In-context reinforcement learning (ICRL) is a frontier paradigm for solving reinforcement learning problems in the foundation model era. While ICRL capabilities have been demonstrated in transformers through task-specific training, the potential of Large Language Models (LLMs) out-of-the-box remains largely unexplored. Recent findings highlight that LLMs often face challenges when dealing with numerical contexts, and limited attention has been paid to evaluating their performance through preference feedback generated by the environment. This paper is the first to investigate LLMs as in-context decision-makers under the problem of Dueling Bandits (DB), a stateless preference-based reinforcement learning setting that extends the classic Multi-Armed Bandit (MAB) model by querying for preference feedback. We compare GPT-3.5 Turbo, GPT-4, GPT-4 Turbo, Llama 3.1, and o1-Preview against nine well-established DB algorithms. Our results reveal that our top-performing LLM, GPT-4 Turbo, has the zero-shot relative decision-making ability to achieve surprisingly low weak regret across all the DB environment instances by quickly including the best arm in duels. However, an optimality gap exists between LLMs and classic DB algorithms in terms of strong regret. LLMs struggle to converge and consistently exploit even when explicitly prompted to do so, and are sensitive to prompt variations. To bridge this gap, we propose an agentic flow framework: LLM with Enhanced Algorithmic Dueling (LEAD), which integrates off-the-shelf DB algorithms with LLM agents through fine-grained adaptive interplay. We show that LEAD has theoretical guarantees inherited from classic DB algorithms on both weak and strong regret. We validate its efficacy and robustness even with noisy and adversarial prompts. The design of our framework sheds light on how to enhance the trustworthiness of LLMs used for in-context decision-making.
Following the success of the in-context learning paradigm in large-scale language and computer vision models, the recently emerging field of in-context reinforcement learning is experiencing a rapid growth. However, its development has been held back by the lack of challenging benchmarks, as all the experiments have been carried out in simple environments and on small-scale datasets. We present \textbf{XLand-100B}, a large-scale dataset for in-context reinforcement learning based on the XLand-MiniGrid environment, as a first step to alleviate this problem. It contains complete learning histories for nearly $30,000$ different tasks, covering $100$B transitions and $2.5$B episodes. It took $50,000$ GPU hours to collect the dataset, which is beyond the reach of most academic labs. Along with the dataset, we provide the utilities to reproduce or expand it even further. With this substantial effort, we aim to democratize research in the rapidly growing field of in-context reinforcement learning and provide a solid foundation for further scaling. The code is open-source and available under Apache 2.0 licence at https://github.com/dunno-lab/xland-minigrid-datasets.
In-context learning refers to the learning ability of a model during inference time without adapting its parameters. The input (i.e., prompt) to the model (e.g., transformers) consists of both a context (i.e., instance-label pairs) and a query instance. The model is then able to output a label for the query instance according to the context during inference. A possible explanation for in-context learning is that the forward pass of (linear) transformers implements iterations of gradient descent on the instance-label pairs in the context. In this paper, we prove by construction that transformers can also implement temporal difference (TD) learning in the forward pass, a phenomenon we refer to as in-context TD. We demonstrate the emergence of in-context TD after training the transformer with a multi-task TD algorithm, accompanied by theoretical analysis. Furthermore, we prove that transformers are expressive enough to implement many other policy evaluation algorithms in the forward pass, including residual gradient, TD with eligibility trace, and average-reward TD.
Structured state space sequence (S4) models have recently achieved state-of-the-art performance on long-range sequence modeling tasks. These models also have fast inference speeds and parallelisable training, making them potentially useful in many reinforcement learning settings. We propose a modification to a variant of S4 that enables us to initialise and reset the hidden state in parallel, allowing us to tackle reinforcement learning tasks. We show that our modified architecture runs asymptotically faster than Transformers in sequence length and performs better than RNN's on a simple memory-based task. We evaluate our modified architecture on a set of partially-observable environments and find that, in practice, our model outperforms RNN's while also running over five times faster. Then, by leveraging the model's ability to handle long-range sequences, we achieve strong performance on a challenging meta-learning task in which the agent is given a randomly-sampled continuous control environment, combined with a randomly-sampled linear projection of the environment's observations and actions. Furthermore, we show the resulting model can adapt to out-of-distribution held-out tasks. Overall, the results presented in this paper show that structured state space models are fast and performant for in-context reinforcement learning tasks. We provide code at https://github.com/luchris429/popjaxrl.
The mathematical formula is the human language to describe nature and is the essence of scientific research. Finding mathematical formulas from observational data is a major demand of scientific research and a major challenge of artificial intelligence. This area is called symbolic regression. Originally symbolic regression was often formulated as a combinatorial optimization problem and solved using GP or reinforcement learning algorithms. These two kinds of algorithms have strong noise robustness ability and good Versatility. However, inference time usually takes a long time, so the search efficiency is relatively low. Later, based on large-scale pre-training data proposed, such methods use a large number of synthetic data points and expression pairs to train a Generative Pre-Trained Transformer(GPT). Then this GPT can only need to perform one forward propagation to obtain the results, the advantage is that the inference speed is very fast. However, its performance is very dependent on the training data and performs poorly on data outside the training set, which leads to poor noise robustness and Versatility of such methods. So, can we combine the advantages of the above two categories of SR algorithms? In this paper, we propose \textbf{FormulaGPT}, which trains a GPT using massive sparse reward learning histories of reinforcement learning-based SR algorithms as training data. After training, the SR algorithm based on reinforcement learning is distilled into a Transformer. When new test data comes, FormulaGPT can directly generate a "reinforcement learning process" and automatically update the learning policy in context. Tested on more than ten datasets including SRBench, formulaGPT achieves the state-of-the-art performance in fitting ability compared with four baselines. In addition, it achieves satisfactory results in noise robustness, versatility, and inference efficiency.
We introduce PokeLLMon, the first LLM-embodied agent that achieves human-parity performance in tactical battle games, as demonstrated in Pokemon battles. The design of PokeLLMon incorporates three key strategies: (i) In-context reinforcement learning that instantly consumes text-based feedback derived from battles to iteratively refine the policy; (ii) Knowledge-augmented generation that retrieves external knowledge to counteract hallucination and enables the agent to act timely and properly; (iii) Consistent action generation to mitigate the panic switching phenomenon when the agent faces a powerful opponent and wants to elude the battle. We show that online battles against human demonstrates PokeLLMon's human-like battle strategies and just-in-time decision making, achieving 49% of win rate in the Ladder competitions and 56% of win rate in the invited battles. Our implementation and playable battle logs are available at: https://github.com/git-disl/PokeLLMon.
Recently, it has been shown that transformers pre-trained on diverse datasets with multi-episode contexts can generalize to new reinforcement learning tasks in-context. A key limitation of previously proposed models is their reliance on a predefined action space size and structure. The introduction of a new action space often requires data re-collection and model re-training, which can be costly for some applications. In our work, we show that it is possible to mitigate this issue by proposing the Headless-AD model that, despite being trained only once, is capable of generalizing to discrete action spaces of variable size, semantic content and order. By experimenting with Bernoulli and contextual bandits, as well as a gridworld environment, we show that Headless-AD exhibits significant capability to generalize to action spaces it has never encountered, even outperforming specialized models trained for a specific set of actions on several environment configurations. Implementation is available at: https://github.com/corl-team/headless-ad.
Recently, extensive studies in Reinforcement Learning have been carried out on the ability of transformers to adapt in-context to various environments and tasks. Current in-context RL methods are limited by their strict requirements for data, which needs to be generated by RL agents or labeled with actions from an optimal policy. In order to address this prevalent problem, we propose AD$^\varepsilon$, a new data acquisition approach that enables in-context Reinforcement Learning from noise-induced curriculum. We show that it is viable to construct a synthetic noise injection curriculum which helps to obtain learning histories. Moreover, we experimentally demonstrate that it is possible to alleviate the need for generation using optimal policies, with in-context RL still able to outperform the best suboptimal policy in a learning dataset by a 2x margin.
We introduce AMAGO, an in-context Reinforcement Learning (RL) agent that uses sequence models to tackle the challenges of generalization, long-term memory, and meta-learning. Recent works have shown that off-policy learning can make in-context RL with recurrent policies viable. Nonetheless, these approaches require extensive tuning and limit scalability by creating key bottlenecks in agents' memory capacity, planning horizon, and model size. AMAGO revisits and redesigns the off-policy in-context approach to successfully train long-sequence Transformers over entire rollouts in parallel with end-to-end RL. Our agent is scalable and applicable to a wide range of problems, and we demonstrate its strong performance empirically in meta-RL and long-term memory domains. AMAGO's focus on sparse rewards and off-policy data also allows in-context learning to extend to goal-conditioned problems with challenging exploration. When combined with a multi-goal hindsight relabeling scheme, AMAGO can solve a previously difficult category of open-world domains, where agents complete many possible instructions in procedurally generated environments.
Large transformer models pretrained on offline reinforcement learning datasets have demonstrated remarkable in-context reinforcement learning (ICRL) capabilities, where they can make good decisions when prompted with interaction trajectories from unseen environments. However, when and how transformers can be trained to perform ICRL have not been theoretically well-understood. In particular, it is unclear which reinforcement-learning algorithms transformers can perform in context, and how distribution mismatch in offline training data affects the learned algorithms. This paper provides a theoretical framework that analyzes supervised pretraining for ICRL. This includes two recently proposed training methods -- algorithm distillation and decision-pretrained transformers. First, assuming model realizability, we prove the supervised-pretrained transformer will imitate the conditional expectation of the expert algorithm given the observed trajectory. The generalization error will scale with model capacity and a distribution divergence factor between the expert and offline algorithms. Second, we show transformers with ReLU attention can efficiently approximate near-optimal online reinforcement learning algorithms like LinUCB and Thompson sampling for stochastic linear bandits, and UCB-VI for tabular Markov decision processes. This provides the first quantitative analysis of the ICRL capabilities of transformers pretrained from offline trajectories.
Large transformer models trained on diverse datasets have shown a remarkable ability to learn in-context, achieving high few-shot performance on tasks they were not explicitly trained to solve. In this paper, we study the in-context learning capabilities of transformers in decision-making problems, i.e., reinforcement learning (RL) for bandits and Markov decision processes. To do so, we introduce and study Decision-Pretrained Transformer (DPT), a supervised pretraining method where the transformer predicts an optimal action given a query state and an in-context dataset of interactions, across a diverse set of tasks. This procedure, while simple, produces a model with several surprising capabilities. We find that the pretrained transformer can be used to solve a range of RL problems in-context, exhibiting both exploration online and conservatism offline, despite not being explicitly trained to do so. The model also generalizes beyond the pretraining distribution to new tasks and automatically adapts its decision-making strategies to unknown structure. Theoretically, we show DPT can be viewed as an efficient implementation of Bayesian posterior sampling, a provably sample-efficient RL algorithm. We further leverage this connection to provide guarantees on the regret of the in-context algorithm yielded by DPT, and prove that it can learn faster than algorithms used to generate the pretraining data. These results suggest a promising yet simple path towards instilling strong in-context decision-making abilities in transformers.
We propose Algorithm Distillation (AD), a method for distilling reinforcement learning (RL) algorithms into neural networks by modeling their training histories with a causal sequence model. Algorithm Distillation treats learning to reinforcement learn as an across-episode sequential prediction problem. A dataset of learning histories is generated by a source RL algorithm, and then a causal transformer is trained by autoregressively predicting actions given their preceding learning histories as context. Unlike sequential policy prediction architectures that distill post-learning or expert sequences, AD is able to improve its policy entirely in-context without updating its network parameters. We demonstrate that AD can reinforcement learn in-context in a variety of environments with sparse rewards, combinatorial task structure, and pixel-based observations, and find that AD learns a more data-efficient RL algorithm than the one that generated the source data.
기존의 오프라인 인컨텍스트 강화학습(in-context reinforcement learning, ICRL) 방법들은 주로 지도 학습 목적 함수에 의존해 왔으며, 이러한 목적 함수는 오프라인 RL 환경에서 한계가 있는 것으로 알려져 있다. 본 연구에서는 오프라인 ICRL 프레임워크 내에 RL 목적 함수를 통합하는 방법을 탐구한다. 150개 이상의 GridWorld 및 MuJoCo 환경 기반 데이터셋에 대한 실험을 통해, RL 목적 함수를 직접 최적화하면 다양한 데이터셋 커버리지, 구조, 전문성 수준, 환경 복잡도 전반에서 널리 사용되는 Algorithm Distillation(AD) 대비 평균 약 30%의 성능 향상을 얻을 수 있음을 보인다. 나아가 난이도가 높은 XLand-MiniGrid 환경에서는 RL 목적 함수가 AD의 성능을 두 배로 높였다. 또한 실험 결과, 가치 학습 과정에 보수성(conservatism)을 추가하면 시험한 거의 모든 설정에서 추가적인 성능 향상이 나타났다. 이러한 결과는 ICRL의 학습 목적 함수를 RL의 보상 최대화 목표와 일치시키는 것이 중요함을 강조하며, 오프라인 RL이 ICRL을 발전시키기 위한 유망한 방향임을 보여준다.