NeurIPS 2024 Papers

In linear bandits, how can a learner effectively learn when facing corrupted rewards? While significant work has explored this question, a holistic understanding across different adversarial models and corruption measures is lacking, as is a full characterization of the minimax regret bounds. In this work, we compare two types of corruptions commonly considered: strong corruption, where the corruption level depends on the learner’s chosen action, and weak corruption, where the corruption level does not depend on the learner’s chosen action. We provide a unified framework to analyze these corruptions. For stochastic linear bandits, we fully characterize the gap between the minimax regret under strong and weak corruptions. We also initiate the study of corrupted adversarial linear bandits, obtaining upper and lower bounds with matching dependencies on the corruption level. Next, we reveal a connection between corruption-robust learning and learning with gap-dependent misspecification—a setting first studied by Liu et al. (2023a), where the misspecification level of an action or policy is proportional to its suboptimality. We present a general reduction that enables any corruption-robust algorithm to handle gap-dependent misspecification. This allows us to recover the results of Liu et al. (2023a) in a black-box manner and significantly generalize them to settings like linear MDPs, yielding the first results for gap-dependent misspecification in reinforcement learning. However, this general reduction does not attain the optimal rate for gap-dependent misspecification. Motivated by this, we develop a specialized algorithm that achieves optimal bounds for gap-dependent misspecification in linear bandits, thus answering an open question posed by Liu et al. (2023a).

The alignment of large language models (LLMs) is critical for developing effective and safe language models. Traditional approaches focus on aligning models during the instruction tuning or reinforcement learning stages, referred to in this paper as `\textit{post alignment}'. We argue that alignment during the pre-training phase, which we term 'native alignment', warrants investigation. Native alignment aims to prevent unaligned content from the beginning, rather than relying on post-hoc processing. This approach leverages extensively aligned pre-training data to enhance the effectiveness and usability of pre-trained models. Our study specifically explores the application of native alignment in the context of Arabic LLMs. We conduct comprehensive experiments and ablation studies to evaluate the impact of native alignment on model performance and alignment stability. Additionally, we release open-source Arabic LLMs that demonstrate state-of-the-art performance on various benchmarks, providing significant benefits to the Arabic LLM community.

Current multimodal Large Language Models (MLLMs) suffer from ''hallucination'', occasionally generating responses that are not grounded in the input images. To tackle this challenge, one promising path is to utilize reinforcement learning from human feedback (RLHF), which steers MLLMs towards learning superior responses while avoiding inferior ones. We rethink the common practice of using binary preferences (*i.e.*, superior, inferior), and find that adopting multi-level preferences (*e.g.*, superior, medium, inferior) is better for two benefits: 1) It narrows the gap between adjacent levels, thereby encouraging MLLMs to discern subtle differences. 2) It further integrates cross-level comparisons (beyond adjacent-level comparisons), thus providing a broader range of comparisons with hallucination examples. To verify our viewpoint, we present the Automated Multi-level Preference (**AMP**) framework for MLLMs. To facilitate this framework, we first develop an automated dataset generation pipeline that provides high-quality multi-level preference datasets without any human annotators. Furthermore, we design the Multi-level Direct Preference Optimization (MDPO) algorithm to robustly conduct complex multi-level preference learning. Additionally, we propose a new hallucination benchmark, MRHal-Bench. Extensive experiments across public hallucination and general benchmarks, as well as our MRHal-Bench, demonstrate the effectiveness of our proposed method. Code is available at https://github.com/takomc/amp.

We study reinforcement learning (RL) problems in which agents observe the reward or transition realizations at their current state _before deciding which action to take_. Such observations are available in many applications, including transactions, navigation and more. When the environment is known, previous work shows that this lookahead information can drastically increase the collected reward. However, outside of specific applications, existing approaches for interacting with unknown environments are not well-adapted to these observations. In this work, we close this gap and design provably-efficient learning algorithms able to incorporate lookahead information. To achieve this, we perform planning using the empirical distribution of the reward and transition observations, in contrast to vanilla approaches that only rely on estimated expectations. We prove that our algorithms achieve tight regret versus a baseline that also has access to lookahead information -- linearly increasing the amount of collected reward compared to agents that cannot handle lookahead information.

Modern vision models are trained on very large noisy datasets. While these models acquire strong capabilities, they may not follow the user's intent to output the desired results in certain aspects, e.g., visual aesthetic, preferred style, and responsibility. In this paper, we target the realm of visual aesthetics and aim to align vision models with human aesthetic standards in a retrieval system. Advanced retrieval systems usually adopt a cascade of aesthetic models as re-rankers or filters, which are limited to low-level features like saturation and perform poorly when stylistic, cultural or knowledge contexts are involved. We find that utilizing the reasoning ability of large language models (LLMs) to rephrase the search query and extend the aesthetic expectations can make up for this shortcoming. Based on the above findings, we propose a preference-based reinforcement learning method that fine-tunes the vision models to distill the knowledge from both LLMs reasoning and the aesthetic models to better align the vision models with human aesthetics. Meanwhile, with rare benchmarks designed for evaluating retrieval systems, we leverage large multi-modality model (LMM) to evaluate the aesthetic performance with their strong abilities. As aesthetic assessment is one of the most subjective tasks, to validate the robustness of LMM, we further propose a novel dataset named HPIR to benchmark the alignment with human aesthetics. Experiments demonstrate that our method significantly enhances the aesthetic behaviors of the vision models, under several metrics. We believe the proposed algorithm can be a general practice for aligning vision models with human values.

Recent works have shown the remarkable superiority of transformer models in reinforcement learning (RL), where the decision-making problem is formulated as sequential generation. Transformer-based agents could emerge with self-improvement in online environments by providing task contexts, such as multiple trajectories, called in-context RL. However, due to the quadratic computation complexity of attention in transformers, current in-context RL methods suffer from huge computational costs as the task horizon increases. In contrast, the Mamba model is renowned for its efficient ability to process long-term dependencies, which provides an opportunity for in-context RL to solve tasks that require long-term memory. To this end, we first implement Decision Mamba (DM) by replacing the backbone of Decision Transformer (DT). Then, we propose a Decision Mamba-Hybrid (DM-H) with the merits of transformers and Mamba in high-quality prediction and long-term memory. Specifically, DM-H first generates high-value sub-goals from long-term memory through the Mamba model. Then, we use sub-goals to prompt the transformer, establishing high-quality predictions. Experimental results demonstrate that DM-H achieves state-of-the-art in long and short-term tasks, such as D4RL, Grid World, and Tmaze benchmarks. Regarding efficiency, the online testing of DM-H in the long-term task is 28$\times$ times faster than the transformer-based baselines.

Generating simulations to train intelligent agents in game-playing and robotics from natural language input, user input, or task documentation remains an open-ended challenge. Existing approaches focus on parts of this challenge, such as generating reward functions or task hyperparameters. Unlike previous work, we introduce FACTORSIM that generates full simulations in code from language input that can be used to train agents. Exploiting the structural modularity specific to coded simulations, we propose to use a factored partially observable Markov decision process representation that allows us to reduce context dependence during each step of the generation. For evaluation, we introduce a generative simulation benchmark that assesses the generated simulation code’s accuracy and effectiveness in facilitating zero-shot transfers in reinforcement learning settings. We show that FACTORSIM outperforms existing methods in generating simulations regarding prompt alignment (i.e., accuracy), zero-shot transfer abilities, and human evaluation. We also demonstrate its effectiveness in generating robotic tasks.

We introduce the Continuous Arcade Learning Environment (CALE), an extension of the well-known Arcade Learning Environment (ALE) [Bellemare et al., 2013]. The CALE uses the same underlying emulator of the Atari 2600 gaming system (Stella), but adds support for continuous actions. This enables the benchmarking and evaluation of continuous-control agents (such as PPO [Schulman et al., 2017] and SAC [Haarnoja et al., 2018]) and value-based agents (such as DQN [Mnih et al., 2015] and Rainbow [Hessel et al., 2018]) on the same environment suite. We provide a series of open questions and research directions that CALE enables, as well as initial baseline results using Soft Actor-Critic. CALE is available as part of the ALE athttps://github.com/Farama-Foundation/Arcade-Learning-Environment.

Deep reinforcement learning (RL) algorithms typically parameterize the policy as a deep network that outputs either a deterministic action or a stochastic one modeled as a Gaussian distribution, hence restricting learning to a single behavioral mode. Meanwhile, diffusion models emerged as a powerful framework for multimodal learning. However, the use of diffusion policies in online RL is hindered by the intractability of policy likelihood approximation, as well as the greedy objective of RL methods that can easily skew the policy to a single mode. This paper presents Deep Diffusion Policy Gradient (DDiffPG), a novel actor-critic algorithm that learns from scratch multimodal policies parameterized as diffusion models while discovering and maintaining versatile behaviors. DDiffPG explores and discovers multiple modes through off-the-shelf unsupervised clustering combined with novelty-based intrinsic motivation. DDiffPG forms a multimodal training batch and utilizes mode-specific Q-learning to mitigate the inherent greediness of the RL objective, ensuring the improvement of the diffusion policy across all modes. Our approach further allows the policy to be conditioned on mode-specific embeddings to explicitly control the learned modes. Empirical studies validate DDiffPG's capability to master multimodal behaviors in complex, high-dimensional continuous control tasks with sparse rewards, also showcasing proof-of-concept dynamic online replanning when navigating mazes with unseen obstacles. Our project page is available at https://supersglzc.github.io/projects/ddiffpg/.

Effective communication requires the ability to refer to specific parts of an observation in relation to others. While emergent communication literature shows success in developing various language properties, no research has shown the emergence of such positional references. This paper demonstrates how agents can communicate about spatial relationships within their observations. The results indicate that agents can develop a language capable of expressing the relationships between parts of their observation, achieving over 90% accuracy when trained in a referential game which requires such communication. Using a collocation measure, we demonstrate how the agents create such references. This analysis suggests that agents use a mixture of non-compositional and compositional messages to convey spatial relationships. We also show that the emergent language is interpretable by humans. The translation accuracy is tested by communicating with the receiver agent, where the receiver achieves over 78% accuracy using parts of this lexicon, confirming that the interpretation of the emergent language was successful.

Large Language Models (LLMs) are becoming a prominent generative AI tool, where the user enters a query and the LLM generates an answer. To reduce harm and misuse, efforts have been made to align these LLMs to human values using advanced training techniques such as Reinforcement Learning from Human Feedback (RLHF). However, recent studies have highlighted the vulnerability of LLMs to adversarial jailbreak attempts aiming at subverting the embedded safety guardrails. To address this challenge, this paper defines and investigates the **Refusal Loss** of LLMs and then proposes a method called **Gradient Cuff** to detect jailbreak attempts. Gradient Cuff exploits the unique properties observed in the refusal loss landscape, including functional values and its smoothness, to design an effective two-step detection strategy. Experimental results on two aligned LLMs (LLaMA-2-7B-Chat and Vicuna-7B-V1.5) and six types of jailbreak attacks (GCG, AutoDAN, PAIR, TAP, Base64, and LRL) show that Gradient Cuff can significantly improve the LLM's rejection capability for malicious jailbreak queries, while maintaining the model's performance for benign user queries by adjusting the detection threshold.

In many reinforcement learning (RL) applications, incorporating heuristic rewards alongside the task reward is crucial for achieving desirable performance. Heuristics encode prior human knowledge about how a task should be done, providing valuable hints for RL algorithms. However, such hints may not be optimal, limiting the performance of learned policies. The currently established way of using heuristics is to modify the heuristic reward in a manner that ensures that the optimal policy learned with it remains the same as the optimal policy for the task reward (i.e., optimal policy invariance). However, these methods often fail in practical scenarios with limited training data. We found that while optimal policy invariance ensures convergence to the best policy based on task rewards, it doesn't guarantee better performance than policies trained with biased heuristics under a finite data regime, which is impractical. In this paper, we introduce a new principle tailored for finite data settings. Instead of enforcing optimal policy invariance, we train a policy that combines task and heuristic rewards and ensures it outperforms the heuristic-trained policy. As such, we prevent policies from merely exploiting heuristic rewards without improving the task reward. Our experiments on robotic locomotion, helicopter control, and manipulation tasks demonstrate that our method consistently outperforms the heuristic policy, regardless of the heuristic rewards' quality. Code is available at https://github.com/Improbable-AI/hepo.

This work presents a novel approach to neural architecture search (NAS) that aims to increase carbon efficiency for the model design process. The proposed framework CE-NAS addresses the key challenge of high carbon cost associated with NAS by exploring the carbon emission variations of energy and energy differences of different NAS algorithms. At the high level, CE-NAS leverages a reinforcement-learning agent to dynamically adjust GPU resources based on carbon intensity, predicted by a time-series transformer, to balance energy-efficient sampling and energy-intensive evaluation tasks. Furthermore, CE-NAS leverages a recently proposed multi-objective optimizer to effectively reduce the NAS search space. We demonstrate the efficacy of CE-NAS in lowering carbon emissions while achieving SOTA results for both NAS datasets and open-domain NAS tasks. For example, on the HW-NasBench dataset, CE-NAS reduces carbon emissions by up to 7.22X while maintaining a search efficiency comparable to vanilla NAS. For open-domain NAS tasks, CE-NAS achieves SOTA results with 97.35% top-1 accuracy on CIFAR-10 with only 1.68M parameters and a carbon consumption of 38.53 lbs of CO2. On ImageNet, our searched model achieves 80.6% top-1 accuracy with a 0.78 ms TensorRT latency using FP16 on NVIDIA V100, consuming only 909.86 lbs of CO2, making it comparable to other one-shot-based NAS baselines. Our code is available at https://github.com/cake-lab/CE-NAS.

Training agents that can coordinate zero-shot with humans is a key mission in multi-agent reinforcement learning (MARL). Current algorithms focus on training simulated human partner policies which are then used to train a Cooperator agent. The simulated human is produced either through behavior cloning over a dataset of human cooperation behavior, or by using MARL to create a population of simulated agents. However, these approaches often struggle to produce a Cooperator that can coordinate well with real humans, since the simulated humans fail to cover the diverse strategies and styles employed by people in the real world. We show \emph{learning a generative model of human partners} can effectively address this issue. Our model learns a latent variable representation of the human that can be regarded as encoding the human's unique strategy, intention, experience, or style. This generative model can be flexibly trained from any (human or neural policy) agent interaction data. By sampling from the latent space, we can use the generative model to produce different partners to train Cooperator agents. We evaluate our method---Generative Agent Modeling for Multi-agent Adaptation (GAMMA)---on Overcooked, a challenging cooperative cooking game that has become a standard benchmark for zero-shot coordination. We conduct an evaluation with real human teammates, and the results show that GAMMA consistently improves performance, whether the generative model is trained on simulated populations or human datasets. Further, we propose a method for posterior sampling from the generative model that is biased towards the human data, enabling us to efficiently improve performance with only a small amount of expensive human interaction data.

Vision-language navigation (VLN) requires an agent to execute actions following human instructions. Existing VLN models are optimized through expert demonstrations by supervised behavioural cloning or incorporating manual reward engineering. While straightforward, these efforts overlook the accumulation of errors in the Markov decision process, and struggle to match the distribution of the expert policy. Going beyond this, we propose an Energy-based Navigation Policy (ENP) to model the joint state-action distribution using an energy-based model. At each step, low energy values correspond to the state-action pairs that the expert is most likely to perform, and vice versa. Theoretically, the optimization objective is equivalent to minimizing the forward divergence between the occupancy measure of the expert and ours. Consequently, ENP learns to globally align with the expert policy by maximizing the likelihood of the actions and modeling the dynamics of the navigation states in a collaborative manner. With a variety of VLN architectures, ENP achieves promising performances on R2R, REVERIE, RxR, and R2R-CE, unleashing the power of existing VLN models.

Datasets often suffer severe selection bias; clinical labels are only available on patients for whom doctors ordered medical exams. To assess model performance outside the support of available data, we present a computational framework for adaptive labeling, providing cost-efficient model evaluations under severe distribution shifts. We formulate the problem as a Markov Decision Process over states defined by posterior beliefs on model performance. Each batch of new labels incurs a “state transition” to sharper beliefs, and we choose batches to minimize uncertainty on model performance at the end of the label collection process. Instead of relying on high-variance REINFORCE policy gradient estimators that do not scale, our adaptive labeling policy is optimized using path-wise policy gradients computed by auto-differentiating through simulated roll-outs. Our framework is agnostic to different uncertainty quantification approaches and highlights the virtue of planning in adaptive labeling. On synthetic and real datasets, we empirically demonstrate even a one-step lookahead policy substantially outperforms active learning-inspired heuristics.

The theory of boosting provides a computational framework for aggregating approximate weak learning algorithms, which perform marginally better than a random predictor, into an accurate strong learner. In the realizable case, the success of the boosting approach is underscored by a remarkable fact that the resultant sample complexity matches that of a computationally demanding alternative, namely Empirical Risk Minimization (ERM). This in particular implies that the realizable boosting methodology has the potential to offer computational relief without compromising on sample efficiency. Despite recent progress, in agnostic boosting, where assumptions on the conditional distribution of labels given feature descriptions are absent, ERM outstrips the agnostic boosting methodology in being quadratically more sample efficient than all known agnostic boosting algorithms. In this paper, we make progress on closing this gap, and give a substantially more sample efficient agnostic boosting algorithm than those known, without compromising on the computational (or oracle) complexity. A key feature of our algorithm is that it leverages the ability to reuse samples across multiple rounds of boosting, while guaranteeing a generalization error strictly better than those obtained by blackbox applications of uniform convergence arguments. We also apply our approach to other previously studied learning problems, including boosting for reinforcement learning, and demonstrate improved results.

Training offline RL models using visual inputs poses two significant challenges, *i.e.*, the overfitting problem in representation learning and the overestimation bias for expected future rewards. Recent work has attempted to alleviate the overestimation bias by encouraging conservative behaviors. This paper, in contrast, tries to build more flexible constraints for value estimation without impeding the exploration of potential advantages. The key idea is to leverage off-the-shelf RL simulators, which can be easily interacted with in an online manner, as the “*test bed*” for offline policies. To enable effective online-to-offline knowledge transfer, we introduce CoWorld, a model-based RL approach that mitigates cross-domain discrepancies in state and reward spaces. Experimental results demonstrate the effectiveness of CoWorld, outperforming existing RL approaches by large margins.

How did humanity coax mathematics from the aether? We explore the Platonic view that mathematics can be discovered from its axioms---a game of conjecture and proof. We describe an agent that jointly learns to pose challenging problems for itself (conjecturing) and solve them (theorem proving). Given a mathematical domain axiomatized in dependent type theory, we first combine methods for constrained decoding and type-directed synthesis to sample valid conjectures from a language model. Our method guarantees well-formed conjectures by construction, even as we start with a randomly initialized model. We use the same model to represent a policy and value function for guiding proof search. Our agent targets generating hard but provable conjectures --- a moving target, since its own theorem proving ability also improves as it trains. We propose novel methods for hindsight relabeling on proof search trees to significantly improve the agent's sample efficiency in both tasks. Experiments on 3 axiomatic domains (propositional logic, arithmetic and group theory) demonstrate that our agent can bootstrap from only the axioms, self-improving in generating true and challenging conjectures and in finding proofs.