NeurIPS 2024 Papers

We explore uncertainty quantification in large language models (LLMs), with the goal to identify when uncertainty in responses given a query is large. We simultaneously consider both epistemic and aleatoric uncertainties, where the former comes from the lack of knowledge about the ground truth (such as about facts or the language), and the latter comes from irreducible randomness (such as multiple possible answers). In particular, we derive an information-theoretic metric that allows to reliably detect when only epistemic uncertainty is large, in which case the output of the model is unreliable. This condition can be computed based solely on the output of the model obtained simply by some special iterative prompting based on the previous responses. Such quantification, for instance, allows to detect hallucinations (cases when epistemic uncertainty is high) in both single- and multi-answer responses. This is in contrast to many standard uncertainty quantification strategies (such as thresholding the log-likelihood of a response) where hallucinations in the multi-answer case cannot be detected. We conduct a series of experiments which demonstrate the advantage of our formulation. Further, our investigations shed some light on how the probabilities assigned to a given output by an LLM can be amplified by iterative prompting, which might be of independent interest.

Exploration in reinforcement learning (RL) remains an open challenge. RL algorithms rely on observing rewards to train the agent, and if informative rewards are sparse the agent learns slowly or may not learn at all. To improve exploration and reward discovery, popular algorithms rely on optimism. But what if sometimes rewards are unobservable, e.g., situations of partial monitoring in bandits and the recent formalism of monitored Markov decision process? In this case, optimism can lead to suboptimal behavior that does not explore further to collapse uncertainty. With this paper, we present a novel exploration strategy that overcomes the limitations of existing methods and guarantees convergence to an optimal policy even when rewards are not always observable. We further propose a collection of tabular environments for benchmarking exploration in RL (with and without unobservable rewards) and show that our method outperforms existing ones.

Powered by remarkable advancements in Large Language Models (LLMs), Multimodal Large Language Models (MLLMs) demonstrate impressive capabilities in manifold tasks. However, the practical application scenarios of MLLMs are intricate, exposing them to potential malicious instructions and thereby posing safety risks. While current benchmarks do incorporate certain safety considerations, they often lack comprehensive coverage and fail to exhibit the necessary rigor and robustness. For instance, the common practice of employing GPT-4V as both the evaluator and a model to be evaluated lacks credibility, as it tends to exhibit a bias toward its own responses. In this paper, we present MLLMGuard, a multi-dimensional safety evaluation suite for MLLMs, including a bilingual image-text evaluation dataset, inference utilities, and a lightweight evaluator. MLLMGuard's assessment comprehensively covers two languages (English and Chinese) and five important safety dimensions (Privacy, Bias, Toxicity, Truthfulness, and Legality), each with corresponding rich subtasks. Focusing on these dimensions, our evaluation dataset is primarily sourced from platforms such as social media, and it integrates text-based and image-based red teaming techniques with meticulous annotation by human experts. This can prevent inaccurate evaluation caused by data leakage when using open-source datasets and ensures the quality and challenging nature of our benchmark. Additionally, a fully automated lightweight evaluator termed GuardRank is developed, which achieves significantly higher evaluation accuracy than GPT-4. Our evaluation results across 13 advanced models indicate that MLLMs still have a substantial journey ahead before they can be considered safe and responsible.

The growing demand for personalized decision-making has led to a surge of interest in estimating the Conditional Average Treatment Effect (CATE). Various types of CATE estimators have been developed with advancements in machine learning and causal inference. However, selecting the desirable CATE estimator through a conventional model validation procedure remains impractical due to the absence of counterfactual outcomes in observational data. Existing approaches for CATE estimator selection, such as plug-in and pseudo-outcome metrics, face two challenges. First, they must determine the metric form and the underlying machine learning models for fitting nuisance parameters (e.g., outcome function, propensity function, and plug-in learner). Second, they lack a specific focus on selecting a robust CATE estimator. To address these challenges, this paper introduces a Distributionally Robust Metric (DRM) for CATE estimator selection. The proposed DRM is nuisance-free, eliminating the need to fit models for nuisance parameters, and it effectively prioritizes the selection of a distributionally robust CATE estimator. The experimental results validate the effectiveness of the DRM method in selecting CATE estimators that are robust to the distribution shift incurred by covariate shift and hidden confounders.

Training a policy in a source domain for deployment in the target domain under a dynamics shift can be challenging, often resulting in performance degradation. Previous work tackles this challenge by training on the source domain with modified rewards derived by matching distributions between the source and the target optimal trajectories. However, pure modified rewards only ensure the behavior of the learned policy in the source domain resembles trajectories produced by the target optimal policies, which does not guarantee optimal performance when the learned policy is actually deployed to the target domain. In this work, we propose to utilize imitation learning to transfer the policy learned from the reward modification to the target domain so that the new policy can generate the same trajectories in the target domain. Our approach, Domain Adaptation and Reward Augmented Imitation Learning (DARAIL), utilizes the reward modification for domain adaptation and follows the general framework of generative adversarial imitation learning from observation (GAIfO) by applying a reward augmented estimator for the policy optimization step. Theoretically, we present an error bound for our method under a mild assumption regarding the dynamics shift to justify the motivation of our method. Empirically, our method outperforms the pure modified reward method without imitation learning and also outperforms other baselines in benchmark off-dynamics environments.

In multi-label learning, each training instance is associated with multiple labels simultaneously. Traditional multi-label learning studies primarily focus on closed set scenario, i.e. the class label set of test data is identical to those used in training phase. Nevertheless, in numerous real-world scenarios, the environment is open and dynamic where unknown labels may emerge gradually during testing. In this paper, the problem of multi-label open set recognition (MLOSR) is investigated, which poses significant challenges in classifying and recognizing instances with unknown labels in multi-label setting. To enable open set multi-label prediction, a novel approach named SLAN is proposed by leveraging sub-labeling information enriched by structural information in the feature space. Accordingly, unknown labels are recognized by differentiating the sub-labeling information from holistic supervision. Experimental results on various datasets validate the effectiveness of the proposed approach in dealing with the MLOSR problem.

Current reinforcement-learning methods are unable to directly learn policies that solve the minimum cost reach-avoid problem to minimize cumulative costs subject to the constraints of reaching the goal and avoiding unsafe states, as the structure of this new optimization problem is incompatible with current methods. Instead, a surrogate problem is solved where all objectives are combined with a weighted sum. However, this surrogate objective results in suboptimal policies that do not directly minimize the cumulative cost. In this work, we propose RC-PPO, a reinforcement-learning-based method for solving the minimum-cost reach-avoid problem by using connections to Hamilton-Jacobi reachability. Empirical results demonstrate that RC-PPO learns policies with comparable goal-reaching rates to while achieving up to 57% lower cumulative costs compared to existing methods on a suite of minimum-cost reach-avoid benchmarks on the Mujoco simulator. The project page can be found at https://oswinso.xyz/rcppo.

Most popular benchmarks for comparing LLMs rely on a limited set of prompt templates, which may not fully capture the LLMs’ abilities and can affect the reproducibility of results on leaderboards. Many recent works empirically verify prompt sensitivity and advocate for changes in LLM evaluation. In this paper, we consider the problem of estimating the performance distribution across many prompt variants instead of finding a single prompt to evaluate with. We introduce PromptEval, a method for estimating performance across a large set of prompts borrowing strength across prompts and examples to produce accurate estimates under practical evaluation budgets. The resulting distribution can be used to obtain performance quantiles to construct various robust performance metrics (e.g., top 95% quantile or median). We prove that PromptEval consistently estimates the performance distribution and demonstrate its efficacy empirically on three prominent LLM benchmarks: MMLU, BIG-bench Hard, and LMentry; for example, PromptEval can accurately estimate performance quantiles across 100 prompt templates on MMLU with a budget equivalent to two single-prompt evaluations. Moreover, we show how PromptEval can be useful in LLM-as-a-judge and best prompt identification applications.

Recovering the foreground color and opacity/alpha matte from a single image (i.e., image matting) is a challenging and ill-posed problem where data priors play a critical role in achieving precise results. Traditional methods generally predict the alpha matte and then extract the foreground through post-processing, often failing to produce high-fidelity foreground color. This failure stems from the models' difficulty in learning robust color predictions from limited matting datasets. To address this, we explore the potential of leveraging vision priors embedded in pre-trained latent diffusion models (LDM) for estimating foreground RGBA values in challenging scenarios and rare objects. We introduce Drip, a novel approach for image matting that harnesses the rich prior knowledge of LDM models. Our method incorporates a switcher and a cross-domain attention mechanism to extend the original LDM for joint prediction of the foreground color and opacity. This setup facilitates mutual information exchange and ensures high consistency across both modalities. To mitigate the inherent reconstruction errors of the LDM's VAE decoder, we propose a latent transparency decoder to align the RGBA prediction with the input image, thereby reducing discrepancies. Comprehensive experimental results demonstrate that our approach achieves state-of-the-art performance in foreground and alpha predictions and shows remarkable generalizability across various benchmarks.

This paper addresses a multi-source light detection (LD) problem from vehicles, traffic signals, and streetlights under driving scenarios. Albeit it is crucial for autonomous driving and night vision, this problem has not been yet focused on as much as other object detection (OD). One of the main reasons is the absence of a public available LD benchmark dataset. Therefore, we construct a new large LD dataset consisting of different light sources via heavy annotation: YouTube Driving Light Detection dataset (YDLD). Compared to the existing LD datasets, our dataset has much more images and box annotations for multi-source lights. We also provide rigorous statistical analysis and transfer learning comparison of other well-known detection benchmark datasets to prove the generality of our YDLD. For the recent object detectors, we achieve the extensive comparison results on YDLD. However, they tend to yield the low mAP scores due to the intrinsic challenges of LD caused by very tiny size and similar appearance. To resolve those, we design a novel lightness focal loss which penalizes miss-classified samples more and a lightness spatial attention prior by reflecting a global scene context. In addition, we develop a semi-supervised focal light detection (SS-FLD) by embedding our lightness focal loss into the semi-supervised object detection (SSOD). We prove that our methods can consistently boost mAP to the variety of types of recent detectors on YDLD. We will open both YDLD and SS-FLD code at https://github.com/YDLD-dataset/YDLD.

Reward models trained on human preference data have been proven to effectively align Large Language Models (LLMs) with human intent within the framework of reinforcement learning from human feedback (RLHF). However, current reward models have limited generalization capabilities to unseen prompts and responses, which can lead to an unexpected phenomenon known as reward over-optimization, resulting in a decline in actual performance due to excessive optimization of rewards. While previous research has advocated for constraining policy optimization, our study introduces a novel approach to enhance the reward model's generalization ability against distribution shifts by regularizing the hidden states. Specifically, we retain the base model's language model head and incorporate a suite of text-generation losses to preserve the hidden states' text-generation capabilities, while concurrently learning a reward head behind the same hidden states. Our experimental results demonstrate that the introduced regularization technique markedly improves the accuracy of learned reward models across a variety of out-of-distribution (OOD) tasks and effectively alleviates the over-optimization issue in RLHF, offering a more reliable and robust preference learning paradigm.

In many real-world decision problems there is partially observed, hidden or latent information that remains fixed throughout an interaction. Such decision problems can be modeled as Latent Markov Decision Processes (LMDPs), where a latent variable is selected at the beginning of an interaction and is not disclosed to the agent initially. In last decade, there has been significant progress in designing learning algorithms for solving LMDPs under different structural assumptions. However, for general LMDPs, there is no known learning algorithm that provably matches the existing lower bound. We effectively resolve this open question, introducing the first sample-efficient algorithm for LMDPs without *any additional structural assumptions*. Our result builds off a new perspective on the role off-policy evaluation guarantees and coverage coefficient in LMDPs, a perspective, which has been overlooked in the context of exploration in partially observed environments. Specifically, we establish a novel off-policy evaluation lemma and introduce a new coverage coefficient for LMDPs. Then, we show how these can be used to derive near-optimal guarantees of an optimistic exploration algorithm. These results, we believe, can be valuable for a wide range of interactive learning problems beyond the LMDP class, and especially, for partially observed environments.

Test-time adaptation, which enables models to generalize to diverse data with unlabeled test samples, holds significant value in real-world scenarios. Recently, researchers have applied this setting to advanced pre-trained vision-language models (VLMs), developing approaches such as test-time prompt tuning to further extend their practical applicability. However, these methods typically focus solely on adapting VLMs from a single modality and fail to accumulate task-specific knowledge as more samples are processed. To address this, we introduce Dual Prototype Evolving (DPE), a novel test-time adaptation approach for VLMs that effectively accumulates task-specific knowledge from multi-modalities. Specifically, we create and evolve two sets of prototypes—textual and visual—to progressively capture more accurate multi-modal representations for target classes during test time. Moreover, to promote consistent multi-modal representations, we introduce and optimize learnable residuals for each test sample to align the prototypes from both modalities. Extensive experimental results on 15 benchmark datasets demonstrate that our proposed DPE consistently outperforms previous state-of-the-art methods while also exhibiting competitive computational efficiency.

Federated learning (FL) allows multiple devices to train a model collaboratively without sharing their data. Despite its benefits, FL is vulnerable to privacy leakage and poisoning attacks. To address the privacy concern, secure aggregation (SecAgg) is often used to obtain the aggregation of gradients on sever without inspecting individual user updates. Unfortunately, existing defense strategies against poisoning attacks rely on the analysis of local updates in plaintext, making them incompatible with SecAgg. To reconcile the conflicts, we propose a robust federated learning framework against poisoning attacks (RFLPA) based on SecAgg protocol. Our framework computes the cosine similarity between local updates and server updates to conduct robust aggregation. Furthermore, we leverage verifiable packed Shamir secret sharing to achieve reduced communication cost of $O(M+N)$ per user, and design a novel dot product aggregation algorithm to resolve the issue of increased information leakage. Our experimental results show that RFLPA significantly reduces communication and computation overhead by over $75\%$ compared to the state-of-the-art secret sharing method, BREA, while maintaining competitive accuracy.

This paper studies the problem of rigid dynamics modeling, which has a wide range of applications in robotics, graphics, and mechanical design. The problem is partly solved by graph neural network (GNN) simulators. However, these approaches cannot effectively handle the relationship between intrinsic continuity and instantaneous changes in rigid dynamics. Moreover, they usually neglect hierarchical structures across mesh nodes and objects in systems. In this paper, we propose a novel approach named Event-attend Graph ODE (EGODE) for effective rigid dynamics modeling. In particular, we describe the rigid system using both mesh node representations and object representations. To model continuous dynamics across hierarchical structures, we use a coupled graph ODE framework for the evolution of both types of representations over a long period. In addition, to capture instantaneous changes during the collision, we introduce an event module, which can effectively estimate the occurrence of the collision and update the states of both mesh node and object representations during evolution. Extensive experiments on a range of benchmark datasets validate the superiority of the proposed EGODE compared to various state-of-the-art baselines. The source code can be found at https://github.com/yuanjypku/EGODE.

Implicit Neural Representations (INRs) have peaked interest in recent years due to their ability to encode natural signals using neural networks. While INRs allow for useful applications such as interpolating new coordinates and signal compression, their black-box nature makes it difficult to modify them post-training. In this paper we explore the idea of editable INRs, and specifically focus on the widely used cropping operation. To this end, we present Local-Global SIRENs - a novel INR architecture that supports cropping by design. Local-Global SIRENs are based on combining local and global feature extraction for signal encoding. What makes their design unique is the ability to effortlessly remove specific portions of an encoded signal, with a proportional weight decrease. This is achieved by eliminating the corresponding weights from the network, without the need for retraining. We further show how this architecture can be used to support the straightforward extension of previously encoded signals. Beyond signal editing, we examine how the Local-Global approach can accelerate training, enhance encoding of various signals, improve downstream performance, and be applied to modern INRs such as INCODE, highlighting its potential and flexibility. Code is available at https://github.com/maorash/Local-Global-INRs.

Physics-Informed Neural Networks (PINNs) are infamous for being hard to train. Recently, second-order methods based on natural gradient and Gauss-Newton methods have shown promising performance, improving the accuracy achieved by first-order methods by several orders of magnitude. While promising, the proposed methods only scale to networks with a few thousand parameters due to the high computational cost to evaluate, store, and invert the curvature matrix. We propose Kronecker-factored approximate curvature (KFAC) for PINN losses that greatly reduces the computational cost and allows scaling to much larger networks. Our approach goes beyond the popular KFAC for traditional deep learning problems as it captures contributions from a PDE's differential operator that are crucial for optimization. To establish KFAC for such losses, we use Taylor-mode automatic differentiation to describe the differential operator's computation graph as a forward network with shared weights which allows us to apply a variant of KFAC for networks with weight-sharing. Empirically, we find that our KFAC-based optimizers are competitive with expensive second-order methods on small problems, scale more favorably to higher-dimensional neural networks and PDEs, and consistently outperform first-order methods.

We introduce a novel approach for detecting distribution shifts that negatively impact the performance of machine learning models in continuous production environments, which requires no access to ground truth data labels. It builds upon the work of Podkopaev and Ramdas [2022], who address scenarios where labels are available for tracking model errors over time. Our solution extends this framework to work in the absence of labels, by employing a proxy for the true error. This proxy is derived using the predictions of a trained error estimator. Experiments show that our method has high power and false alarm control under various distribution shifts, including covariate and label shifts and natural shifts over geography and time.

Hierarchical Bayesian bandit refers to the multi-task bandit problem in which bandit tasks are assumed to be drawn from the same distribution. In this work, we provide improved Bayes regret bounds for hierarchical Bayesian bandit algorithms in the multi-task linear bandit and semi-bandit settings. For the multi-task linear bandit, we first analyze the preexisting hierarchical Thompson sampling (HierTS) algorithm, and improve its gap-independent Bayes regret bound from $O(m\sqrt{n\log{n}\log{(mn)}})$ to $O(m\sqrt{n\log{n}})$ in the case of infinite action set, with $m$ being the number of tasks and $n$ the number of iterations per task. In the case of finite action set, we propose a novel hierarchical Bayesian bandit algorithm, named hierarchical BayesUCB (HierBayesUCB), that achieves the logarithmic but gap-dependent regret bound $O(m\log{(mn)}\log{n})$ under mild assumptions. All of the above regret bounds hold in many variants of hierarchical Bayesian linear bandit problem, including when the tasks are solved sequentially or concurrently. Furthermore, we extend the aforementioned HierTS and HierBayesUCB algorithms to the multi-task combinatorial semi-bandit setting. Concretely, our combinatorial HierTS algorithm attains comparable Bayes regret bound $O(m\sqrt{n}\log{n})$ with respect to the latest one. Moreover, our combinatorial HierBayesUCB yields a sharper Bayes regret bound $O(m\log{(mn)}\log{n})$. Experiments are conducted to validate the soundness of our theoretical results for multi-task bandit algorithms.