of nodes, and generate a random DAG $G=(V,E)$ where $(v_a, v_b) \in E$ with i.i.d. probability $p_1$ if $a<b$ and $0$ if $a > b$. We provide upper bounds on the probability that a subset of $V-\{x,y\}$ d-separates $x$ and $y$, conditional on $x$ and $y$ being d-separable; our upper bounds decay exponentially fast to $0$ as $|V| \rightarrow \infty$. For the PC Algorithm, while it is known that its worst-case guarantees fail on non-sparse graphs, we show that the same is true for the average case, and that the sparsity requirement is quite demanding: for good performance, the density must go to $0$ as $|V| \rightarrow \infty$ even in the average case. For UniformSGS, while it is known that the running time is exponential for existing edges, we show that in the average case, that is the expected running time for most non-existing edges as well.

Jianguo Zhang, Stephen Roller, Kun Qian et al. · salesforce

End-to-end task-oriented dialogue (TOD) systems have achieved promising performance by leveraging sophisticated natural language understanding and natural language generation capabilities of pre-trained models. This work enables the TOD systems with more flexibility through a simple cache. The cache provides the flexibility to dynamically update the TOD systems and handle both existing and unseen dialogue scenarios. Towards this end, we first fine-tune a retrieval module to effectively retrieve the most relevant information entries from the cache. We then train end-to-end TOD models that can refer to and ground on both dialogue history and retrieved information during TOD generation. The cache is straightforward to construct, and the backbone models of TOD systems are compatible with existing pre-trained generative models. Extensive experiments demonstrate the superior performance of our framework, with a notable improvement in non-empty joint goal accuracy by 6.7% compared to strong baselines.

Jiacheng Guo, Minshuo Chen, Huan Wang et al. · salesforce

This paper studies the sample-efficiency of learning in Partially Observable Markov Decision Processes (POMDPs), a challenging problem in reinforcement learning that is known to be exponentially hard in the worst-case. Motivated by real-world settings such as loading in game playing, we propose an enhanced feedback model called ``multiple observations in hindsight'', where after each episode of interaction with the POMDP, the learner may collect multiple additional observations emitted from the encountered latent states, but may not observe the latent states themselves. We show that sample-efficient learning under this feedback model is possible for two new subclasses of POMDPs: \emph{multi-observation revealing POMDPs} and \emph{distinguishable POMDPs}. Both subclasses generalize and substantially relax \emph{revealing POMDPs} -- a widely studied subclass for which sample-efficient learning is possible under standard trajectory feedback. Notably, distinguishable POMDPs only require the emission distributions from different latent states to be \emph{different} instead of \emph{linearly independent} as required in revealing POMDPs.

Yu Bai, Fan Chen, Huan Wang et al.

Neural sequence models based on the transformer architecture have demonstrated remarkable \emph{in-context learning} (ICL) abilities, where they can perform new tasks when prompted with training and test examples, without any parameter update to the model. This work first provides a comprehensive statistical theory for transformers to perform ICL. Concretely, we show that transformers can implement a broad class of standard machine learning algorithms in context, such as least squares, ridge regression, Lasso, learning generalized linear models, and gradient descent on two-layer neural networks, with near-optimal predictive power on various in-context data distributions. Using an efficient implementation of in-context gradient descent as the underlying mechanism, our transformer constructions admit mild size bounds, and can be learned with polynomially many pretraining sequences. Building on these ``base'' ICL algorithms, intriguingly, we show that transformers can implement more complex ICL procedures involving \emph{in-context algorithm selection}, akin to what a statistician can do in real life -- A \emph{single} transformer can adaptively select different base ICL algorithms -- or even perform qualitatively different tasks -- on different input sequences, without any explicit prompting of the right algorithm or task. We both establish this in theory by explicit constructions, and also observe this phenomenon experimentally. In theory, we construct two general mechanisms for algorithm selection with concrete examples: pre-ICL testing, and post-ICL validation. As an example, we use the post-ICL validation mechanism to construct a transformer that can perform nearly Bayes-optimal ICL on a challenging task -- noisy linear models with mixed noise levels. Experimentally, we demonstrate the strong in-context algorithm selection capabilities of standard transformer architectures.

Tianyu Guo, Wei Hu, Song Mei et al.

While large language models based on the transformer architecture have demonstrated remarkable in-context learning (ICL) capabilities, understandings of such capabilities are still in an early stage, where existing theory and mechanistic understanding focus mostly on simple scenarios such as learning simple function classes. This paper takes initial steps on understanding ICL in more complex scenarios, by studying learning with representations. Concretely, we construct synthetic in-context learning problems with a compositional structure, where the label depends on the input through a possibly complex but fixed representation function, composed with a linear function that differs in each instance. By construction, the optimal ICL algorithm first transforms the inputs by the representation function, and then performs linear ICL on top of the transformed dataset. We show theoretically the existence of transformers that approximately implement such algorithms with mild depth and size. Empirically, we find trained transformers consistently achieve near-optimal ICL performance in this setting, and exhibit the desired dissection where lower layers transforms the dataset and upper layers perform linear ICL. Through extensive probing and a new pasting experiment, we further reveal several mechanisms within the trained transformers, such as concrete copying behaviors on both the inputs and the representations, linear ICL capability of the upper layers alone, and a post-ICL representation selection mechanism in a harder mixture setting. These observed mechanisms align well with our theory and may shed light on how transformers perform ICL in more realistic scenarios.

Tian Lan, Huan Wang, Caiming Xiong et al.

We introduce WarpSci, a domain agnostic framework designed to overcome crucial system bottlenecks encountered in the application of reinforcement learning to intricate environments with vast datasets featuring high-dimensional observation or action spaces. Notably, our framework eliminates the need for data transfer between the CPU and GPU, enabling the concurrent execution of thousands of simulations on a single or multiple GPUs. This high data throughput architecture proves particularly advantageous for data-driven scientific research, where intricate environment models are commonly essential.

Haolin Chen, Shiyu Wang, Can Qin et al.

Diffusion language models promise bidirectional context and infilling capabilities that autoregressive coders lack, yet practical systems remain heavyweight. We introduce CoDA, a 1.7B-parameter diffusion coder trained on TPU with a fully open-source training pipeline. CoDA pairs large-scale diffusion pre-training with code-centric mid-training and instruction tuning, enabling confidence-guided sampling that keeps inference latency competitive. On Humaneval, MBPP, and EvalPlus, CoDA-1.7B-Instruct matches or surpasses diffusion models up to 7B parameters. Our release includes model checkpoints, evaluation harnesses, and TPU training pipelines to accelerate research on lightweight diffusion-based coding assistants.

Devansh Arpit, Huan Wang, Yingbo Zhou et al.

In Domain Generalization (DG) settings, models trained independently on a given set of training domains have notoriously chaotic performance on distribution shifted test domains, and stochasticity in optimization (e.g. seed) plays a big role. This makes deep learning models unreliable in real world settings. We first show that this chaotic behavior exists even along the training optimization trajectory of a single model, and propose a simple model averaging protocol that both significantly boosts domain generalization and diminishes the impact of stochasticity by improving the rank correlation between the in-domain validation accuracy and out-domain test accuracy, which is crucial for reliable early stopping. Taking advantage of our observation, we show that instead of ensembling unaveraged models (that is typical in practice), ensembling moving average models (EoA) from independent runs further boosts performance. We theoretically explain the boost in performance of ensembling and model averaging by adapting the well known Bias-Variance trade-off to the domain generalization setting. On the DomainBed benchmark, when using a pre-trained ResNet-50, this ensemble of averages achieves an average of $68.0\%$, beating vanilla ERM (w/o averaging/ensembling) by $\sim 4\%$, and when using a pre-trained RegNetY-16GF, achieves an average of $76.6\%$, beating vanilla ERM by $6\%$. Our code is available at https://github.com/salesforce/ensemble-of-averages.

Devansh Arpit, Huan Wang, Caiming Xiong et al.

We introduce a parameterization method called Neural Bayes which allows computing statistical quantities that are in general difficult to compute and opens avenues for formulating new objectives for unsupervised representation learning. Specifically, given an observed random variable $\mathbf{x}$ and a latent discrete variable $z$, we can express $p(\mathbf{x}|z)$, $p(z|\mathbf{x})$ and $p(z)$ in closed form in terms of a sufficiently expressive function (Eg. neural network) using our parameterization without restricting the class of these distributions. To demonstrate its usefulness, we develop two independent use cases for this parameterization: 1. Mutual Information Maximization (MIM): MIM has become a popular means for self-supervised representation learning. Neural Bayes allows us to compute mutual information between observed random variables $\mathbf{x}$ and latent discrete random variables $z$ in closed form. We use this for learning image representations and show its usefulness on downstream classification tasks. 2. Disjoint Manifold Labeling: Neural Bayes allows us to formulate an objective which can optimally label samples from disjoint manifolds present in the support of a continuous distribution. This can be seen as a specific form of clustering where each disjoint manifold in the support is a separate cluster. We design clustering tasks that obey this formulation and empirically show that the model optimally labels the disjoint manifolds. Our code is available at \url{https://github.com/salesforce/NeuralBayes}

Itai Feigenbaum, Devansh Arpit, Huan Wang et al. · salesforce, stanford

Large Language Model (LLM) editing modifies factual information in LLMs. Locate-and-Edit (L\&E) methods accomplish this by finding where relevant information is stored within the neural network, and editing the weights at that location. The goal of editing is to modify the response of an LLM to a proposition independently of its phrasing, while not modifying its response to other related propositions. Existing methods are limited to binary propositions, which represent straightforward binary relations between a subject and an object. Furthermore, existing methods rely on semantic subject labels, which may not be available or even be well-defined in practice. In this paper, we show that both of these issues can be effectively skirted with a simple and fast localization method called Gradient Tracing (GT). This localization method allows editing arbitrary propositions instead of just binary ones, and does so without the need for subject labels. As propositions always have a truth value, our experiments prompt an LLM as a boolean classifier, and edit its T/F response to propositions. Our method applies GT for location tracing, and then edit the model at that location using a mild variant of Rank-One Model Editing (ROME). On datasets of binary propositions derived from the CounterFact dataset, we show that our method -- without access to subject labels -- performs close to state-of-the-art L\&E methods which has access subject labels. We then introduce a new dataset, Factual Accuracy Classification Test (FACT), which includes non-binary propositions and for which subject labels are not generally applicable, and therefore is beyond the scope of existing L\&E methods. Nevertheless, we show that with our method editing is possible on FACT.

Itai Feigenbaum, Devansh Arpit, Huan Wang et al.

Causal discovery aims to recover information about an unobserved causal graph from the observable data it generates. Layerings are orderings of the variables which place causes before effects. In this paper, we provide ways to recover layerings of a graph by accessing the data via a conditional entropy oracle, when distributions are discrete. Our algorithms work by repeatedly removing sources or sinks from the graph. Under appropriate assumptions and conditioning, we can separate the sources or sinks from the remainder of the nodes by comparing their conditional entropy to the unconditional entropy of their noise. Our algorithms are provably correct and run in worst-case quadratic time. The main assumptions are faithfulness and injective noise, and either known noise entropies or weakly monotonically increasing noise entropies along directed paths. In addition, we require one of either a very mild extension of faithfulness, or strictly monotonically increasing noise entropies, or expanding noise injectivity to include an additional single argument in the structural functions.

Yu Bai, Song Mei, Huan Wang et al.

Quantifying the data uncertainty in learning tasks is often done by learning a prediction interval or prediction set of the label given the input. Two commonly desired properties for learned prediction sets are \emph{valid coverage} and \emph{good efficiency} (such as low length or low cardinality). Conformal prediction is a powerful technique for learning prediction sets with valid coverage, yet by default its conformalization step only learns a single parameter, and does not optimize the efficiency over more expressive function classes. In this paper, we propose a generalization of conformal prediction to multiple learnable parameters, by considering the constrained empirical risk minimization (ERM) problem of finding the most efficient prediction set subject to valid empirical coverage. This meta-algorithm generalizes existing conformal prediction algorithms, and we show that it achieves approximate valid population coverage and near-optimal efficiency within class, whenever the function class in the conformalization step is low-capacity in a certain sense. Next, this ERM problem is challenging to optimize as it involves a non-differentiable coverage constraint. We develop a gradient-based algorithm for it by approximating the original constrained ERM using differentiable surrogate losses and Lagrangians. Experiments show that our algorithm is able to learn valid prediction sets and improve the efficiency significantly over existing approaches in several applications such as prediction intervals with improved length, minimum-volume prediction sets for multi-output regression, and label prediction sets for image classification.

Tengyang Xie, Nan Jiang, Huan Wang et al.

Recent theoretical work studies sample-efficient reinforcement learning (RL) extensively in two settings: learning interactively in the environment (online RL), or learning from an offline dataset (offline RL). However, existing algorithms and theories for learning near-optimal policies in these two settings are rather different and disconnected. Towards bridging this gap, this paper initiates the theoretical study of policy finetuning, that is, online RL where the learner has additional access to a "reference policy" $μ$ close to the optimal policy $π_\star$ in a certain sense. We consider the policy finetuning problem in episodic Markov Decision Processes (MDPs) with $S$ states, $A$ actions, and horizon length $H$. We first design a sharp offline reduction algorithm -- which simply executes $μ$ and runs offline policy optimization on the collected dataset -- that finds an $\varepsilon$ near-optimal policy within $\widetilde{O}(H^3SC^\star/\varepsilon^2)$ episodes, where $C^\star$ is the single-policy concentrability coefficient between $μ$ and $π_\star$. This offline result is the first that matches the sample complexity lower bound in this setting, and resolves a recent open question in offline RL. We then establish an $Ω(H^3S\min\{C^\star, A\}/\varepsilon^2)$ sample complexity lower bound for any policy finetuning algorithm, including those that can adaptively explore the environment. This implies that -- perhaps surprisingly -- the optimal policy finetuning algorithm is either offline reduction or a purely online RL algorithm that does not use $μ$. Finally, we design a new hybrid offline/online algorithm for policy finetuning that achieves better sample complexity than both vanilla offline reduction and purely online RL algorithms, in a relaxed setting where $μ$ only satisfies concentrability partially up to a certain time step.

Rachel Luo, Aadyot Bhatnagar, Yu Bai et al.

Probabilistic classifiers output confidence scores along with their predictions, and these confidence scores should be calibrated, i.e., they should reflect the reliability of the prediction. Confidence scores that minimize standard metrics such as the expected calibration error (ECE) accurately measure the reliability on average across the entire population. However, it is in general impossible to measure the reliability of an individual prediction. In this work, we propose the local calibration error (LCE) to span the gap between average and individual reliability. For each individual prediction, the LCE measures the average reliability of a set of similar predictions, where similarity is quantified by a kernel function on a pretrained feature space and by a binning scheme over predicted model confidences. We show theoretically that the LCE can be estimated sample-efficiently from data, and empirically find that it reveals miscalibration modes that are more fine-grained than the ECE can detect. Our key result is a novel local recalibration method LoRe, to improve confidence scores for individual predictions and decrease the LCE. Experimentally, we show that our recalibration method produces more accurate confidence scores, which improves downstream fairness and decision making on classification tasks with both image and tabular data.

Stanislaw Jastrzebski, Devansh Arpit, Oliver Astrand et al.

The early phase of training a deep neural network has a dramatic effect on the local curvature of the loss function. For instance, using a small learning rate does not guarantee stable optimization because the optimization trajectory has a tendency to steer towards regions of the loss surface with increasing local curvature. We ask whether this tendency is connected to the widely observed phenomenon that the choice of the learning rate strongly influences generalization. We first show that stochastic gradient descent (SGD) implicitly penalizes the trace of the Fisher Information Matrix (FIM), a measure of the local curvature, from the start of training. We argue it is an implicit regularizer in SGD by showing that explicitly penalizing the trace of the FIM can significantly improve generalization. We highlight that poor final generalization coincides with the trace of the FIM attaining a large value early in training, to which we refer as catastrophic Fisher explosion. Finally, to gain insight into the regularization effect of penalizing the trace of the FIM, we show that it limits memorization by reducing the learning speed of examples with noisy labels more than that of the examples with clean labels.

Yu Bai, Minshuo Chen, Pan Zhou et al.

Meta-learning aims to perform fast adaptation on a new task through learning a "prior" from multiple existing tasks. A common practice in meta-learning is to perform a train-validation split (\emph{train-val method}) where the prior adapts to the task on one split of the data, and the resulting predictor is evaluated on another split. Despite its prevalence, the importance of the train-validation split is not well understood either in theory or in practice, particularly in comparison to the more direct \emph{train-train method}, which uses all the per-task data for both training and evaluation. We provide a detailed theoretical study on whether and when the train-validation split is helpful in the linear centroid meta-learning problem. In the agnostic case, we show that the expected loss of the train-val method is minimized at the optimal prior for meta testing, and this is not the case for the train-train method in general without structural assumptions on the data. In contrast, in the realizable case where the data are generated from linear models, we show that both the train-val and train-train losses are minimized at the optimal prior in expectation. Further, perhaps surprisingly, our main result shows that the train-train method achieves a \emph{strictly better} excess loss in this realizable case, even when the regularization parameter and split ratio are optimally tuned for both methods. Our results highlight that sample splitting may not always be preferable, especially when the data is realizable by the model. We validate our theories by experimentally showing that the train-train method can indeed outperform the train-val method, on both simulations and real meta-learning tasks.