ICML 2024 Papers

We present an algorithm for skill discovery from expert demonstrations. The algorithm first utilizes Large Language Models (LLMs) to propose an initial segmentation of the trajectories. Following that, a hierarchical variational inference framework incorporates the LLM-generated segmentation information to discover reusable skills by merging trajectory segments. To further control the trade-off between compression and reusability, we introduce a novel auxiliary objective based on the Minimum Description Length principle that helps guide this skill discovery process. Our results demonstrate that agents equipped with our method are able to discover skills that help accelerate learning and outperform baseline skill learning approaches on new long-horizon tasks in BabyAI, a grid world navigation environment, as well as ALFRED, a household simulation environment.

Learning new knowledge frequently occurs in our dynamically changing world, e.g., humans culturally evolve by continuously acquiring new abilities to sustain their survival, leveraging collective intelligence rather than a large number of individual attempts. The effective learning paradigm during cultural evolution is termed socialized learning (SL). Consequently, a straightforward question arises: Can multi-agent systems acquire more new abilities like humans? In contrast to most existing methods that address continual learning and multi-agent collaboration, our emphasis lies in a more challenging problem: we prioritize the knowledge in the original expert classes, and as we adeptly learn new ones, the accuracy in the original expert classes stays superior among all in a directional manner. Inspired by population genetics and cognitive science, leading to unique and complete development, we propose Multi-Agent Socialized Collaboration (MASC), which achieves SL through interactions among multiple agents. Specifically, we introduce collective collaboration and reciprocal altruism modules, organizing collaborative behaviors, promoting information sharing, and facilitating learning and knowledge interaction among individuals. We demonstrate the effectiveness of multi-agent collaboration in an extensive empirical study. Our code will be publicly available at https://github.com/yxjdarren/SL.

The recent embrace of machine learning (ML) in the development of autonomous weapons systems (AWS) creates serious risks to geopolitical stability and the free exchange of ideas in AI research. This topic has received comparatively little attention of late compared to risks stemming from superintelligent artificial general intelligence (AGI), but requires fewer assumptions about the course of technological development and is thus a nearer-future issue. ML is already enabling the substitution of AWS for human soldiers in many battlefield roles, reducing the upfront human cost, and thus political cost, of waging offensive war. In the case of peer adversaries, this increases the likelihood of "low intensity" conflicts which risk escalation to broader warfare. In the case of non-peer adversaries, it reduces the domestic blowback to wars of aggression. This effect can occur regardless of other ethical issues around the use of military AI such as the risk of civilian casualties, and does not require any superhuman AI capabilities. Further, the military value of AWS raises the specter of an AI-powered arms race and the misguided imposition of national security restrictions on AI research. Our goal in this paper is to raise awareness among the public and ML researchers on the near-future risks posed by full or near-full autonomy in military technology, and we provide regulatory suggestions to mitigate these risks. We call upon AI policy experts and the defense AI community in particular to embrace transparency and caution in their development and deployment of AWS to avoid the negative effects on global stability and AI research that we highlight here.

The problem of two-player zero-sum Markov games has recently attracted increasing interests in theoretical studies of multi-agent reinforcement learning (RL). In particular, for finite-horizon episodic Markov decision processes (MDPs), it has been shown that model-based algorithms can find an $\epsilon$-optimal Nash Equilibrium (NE) with the sample complexity of $O(H^3SAB/\epsilon^2)$, which is optimal in the dependence of the horizon $H$ and the number of states $S$ (where $A$ and $B$ denote the number of actions of the two players, respectively). However, none of the existing model-free algorithms can achieve such an optimality. In this work, we propose a model-free stage-based algorithm and show that it achieves the same sample complexity as the best model-based algorithm, and hence for the first time demonstrate that model-free algorithms can enjoy the same optimality in the $H$ dependence as model-based algorithms. The main improvement of the dependency on $H$ arises by leveraging the popular variance reduction technique based on the reference-advantage decomposition previously used only for single-agent RL. However, such a technique relies on a critical monotonicity property of the value function, which does not hold in Markov games due to the update of the policy via the coarse correlated equilibrium (CCE) oracle. Thus, to extend such a technique to Markov games, our algorithm features a key novel design of updating the reference value functions as the pair of optimistic and pessimistic value functions whose value difference is the smallest in the history in order to achieve the desired improvement in the sample efficiency.

Following natural language instructions by executing actions in digital environments (e.g. web-browsers and REST APIs) is a challenging task for language model (LM) agents. Unfortunately, LM agents often fail to generalize to new environments without human demonstrations. This work presents BAGEL, a method for bootstrapping LM agents without human supervision. BAGEL converts a seed set of randomly explored trajectories to synthetic demonstrations via round-trips between two noisy LM components: an LM labeler which converts a trajectory into a synthetic instruction, and a zero-shot LM agent which maps the synthetic instruction into a refined trajectory. By performing these round-trips iteratively, BAGEL quickly converts the initial distribution of trajectories towards those that are well-described by natural language. We adapt the base LM agent at test time with in-context learning by retrieving relevant BAGEL demonstrations based on the instruction, and find improvements of over 2-13% absolute on ToolQA and MiniWob++, with up to 13x reduction in execution failures.

Robust Reinforcement Learning (RRL) is a promising Reinforcement Learning (RL) paradigm aimed at training robust to uncertainty or disturbances models, making them more efficient for real-world applications. Following this paradigm, uncertainty or disturbances are interpreted as actions of a second adversarial agent, and thus, the problem is reduced to seeking the agents' policies robust to any opponent's actions. This paper is the first to propose considering the RRL problems within the positional differential game theory, which helps us to obtain theoretically justified intuition to develop a centralized Q-learning approach. Namely, we prove that under Isaacs's condition (sufficiently general for real-world dynamical systems), the same Q-function can be utilized as an approximate solution of both minimax and maximin Bellman equations. Based on these results, we present the Isaacs Deep Q-Network algorithms and demonstrate their superiority compared to other baseline RRL and Multi-Agent RL algorithms in various environments.

Decision-making problems, categorized as single-agent, e.g., Atari, cooperative multi-agent, e.g., Hanabi, competitive multi-agent, e.g., Hold'em poker, and mixed cooperative and competitive, e.g., football, are ubiquitous in the real world. Although various methods have been proposed to address the specific decision-making categories, these methods typically evolve independently and cannot generalize to other categories. Therefore, a fundamental question for decision-making is: *Can we develop **a single algorithm** to tackle **ALL** categories of decision-making problems?* There are several main challenges to address this question: i) different decision-making categories involve different numbers of agents and different relationships between agents, ii) different categories have different solution concepts and evaluation measures, and iii) there lacks a comprehensive benchmark covering all the categories. This work presents a preliminary attempt to address the question with three main contributions. i) We propose the generalized mirror descent (GMD), a generalization of MD variants, which considers multiple historical policies and works with a broader class of Bregman divergences. ii) We propose the configurable mirror descent (CMD) where a meta-controller is introduced to dynamically adjust the hyper-parameters in GMD conditional on the evaluation measures. iii) We construct the GameBench with 15 academic-friendly games across different decision-making categories. Extensive experiments demonstrate that CMD achieves empirically competitive or better outcomes compared to baselines while providing the capability of exploring diverse dimensions of decision making.

An emergency responder management (ERM) system dispatches responders, such as ambulances, when it receives requests for medical aid. ERM systems can also proactively reposition responders between predesignated waiting locations to cover any gaps that arise due to the prior dispatch of responders or significant changes in the distribution of anticipated requests. Optimal repositioning is computationally challenging due to the exponential number of ways to allocate responders between locations and the uncertainty in future requests. The state-of-the-art approach in proactive repositioning is a hierarchical approach based on spatial decomposition and online Monte Carlo tree search, which may require minutes of computation for each decision in a domain where seconds can save lives. We address the issue of long decision times by introducing a novel reinforcement learning (RL) approach, based on the same hierarchical decomposition, but replacing online search with learning. To address the computational challenges posed by large, variable-dimensional, and discrete state and action spaces, we propose: (1) actor-critic based agents that incorporate transformers to handle variable-dimensional states and actions, (2) projections to fixed-dimensional observations to handle complex states, and (3) combinatorial techniques to map continuous actions to discrete allocations. We evaluate our approach using real-world data from two U.S. cities, Nashville, TN and Seattle, WA. Our experiments show that compared to the state of the art, our approach reduces computation time per decision by three orders of magnitude, while also slightly reducing average ambulance response time by 5 seconds.

Neural Language Models (LMs) offer an exciting solution for general-purpose embodied control. However, a key technical issue arises when using an LM-based controller: environment observations must be converted to text, which coupled with history, results in long and verbose textual prompts. As a result, prior work in LM agents is limited to restricted domains with small observation size as well as minimal needs for interaction history or instruction finetuning. In this paper, we introduce diff history, a simple and highly effective solution to these issues. By applying the Unix diff command on consecutive text observations in the interaction histories used to prompt LM policies, we can both abstract away redundant information and focus the content of textual inputs on the salient changes in the environment. On NetHack, an unsolved video game that requires long-horizon reasoning for decision-making, LMs tuned with diff history match state-of-the-art performance for neural agents while needing 1800X fewer training examples compared to prior work. Even on the simpler BabyAI-Text environment with concise text observations, we find that although diff history increases the length of prompts, the representation it provides offers a 25% improvement in the efficiency of low-sample instruction finetuning. Further, we show that diff history scales favorably across different finetuning dataset sizes. We open-source our code and data to https://diffhistory.github.io.

In this work, we focus on safe policy improvement in multi-agent domains where current state-of-the-art methods cannot be effectively applied because of large state and action spaces. We consider recent results using Monte Carlo Tree Search for Safe Policy Improvement with Baseline Bootstrapping and propose a novel algorithm that scales this approach to multi-agent domains, exploiting the factorization of the transition model and value function. Given a centralized behavior policy and a dataset of trajectories, our algorithm generates an improved policy by selecting joint actions using a novel extension of Max-Plus (or Variable Elimination) that constrains local actions to guarantee safety criteria. An empirical evaluation on multi-agent SysAdmin and multi-UAV Delivery shows that the approach scales to very large domains where state-of-the-art methods cannot work.

Large Language Models (LLMs) employ auto-regressive decoding that requires sequential computation, with each step reliant on the previous one's output. This creates a bottleneck as each step necessitates moving the full model parameters from High-Bandwidth Memory (HBM) to the accelerator's cache. While methods such as speculative decoding have been suggested to address this issue, their implementation is impeded by the challenges associated with acquiring and maintaining a separate draft model. In this paper, we present Medusa, an efficient method that augments LLM inference by adding extra decoding heads to predict multiple subsequent tokens in parallel. Using a tree-based attention mechanism, Medusa constructs multiple candidate continuations and verifies them simultaneously in each decoding step. By leveraging parallel processing, Medusa reduces the number of decoding steps required. We present two levels of fine-tuning procedures for Medusa to meet the needs of different use cases: Medusa-1: Medusa is directly fine-tuned on top of a frozen backbone LLM, enabling lossless inference acceleration. Medusa-2: Medusa is fine-tuned together with the backbone LLM, enabling better prediction accuracy of Medusa heads and higher speedup but needing a special training recipe that preserves the model's capabilities. Moreover, we propose several extensions that improve or expand the utility of Medusa, including a self-distillation to handle situations where no training data is available and a typical acceptance scheme to boost the acceptance rate while maintaining generation quality. We evaluate Medusa on models of various sizes and training procedures. Our experiments demonstrate that Medusa-1 can achieve over 2.2$\times$ speedup without compromising generation quality, while Medusa-2 further improves the speedup to 2.3-2.8$\times$.

Current Large Language Models (LLMs) are not only limited to some maximum context length, but also are not able to robustly consume long inputs. To address these limitations, we propose ReadAgent, an LLM agent system that increases effective context length up to 20x in our experiments. Inspired by how humans interactively read long documents, we implement ReadAgent as a simple prompting system that uses the advanced language capabilities of LLMs to (1) decide what content to store together in a memory episode, (2) compress those memory episodes into short episodic memories called *gist memories*, and (3) take actions to look up passages in the original text if ReadAgent needs to remind itself of relevant details to complete a task. We evaluate ReadAgent against baselines using retrieval methods, using the original long contexts, and using the gist memories. These evaluations are performed on three long-document reading comprehension tasks: QuALITY, NarrativeQA, and QMSum. ReadAgent outperforms the baselines on all three tasks while extending the effective context window by 3.5-20x.

Incomplete multi-view clustering (IMVC) methods typically encounter three drawbacks: (1) intense time and/or space overheads; (2) intractable hyper-parameters; (3) non-zero variance results. With these concerns in mind, we give a simple yet effective IMVC scheme, termed as ToRES. Concretely, instead of self-expression affinity, we manage to construct prototype-sample affinity for incomplete data so as to decrease the memory requirements. To eliminate hyper-parameters, besides mining complementary features among views by view-wise prototypes, we also attempt to devise cross-view prototypes to capture consensus features for jointly forming high-quality clustering representation. To avoid the variance, we successfully unify representation learning and clustering operation, and directly optimize the discrete cluster indicators from incomplete data. Then, for the resulting objective function, we provide two equivalent solutions from perspectives of feasible region partitioning and objective transformation. Many results suggest that ToRES exhibits advantages against 20 SOTA algorithms, even in scenarios with a higher ratio of incomplete data.

Vector quantization is a fundamental operation for data compression and vector search. To obtain high accuracy, multi-codebook methods represent each vector using codewords across several codebooks. Residual quantization (RQ) is one such method, which iteratively quantizes the error of the previous step. While the error distribution is dependent on previously-selected codewords, this dependency is not accounted for in conventional RQ as it uses a fixed codebook per quantization step. In this paper, we propose QINCo, a neural RQ variant that constructs specialized codebooks per step that depend on the approximation of the vector from previous steps. Experiments show that QINCo outperforms state-of-the-art methods by a large margin on several datasets and code sizes. For example, QINCo achieves better nearest-neighbor search accuracy using 12-byte codes than the state-of-the-art UNQ using 16 bytes on the BigANN1M and Deep1M datasets.

We examine two types of binary betting markets, whose primary goal is for profit (such as sports gambling) or to gain information (such as prediction markets). We articulate the interplay between belief and price-setting to analyse both types of markets, and show that the goals of maximising bookmaker profit and eliciting information are fundamentally incompatible. A key insight is that profit hinges on the deviation between (the distribution of) bettor and true beliefs, and that heavier tails in bettor belief distribution implies higher profit. Our algorithmic contribution is to introduce online learning methods for price-setting. Traditionally bookmakers update their prices rather infrequently, we present two algorithms that guide price updates upon seeing each bet, assuming very little of bettor belief distributions. The online pricing algorithm achieves stochastic regret of $\mathcal{O}(\sqrt{T})$ against the worst local maximum, or $\mathcal{O}(\sqrt{T \log T})$ with high probability against the global maximum under fair odds. More broadly, the inherent tradeoff between profit and information-seeking in binary betting may inspire new understandings of large-scale multi-agent behaviour.

*Relative overgeneralization* (RO) occurs in cooperative multi-agent learning tasks when agents converge towards a suboptimal joint policy due to overfitting to suboptimal behaviors of other agents. No methods have been proposed for addressing RO in multi-agent policy gradient (MAPG) methods although these methods produce state-of-the-art results. To address this gap, we propose a general, yet simple, framework to enable optimistic updates in MAPG methods that alleviate the RO problem. Our approach involves clipping the advantage to eliminate negative values, thereby facilitating optimistic updates in MAPG. The optimism prevents individual agents from quickly converging to a local optimum. Additionally, we provide a formal analysis to show that the proposed method retains optimality at a fixed point. In extensive evaluations on a diverse set of tasks including the *Multi-agent MuJoCo* and *Overcooked* benchmarks, our method outperforms strong baselines on 13 out of 19 tested tasks and matches the performance on the rest.

Large language models (LLMs) demonstrate impressive reasoning abilities, but translating reasoning into actions in the real world remains challenging. In particular, it is unclear how to complete a given task provably within a minimum number of interactions with the external environment, e.g., through an internal mechanism of reasoning. To this end, we propose the first framework with provable regret guarantees to orchestrate reasoning and acting, which we call *reason for future, act for now* (**RAFA**). Specifically, we design a prompt template for reasoning that learns from the memory buffer and plans a future trajectory over a long horizon (*reason for future*). At each step, the LLM agent takes the initial action of the planned trajectory (*act for now*), stores the collected feedback in the memory buffer, and reinvokes the reasoning routine to replan the future trajectory from the new state. The key idea is to cast reasoning in LLMs as learning and planning in Bayesian adaptive Markov decision processes (MDPs). Correspondingly, we prompt LLMs with the memory buffer to estimate the unknown environment (learning) and generate an optimal trajectory for multiple future steps that maximize a value function (planning). The learning and planning subroutines are performed in an in-context manner to emulate the actor-critic update for MDPs. Our theoretical analysis establishes a $\sqrt{T}$ regret, while our experimental validation demonstrates superior empirical performance.

We address the challenge of utilizing large language models (LLMs) for complex embodied tasks, in the environment where decision-making systems operate timely on capacity-limited, off-the-shelf devices. We present DeDer, a framework for decomposing and distilling the embodied reasoning capabilities from LLMs to efficient, small language model (sLM)-based policies. In DeDer, the decision-making process of LLM-based strategies is restructured into a hierarchy with a reasoning-policy and planning-policy. The reasoning-policy is distilled from the data that is generated through the embodied in-context learning and self-verification of an LLM, so it can produce effective rationales. The planning-policy, guided by the rationales, can render optimized plans efficiently. In turn, DeDer allows for adopting sLMs for both policies, deployed on off-the-shelf devices. Furthermore, to enhance the quality of intermediate rationales, specific to embodied tasks, we devise the embodied knowledge graph, and to generate multiple rationales timely through a single inference, we also use the contrastively prompted attention model. Our experiments with the ALFRED benchmark demonstrate that DeDer surpasses leading language planning and distillation approaches, indicating the applicability and efficiency of sLM-based embodied policies derived through DeDer.

Asynchronous action coordination presents a pervasive challenge in Multi-Agent Systems (MAS), which can be represented as a Stackelberg game (SG). However, the scalability of existing Multi-Agent Reinforcement Learning (MARL) methods based on SG is severely restricted by network architectures or environmental settings. To address this issue, we propose the Stackelberg Decision Transformer (STEER). It efficiently manages decision-making processes by incorporating the hierarchical decision structure of SG, the modeling capability of autoregressive sequence models, and the exploratory learning methodology of MARL. Our approach exhibits broad applicability across diverse task types and environmental configurations in MAS. Experimental results demonstrate both the convergence of our method towards Stackelberg equilibrium strategies and its superiority over strong baselines in complex scenarios.